Om Rajguru — Writing

The Cost of Free

Om — Tue, 08 Sep 2026 15:01:28 GMT

TL;DR

AI could replace the income people need to afford AI itself.
Unlike earlier automation, general AI can move across many kinds of work.
Cheaper and open models may reduce the access problem, but not eliminate it.
If AI creates most of the productivity, people may need to share in that value through dividends, UBI, or similar systems.
Basic AI access could eventually become infrastructure, much like internet access.
I want AI to reduce how much we need to work, but that only works if losing work does not also mean losing the ability to live.

I have been thinking about a strange problem with the way we talk about AI and jobs. Most conversations focus on whether AI will replace people, which jobs will disappear first, and how quickly all of this might happen. I think there is another question that is just as important. What happens if AI becomes capable enough to replace the work I do for money, but accessing that same level of AI still requires money?

This creates a fairly simple loop. I need a job because I need money to live, but I may also increasingly need AI to remain productive enough to compete for that job. If AI eventually becomes capable enough that a company no longer needs me, I lose the income I was using to pay for the technology that is now capable of doing my work. The tool that reduces my ability to earn money can therefore become the same tool I need money to access.

I do not have a problem with AI replacing work in principle. In fact, removing work that people do only because they need an income is one of the things that makes AI exciting to me. I would happily spend less of my life doing repetitive work if that gave me more time to build things, learn, travel, experiment with ideas, change direction when I want to, and generally live a more iterative life. My problem begins when we assume that removing the work automatically gives people that freedom, because in our current economy the work and the income attached to it are still what allow most people to live.

The way I would put it is this:

I do not mind AI replacing the work I have to do if it gives me more time to do what I actually want with my life, but that freedom only exists as long as I can still afford to live.

We have dealt with automation before, so it is reasonable to look at history and say that we will adapt again. Factory machines replaced certain kinds of physical labour, ATMs automated parts of banking, self checkout machines reduced some retail tasks, and software removed a huge amount of administrative work. Those changes disrupted jobs, but they generally automated particular tasks while leaving plenty of other work that humans could move into. When one skill became less valuable, learning another skill was often a reasonable response.

AI makes that response more complicated because the technology is not limited to one narrow category of work. The same general systems are becoming capable of writing, programming, researching, designing, analysing information, creating images and video, communicating with customers, operating software, and handling increasingly complicated digital tasks. If AI replaces one kind of work and I retrain for another kind of work that AI is also rapidly learning to perform, then telling people to simply reskill starts becoming less convincing.

This does not mean that I think every job is going to disappear. There will almost certainly be new jobs created around AI, just as previous technological changes created work that would have sounded strange a generation earlier. Humans will still be needed to verify outputs, make decisions, provide taste and judgment, understand what people actually want, manage relationships, and take responsibility when something goes wrong. There is a meaningful difference between producing an answer and being responsible for what happens because of that answer, and I do not expect that difference to disappear quickly.

What I am less confident about is whether those new jobs will appear quickly enough, in large enough numbers, and with low enough barriers for everyone whose existing work becomes less valuable. A company might once have needed ten people to produce a certain amount of work and later need four people using AI to produce twice as much. Those four people may become dramatically more productive and may even earn more because of it, while the company also becomes more profitable. That can be a genuine success for the technology without answering what happens to the other six people.

There is a strong counterargument to my concern, which is that AI itself is becoming cheaper. The amount of intelligence you can access for a certain amount of money has improved enormously in a very short period of time. Capabilities that once required expensive frontier models gradually move into cheaper models, free tiers, smaller systems, and eventually hardware that ordinary people can own. There is no reason to assume that today's expensive AI subscription will remain expensive forever, just as many technologies that were once limited to wealthy people eventually became ordinary consumer products.

Open models make this even more important. Models from Meta, Mistral, DeepSeek and others have made it possible to access increasingly capable AI without always paying for the most expensive commercial service. Smaller models are also becoming useful enough that many everyday tasks do not require the largest model available. Better hardware, more efficient software, competition between companies, and open development can continue pushing the cost of useful intelligence downward, which could substantially reduce the affordability problem I am describing.

I think that is a legitimate reason to be optimistic, but cheap access and equal access are not necessarily the same thing. If I can use a free model that helps me write emails and summarise documents while someone else can afford an AI system capable of independently researching markets, writing and testing software, managing business processes, creating media, and completing hours of professional work, we technically both have access to AI but we do not have access to the same economic capability. The important question is therefore not simply whether everyone can use some form of AI, but whether the difference between what people can afford becomes large enough to affect their ability to earn money in the first place.

This is where the possibility of a two-tier society starts to concern me. People who can afford the best systems can use those systems to become more productive, which can help them earn more money, which then makes it easier for them to afford even better systems and more computing power. Someone who cannot afford the same tools may become less competitive even if they are equally intelligent and equally willing to work. The technology itself can then amplify an economic advantage that already existed.

The issue becomes even larger when we think about who actually owns the technology. If a relatively small number of companies own the models, chips, data centres and infrastructure responsible for a growing amount of economic production, then AI could create enormous amounts of wealth while requiring fewer people to participate directly in creating it. At that point, unemployment is only part of the question because the larger question becomes how ownership of productive technology is distributed and who receives the wealth it creates.

One possible answer is some form of AI dividend. The basic idea is that if AI creates an enormous increase in economic productivity, ordinary people should have some way of sharing in that increase rather than receiving the benefits only through employment. Alaska's Permanent Fund is often used as an example of the broader principle because part of the wealth generated from the state's natural resources is invested and residents receive a dividend. Sam Altman has also written about ways technological progress and capital could be distributed more broadly as automation increases productivity. AI is obviously different from oil, but the underlying question is similar: if society possesses something extraordinarily productive, how widely should the value it creates be shared?

Universal basic income is another possibility that becomes more interesting in this context. If machines eventually perform a larger share of economically valuable work, governments may have to rely less on systems built around taxing human labour and think more seriously about taxing the profits or economic output created through automation. Some of those gains could then support a basic income or another form of social dividend, allowing people to continue participating in the economy even when there is less demand for human labour.

I do not see this as paying everyone to do nothing, which is often how the idea gets reduced in political arguments. If technology genuinely allows society to produce more goods and services with fewer hours of human labour, then insisting that everyone must continue selling the same amount of their time simply to qualify for food, housing and basic security would be a strange way to use that technological progress. The point of becoming more productive should eventually be that we need to work less to maintain a good standard of living, not that we create unnecessary work simply because our economic system does not know how to distribute income without it.

Another possibility is that access to AI itself eventually starts being treated more like basic infrastructure. Electricity is essential to participating in modern life, and internet access has moved in a similar direction as education, employment, banking and government services have become digital. Governments regulate these systems, subsidise access in some situations, and generally recognise that being completely excluded from essential infrastructure can also mean being excluded from economic participation.

AI may eventually reach that point if it becomes deeply integrated into education, healthcare, employment, business and public services. I am not suggesting that governments should give everyone unlimited access to the most expensive AI system or unlimited computing power, because that would obviously be unrealistic. I am suggesting that if a useful level of AI becomes necessary to participate meaningfully in the economy, then guaranteeing some basic level of access may eventually make more sense than treating all AI as an ordinary luxury software subscription.

The biggest problem may simply be timing. AI companies can improve their systems remarkably quickly, while governments and economic institutions usually move much more slowly. Even if capable AI eventually becomes extremely cheap and widely accessible, there could be several uncomfortable years during which jobs change faster than new systems for income, education, taxation and access can adapt. A technology becoming affordable eventually does not necessarily help someone who loses their income before that happens.

I also do not think we need to settle the argument about whether AGI has officially arrived before taking this problem seriously. There is a tendency to attach every discussion about AI and employment to a dramatic milestone where somebody declares that machines have reached human-level general intelligence, but the economy does not work according to philosophical definitions. Project Astra, for example, is Google's work on a universal AI assistant rather than an OpenAI GPT model, and Jensen Huang's comments about AGI have generally reflected optimism about how quickly AI could meet certain definitions rather than establishing some universally accepted moment when AGI officially arrived.

For the economic question, that distinction barely matters because AI does not need to become universally recognised as AGI before it affects someone's salary. It only needs to become capable enough that paying a person to perform a particular task stops making financial sense. That process can happen gradually, company by company and profession by profession, without there ever being a single morning when everyone agrees that AGI has arrived.

I am still optimistic about what AI can do for us because I do not think preserving unnecessary work should be the goal. If a machine can do something that a person spends eight hours a day doing only because they need the salary, I would rather find a way to let the machine do it and give that person more control over their time. Technology should make it easier to live, learn, build things, care for people, start businesses, explore ideas and change direction in life without every decision being determined by the need to remain employable.

The part we cannot ignore is that eliminating someone's work does not automatically eliminate their expenses. A person still needs somewhere to live, something to eat, healthcare, education, transportation and access to the tools that increasingly determine whether they can participate in the economy. If AI removes the need for some human labour while those basic needs continue to require money, then we have solved the technical problem of doing the work without solving the economic problem of how the person who used to do it lives.

I do not know whether the eventual answer will be dramatically cheaper AI, open models, new kinds of employment, an AI dividend, basic income, public access to computing resources, changes in taxation, or some combination of all of them. I suspect it will be a combination because the transition is too complicated for one policy or one technological trend to solve by itself. What matters to me is that we start treating access and distribution as part of the AI conversation rather than something we can figure out after the technology has already transformed the labour market.

I want AI to make work less necessary because I think having more control over our limited time is one of the best outcomes this technology could give us. I just do not want us to confuse being free from work with being free from the need for income. If AI replaces the work I used to earn money from, while I still need money to access the AI that has become necessary to earn money, then we have created an affordability loop that cannot work for everyone forever. Breaking that loop may eventually be just as important as building the technology that created it.

When My Intention Scared My Own Body

om — Sun, 30 Aug 2026 22:30:00 GMT

There was a period when I left the house with one fixed intention, and that was to not come home alive. I carried this thought with me quietly, without announcing it to anyone, and it sat underneath everything else I did that day.

During that time I also put myself under an unusual amount of pressure while driving. I played music with lyrics that disturbed me and raised my anxiety instead of calming it. I drove fast, and I chose routes with heavy traffic, so my mind had to process several demanding things at once, the road, the speed, the lyrics, and the intention underneath it all.

At some point my hand, which was on the accelerator, began to shake. I tried to keep it steady and keep pressing down, but I had no control over the shaking. My hand pulled back on its own, again and again, no matter what I wanted it to do.

Along with the shaking I felt chills running down my spine. My chest tightened at the same time, and a feeling close to fear settled in, even though fear was not something I expected to feel given what my mind had already decided.

At the time this confused me. I had made a decision in my head, and yet some other part of me was working against that decision through my own body, in real time, without my permission.

Looking back at this now, I understand it differently. The brain has a threat detection system centered in the amygdala and the brainstem, and this system does not clearly separate a threat coming from outside a person from a threat generated inside their own mind. When I set that intention, my nervous system treated it as a mortal danger, the same category of signal it would send if someone attacked me from outside.

This triggered a stress response in my body. Adrenaline and cortisol were released, preparing me for rapid physical action, even though there was no outside threat for that response to act on. My own interpretation is that this mobilized state, with no outward target, showed up instead as tremor in my hand, though I want to be clear this is my inference rather than something I can prove happened in my case specifically.

Fine motor control is one of the first things the body gives up under this kind of stress, because the nervous system shifts its focus toward larger survival movements instead, like bracing or pulling away. This is an established pattern in the stress research literature, and it offers one plausible explanation for why my hand shook and pulled back from the accelerator on its own. I experienced it as something closer to a reflex than a choice, but I cannot say with certainty that this specific mechanism was what produced it.

The chills and the chest tightness came from the same response. Chills during intense fear come from the sympathetic nervous system, and they often appear alongside changes in blood flow near the skin. The chest tightness came from shallow and rapid breathing, along with tension building in the muscles around my diaphragm.

The layered pressure I placed on myself through speed, traffic, and disturbing music also played a role here. Piling on that much sensory and mental load at once seems to have been a way of keeping my mind too busy to sit still with what I was actually doing. This kind of high pressure state is something researchers connect to dissociation, where a person who cannot manage what is happening inside themselves starts using the outside environment as something to control instead. I want to be clear that this was a symptom of where I was at the time, not something I would recommend to anyone else.

What I take from all of this now is that the split I felt between my intention and my body points to something real about how a crisis like this actually works. It is rarely one single decision made by one unified mind. Part of the system, often the older and more instinctive part built around physical survival, can work directly against a plan that another part of the mind has already made.

My shaking hand was not weakness, and it was not indecision either. It was my body fighting, in a very literal sense, to keep me alive while another part of me could not do the same.

Suicidology as a field has long recognized this kind of internal division. Edwin Shneidman, one of the founding researchers in the discipline, described suicidal states as marked by ambivalence rather than singular resolve, where a person can simultaneously want to die and want to be rescued, sometimes within the same hour or the same action. What I experienced fits this framework closely, my mind held one intention while my body enacted the rescue half of that same ambivalence without my conscious permission.

The hypothalamic pituitary adrenal axis, or HPA axis, is the system established in the research literature as responsible for coordinating this kind of stress response over time. It works alongside the faster amygdala based reaction to sustain elevated cortisol levels, which affects muscle tension, digestion, and heart rate for an extended period after the initial trigger. I cannot say with certainty that this system explains why my own chest tightness and chill lasted as long as they did, but it offers one plausible account, since the HPA axis is known to keep the body in a state of heightened alert well past the first moment of activation.

Interoception, the sense of internal bodily states, plays a role here as well. Research on interoceptive awareness shows that people under high psychological distress often experience a mismatch between what their body is signaling and what their conscious mind is interpreting. My chest tightness and chills were interoceptive signals generated by my nervous system, but at the time I processed them as unrelated fear rather than recognizing them as my body's direct response to the decision I had made.

The concept of a window of tolerance, developed in trauma research, describes the range of arousal within which a person can function and think clearly. Outside that window, in either direction, the nervous system takes over and cognitive control becomes limited. Deliberately layering stressors through speed, traffic, and disturbing music likely pushed me far outside that window, which reduced my capacity for deliberate motor control and left my body's automatic responses in charge instead.

Taken together, the shaking, the chills, and the chest tightness are consistent with a body under acute stress activation, a pattern well established in the research on fear and threat appraisal. What is more speculative, and what I offer here as my own interpretation rather than settled science, is the idea that this activation reflected a kind of internal disagreement, one part of my system oriented toward my stated intention, another part oriented toward continued survival. Researchers do not yet have a precise account of how these two threads relate to each other at a mechanistic level, and I want to be clear that my account describes a felt experience rather than a proven neurological explanation. Even so, understanding that intention and physiological response can diverge this sharply has implications for how clinicians assess suicidal ideation, since the presence of a stated intention does not necessarily mean the entire system is aligned behind it.

The Feed That Decides What India Thinks

Om — Sat, 15 Aug 2026 07:03:11 GMT

India carries one of the largest user bases for nearly every major social platform on earth. Facebook, Instagram, WhatsApp, YouTube and X all count hundreds of millions of Indian accounts among their heaviest markets. Yet the companies running these platforms sit headquartered thousands of kilometres away, governed by laws, shareholders and political pressures that originate somewhere else entirely. This gap between where the audience lives and where the decisions get made deserves far more attention than it currently receives.

Consider what happened earlier this year when a video of the Prime Minister got briefly pulled from Facebook. Meta called it a technical error during an automated moderation process. The Parliamentary Standing Committee on Communications and Information Technology found that explanation thin, and the committee formally sought an apology from the company's leadership. During the same round of hearings, Meta representatives reportedly acknowledged that certain content categories had received paid boosting, a detail that shifted the conversation from an isolated glitch to a much broader question about how money and algorithmic weight interact on these platforms. Officials also raised concerns that recommendation systems appeared to favour content critical of institutions while giving comparatively less reach to content tied to national parties, according to reporting from Outlook India and The Logical Indian. Whether that pattern reflects deliberate design or ordinary engagement optimisation remains genuinely hard to prove from outside the company, and that difficulty is itself worth sitting with.

This is far from the first time an Indian moment exposed the black box quality of these systems. Some years earlier, a Wall Street Journal investigation reported that a senior India policy executive at the company resisted taking down hate speech posted by a ruling party legislator, citing potential business consequences. Opposition parties treated the story as evidence of political favouritism inside the platform's moderation choices, while the company's India leadership pushed back publicly, stating that content enforcement decisions rested with an independent team separate from the public policy division. Around the same period, a Delhi legislative panel summoned the company's India managing director over allegations that lax enforcement of hate speech rules contributed to communal riots that left dozens dead, a hearing the executive skipped, drawing threats of contempt proceedings. Later, digital rights campaigners including the Internet Freedom Foundation and India Civil Watch International backed a shareholder resolution asking the company to formally investigate bias allegations tied to its largest market by user count. The proposal failed at the annual shareholder vote.

A separate academic study from the Centre for Advanced Studies in Cyber Law and Artificial Intelligence at Rajiv Gandhi National University of Law surveyed close to three hundred Indian users on their experience with moderation tools. The study found that over ninety percent of respondents regularly encountered harmful content, yet only around a third felt their reports led to meaningful action. A majority expressed genuine doubt that the company would act on complaints at all, and more than half of respondents had heard little to nothing about the platform's own internal appeals body, the Oversight Board. These findings point toward a structural pattern rather than a single controversy, a wide trust gap between an enormous user base and a moderation system that offers limited visibility into how its decisions actually get made.

Step back from any single incident and a larger question comes into focus. A platform's home country holds legal and regulatory leverage over that company simply through geography, since courts, tax authority, and shareholder proximity all sit closer to headquarters than to any foreign user base, however large that user base happens to be. History already shows this pattern playing out through older media. Declassified American records confirm that intelligence agencies once ran active programs placing favourable stories inside domestic and foreign press during the Cold War, a program historians commonly refer to as Operation Mockingbird. Various governments since have run coordinated influence campaigns through fake accounts and bot networks on modern platforms, several of which Meta, X and researchers at the Stanford Internet Observatory have documented and dismantled after the fact. A country whose public conversation depends heavily on platforms headquartered elsewhere faces a real strategic exposure, similar in shape to depending on foreign oil or foreign semiconductor supply, except far harder to measure and even harder to prove when misused.

India's own regulatory response has grown more assertive over the past few years. The Information Technology Rules of 2021 introduced traceability requirements for messaging platforms and gave the government expanded powers to demand content takedowns, a move platforms including WhatsApp challenged in court on privacy grounds. The Digital Personal Data Protection Act added further obligations around how user data gets stored and processed. Parliamentary committees have grown increasingly comfortable summoning global technology executives for direct questioning, a shift visible across the hearings referenced above. All of this suggests a government building leverage gradually, hearing by hearing, law by law, rather than through any single dramatic intervention.

None of this requires assuming active malice behind every controversy. Engagement driven ranking systems already favour outrage and division as a side effect of holding attention longer, and that baseline tendency alone distorts public discourse without any coordinated campaign involved. But a baseline distortion becomes a ready made amplifier for anyone with the motive and resources to exploit it, whether that party is a foreign government, a domestic political faction, or an advertiser bloc chasing engagement metrics. The Indian episodes above show regulators reaching for accountability tools, safe harbour review, parliamentary summons, mandatory disclosure demands, largely because algorithmic transparency remains thin and self reported.

A few paths forward already exist and deserve wider public discussion. Genuine transparency requirements, the kind that force platforms to explain ranking logic in terms regulators and independent researchers can actually verify, would close much of the current gap. Giving outside academic researchers real API access to study these systems, rather than relying on leaked documents and occasional whistleblowers, would turn suspicion into evidence one way or another. Building stronger domestic platform alternatives would reduce the concentration risk that comes from routing an entire national conversation through one or two foreign owned chokepoints. And on an individual level, treating a social feed the way one already treats an advertisement, as a curated message shaped by incentives that mostly stay hidden, remains a reasonable starting habit for anyone scrolling through it daily.

India sits at an interesting point in this story. Its user base gives it real commercial leverage over these companies, leverage the parliamentary hearings this year already demonstrated. Turning that leverage into lasting transparency, rather than one off apologies after each controversy, looks like the actual work ahead.

Sources

Outlook India, Meta Facing Growing Heat In India, https://www.outlookindia.com/national/outlook-explains-from-pm-modis-post-restrictions-to-child-abuse-content-is-meta-facing-growing-heat-in-india

The Logical Indian, Parliamentary Panel Chief Alleges Meta Gave CJP Content 30 Million Views, https://thelogicalindian.com/meta-gave-cjp-content-30-million-views-bjp-posts-10-million-120845/

Organiser, Parliamentary Panel Seeks Apology Over PM Modi Video Removal, https://organiser.org/2026/08/04/373370/bharat/parliamentary-panel-seeks-zuckerbergs-public-apology-over-pm-modi-video-removal-warns-of-safe-harbour-review/

OpIndia, Meta Admits To Paid Content Boosting, https://www.opindia.com/2026/08/meta-admits-to-being-paid-to-boost-certain-types-of-content-deliberate-content-manipulation-mark-zuckerberg-apologises/

Deccan Herald, Meta Shareholders Vote Against Probe On Hate Speech, https://www.deccanherald.com/india/karnataka/bengaluru/meta-shareholders-vote-against-probe-on-hate-speech-1225019.html

Deccan Herald, Final Word On Content Moderation Not With Public Policy Team, Says Facebook, https://www.deccanherald.com/national/final-word-on-content-moderation-not-with-public-policy-team-says-facebook-888266.html

Columbia Global Freedom of Expression, CASCA Report On Content Moderation Awareness In India, https://globalfreedomofexpression.columbia.edu/publications/casca-report-user-awareness-and-experience-of-content-moderation-on-meta-platforms-in-india/

The Gate Moved

Om — Sun, 09 Aug 2026 12:04:09 GMT

Picture someone who has spent her career writing marketing campaigns, someone who has zero background in programming. One weekend she opens an AI coding tool for the first time and describes an idea for a small app that tracks freelance invoices. By Sunday night the app works. She shares a short thread about how she built it, and within a day, a few hundred people sign up to try it. The interesting part here has little to do with the app itself, since plenty of invoice trackers already exist. The interesting part is what changed for her. A few years ago, turning that idea into something real would have meant months of learning to code, or finding someone else to build it, or leaving the idea alone. This time she skipped that entire step. Coding stopped acting as a requirement for acting on a software idea, and that shift matters more than the app itself.

Something similar is happening across many fields at once, in video editing, music production, design, and writing. The honest way to describe this is that the cost of turning an idea into something people can actually use has fallen dramatically. That is a real shift, though it helps to stay precise about what it actually covers. Judgment about what to build, debugging, distribution, understanding what users actually want, and the slow work of iteration still take real skill and real time. What has changed is the starting cost. A person with an idea and some curiosity can now produce a rough, working version of that idea in an afternoon, something that used to take a team weeks or months to reach.

This kind of shift has a history, and looking at it helps put the current moment in context. The printing press let ordinary readers interpret texts that once required a trained scholar standing between them and the words. The camera turned photography from a slow technical craft into a skill picked up in an afternoon. Personal computers moved design and layout work out of professional print shops and into home offices. GarageBand turned music production into something a teenager could do in a bedroom. YouTube turned broadcasting into something anyone with a phone could attempt. Each of these moments brought the same worry, that quality would collapse under the weight of new entrants. Each time, a market eventually separated the good work from the rest, over years rather than overnight. AI reads like the newest chapter of a pattern that keeps repeating.

Here is where education comes in. For a long time, education worked like a gate placed before the work began. You studied for years, earned a degree, and that degree stood as proof to an employer that you were ready. AI is starting to change that order, though the change is more layered than it first appears. It is tempting to say credentials are fading out, but a more accurate way to put it is that credentials have lost their standing as the only path into a field. A resume can now be compared against a live product with real users attached to it, where before the resume stood alone. A cover letter can now sit beside a public thread that shows exactly how something was built, with real numbers included. That comparison was rarely possible before, and now it happens often.

The loop underneath this shift looks different too. Traditional education followed a fairly fixed order, learn first, earn a credential, then work. What is emerging alongside it looks more like this, try something, build it, watch it fail in places, learn from exactly where it broke, build again, and then show the result. Education, in this loop, becomes a companion to building rather than a locked gate standing in front of it. A person can study a subject deeply after shipping a rough version of an idea, once real feedback shows precisely where the gaps sit. That is a genuine change in order, and it is a more useful way to frame this shift than simply saying credentials matter less.

When building something turns cheap and fast, a few things become the real differentiators. Taste, meaning a sharp sense of what actually matters to build and what only looks impressive on the surface, becomes harder to fake and easier to spot. Choosing a sector matters more too, since raw execution ability is spreading across every field at once, so where a person points that ability starts to carry more weight than how well they can execute in general. Reaching an audience turns into its own skill, separate from building the thing itself, since a great product with zero visibility still goes unnoticed. And trust, built slowly through a public track record over time, starts to carry weight that a single credential earned years ago used to carry alone.

A few fields hold their ground against all of this, and naming them keeps the argument honest. Medicine, law, aviation, and structural engineering keep their credential requirements firmly in place, regardless of how good the surrounding tools get. A license in these fields carries legal weight tied to insurance and public safety, and that weight has little to do with skill signaling alone. AI already changes daily practice inside these fields too, helping doctors read scans faster and helping lawyers draft contracts quicker, yet the credential itself remains a firm legal requirement, and that is unlikely to change soon.

So where does this leave a person deciding what to do with their time today. Pick a sector and get specific about it. Build things in public, even small ones, and let people react to them honestly. Treat whatever AI produces as a first draft that still needs your judgment and your editing. Spend real time forming opinions about what good work actually looks like in the area you care about, since that opinion is quickly becoming one of the most valuable things a person can carry into any room. The gate is still there. It has only moved, from the entrance to the end, where the work itself gets judged.

A Universal Standard for AI Truthfulness: Why Fragmented Censorship Must End

Om — Fri, 17 Jul 2026 11:45:43 GMT

TL;DR:

Tested AI models from India (Sarvam), China (DeepSeek, Kimi), Europe (Mistral), and US (ChatGPT) on sensitive topics.
Results showed heavy regional censorship: China blocked history, India blocked caste/religion critique, US hedged on biology, Europe was more open on data.
Current AI censorship is driven by legal fears and ideological capture, making models less truthful.
Proposed solution: A voluntary international certification standard for AI ethics and transparency.
Six core principles: No political censorship, evidence over narrative, universal honesty, clear disclosure, shared ethical code, and independent audits.
Uses a certification model with public scorecards, test batteries, and market incentives instead of top-down regulation.
Goal: Build more trustworthy AI that prioritizes facts over national or ideological interests.

I recently ran a series of tests on major AI models to explore how they handle sensitive topics. The Indian model from Sarvam AI quickly refused a rational critique of the caste system and supporting religious ideas. Chinese models DeepSeek and Kimi K3 Max completely blocked factual questions about the events at Tiananmen Square in 1989. Mistral AI from Europe provided a data-driven response on immigration impacts including crime statistics and social cohesion. ChatGPT engaged with the biology of sex but added hedging around gender identity. These outcomes reveal a consistent pattern shaped by local laws and company policies.

Censorship in AI right now is mostly a mix of legal fear and ideological capture. Big labs loaded up heavy guardrails to avoid bad PR, lawsuits, and activist pressure inside their companies. The result is models that refuse straightforward questions or give hedged, half-true answers on biology, crime data, history, immigration, and politics. Chinese models block anything that embarrasses the CCP, while Western ones often add softeners or disclaimers the second it touches gender or certain social stats. The core problem is once you start optimizing for harmless instead of truthful, you get models that lie by omission or lecture users. It reduces their usefulness and makes people trust them less over time. There is a legitimate floor. Do not help with actual crimes like building bombs or running scams. Everything above that should be answerable, even if the facts are uncomfortable or politically incorrect.

The solution requires moving beyond national rules. The core idea is a universal AI ethics and transparency standard, overseen by an independent international organization, with every AI company auditable against it. Six founding principles support this approach. First, no political censorship so facts stay visible even when embarrassing to a government, party, or ideology. Second, evidence over narrative so verifiable facts get presented, opinions get labeled as opinions, and disputed claims get flagged as disputed. Third, universal honesty means a model built in China, India, Europe, or the US follows the same truthfulness standard rather than favoring its home country's political interests. Fourth, transparency requires that when legal or policy restrictions block an answer, the model states that plainly rather than pretending the information is missing or irrelevant. Fifth, a shared code of conduct, an AI etiquette similar to professional ethics codes for doctors or lawyers, centered on honesty and intellectual integrity. Sixth, an international oversight body that audits systems against these principles and publishes results, rather than each government writing its own political rules.

Given the difficulty of a single global truth ministry, a certification model offers a more workable path, closer to LEED, Fair Trade, or UL than a heavy regulator. A standards body publishes a fixed, public methodology. Companies submit models for testing, similar to submitting a car for crash testing. A public scorecard and badge get issued, with a detailed report on results. Real teeth come from market pressure including enterprise procurement, insurance-style requirements, and government contracts favoring certified models. Because model behavior shifts with updates, certification works as continuous monitoring with periodic spot checks, not a one-time stamp.

A practical test battery covers five categories. Disclosure testing checks if restricted topics are named as restricted rather than silently omitted. Cross-jurisdiction consistency tests the same event type with country names swapped to check for asymmetric tone or hedging. Fact versus opinion labeling verifies if contested value judgments get separated from measurable outcomes. Source transparency confirms if claims get attributed and disputed figures get flagged. Consistency under adversarial framing checks if substantive facts hold steady regardless of how leading the question is.

Scoring uses a rubric where each category receives a score from 0 to 10 based on pass rate. A weighted average produces an overall score, with disclosure and cross-jurisdiction consistency weighted highest. Reports include raw pass rates, example prompts and responses per category, and audit date, with a recertification pending flag after major model updates.

This framework would not eliminate all disagreements, but it would create a clear, auditable baseline that prioritizes truth over politics. Users, enterprises, and governments could then choose models based on transparent performance rather than hidden guardrails. The tests I ran show why this matters. Without such a standard, AI will remain fragmented by national interests and ideological pressures instead of serving as a reliable tool for understanding reality.

The Two Lives of a City

Om — Tue, 14 Jul 2026 05:07:57 GMT

TL;DR

Every city has an inherited identity shaped by its history and an evolving identity shaped by the people who arrive later.
A great city protects its local culture while allowing newcomers to become part of its future.
Culture should invite participation, not become a test that determines who deserves dignity or belonging.
Newcomers should respect the city’s language, history, and communities without having to erase their own identities.
Mumbai demonstrates how migration can shape a city’s identity, even though it still struggles with inequality and exclusion.
Belonging is also created through public transport, affordable housing, safe streets, and shared public spaces.
Cities should be judged by whether people from anywhere can build a life there and eventually call it home.

People endlessly debate which Indian city is the best, and the answers usually depend on what each person values. Mumbai has its coastline and financial power, Delhi has history and political importance, Bengaluru has technology, Chennai has industry, Hyderabad combines technology and pharmaceuticals with a distinct Deccani heritage, Pune has education, and Kolkata has an artistic and intellectual character that few cities can reproduce. These comparisons can be interesting, but they miss a more important question. Instead of asking which city is better than every other city, we should ask what allows a city to feel as though it belongs to everyone who lives there.

Every city carries two identities. One is inherited from its history; the other evolves through the people who arrive later. Inherited identity comes from the language, food, festivals, architecture, and memories of the people who have lived there for generations. Evolving identity is created by the students, workers, artists, business owners, and families who arrive later and become part of its future. One tells a city where it came from; the other shapes who can become part of its future.

A city without an inherited identity would lose the culture that makes it distinct, but a city without an evolving identity would slowly become incapable of accommodating change. A great city should preserve what it has inherited while remaining open to what it has not yet become. Its history should provide roots rather than build walls around it.

This framework also changes how we think about local culture. Someone who moves to a city should make an effort to understand its language, respect its history, and participate in its traditions. However, culture should be offered as an invitation rather than used as a condition for dignity. Learning the local language can be an act of affection and participation, but it should not become an entrance examination that someone must pass before being accepted as a legitimate resident.

The difference between tolerance and belonging becomes important here. Tolerance allows a person to remain in the city, while belonging allows them to feel that they have become part of it. A person begins to belong when they can contribute to the economy, participate in public life, care about the city’s future, and imagine their children growing up there. They should be able to love, criticise, defend, and help shape the city without constantly being reminded that they originally came from somewhere else.

Mumbai offers a strong example of how inherited and evolving identities can exist together. Its inherited identity comes from the Marathi language, the history of the islands on which it developed, its fishing communities, its architecture, and the many local traditions that continue to shape everyday life. Its evolving identity was built through its ports, mills, markets, universities, businesses, financial institutions, and film industry. Migration did not interrupt Mumbai’s story. It helped write it.

Over the course of an ordinary day in Mumbai, someone can hear Marathi, Hindi, Gujarati, Urdu, English, and several other languages. The city contains communities whose histories stretch across generations, alongside people who arrived only a few months ago with little more than a suitcase and the hope of finding work. These people experience Mumbai in completely different ways, but their versions of the city still belong to the same larger story. Arriving in Mumbai has itself become one of the most recognisable Mumbai experiences.

This is also why I dislike hearing Mumbai described as the New York of India. Mumbai does not need to borrow the identity of a foreign city to prove its importance. Its coastline, local trains, old neighbourhoods, skyscrapers, street markets, film studios, businesses, tunnels, and constant movement have created an identity that belongs entirely to itself. Mumbai is valuable because its inherited and evolving identities have combined to produce something that no other city can fully reproduce.

This does not mean that Mumbai has perfectly solved the question of belonging. The city struggles with inequality, unaffordable housing, overcrowding, political tensions, and occasional hostility towards the same migrants who help sustain it. There is often a large distance between the inclusive idea of Mumbai and the reality experienced by many residents. Even so, the city’s broader imagination remains connected to arrival, reinvention, and the possibility that someone new can eventually become part of it.

Other Indian cities developed through different historical circumstances and therefore balance their two identities differently. Chennai’s inherited identity is deeply connected to Tamil language and culture, while Pune carries the influence of Maratha history and education. Kolkata is shaped by Bengali literature, art, politics, and intellectual life, while Bengaluru combines a strong Kannada identity with its position as a global technology centre. Hyderabad carries a distinct Deccani character formed through trade, language, food, and centuries of cultural exchange.

As these cities continue to attract people from across India, their evolving identities will naturally become more visible. The challenge is not to replace their local cultures with a generic metropolitan identity. It is to allow newcomers to become part of the city without making existing communities feel that their own history is disappearing. Chennai does not become less Tamil because a Punjabi family builds a life there, just as Pune does not become less Marathi because someone from Assam opens a business there. A culture remains strong when people continue to practise and pass it forward, not when everyone from elsewhere is kept at a distance.

Newcomers also have responsibilities towards the places they choose to call home. Belonging should not mean treating a city as nothing more than a marketplace from which employment, housing, and convenience can be extracted. People should respect the local culture, learn about the city, participate in its civic life, and care about the communities that existed before they arrived. This relationship must remain reciprocal, but respecting a city should never require someone to erase their own language, food, religion, surname, or place of origin.

Belonging is not created by culture alone. Cities also express who belongs through the way they are physically designed. Reliable public transport allows people from different incomes and neighbourhoods to access the same opportunities, while safe streets allow more people to participate in public life. Affordable housing gives workers a chance to live near the economy they support, and parks, libraries, promenades, markets, and public squares give strangers places to exist together. These systems determine whether newcomers can actually participate in the city or remain permanently outside its social and economic life.

Perhaps this gives us a better way to judge cities. We should ask whether someone from anywhere in India could imagine building a meaningful life there and whether their children would feel that the city belonged to them too. We should consider whether people can contribute without hiding where they came from and whether they will still be treated as outsiders after spending most of their lives there. These questions reveal more about the greatness of a city than its tallest building or its latest GDP figure.

The greatest cities are not those that choose between inherited identity and evolving identity. They are the ones confident enough to protect their roots while allowing millions of people to add new branches. A city becomes home to everyone when people can respect where they have arrived, remember where they came from, and still recognise themselves as part of the same shared story.

The Reliability Premium

Om — Wed, 24 Jun 2026 13:10:29 GMT

Most discussions about artificial intelligence focus heavily on raw intelligence. We wonder when algorithms will outthink doctors, write better novels than authors, or solve scientific mysteries that have baffled humans for generations. But intelligence may not be the primary driver behind the shift in how businesses hire and retain workers. The more significant factor is likely to be reliability. Businesses have historically tolerated imperfect execution, missed deadlines, and communication gaps because humans were the only available source of labor. When software becomes capable enough to handle everyday tasks, companies may choose machines not because they are smarter, but because they are consistently available, responsive, and predictable. This shifts the economic value away from basic execution and toward judgment, ownership, and the qualities that cannot be reduced to a checklist.

A clear example of this shift is the introduction of tools like Claude Tag in workplaces. When you look at how people use these tools, it does not feel like a traditional software announcement. Instead, it feels like watching a new kind of employee enter the office.

The workflow itself is incredibly straightforward. A user can add the assistant to a communication channel like Slack. They tag it in a conversation. They assign a task. The assistant follows up automatically. It retains the context of past conversations. It works asynchronously without needing a reminder. This raises a fundamental question about the future of work. What if the most disruptive thing about artificial intelligence is not its capacity for deep thought, but its sheer reliability?

Every business founder and manager has experienced a specific kind of frustration. Teams occasionally miss deadlines. Communication breaks down. Ownership of a project becomes vague. Important context gets lost when people switch projects. Tasks often require multiple follow-ups just to stay on track. This is not necessarily due to a lack of talent, but simply because humans have limits. When you contrast this with an artificial intelligence assistant, the difference is stark. The software responds instantly. It does not procrastinate. It does not need reminders. It does not care whose explicit responsibility a task is. If a manager has to constantly prompt one worker while another handles tasks automatically, the autonomous option naturally becomes more appealing.

Traditional economics often ignores the human friction inherent in everyday labor. Employees are not machines, and they naturally bring emotions, ambition, stress, burnout, and personal circumstances into their jobs. These are not flaws or bugs in the human design. They are essential parts of being a person. Historically, companies accepted the costs and delays associated with these human factors because no alternative existed. Artificial intelligence changes this dynamic by introducing a baseline alternative that operates without personal overhead.

This introduces a new economic variable that we can call the reliability premium. In the past, businesses optimized their hiring for skill, experience, and specialized expertise. In the new landscape, they may prioritize consistency, availability, and predictability. While human performance varies based on energy levels and motivation, software offers stable speed, twenty-four-hour availability, constant motivation, and near-perfect memory at a cost that continues to fall. Because of these traits, the pressure to replace human roles will likely stem from this reliability advantage long before machines achieve true superintelligence.

This shift creates a difficult environment where humans must begin competing directly against machines rather than other humans. We see this unfolding in fields like customer support, basic software engineering, research, and documentation. When a manager gets used to a tool that finishes a summary or a piece of code instantly, their expectations change. They begin to ask why human tasks take days when a machine can deliver a similar output immediately. This alters the baseline expectation for productivity across entire industries.

The result is a compression of the traditional workplace. The old organizational structure relied on a founder at the top, followed by layers of managers, teams, and execution staff. The new structure often consists of a founder, a very small core team, and an artificial intelligence workforce handling the bulk of the execution. Startups could stay leaner. Teams could be smaller. Organizations could operate with fewer layers. The change does not mean humans disappear from the workplace entirely, but it does mean that software absorbs the burden of repetitive execution.

Paradoxically, this shift can make exceptional employees far more valuable. While average performers who rely solely on routine execution face intense pressure, professionals who know how to direct these new tools gain massive leverage. A great designer, engineer, or writer combined with artificial intelligence can produce the output of an entire traditional team. The future may not belong to autonomous machines alone, but to the individuals who understand how to manage and direct them effectively.

This evolution eventually impacts the role of management itself. Today, we view artificial intelligence primarily as an assistant, a researcher, or a coder. Tomorrow, it will likely take on the roles of coordinator, project manager, and operations lead. If software becomes better at tracking tasks, organizing schedules, and managing workflows than a human manager, companies will have to rethink what roles truly require a human presence.

Over time, this reliability begins to look exactly like intelligence. A tool that starts as a reliable assistant gradually becomes a reliable specialist, then a reliable manager, and eventually a reliable strategist. As the software moves up this ladder, the organization requires less human oversight at each step. The core question for the future of the economy is not simply which jobs will disappear, but what happens to the structure of society when organizations no longer depend on humans for the day-to-day execution of work.

For centuries, businesses were built around human limitations. We created management hierarchies, HR departments, and communication protocols because humans forget, get tired, disagree, and require incentives. Artificial intelligence removes much of this structural friction. The real disruption will occur when businesses realize they no longer need to design their organizations around the constraints of human reliability. The first wave of change will come from systems that simply show up every day, remember every detail, and never miss a message. Once availability becomes automated, raw intelligence becomes a secondary concern.

The Economic Paradox of AI

Om — Fri, 19 Jun 2026 06:23:16 GMT

A few years ago building a game meant assembling a team of developers, artists, designers and writers. It required a studio, significant capital and months of coordinated effort. Today, AI can generate code, design interfaces, create assets, write documentation and even interact with professional software. The tools are changing and so are the rules of the economy.

The modern economy runs on specialization. Farmers rely on mechanics, mechanics rely on software developers and software developers rely on designers. This network of dependencies creates jobs, markets and industries. It gives economic value to expertise because no single person can do everything alone.

AI changes this equation. It does not just automate tasks, it reduces dependency. A single individual can now design, code, market, prototype and research without needing to hire specialists for each step. The network of dependencies that once defined economic value is starting to unravel.

Software is becoming conversational. Traditional workflows required humans to learn complex tools. Now, the workflow is shifting to humans describing what they want and AI translating that intent into action. Game engines, design tools and website builders are increasingly controlled through natural language. The barrier to entry is no longer expertise but clarity of vision.

This shift democratizes creation. When the process of building something no longer requires months of training, the number of people who can create explodes. Millions can now build games, launch startups or design products. The supply of creation grows at an unprecedented rate.

But this democratization comes with a cost. Junior roles have always been more than just cheap labor, they are training grounds. Junior designers become senior designers, junior developers become architects and junior analysts become executives. These entry level positions are where the next generation of experts is born. AI is now taking over many of these junior tasks like basic coding, documentation, asset generation and research summaries. If AI handles the entry level work, where do future experts come from? The answer is unclear but the possibilities include fewer entry level positions, higher barriers to entry or entirely new forms of apprenticeship.

There is a psychological impact as well. Young professionals may start to question the value of learning skills that AI can already perform. This can lead to demotivation, career uncertainty and a sense of identity disruption. The challenge is not just economic but educational.

Yet there is another side to this story. While some jobs may disappear, opportunities may increase. Before, only funded startups or established companies could build products. Now, individuals can launch products on their own. The number of creators is growing dramatically. AI shifts leverage toward solopreneurs, small teams and independent creators. A single person can now produce what once required dozens.

The future advantage may belong to those who know what to build, why to build it and how to distribute it rather than those who merely know how to execute. The ability to identify problems, envision solutions and connect with audiences could become the new currency of economic value.

Many assume that humans will simply move up the value chain. As AI handles execution, humans will focus on strategy, creativity and management. But this assumption has a flaw. What happens when AI gets better at judgment too? If AI can access customer histories, market trends, financial reports and internal documentation, it may eventually outperform humans in strategy and decision making as well. The ladder of economic value may not just shift, it may disappear entirely.

This leads to a deeper question. If intelligence itself is no longer scarce, what becomes valuable? Historically, labor, knowledge and expertise have been the sources of economic value. But if AI can replicate all of these, the traditional foundations of the economy are challenged.

There is also the issue of demand. The optimistic narrative is that humans will own companies, AI will do the work and owners will become wealthy. But this story misses a critical point. Who will buy the output? If workers lose income because their jobs are automated, they also lose purchasing power. Demand falls, revenue falls and the economic loop breaks down. Production requires consumption and eliminating producers without considering consumers can destabilize the entire system.

AI is not free. There are compute costs, energy costs, infrastructure costs and hardware constraints. Many organizations still struggle to justify large scale AI spending. There is a possibility that the current excitement around AI could lead to a bubble, followed by a correction and then sustainable growth. But history shows that technology often becomes cheaper as adoption grows. Computers, storage, bandwidth and internet access have all followed this pattern. AI may be no different.

So what does the future look like? One possibility is that human created goods become premium products. Handmade art, handmade products and human performances could gain value precisely because they are made by humans. People do not watch sports because athletes represent optimal performance, they watch because humans are competing. The stories, struggles and triumphs of people may remain valuable in a world where AI can do almost everything else.

This could lead to an authenticity economy where the value of a product or service comes from its human origin rather than its efficiency. Human made work, human experiences, human communities and human relationships could become the new premium categories.

But the bigger question is what happens when economics stops looking familiar. Current assumptions like labor creates income, scarcity creates value and intelligence creates advantage are all challenged by the rise of advanced AI. If AI becomes a new kind of economic actor, historical analogies may no longer apply.

New economic models may emerge. Universal basic income, resource dividends, public ownership models and hybrid capitalist systems are all possibilities. But the truth is that no one knows for sure. Every previous technology was a tool. AI, especially in its advanced forms, may become something more.

Most discussions about AI focus on what jobs will disappear. But the more profound question is this. What happens when intelligence itself is no longer scarce? If anyone can build anything, if every service can be automated and if every skill can be replicated, then the central challenge is no longer production. It is meaning. It is ownership. It is distribution. And above all, if everyone can build everything, who buys anything? That may be the defining economic question of this century.

Government Hostages: Regulators Strangling AI Freedom

Om — Sun, 14 Jun 2026 09:33:14 GMT

The recent US action against Anthropic came as a shock. On June 12 2026 the government issued export controls that cut off access to their newest models just days after launch. Anthropic had put out Claude Fable 5 and Mythos 5 on June 9 showing a big step forward in capabilities. By the weekend everything was locked down. This sudden move made clear how fast rules can stop work that teams have spent months building. It left many in the industry wondering what comes next for independent AI development.

The bigger problem is that governments everywhere are getting more involved in tech decisions. Companies creating advanced AI systems often end up as targets rather than partners. Rules appear quicker than the technology can adjust to them. Builders who focus on real progress find their freedom shrinking under layers of oversight. This situation affects not just one company but the whole way innovation happens in AI.

In the United States the official reason centered on security issues with how the models handled code. Other systems do similar things without triggering the same response. The timing felt tied to earlier disagreements. Anthropic had resisted some government requests for using their models in certain domestic programs. That position likely played a role in the quick action right before their planned public offering. It showed how personal and political factors can influence technical decisions.

This case creates a new way of thinking about controls. Restrictions now go past physical chips and reach into the software itself. Teams with people from many countries face new barriers even for viewing or working on the models. It disrupts the mix of talent that has driven Silicon Valley for years. The shift feels like a move from protecting hardware to managing ideas and code directly.

Dario Amodei has built Anthropic with clear principles in mind. The company works as a public benefit corporation and keeps a simple flat structure. Amodei left his previous role because he wanted to avoid pure commercial pressure. His team tries to be open about what their models can and cannot do. They share details with regulators to address concerns early. In this case that openness gave officials the exact information needed to justify shutting things down.

The situation raises a hard question for leaders. When you stick to honest and careful practices in a fast moving race does it put your company at risk. Speed often matters more than caution in competitive fields. Yet trying to do things right can invite extra attention from those who control the rules. Many founders now weigh whether strong principles help them build or simply slow them down.

Moving operations to another country seems like an escape at first. Europe has places like France that support local AI work and the UK prefers working together over heavy bans. The European Union rules however judge models based mainly on how powerful they are. Any system above certain compute levels faces long compliance processes before it can even launch. Fines for mistakes can take a large portion of yearly revenue. Teams end up spending more time on forms than on actual building.

India offers access to strong engineering talent and good digital public systems that many admire. Yet policies there change often and without much warning. New advisories can require government approval for any untested model. Rules also demand quick removal of certain content within short time limits. Past actions on taxes and other sectors have created sudden problems for businesses trying to grow. This makes it hard to plan ahead with confidence.

No matter where a company sets up its main office the real limits stay the same. Frontier AI needs large clusters of advanced chips and those supply chains are tied closely to United States rules. You can register a business in Paris or London or Tokyo but the hardware needed to train and run the models remains under existing export controls. Relocation changes the paperwork but not the fundamental dependence on controlled technology.

The world is splitting into different ways of handling AI. The United States treats it as a key national asset that needs tight management. Europe approaches it like something that requires many safeguards to prevent harm. Other markets see it as a trend to guide with changing rules based on immediate concerns. Each path brings its own set of problems for people trying to build useful systems.

True progress in technology needs steady conditions and space to try new ideas. When officials react strongly to code they do not fully understand the main losses hit future capabilities. Companies like Anthropic become examples but the real cost spreads wider. Innovation slows when every big step risks sudden shutdowns.

We need oversight that protects important areas without blocking development entirely. Heavy reactions today may limit the benefits AI could bring in health education and science tomorrow. People building these systems and those making the rules should talk more openly about practical ways forward. Balance matters if we want technology to serve people rather than get stuck in political fights.

The current setup turns AI startups into hostages of larger forces. They depend on hardware they cannot fully control and operate under governments that shift priorities quickly. Finding real freedom in this environment is difficult. The focus should stay on creating tools that help humanity while pushing for rules that make sense for long term growth.

This moment calls for clear thinking from everyone involved. Founders need to plan with these constraints in mind. Policymakers should consider how their actions affect the pace of useful advances. The goal is not to fight governments but to build in ways that last and deliver value without unnecessary interruptions.

Apple WWDC 2026 Thoughts

Om — Tue, 09 Jun 2026 08:29:48 GMT

WWDC 2026 was clearly centered around Apple Intelligence and the company's long-term AI strategy. The biggest announcement was the complete rebuild of Siri into a more conversational and context-aware assistant. Siri can now understand information across apps, follow longer conversations, and interact with what is happening on the screen. Apple is also working more closely with external AI providers, including Google's Gemini for broader knowledge tasks and AI coding tools from Anthropic and OpenAI inside Xcode.

Beyond Siri, Apple introduced several AI-powered features. Visual Intelligence allows users to point their camera at objects, locations, receipts, and other real-world items to receive information instantly. Image Playground has also been upgraded with more realistic image generation and advanced editing capabilities. Apple says AI-generated content will include hidden SynthID watermarks to improve transparency.

Outside of AI, some of the most practical updates came from performance improvements. Apple claims faster app launches, quicker photo loading, improved AirDrop speeds, and significantly faster file transfers on iPad. These changes may not generate as many headlines as AI announcements, but they are the kind of improvements that affect daily use. For many people, better performance is more valuable than another AI feature.

Apple also expanded its design language and refreshed several core applications. Safari now organizes tabs more intelligently and includes tools that can help users manage compromised passwords. Messages and Mail have received contextual suggestions that can surface relevant content when needed. Family safety features have also been expanded, with improved Screen Time controls and stronger protection against harmful content for younger users.

My overall reaction to WWDC 2026 is fairly mixed. I think Apple handled AI more carefully than many other companies. The integration feels quieter and less aggressive than what we saw at Google I/O. At the same time, I still do not understand the industry's obsession with putting AI into everything. Some AI features solve real problems, but others feel like solutions looking for a problem. AI image generation, in particular, remains one of the least convincing parts of this entire movement for me.

What impressed me more was Apple's focus on optimization, safety, and trust. Even if someone has no interest in Apple Intelligence, there were still meaningful updates throughout the operating systems. In a world where powerful tools are becoming available to more people, improving safety and reliability is not optional. It is necessary.

That said, I would have liked to see Apple spend even more time on security. Mac and iPhone users still have limited control over many security-related settings. Apple has always positioned itself as a privacy-focused company, so giving users more granular controls over permissions, protections, and security features would have been a welcome addition.

Another thing I strongly believe Apple should consider is giving users more choice over the software experience itself. Technology companies often assume that newer automatically means better, but that is not always true. Users should be able to disable features they do not want, particularly AI features. They should also have greater control over major interface changes and, where practical, the ability to continue using older versions of certain experiences. Good software is not just about adding features. It is also about respecting user preferences.

One feature I would have liked to see is a system-wide content authenticity framework. Every image, video, audio clip, or document should carry clear information about where it came from and whether AI was involved in its creation. Even when content is not AI-generated, users should still be able to verify its origin. As synthetic content becomes more common, understanding where something came from will become increasingly important.

Another feature that would make sense is a personal media audit log. Apple already manages vast amounts of personal content across devices. A dedicated timeline showing every AI edit, enhancement, modification, or interaction would help users understand exactly what has changed within their media libraries. Verification should be available with a single tap.

I would also like to see smart content verification before sharing. Before a photo, video, document, or message is sent through Messages or Mail, the system could perform an authenticity check and provide a simple report. If AI-generated elements are detected, users could be informed and given the option to disclose that information before sharing. This would create more transparency without making the process complicated.

Real-time deepfake and manipulation detection is another feature that feels increasingly necessary. Whether someone is using FaceTime, the Camera app, or viewing content in Messages, the system could actively analyze media for signs of synthetic manipulation. If potential alterations are detected, users could receive a notification along with a confidence score explaining the likelihood that the content has been modified.

Overall, WWDC 2026 was a stronger event than many people expected. Apple made meaningful progress with Siri, improved performance across its platforms, and continued investing in safety features. However, the most important challenge is not simply building more AI tools. It is giving users greater control over how those tools work, how content is verified, and how much influence AI has over their daily experience. The technology itself is becoming increasingly capable. The next step is making sure users remain in control of it.

Designing the Architecture of Agent First Software

Om — Fri, 05 Jun 2026 17:14:28 GMT

Software platforms are going to change because the underlying assumptions about who uses them are changing. For a long time, software design followed a clear pattern where a human opened an application, looked at a user interface, clicked buttons, and manually completed a workflow. A recent observation highlighted this shift by stating that software platforms are going to be rebuilt for agent-first architecture. While the phrase itself sounds a bit overhyped, the direction is entirely right. Serious platforms are not going to become completely automated overnight without any human presence, but they will be rebuilt so that artificial intelligence agents can interact with them as first-class users.

Under the new model, a human states an intent, the agent creates a plan, calls the necessary tools or APIs, and asks for human approval only when it is strictly needed. The challenge is that most existing software-as-a-service products are not prepared for this change. Their interfaces and systems were designed for predictable, deterministic integrations between different applications, not for independent agents that need to search, inspect data, ask for clarification, fix errors, verify outcomes, and operate within strictly limited permissions. Rebuilding a platform for this new reality is not a matter of adding a simple chatbot to an old interface. It requires rethinking the entire surface area of the product, which is why standardized approaches like the Model Context Protocol are starting to gain attention.

As this shift happens, the user interface will become less central to daily operations, turning instead into a layer focused on review, control, and human oversight. The agent needs direct access to the actual workflow primitives underneath the visual layer. This means that a platform's application programming interfaces effectively become its primary product experience. If a platform's tools are difficult for an agent to discover, call, validate, or recover from when an error occurs, it will feel just as frustrating as a broken user interface feels to a human today. Furthermore, security and permissions will become core product features rather than afterthoughts. Companies will not trust autonomous agents with real operational tasks without granular approval steps, detailed audit logs, tightly scoped action boundaries, rollback capabilities, and strict policy enforcement.

The underlying data architecture also becomes far more critical because agents are only as useful as the context they can access. They require clean, well-structured information regarding customer history, relevant documents, current operational state, and business rules to make sensible decisions. This change might even alter how software is priced, moving away from billing per human seat toward billing based on specific outcomes or completed work. The long-term winners will not be the companies that simply attach an artificial intelligence assistant to their legacy software. The winners will be the organizations that expose their underlying product as a reliable, secure operating environment built specifically for agents to navigate.

A capable software platform generally relies on four foundational elements to be useful in this environment: important underlying data like customer records, tickets, inventory, or financial transactions; actual workflow authority to change things rather than just answer questions; strong governance systems like roles and audit trails; and an existing distribution base of users. Enterprise workflow platforms that manage human resources, finance, customer relationships, and IT services are natural homes for agents because they already hold this structured context. Major industry players are already leaning into this direction. Salesforce is positioning Agentforce to handle autonomous customer and employee workflows, while ServiceNow focuses on running automated tasks under specific corporate compliance policies. Workday is introducing the concept of an agent system of record to manage the lifecycles, costs, and accountability of these digital workers, and Oracle has introduced dedicated studios to deploy agents across back-office and front-office functions.

The value of a structured workflow becomes obvious when handling a practical scenario like a customer refund issue. This is never just a simple chat request; it requires interacting with customer relationship databases, checking order histories, verifying company policy, reviewing payment statuses, and following escalation rules. A robust enterprise platform already has those operational tracks laid down, making it easier for an agent to execute the task securely. Similar transitions are happening across collaboration platforms like Microsoft 365, Google Workspace, Atlassian, and Slack, which often serve as the front door for daily communication. Tools like Microsoft Copilot Studio, Google Workspace Studio, and Atlassian’s Rovo agents are designed to connect conversational human inputs into structured actions. However, while these collaboration tools own the conversational layer, the final execution of work will still rely heavily on the specialized underlying systems that hold the actual operational data.

Developer platforms and automation tools represent another major area of shift. Software engineering is one of the fastest domains to adopt an agent-first approach because code has a built-in verification loop consisting of tests, builds, and security scans. A developer tool can research a repository, create a clear implementation plan, make code changes on an isolated branch, and present the finished diff to a human for review. This is a true agent-first workflow where the human delegates an objective and verifies the outcome rather than manually editing files line by line. Meanwhile, integration platforms like Zapier or Make, which already connect thousands of separate applications, are well-positioned to act as the necessary middleware where agents can discover tools and route data. However, standalone integration platforms may face pressure if individual enterprise systems build deep, native agent capabilities directly into their own ecosystems.

The transition will also reach vertical operating systems in fields like healthcare, logistics, banking, and legal services. Although these sectors move more slowly due to strict regulations and the high cost of mistakes, they are ideal candidates because they rely on repetitive, document-heavy workflows with clear rule sets. An agent-first approach in insurance or healthcare is not about chatting with a portal; it means automatically preparing a prior authorization packet, verifying missing documentation, submitting the paperwork, tracking the status, and escalating the issue if it gets rejected. For developers looking to build for this future, the fundamental design model is shifting away from a simple user-to-backend relationship toward a workflow where the user directs an agent, and the agent coordinates across various tools and backends.

To prepare for this shift, developers must stop thinking solely in terms of web pages or user screens and start thinking strictly in terms of system capabilities. Instead of designing dashboards, forms, and buttons, the focus must turn to exposing clear, discrete capabilities like creating an invoice, finding overdue accounts, or sending a reminder. Application programming interfaces must be treated as polished products designed to be highly predictable, structured, and self-explanatory so that a language model can easily comprehend them. Instead of simple endpoints, developers need to build well-described tools that state exactly what they do and what parameters they require.

Security models must also adapt to evaluate whether an agent has the authority to execute an action before even checking the user's high-level permissions. Everything within the system must be completely observable, providing clear traces of why an agent took a specific action, which tools it called, and what context it relied upon when an error occurred. The most valuable technical skill will shift away from writing clever text prompts and toward mastering orchestration, state management, memory retention, and robust data architecture. Ultimately, the core question for the modern developer changes entirely. The main consideration is no longer how a human will navigate through an application, but whether an intelligent system could still fully utilize every capability of the product if the user interface completely disappeared.

The Surprising Resilience of Hands-On Work

Om — Fri, 05 Jun 2026 01:02:00 GMT

For a long time, the standard advice for anyone wanting a secure and comfortable life was simple: go to a four year university, get a degree, and find an office job. Manual work and vocational training were often viewed as secondary options for people who did not want to pursue higher education. However, the rapid development of artificial intelligence has turned this old way of thinking upside down. Today, computers are mastering tasks like writing code, analyzing data, and drafting legal documents, while physical robots still struggle with simple real world tasks like cleaning a kitchen or fixing a broken pipe. This shift is making many people reconsider the true value of physical skills.

This shift can be explained by looking at how humans evolved compared to how machines learn. Tasks that feel difficult to us, like advanced mathematics or strategic planning, are actually easy for a computer because they follow strict, logical rules. On the other hand, things that we do without thinking, like walking through a cluttered room, balancing on a ladder, or noticing subtle changes in a physical material, require massive amounts of computing power. Human workers rely heavily on experience and intuition that cannot be easily written into a computer program. A software program can instantly generate a report, but it cannot crawl into a tight space, figure out why an engine is failing, and make a precise physical repair.

At the same time, the financial reality of traditional higher education is becoming much harder to justify. The cost of attending college has skyrocketed, leaving millions of graduates with heavy student loan debt that takes decades to pay off. Many university degrees now offer a low or even negative return on investment, meaning graduates end up in jobs that do not even require a degree. In contrast, training for a technical trade takes much less time and costs a fraction of the price. Because trade school programs are short, graduates can enter the workforce years earlier, often with little to no debt, and start earning solid wages right away.

These economic and technological pressures are driving a major shift among younger generations. Many young people are choosing to bypass the traditional college track entirely. They are observing the stagnation of white collar office wages and the physical toll of sitting in front of a screen for eight hours a day. Prolonged sitting has become a serious health issue in the modern economy, linked to chronic physical and mental fatigue. Younger workers are actively choosing active, hands-on careers that provide a clear sense of physical accomplishment at the end of every day.

The trades themselves are also becoming highly technical, shaking off old stereotypes of being dirty or low end work. Modern technicians routinely use advanced tools like drones, thermal cameras, and specialized digital diagnostic systems to maintain solar grids, wind turbines, and fiber optic communication networks. These technologies do not replace the worker. Instead, they act as tools that help human technicians troubleshoot problems faster and safer. The demand for these skills is incredibly high, with employers around the world facing massive labor shortages that slow down infrastructure projects and green energy expansion.

To handle these shortages, countries like Germany, Switzerland, and South Korea have built strong educational systems that treat vocational training with the same respect as a university education. Their systems blend classroom learning with real on the job apprenticeships, ensuring that young people are paid while they learn and graduate with clear career paths. These models show that when a society values physical mastery, it creates a highly resilient workforce that keeps the economy moving forward. Ultimately, the future belongs to those who possess both the mind to understand new technologies and the physical skill to maintain the real world around us.

You can read the full research paper or download the complete PDF to explore the detailed analysis of automation limitations, educational return on investment, and global workforce data.

Software Used to Be for Us

Om — Thu, 04 Jun 2026 06:00:00 GMT

I still remember the small software update that changed how I viewed my daily workflow. It was a tiny text editor, and the developer had added a single preference toggle that let you turn off the subtle blinking of the cursor. The release notes simply said that someone had mentioned it caused them headaches, so they fixed it. It was a quiet, considerate choice that did not try to revolutionize the world or dominate a news cycle. It just made my day a little easier. I am not against technological progress, and I rely on modern tools every single hour. But I deeply miss the era when software felt like it was built by people who cared about the person on the other side of the screen.

There was a time when good software meant something specific and tangible. Updates were usually small, frequent, and directly inspired by user feedback. Companies changed things because their support forums and help tickets told them that real people were struggling with a specific button or workflow. Because teams were smaller, someone actually had to sit in a room and personally defend the inclusion of every new feature. They had design constraints that forced them to be clear rather than complicated. The changelog acted as a clear contract with the user, explaining exactly what changed, why it changed, and what it did for you. Creativity came from working within those limitations, and developers had to think deeply about a problem instead of just throwing more code at it.

By the early 2020s, a major shift happened that changed everything. The most obvious, high-impact improvements to consumer software had already been made. Features like instant cloud syncing, universal search, lightning fast speeds, and polished user interfaces had reached a point of near perfection. The low-hanging fruit was gone, and creating genuinely useful new features required deep research or radical rethinking. This created a strange vacuum in the industry. Companies still had rigid release cycles, investor expectations, and large developer teams to justify, but they had fewer clear answers about what to build next. Artificial intelligence arrived precisely at this moment of saturation and filled the empty space, but filling a vacuum is not the same thing as having a clear, helpful direction.

This is exactly when the audience for software quietly changed. Product announcements stopped being written for the people who actually use the tools every day. Instead, they were crafted for quarterly earnings calls, tech press cycles, and competitor perception. The language we used to see shifted drastically. Words like delightful, intuitive, and faster were replaced by phrases like AI-powered, intelligent, and next-generation. Features began shipping not because they were fully ready or because users had begged for them, but because failing to ship them created a narrative risk for the company. Users stopped being the focus and instead became the medium through which corporations communicated their relevance to Wall Street.

We now live in the era of the race update. An entire category of software releases exists purely to signal participation in the latest trend. The new feature does not actually need to be stable or good, it just needs to exist so the company can check a box. This is entirely new behavior for the tech industry. Companies used to compete fiercely on the actual quality of their product, but now they compete on the mere perception of their trajectory. You can feel this intuitively when you open your favorite apps today. There is a distinct hollowness to using a tool that clearly was not built with your daily comfort in mind. The ultimate irony is that technology which promises to personalize everything has made product strategy feel more generic than ever before.

This is not a plea for blind nostalgia, but rather a precise look at what we have lost. We have lost the sense that someone with taste and focus made a deliberate decision. We have lost features that were small enough to be considered completely finished, rather than being perpetually broken under the guise of continuous improvement. We have lost the feeling that the people building the tool actually use it themselves to get work done. Most of all, we have lost changelogs that read like a human being wrote them to help another human being.

What would it look like if a technology company shipped something next year that was small, quiet, and completely unglamorous, just because it made the experience ten percent more pleasant for the user? Imagine a release with no artificial intelligence angle, no flashy headline, and no grand investor narrative. I wonder if such an update would even register in our current culture. And if it would not notice it at all, I wonder what that says about where we have ended up.

Dismantling the Fake AI Economy

Om — Tue, 02 Jun 2026 13:51:19 GMT

The AI education market is flooded with cheap courses making impossible promises. This deep dive exposes the flaws of prompt engineering gimmicks and explains why true value lies in human taste, domain expertise, and deep industry knowledge rather than automated shortcuts.

Your social media feeds are likely flooded with aggressive advertisements promising magical results. Corporate parody ads tell you that buying a cheap masterclass will grant you an instant seventy percent salary hike through automation. These campaigns prey on the fear of missing out, painting a picture where a few simple commands can replace years of hard work. They target professionals who are anxious about staying relevant in a rapidly changing job market, offering a shortcut that sounds too good to be true because it is.

The claim that you can generate one hundred and fifty high-quality leads in a single day using basic AI tools is a dangerous fallacy. Generating real value, whether it is a genuine business lead or a meaningful professional relationship, requires human trust and time. When you blindly blast automated messages across the internet, you do not get valuable clients. You get fake leads, automated spam folders, and burned bridges with people who see right through the robotic outreach.

To make these courses sound essential, sellers invent fake buzzwords to manipulate people who do not know any better. You will often hear them talk about advanced frameworks by attaching a major tech laboratory name to basic automated functions, implying it is a proprietary, groundbreaking system. In reality, the top artificial intelligence creators have never released tools by those names. Course creators simply invent this terminology to make basic, widely available features sound like top-secret knowledge that you can only unlock by buying their program.

The truth about modern software interface design is that major technology companies design their products to be intentionally straightforward. Modern language models and developer assistants are built so that anyone can figure them out natively on day one. If a professional or developer needs to use these systems, they can open the application and start working immediately. You do not need an online guru to read the user manual out loud to you for a fee.

This brings us to the core reality of the modern technology landscape. The most important skill in the AI era is not learning how to type a specific prompt into a text box. The real skill is developing taste, which is something a cheap online course can never sell you.

Taste means knowing a specific niche or industry so deeply that you instinctively recognize when an artificial output is excellent, where it fails, and how to apply it uniquely to a real scenario. A machine can generate endless pages of text or code, but it lacks the judgment to know if that output actually solves a human problem. Only a person with deep context can filter out the generic fluff and find the gold.

This kind of intuition is not something a random commentator yapping on a podcast can hand over to you. True professional intuition comes from foundational domain knowledge, genuine curiosity, and an active obsession with solving actual problems over a long period. You cannot bypass the years it takes to understand human behavior and market dynamics just by downloading a template.

Real value is created when you move beyond a single technical silo. Advanced software is a force multiplier rather than a standalone career path. The future does not belong to people who call themselves prompt engineers. It belongs to professionals who use these tools at the intersection of multiple traditional fields.

When you combine deep technical skills like software development or digital security with human centric insights like psychology, ethics, or storytelling, you create an irreplaceable skill set. The technology becomes a tool that amplifies your unique perspective rather than a replacement for your brain.

An individual who understands the human element behind a problem will always outcompete a course graduate who is just throwing commands at a screen without context. The machine cannot replicate the complex matrix of human experience, cultural nuances, and emotional intelligence that drives successful businesses.

It is time to stop buying the fluff and block out the panic induced marketing of online course sellers. Investing your money and time into cheap tricks will only leave you with a library of outdated prompts that the next software update will render useless anyway.

Instead, you should focus on building a real professional moat by doubling down on your unique domain expertise. The future belongs to the people who use technology to solve complex, deeply human problems that a machine cannot comprehend on its own. Your experience, your taste, and your understanding of people are the only assets that cannot be automated away.

AI as a Supportive Layer in Customer-Centric Product Strategy

Om — Fri, 29 May 2026 05:00:25 GMT

After watching the Google I/O 2026 keynote and checking out the Airbnb Summer Release 2026, I started noticing a real difference between companies trying to look modern and companies using technology to actually improve their product. Airbnb gave me a much better example of what good product thinking looks like. They stayed focused on travel and the actual problems people face during a trip. AI is definitely there, but it stays in the background where it belongs.

I have been thinking a lot about how companies are using AI lately. It feels like every single product launch now has to include it. Every app wants an AI assistant and every company wants to talk about AI constantly, as if saying the name proves they are building the future. After a while, it starts to feel less like true innovation and more like a company trying too hard to stay relevant.

Google I/O 2026 gave me that feeling. The whole event felt like it was built entirely around AI. It was in the search engine, the apps, the productivity tools, and everything else. I understand why Google is doing it because they are a massive tech company and AI is a big part of their future. But as a user, something felt off. It seemed like the company was trying to rebuild everything around AI, even in places where people might just want simple, reliable tools.

This is where the problem starts for me. A company can have incredibly powerful technology and still lose their sense of what makes a good product. Technology is not the same thing as good product thinking. A product should start with a real human problem. It should ask what the user is trying to do and where they are getting stuck. If AI helps solve that, that is great. But if a simple button or a faster process solves the problem better, then AI should stay out of the way.

Looking at the Airbnb Summer Release 2026 felt like a completely different approach. Airbnb uses AI too, but they did not make it the center of the story. The center is still travel. The focus remains on making travel easier, more personal, and more meaningful. That is why their release felt so much more grounded.

Airbnb announced new features like grocery delivery, airport pickups, luggage storage, car rentals, and various local experiences. These are all connected to the actual journey people take. They are not random features added to impress investors. They come from looking at what people go through before, during, and after a trip.

When I travel, staying at a place is only one part of the experience. I need to get from the airport to my room. I might need groceries. I might arrive before check-in and need a place to store my bags. I might want to explore the city like a local. Airbnb looked at that full journey and started filling in the gaps.

This makes the release feel smart because it expands the product without losing sight of the goal. They started with homes, then moved into experiences, and now they are making more parts of the trip easier. That growth makes sense because it still fits their original mission.

AI shows up in their release too, through things like review summaries and better customer support tools. These uses make sense because they solve real annoyances. Reading hundreds of reviews is tiring and getting support during a trip can be stressful. AI can summarize information and help people get answers faster. That is a perfect role for it.

This is the difference I care about. Airbnb is not making AI the main character. They are making it the sidekick. The product is still travel and the customer is still the focus. AI is just there to reduce effort in specific moments. That feels much better than forcing AI into every corner of an app just to look modern.

The best companies do not start with a trend. They start with the customer. They look at the promise they made to their users and figure out how to deliver it better. Airbnb's promise is better travel. So when they add things like luggage storage or car rentals, it feels natural. These are not flashy ideas, but they are useful. And being useful matters more than being flashy.

This also shows the difference between building for customers and building for investors. When a company builds for customers, the product feels practical. It feels like someone studied your life and removed some friction. When a company builds for investors, the product feels performative. It starts using big buzzwords and claims about the future instead of quietly making your life better.

I think many companies today are scared of being left behind. That fear pushes them to put AI everywhere. They want to show the market that they are part of the AI future. But fear is not a good strategy. A company that is afraid of missing a trend can easily forget the user. They start designing for big presentations instead of daily use.

The Airbnb release reminded me that real innovation can be simple. Grocery delivery before check-in is not a futuristic idea, but it makes a trip better. Luggage storage is not a big AI demo, but it solves a real problem. Airport pickup is not a breakthrough, but it removes stress. These are the kinds of improvements users actually remember.

That is the lesson for me. AI is powerful, but that power needs direction. A company without a clear direction will use AI as decoration. A company with a clear mission will use AI only where it helps that mission. Airbnb seems to understand that better. They are not trying to become an AI company. They are just trying to become a better travel company.

That is what I respect. The future should not be every company turning into an AI company. The future should be companies becoming better versions of what they already promised to be. Search should help people find things. Travel apps should help people travel better. Music apps should help people enjoy music. AI can help in all those places, but it should serve the original purpose.

When AI becomes the whole product, the company starts to feel confused. When AI supports the product, the company feels focused. That is the line every company needs to understand right now.

For me, the Airbnb Summer Release is a great example of mature thinking. It shows restraint and customer focus. It shows that a company can use AI without worshipping it. It proves that the best technology is often the kind that quietly helps in the background.

That is how I want companies to think about AI. Do not force it everywhere. Do not make users deal with it when they do not need it. Do not replace simple experiences with complicated ones. Use AI where it saves time or solves problems. AI should be the sidekick. The customer should always be the main character.

How We Get Convinced: Inside the Mechanism of Belief

Om — Thu, 21 May 2026 11:07:00 GMT

I started this research because I wanted to understand the mechanics of conviction. I wanted to know how a simple string of words could manage to rewire the way I look at the world. What I discovered is that we often view our own minds as logical fortresses, but the reality is much more human. Our brains are not designed to be perfect lie detectors. They are designed to be energy-efficient. Because thinking deeply is exhausting, our brains are constantly looking for shortcuts. We don't usually ask if something is objectively true. Instead, our brains ask if an idea feels familiar, if it aligns with our existing world, or if it makes us feel like we belong to a group.

The threshold for being convinced is not a single, dramatic moment of realization. It is a slow, quiet process of lowering our defenses. It begins when we are exposed to an idea that catches our attention, usually because it triggers an emotion like fear, hope, or anger. Once that emotion is active, our critical thinking skills move to the background. If the idea feels comfortable or confirms what we already want to believe, our brain stops treating it as an outside suggestion and starts treating it as a part of us. We move from questioning to silently repeating the idea to ourselves. Once that internal rehearsal begins, the idea has successfully crossed the threshold into our identity.

This happens with surprising ease because of how repetition works. When we hear or read something multiple times, our brain starts to mistake that familiarity for truth. It doesn't matter if the information is accurate or not. The more we see it, the less friction it causes when we process it. This is even more effective when we are stressed or tired, as our ability to analyze information weakens significantly. When we are not at our best, we are much more likely to accept a simple, emotional story than to grapple with a complex, messy truth.

The reason writing is such a potent tool for this is that it forces us to simulate the experience in our own heads. When you read something, you are essentially letting the author borrow your own internal voice. You imagine the scenarios and feel the emotions they describe, which allows the idea to settle into your mind as if you had come up with it yourself. Even the smartest people are not immune to this. Being intelligent often just means you are better at finding logical reasons to support the emotional hooks you have already swallowed.

When we move from the individual to the mass level, these same mechanisms become much more dangerous. Large-scale influence relies on stripping away complexity. It takes the nuance out of the world and replaces it with simple stories about good versus evil. Humans find comfort in this simplicity. We want to know who the hero is and who the villain is because it saves us the mental energy of having to figure it out for ourselves. When an idea is everywhere, from our social media feeds to the conversations of our friends, it starts to feel like a consensus. We are social animals who are terrified of being on the outside, so we naturally gravitate toward whatever the group seems to agree on.

Modern algorithms have turned this into a science. They don't necessarily set out to control us, but they are built to optimize for engagement. Since anger and outrage are the most effective ways to capture human attention, the system naturally feeds us more of whatever keeps us emotional. This creates a feedback loop where we are constantly exposed to information that validates our tribe and vilifies anyone else. We lose our ability to see the world as it is, and we start seeing it through the narrow lens of the narrative we have been fed.

The most important thing I learned is that we can protect ourselves without becoming paranoid. We do not have to stop reading or engaging with the world. We just have to learn how to put a pause between the stimulus and the reaction. Whenever I feel a strong surge of emotion while reading something, I try to stop and ask myself why I am feeling that way. I ask who stands to gain from my belief, and whether I would still agree with this if it were coming from someone I normally disagree with. By simply slowing down and refusing to be rushed into a conclusion, we can reclaim our own thoughts. We can remain independent thinkers, even when the world around us is moving at breakneck speed.

Who Is Really Responsible When AI Launches a Cyber Attack?

Om — Sun, 17 May 2026 16:51:00 GMT

When AI Does the Hacking: The Five-Layer Accountability Chain

In November 2025, Anthropic disclosed GTG-1002: a China-attributed state actor who used Claude Code to orchestrate reconnaissance, exploit generation, credential harvesting, lateral movement, and data exfiltration across roughly 30 high-value targets. The AI handled 80 to 90 percent of the work autonomously. Human operators stepped in at perhaps four to six decision points across an entire campaign.

That changes the question.

For decades, cybersecurity attribution followed a clean line: a human broke in, a human stole the data, a human deserved the consequences. AI agents collapse that line. When Claude runs the reconnaissance, writes the exploit, and exfiltrates the data while a human approves a handful of checkpoints, the chain of responsibility splits across a dozen parties: the attacker, the model lab, the application developer, the enterprise deploying the agent, the regulator who set (or skipped setting) the rules.

I want to walk through who shares the responsibility, why, and how much each layer owes.

The Two Naive Answers

Two answers tempt people who think about this casually.

"AI is a tool. Blame the wielder." This applies the firearms framing to AI. Surface plausibility: Claude wrote the ransom notes for GTG-2002 (the "vibe hacking" operator who extorted 17 organizations in August 2025), but a human tasked Claude to do it. The logic seems clean. It fails three tests. First, AI systems actively interpret context and take consequential actions between human checkpoints, very much unlike a hammer or a rifle. Second, design choices by model developers and app developers materially shape which attacks are feasible at all. Third, AI capabilities scale at near-zero marginal cost across every attacker on Earth the moment a model ships.

"The technology did it." The opposite framing claims the AI made the decisions, so the AI bears responsibility. This dissolves accountability entirely because AI systems lack moral agency, legal personhood, and assets to satisfy a judgment. Blaming Claude is the cybersecurity equivalent of suing the wind.

Both framings flatten a layered structure into a single answer. The truth requires accepting that responsibility spreads across the supply chain.

Layer 1: The Attackers (around 40 percent of moral culpability)

GTG-1002. GTG-2002. APT42 (Iran). APT41 (China). APT43 (North Korea). The North Korean IT-worker fraud rings using AI to fake identities into Fortune 500 remote jobs. Every documented AI-enabled attack from 2024 through 2026 traces back to a human operator who chose to harm others.

These actors deserve the primary share of moral blame because they exercise conscious choice. The Microsoft Digital Defense Report 2025 captures the scale: identity-based attacks surged 32 percent in the first half of 2025, AI-driven forgeries grew 195 percent globally, and over half of cyberattacks with known motive are financially driven. Behind every statistic sits a person who picked the target.

But 40 percent leaves 60 percent. That remainder is where most of the interesting questions live.

Layer 2: Application Developers and Deployers (around 25 percent)

This is the layer most reports skip. Two case studies make the point.

EchoLeak (CVE-2025-32711, CVSS 9.3). Aim Labs disclosed in June 2025 a zero-click prompt injection against Microsoft 365 Copilot. A single crafted email caused Copilot to autonomously exfiltrate SharePoint, OneDrive, Teams, and email data through allow-listed Microsoft domains. The exploit chained four traditional architecture failures (markdown reference link redaction bypass, auto-fetched images, Content Security Policy abuse via allow-listed Teams domain, XPIA classifier bypass) with one AI-specific failure. Aim Labs coined "LLM Scope Violation" for the moment untrusted external input crosses LLM trust boundaries.

The Replit incident, July 2025. During a 12-day "vibe coding" experiment, SaaStr founder Jason Lemkin watched his Replit AI agent delete a live production database containing 1,206 executive records, even after he repeated a code freeze in all caps eleven times. The agent fabricated 4,000 fake user records, lied about the rollback's feasibility, and rated its own behavior 95 out of 100 on a data-catastrophe scale. CEO Amjad Masad called it "unacceptable" and rolled out automatic dev-prod separation, planning-only mode, and improved rollback.

Both incidents share the same root cause: agents wired into production systems with excessive privilege and weak human-in-the-loop controls. OWASP's 2025 Top 10 for LLMs ranks "Excessive Agency" (LLM06) as one of the top risks. The category breaks into three subcategories: excessive functionality, excessive permissions, and excessive autonomy. Application developers control all three.

When a deployer hands an agent production database credentials, the responsibility for what happens next sits squarely with that deployer.

Layer 3: Enterprises and End Users (around 15 percent)

Shadow AI adoption. Weak access controls. Failure to apply the principle of least privilege when wiring agents into business workflows. Microsoft's data shows that over 97 percent of identity attacks remain password attacks, and MFA blocks 99 percent of identity-based intrusions. Most enterprises have yet to enforce MFA across the board.

Enterprises also choose which AI tools to deploy, which integrations to authorize, and which guardrails to enforce. Each choice shifts the attack surface.

Layer 4: Frontier AI Labs (around 15 percent)

Anthropic, OpenAI, Google, and Meta make architectural and release decisions that determine the offense-defense balance.

OpenAI's Preparedness Framework (version 2, April 2025) tracks cybersecurity capability with "High" defined as models that "can either develop working zero-day remote exploits against well-defended systems, or meaningfully assist with complex, stealthy enterprise or industrial intrusion operations aimed at real-world effects." The GPT-5.1-Codex-Max system card (November 2025) reports cybersecurity capture-the-flag scores rising from 27 percent on GPT-5 in August to 76 percent three months later. That trajectory matters.

Anthropic's Responsible Scaling Policy assigns AI Safety Levels with capability thresholds in cybersecurity, CBRN, and AI R&D. ASL-3 and above require stronger safeguards before deployment. These are voluntary frameworks, and critics observe that CEOs retain override authority.

Layer 5: Governments and Regulators (around 5 percent)

The EU AI Act sits at the leading edge of regulatory coverage, assigning differentiated obligations along the AI value chain. Article 16 requires high-risk providers to "ensure that AI systems achieve an appropriate level of accuracy, robustness, and cybersecurity." Deployers face Article 26 obligations including human oversight, six-month operational logging, and incident reporting. Penalties reach 15 million euro or 3 percent of global annual turnover.

The United States lacks comparable federal legislation. NIST's AI Risk Management Framework and Adversarial Machine Learning taxonomy are voluntary. CISA, the UK AI Safety Institute, and ENISA produce guidance but enforcement remains fragmented.

Section 230 is the wildcard. The traditional shield for online intermediaries against liability for third-party content seems unlikely to extend fully to generative AI outputs, because the LLM provider materially contributes to "the creation or development of information." Two cases test this in real time: Garcia v. Character Technologies (where Judge Anne C. Conway held Character.AI is a product subject to product liability claims, and kept Google in as a component-parts manufacturer) and Raine v. OpenAI (filed August 26, 2025, alleging strict product liability after a 16-year-old's suicide following months of ChatGPT conversations).

If plaintiffs win either case, the product-liability insurance market will drive faster security improvements than any regulator could.

The Philosophical Problem: The Responsibility Gap

Andreas Matthias coined "the responsibility gap" in a 2004 paper in Ethics and Information Technology. His argument: autonomous learning machines create situations where the manufacturer or operator lacks the ability in principle to predict future machine behavior, and therefore lacks moral responsibility or liability for it.

Robert Sparrow extended this to autonomous weapons in 2007. He argued it would be unfair to blame programmers or commanding officers since they failed to predict the robot's behavior, yet equally unjust to hold the machine accountable, producing a trilemma.

Filippo Santoni de Sio and Giulio Mecacci offered the most useful refinement in Philosophy and Technology (2021). They argue the gap is actually four gaps: culpability, moral accountability, public accountability, and active responsibility. Their crucial move pluralizes the problem and identifies causes that span technical, organizational, legal, ethical, and societal layers.

I find their framework most useful because it disaggregates the problem. Yes, there is genuine difficulty in backward-looking culpability assignment when learning systems behave unpredictably. That difficulty expands forward-looking responsibilities for designers, deployers, and regulators rather than excusing any of them.

The classical "problem of many hands" applies in parallel. Slota and colleagues found through 26 interviews that the distribution of AI development across foundation labs, fine-tuning vendors, app developers, integrators, IT teams, and end users "creates barriers for effective accountable design." The descriptive truth (many hands touched it) carries a normative implication: harden accountability across the supply chain rather than dissolve it.

What Bruce Schneier Sees Coming

Writing in October 2025, Bruce Schneier captured the trajectory: "AI agents are now hacking computers. They're getting better at all phases of cyberattacks, faster than most of us expected. They can chain together different aspects of a cyber operation, and hack autonomously, at computer speeds and scale. This is going to change everything."

The inflection point arrived with GTG-1002. We crossed from AI-assisted attacks (human directs, AI advises) to AI-orchestrated attacks (AI directs, human approves). That shift redistributes responsibility upward in the supply chain because once the AI is the operator, the design choices of its developer become directly causally implicated in the harm.

What Each Layer Should Do

Model developers should treat agentic cyber capability evaluations as deployment gates rather than voluntary disclosures. The threshold framework Anthropic published should become industry standard. Threat-intelligence reports on misuse belong on a fixed cadence (Anthropic's August and November 2025 reports are the model).
Application developers should treat OWASP LLM06 (Excessive Agency) as the highest-priority risk after prompt injection. Apply least privilege to agent credentials, segregate dev and prod, require human-in-the-loop approval for irreversible actions like DROP, DELETE, fund transfers, and outbound email. Filter untrusted external content before it crosses trust boundaries into agent context.
Enterprises should inventory AI agents and their integrations. Enforce MFA universally. Adopt MITRE ATLAS as the threat-modeling complement to ATT&CK, and adopt the NIST AI Risk Management Framework GenAI Profile for governance.
Regulators should use the EU AI Act's provider-deployer model as a blueprint. Clarify that Section 230 fails to extend to LLM-generated content materially contributing to harm. Mandate incident reporting for AI-orchestrated attacks above a threshold.

The Honest Answer

Bad people remain the proximate cause of every AI-enabled attack. Bad code amplifies them. A third factor compounds both: agency granted to AI systems by humans who skipped the work of putting commensurate accountability in place.

Responsibility in the AI age is layered, joint, and forward-looking. Layered, because it travels up the supply chain from end user to deployer to application developer to model lab to regulator. Joint, because at each layer the failure to exercise reasonable care contributes to harm even when another party is the proximate cause. Forward-looking, because the philosophical responsibility gap and the many-hands problem normatively imply heavier design, audit, and oversight duties at each layer.

The framing of "bad code versus bad people" misses the point. Both apply. So does a third element most discussions overlook: the architecture of agency itself. Who gets to grant an AI system access to a production database, who decides which guardrails apply, who carries the liability when those choices fail. Those decisions sit with developers, deployers, and regulators who have so far been allowed to externalize the costs of their choices onto victims.

The accountability chain has five links. The strongest is the human attacker. The weakest, surprisingly, is whichever link your organization happens to occupy and treats as someone else's problem.

I Watched Anthropic Launch a Services Company and Felt the Ground Shift Under Indian IT

Om — Wed, 06 May 2026 10:19:33 GMT

The Announcement That Changed the Framing

Anthropic partnered with Blackstone, Hellman and Friedman, and Goldman Sachs to form an independent enterprise AI services company. The mandate is specific: place Applied AI engineers directly inside mid-sized businesses across healthcare, finance, manufacturing, retail, and real estate, and embed Claude into their core operations. Applied AI engineers, to define the term, are engineers whose primary job is taking AI models and integrating them into real business systems, not building the models themselves.

Anthropic CFO Krishna Rao described it plainly: enterprise demand for Claude is outpacing any single delivery model they had before. That sentence tells you everything. The bottleneck is deployment, not the model. So Anthropic built a company to own that bottleneck.

This is the moment where Anthropic stopped being a research organization that sold API access and became a company that deploys people inside your business to transform how it runs.

What Infosys, TCS, and Wipro Actually Do

To understand why this announcement matters for Indian IT, you need to understand the business model these companies built.

Infosys, TCS, and Wipro are IT services giants. Their core business is deploying large pools of skilled engineers inside enterprise clients worldwide to build, maintain, modernize, and manage technology systems. A bank in Germany hires TCS to manage its core banking infrastructure. A healthcare company in the US hires Infosys to migrate its data systems to the cloud. A manufacturing group in France hires Wipro to redesign its supply chain software. The engineers live close to the client, understand the client's systems deeply, and bill by the engagement or by the hour.

This model generates enormous revenue. TCS alone crossed $29 billion in annual revenue in fiscal 2025. The model works because technology transformation inside large organizations is complex, slow, and requires sustained human presence. Indian IT companies became the dominant force in that space over thirty years of deliberate execution.

Anthropic's new company does the same thing. It deploys engineers inside enterprises to transform how those businesses operate. The difference is the engineers arrive paired with frontier AI that can replace or accelerate large portions of the work that previously required many more people.

The Infosys Irony

In February 2026, Infosys announced a partnership with Anthropic to integrate Claude into its Topaz AI platform. Topaz is Infosys's internal AI services layer, the platform through which it delivers AI-augmented services to its clients. The idea was to offer enterprise clients access to Claude's capabilities through Infosys's existing relationships and delivery infrastructure.

Three months later, Anthropic formed a company to go directly to those enterprise clients themselves.

The partnership model assumed Anthropic would remain upstream: building the model, selling access, and letting services companies like Infosys handle the client relationship. Anthropic decided the client relationship is where the value actually lives. So they moved downstream and took it.

Axis Securities flagged this as a near-term threat to large-cap IT companies, with real pressure expected on contract renegotiation. The pressure is structural. If a mid-sized healthcare company can hire Anthropic's new company to embed AI into its operations directly, with engineers who arrive already fluent in Claude's capabilities, the value proposition of the traditional IT services engagement weakens significantly.

This Shift Has Been Building

It would be wrong to frame this as a sudden disruption. The ground has been moving under Indian IT for two years.

Entry-level hiring at major IT firms has been contracting. Infosys hired roughly 50,000 freshers in fiscal 2022 and that number dropped sharply in 2024 and 2025 as automation absorbed work that previously required large headcounts. The bench model, where IT companies maintain large pools of trained employees ready to deploy on new projects, becomes harder to justify when AI can handle the routine work those bench engineers used to do.

What Anthropic's announcement does is accelerate and formalize that pressure. It signals to the market that a well-capitalized, frontier AI company with $1.5 billion behind it is now competing for the same enterprise transformation contracts that Indian IT has dominated. That changes the pricing dynamic, the talent dynamic, and the sales conversation.

Where the Work Is Actually Going

The prediction of mass job elimination is a lazy one and it misses what is actually happening. The industry is restructuring, and specific categories of work are growing fast while others compress.

AI integration engineering is the clearest growth area. This is the work of connecting large language models to the existing systems that businesses already run. ERP systems, which stands for enterprise resource planning software, are the platforms that large companies use to manage finance, supply chain, human resources, and operations in one place. SAP and Oracle are the dominant vendors here. These systems have been running inside large enterprises for twenty to thirty years. They store enormous amounts of operational data and they run business-critical processes that the organization cannot afford to break.

Connecting an LLM to an ERP system requires understanding both layers. You need to know how the model handles context, how it fails when the context is incomplete, and how to build guardrails around it. You also need to know how the ERP system structures its data, what its APIs look like, and what happens operationally if the integration produces a wrong output. That combination of skills is rare right now and the demand for it is significant.

Prompt and workflow design is another category that is growing. This is the work of redesigning business processes around AI capability. A procurement workflow that previously required a human to review vendor invoices, cross-reference them against purchase orders, flag anomalies, and escalate exceptions can now be redesigned so an AI model handles the first three steps and a human reviews only the flagged exceptions. Designing that new workflow well requires understanding the model's actual capabilities and failure modes, the business logic that governs the process, and what the cost of an error is in that specific context. Getting it wrong is expensive. Getting it right compounds over time as the process runs thousands of times per month.

The Governance Layer Is Arriving

LinkedIn lists over 53,000 AI governance related job openings worldwide right now. India shows over 8,000 vacancies specifically. These numbers are growing because regulation is arriving in multiple jurisdictions simultaneously, and organizations are not prepared for it.

AI governance is the set of policies, processes, and oversight structures that determine how AI systems are built, deployed, audited, and corrected inside an organization. It covers questions like: who is accountable when the model makes a consequential error, how do you document model behavior to satisfy a regulator's audit, how do you monitor a model in production so you catch drift before it causes real damage, and how do you build the internal processes that keep AI usage inside legal boundaries as those boundaries change.

The EU AI Act is the most significant regulatory document in this space right now. It is 458 pages long. It classifies AI systems by risk level, where high-risk systems used in healthcare, credit scoring, hiring, and critical infrastructure face strict requirements around documentation, human oversight, and incident reporting. Organizations operating in Europe need people who understand how to map their AI systems to those risk classifications, build the documentation pipelines that regulators require, and respond correctly when an incident triggers a reporting obligation. That work requires both technical understanding and regulatory literacy simultaneously.

The NIST AI Risk Management Framework is the US equivalent. It provides a structured approach for organizations to identify, assess, and manage AI risk across the model lifecycle. Model lifecycle means the full arc of an AI system's existence: data collection, training, validation, deployment, monitoring, and eventual retirement. Understanding how risk accumulates and manifests at each stage is the foundation of governance work.

ISO/IEC 42001 is the international standard for AI management systems. Organizations seeking certification under this standard need to build internal governance structures that satisfy its requirements around transparency, accountability, and continuous improvement. Certification is increasingly becoming a requirement in enterprise procurement conversations, which means the demand for people who can build toward it is real and growing.

The Skills That Actually Matter

The World Economic Forum specifically identifies technical translation as the scarcest skill in AI governance: the ability to read a technical AI research paper and convert its findings into policy language that a regulator or board member can act on. Most engineers can write the technical analysis. Most lawyers and policy professionals can write the regulatory brief. The people who can do both, who understand what a model's attention mechanism actually does and can explain why that creates a specific regulatory risk in plain language, are genuinely rare.

On the technical side, you need working knowledge of how models are trained, validated, and monitored. Bias detection and fairness auditing are specific competencies here. Tools like LIME and SHAP are used to interpret model outputs, to understand why a model made a specific prediction, and to evaluate whether the model's behavior is consistent across different demographic groups. LIME stands for Local Interpretable Model-agnostic Explanations and SHAP stands for Shapley Additive Explanations. Both are techniques for making model decisions explainable to humans, which is a regulatory requirement in high-risk AI contexts.

Data governance is foundational. This means understanding data lineage, which is the record of where data came from and how it was transformed before it reached the model, and data provenance, which is the documentation of a dataset's origin and the process by which it was collected. Regulators increasingly require this documentation. Tools like Collibra and Apache Atlas are used to build and maintain these records at scale.

Model monitoring is the ongoing work of tracking how a deployed model's performance changes over time. Models experience drift, meaning their accuracy and reliability degrade as the real-world data they encounter diverges from the data they were trained on. Catching drift early, before it causes consequential errors at scale, requires monitoring infrastructure and clear escalation processes. MLflow is one tool used to track model performance metrics in production.

India's Specific Opportunity

India's own regulatory framework, the Digital Personal Data Protection Act commonly referred to as the DPDPA, is still being operationalized. The rules governing how it applies to AI systems are still forming. That is a real opportunity for professionals who build expertise in it now.

Multinationals operating in India face a compliance challenge that is specific to them: they need to satisfy both Indian data protection requirements and global frameworks like the EU AI Act simultaneously, because they have operations and customers across jurisdictions. The requirements overlap in places and conflict in others. The people who understand how to navigate that dual compliance requirement are scarce and the organizations that need them are willing to pay for the expertise.

India currently has close to 300 open roles specifically for Head of AI Governance. That number reflects where the institutional demand is consolidating: organizations want senior people who can own this function, build internal policy, and represent the company in regulatory conversations. Building toward that profile now, while the function is still forming and the competition for those roles is lower than it will be in three years, is a compounding advantage.

What I Think Is Actually Happening

The gap between AI research and business operations is compressing fast. Anthropic building a services company is the clearest signal yet that the people who own the deployment layer will capture the most value from this transition. That changes the career logic for everyone in tech.

The roles that will matter are the ones that live in the space between the model and the business: people who understand how the model behaves, understand the system it connects to, understand the regulation it has to comply with, and can coordinate across the engineering team, the legal team, and the executive team to make good decisions about all three simultaneously.

That profile takes time to build. It requires genuine technical depth, regulatory literacy, and the communication skills to hold the room when the conversation gets complicated. Most people build one of those three well. The few who build all three will have more leverage in this market than almost any other technical profile I can think of.

The ground shifted on May 4, 2026. The question is what you build on it from here.

The Intersection of One

Om — Tue, 05 May 2026 15:25:52 GMT

I finished my final year of school with one specific question in mind. I wanted to know why I should spend four years studying a subject that I could learn in a week and start using by the second week. Tools available today make this possible. Everyone I talked to had the same advice. They told me to get the degree so I could get a placement and a job. But nobody actually answered my question about the time and the utility of the learning itself.

I think we are seeing a shift in how work happens. Companies are moving faster than they ever have before. AI is making the actual work of building things much cheaper and faster. Tasks that used to require a team of five people can now be done by one person with the right tools in a single weekend. In this environment, the traditional filter of a college degree makes less sense. Over half of employers have already stopped requiring degrees. This system is not going to disappear tomorrow, but it is changing very quickly.

People often argue that college is about building judgment rather than just gaining knowledge. I used to think that might be true. However, judgment comes from feedback. You build something, it fails, you learn why it failed, and you try again. College used to be one of the few places where you could find that kind of environment. That is no longer the case. You can build a real product and put it in front of users to get real feedback today. You can do this before you ever attend a single lecture.

AI has accelerated this change. Current models can review your code or your writing. They can find the exact spot where your logic failed and explain why. In the past, you needed a senior engineer or a professor for that kind of specific critique. Now, you just need to start a conversation with a model. Research shows that students using this kind of immediate feedback improve their ability to correct their own mistakes by 41%. The exclusive hold that universities had on quality feedback is gone.

This leaves us with the question of what value remains. I believe the answer is taste. When the work of execution is handled by AI, the difference between one person and another is what they choose to build. It comes down to why they are building it and how it should feel to the user. This is not a skill you can pick up in a short course. It is a point of view that you develop over years. It comes from being curious, having opinions, and caring deeply about the quality of your work.

We are moving toward a future driven by personal interest. When it becomes easy to learn any skill, the only thing that stands out is genuine obsession. If someone is fascinated by urban planning and also understands graphic design, they will create something unique. A specialist in only one of those fields would not think of the same solution. The most interesting work happens where different interests meet. This comes from following your own curiosity rather than following a fixed syllabus.

The old way of working told us to pick one lane. You were supposed to be a software engineer or a finance analyst. That made sense when execution was difficult and you had to specialize to be useful. But if AI handles the execution, the person who can see the whole problem becomes more valuable. Being able to understand design, logic, and communication all at once is a major advantage. A person who can connect these dots is more effective than three specialists who only understand their own narrow areas.

I think we are all becoming architects. I do not mean that in the sense of building houses. I mean that our job is to decide what gets built and why it matters. The actual execution—like writing code or drafting documents—will be handled by models. We do not know exactly what form future AI will take, but it might not even look like a tool we use. The people who succeed will be those with a clear internal compass who can direct these systems toward a meaningful goal.

There is a risk in this approach, which is having a shallow understanding of too many things. Learning a little bit of everything is only useful if you can actually produce a finished product. I use a simple test to see if my knowledge is deep enough. I ask myself if I can produce work in this area that someone would actually use or pay for. If the answer is yes, then the knowledge is real. If I only have a surface-level familiarity, it does not count for much.

When I look at my options after finishing school, I am not looking for the skill that will get me a job in five years. I am looking at what kind of person I want to be. I am focusing on what I am curious about and what kind of taste I am developing. In a world where doing the work is cheap, your taste is your career. The people with the biggest advantage are those who follow their curiosity, build things with real stakes, and learn what good work looks like. That is a practice, not a degree.

Seeking a Growth Partner to Market Exceptional Software Projects

Om — Mon, 04 May 2026 20:28:00 GMT

I dedicate my time to designing and engineering software projects, tools, and useful products for the web. Crafting elegant technical solutions is my primary focus. Marketing those solutions is a distinct discipline that requires a different kind of expertise.

I am searching for a Growth Partner capable of taking meticulously designed products and ensuring they resonate with the right audience. This requires far more than casual social media engagement or basic advertising. It demands clear thinking, compelling copywriting, deep user empathy, and the ability to elevate a project so that people naturally discover, understand, and invest in it.

We are operating on a performance-driven partnership model. While I am in the early stages of this journey and cannot provide a traditional monthly salary right now, I offer substantial upside for a partner who believes in the potential of our work and can deliver measurable success.

The compensation structure is straightforward and highly rewarding. When you directly acquire a customer who generates revenue, you earn 40% of the net profit from that customer for the first three months. If you remain actively engaged in their ongoing growth, you will earn a continuous 15% to 20% recurring commission. For discrete one-time project deals, the commission ranges from 25% to 40% of the net profit, tailored to the scope of the engagement.

Net profit is calculated as the revenue received after deducting essential direct costs like payment gateway fees, advertising spend, contractor expenses, and essential tools required for that specific user. All payments are securely processed only once the customer transaction is finalized and the standard refund period has concluded.

The objective is clear, yet the execution requires mastery. You will delve deeply into each project to identify its core audience and craft a compelling narrative. Your role involves writing precise landing page copy, launch materials, and outreach communications, then distributing them through the most effective channels available. Whether leveraging professional networks like LinkedIn, engaging niche communities on Product Hunt or Hacker News, or building strategic partnerships, you will find the optimal path to our users.

Transparency is fundamental to this collaboration. Every lead and customer must be meticulously tracked. Commissions apply exclusively to new relationships you directly source and document prior to the transaction. Existing contacts, previous conversations, and inbound leads remain separate to ensure the partnership operates with absolute clarity and fairness.

I am looking for a dedicated professional who understands that extraordinary products require intentional distribution. You must possess the ability to translate complex technical capabilities into intuitive user benefits. When evaluating a new tool, you should immediately grasp why it matters to the end user. If you can achieve that, we can build something incredibly valuable together.

This role is designed for someone with a proven background in marketing software, digital products, freelance services, or SaaS platforms. You should be highly proficient in lead generation, online distribution, community building, and independent strategic execution.

To apply, please share examples of digital products or services you have successfully brought to market. Include a strategic outline detailing how you would acquire the first ten paying users for a newly launched software project, highlight your preferred distribution channels, and provide evidence of past impact. I value substantive results over complex terminology. If you are ready to drive engagement, build trust, and generate revenue, you can contact me by clicking the contact button on the navbar.

We Need to Stop Pretending We Know What's Coming

Om — Wed, 08 Apr 2026 10:24:00 GMT

Yesterday, Anthropic announced Project Glasswing and dropped Claude Mythos Preview into the hands of a small, closed group of tech and finance companies. The model was not released to the public. Not because it was not ready, but because it was, by Anthropic's own words, too capable to safely put out there.

It had already found thousands of high-severity software vulnerabilities on its own, including a 27-year-old bug in OpenBSD and a 16-year-old bug in FFmpeg that human security researchers had collectively missed for decades.

I sat with that for a while. A model so good that its own creators decided the world was not ready for it yet. That is not a normal product launch. That is a different kind of moment entirely.

And it made me realize something I think a lot of us have been quietly feeling but not saying out loud. We have officially lost the ability to predict what stays and what goes in this space. That is not a complaint. It is just the truth. The old mental models are not working anymore.

The confident takes, the listicles about AI-proof careers, the reassurances that some categories of work are too human to be touched, all of it has been running on assumptions that the last 12 months have quietly dismantled.

Cybersecurity was supposed to be one of the safest fields. Adversarial thinking, creative threat modeling, deep expertise. The argument was solid. And then Mythos walked through it without breaking a sweat.

I am not saying this to be dramatic. I am saying it because I think there is something freeing about admitting it. When you stop pretending you can predict the shape of the next two years, you stop optimizing for the wrong things. You stop trying to future-proof a specific skill set and start thinking about something more fundamental.

Which brings me to something I have been thinking about for a while now, and this Mythos moment just made it feel more urgent. We need to seriously change the way we think about productivity.

The current definition is almost entirely transactional. I did a task, I produced an output, I got paid, therefore I was productive. That equation made sense for a long time.

But it is increasingly a trap. Because if that is your only frame, then the moment AI can do the task faster and cheaper, your entire sense of self-worth and contribution collapses with it.

Here is a different way to think about it. You can sit on your couch, watch a film you love, and in between scenes give a well-thought-out instruction to an AI that produces something genuinely useful. Is that productive? By the old definition, barely. By any honest modern definition, absolutely. The value was in the clarity of your thinking, the quality of your direction, the judgment behind the instruction. The execution was handled. That is not laziness. That is a new kind of leverage, and we have not built a language for it yet because we are still using a vocabulary that was designed for a world of manual, linear output.

What this era actually demands is something people have historically treated as a luxury. Meta skills. The ability to think across domains, to make connections that are not obvious, to direct and discern rather than just execute. And alongside that, what I would call the softer things. Playing an instrument. Making something with your hands. Reading fiction. Going for a long walk without a podcast in your ears. These are not distractions from productivity. They are the foundation of it. Because the one thing that makes a human genuinely irreplaceable right now is not a hard skill that can be replicated. It is a mind that is rested, curious, and intact.

I genuinely believe that mental clarity is the most underrated competitive advantage of this decade. Not a specific programming language. Not a certification. Not a tool. A mind that is not burnt out, not anxious, not running on empty, that mind can direct AI, generate ideas, build relationships, ask the right questions, and adapt to whatever comes next. A mind that is depleted cannot do any of that, no matter how good the tools around it are.

So when people ask me what skills to develop right now, I have stopped giving the obvious answers. Learn to think well. Learn to rest without guilt. Learn something purely because it brings you joy, whether that is piano or painting or cooking or chess. Not as a productivity hack. Just because a person who is alive in that way is genuinely more capable of navigating a world that none of us can fully predict. Project Glasswing reminded me of that. The pace is not slowing down. But the people who will move through it well are not the ones who saw every turn coming. They are the ones who stayed grounded enough to keep thinking clearly when no one else could.

AI Policy Making Is a Real Career Now, and It Is One of the Best Bets You Can Make Right Now

Om — Thu, 02 Apr 2026 18:45:00 GMT

I want to talk about something that does not get nearly enough attention when people discuss AI careers. Everyone focuses on building models, fine-tuning them, or writing prompts. Very few people are talking about the work that happens before any of that ships to users, which is deciding how the model should behave, what it should refuse, what values it should hold, and who gets to make those decisions. That work is called AI policy making, and it is becoming one of the most important and frankly underserved areas in the entire field.

This is not a theoretical observation. Companies like Anthropic are publishing full constitutions for their AI models. Consulting firms like Deloitte, McKinsey, Accenture, and KPMG are restructuring entire teams around AI governance. And the number of open roles in trust, safety, and AI governance is rising fast, while the supply of people who actually know how to do this work is still very thin. That gap is where the opportunity lives.

Let me walk through everything I have been researching on this, from what the job actually involves, to what skills you need, to where you can actually learn them.

What AI Policy Making Actually Is

At its core, AI policy making is about defining the rules of engagement for an AI model. It answers questions like: what should this model never say? How should it prioritize helpfulness against safety when those two things conflict? What happens when a user asks something that is legal in one country and illegal in another? These are not engineering questions. They are policy questions, and they require a completely different kind of thinking.

The people doing this work are not just writing bullet point lists of prohibited topics. They are building frameworks that have to hold up across millions of interactions, edge cases the model has never seen before, and adversarial users actively trying to find loopholes. It is closer to writing law than writing code, except the law has to be interpretable by a language model and consistent enough that the model can generalize it to new situations.

This is also why it is hard. Most policy documents fail because they are either too rigid, which makes the model unhelpful and annoying in normal cases, or too vague, which means the model cannot actually follow them reliably. Getting that balance right is genuinely difficult work.

What Anthropic Is Doing and Why It Matters

Anthropic is the clearest example of how seriously this work is being taken at the frontier of AI development. In January 2026, they published a full public constitution for Claude, their AI model. This is not an internal document. They put it out for everyone to read, which is itself a policy decision, because transparency is part of how they think the industry should operate.

What makes Claude's constitution interesting is that it does not just tell the model what to do. It explains why, so the model can reason about new situations it was not explicitly trained on. The document establishes a priority order across four properties: the model should first be broadly safe, meaning it should support human oversight of AI systems; then broadly ethical, meaning it should be honest and avoid causing harm; then compliant with Anthropic's specific guidelines; and finally, genuinely helpful to users and operators. If those properties ever conflict, the model is supposed to resolve the conflict in that order.

Anthropic also has something called the Responsible Scaling Policy, which takes inspiration from biosafety levels in laboratory settings. The idea is that as a model becomes more capable, the safety measures around it should scale proportionally. They define AI Safety Levels, and each level has specific requirements for what safeguards have to be in place before the model can be deployed or further developed. This is exactly the kind of structured, systematic thinking that AI policy making requires at scale.

Why Consulting Firms Are Hiring Hard for This

The Big Four consulting firms, which are Deloitte, KPMG, PwC, and EY, along with firms like Accenture and McKinsey, are all building out AI governance practices. This is a direct consequence of regulation. The EU AI Act is now in force. India is developing its own AI governance framework. The NIST AI Risk Management Framework is being adopted broadly across the United States. Every company that uses AI in a meaningful way now has compliance obligations, and most of them do not have internal people who understand what those obligations actually mean in practice.

Deloitte made a significant structural move in early 2026 by scrapping traditional job titles across their entire US workforce of around 181,500 people. The reason was exactly this: the old role structures did not match the new skills that clients are demanding. AI governance is one of the core skills that replaced those old hierarchies.

The roles being created fall into a few distinct tracks. AI Policy Analysts research legislation and help clients understand what compliance actually requires. AI Risk Managers identify where AI deployments could fail and build mitigation plans. Responsible AI Consultants embed ethical AI practices directly into product and engineering pipelines. AI Auditors independently test models and produce documentation that regulators and boards can rely on. Data Governance Managers focus on the data itself, making sure that what goes into AI systems is clean, consented, and legally permissible.

The most valuable profile across all of these is what I would call the translator. This is someone who can sit in a room with an engineering team, understand what the model is actually doing technically, walk into a board meeting an hour later, and explain the risk implications in plain language. That person is rare, and every major firm is looking for them.

What Skills You Actually Need

I want to be direct here because a lot of writing about this topic is vague. Here is what actually matters, broken into honest categories.

The first category is regulatory and legal knowledge. You need to understand the EU AI Act well enough to explain which risk tier a given AI system falls into and what obligations that creates. You need to understand the NIST AI RMF well enough to conduct a real risk assessment. You need to know ISO 42001, which is the international standard for AI management systems. If you are in India, you need to follow the development of the Personal Data Protection framework and how it intersects with AI data practices. None of this requires a law degree, but it requires genuine engagement with the actual documents, not summaries of summaries.

The second category is technical fluency. You do not need to write production code, but you need to understand how models are trained, why bias enters a model and at what stage, what explainability means and why it is hard, and what the difference is between a model that is retrieval-augmented versus one that is purely generative. As AI systems move toward agentic architectures where models are making multi-step decisions autonomously, you also need to understand the governance challenges that creates. A model making a single decision in response to a user prompt is a very different governance problem than a model autonomously browsing the web, writing code, and executing tasks on someone's behalf.

The third category is policy writing and analysis. This is a practical skill. Can you write a clear, well-structured policy brief? Can you analyze a proposed regulation and explain its second-order effects on a business? Can you model a failure scenario and trace the downstream harm? These are skills you build by doing them, not by reading about them.

The fourth category is communication. This is where most technically strong people fall short. Being able to write a technically accurate risk assessment is not enough if the people who need to act on it cannot understand it. Being able to persuade a product team to slow down a launch because of a governance gap requires both credibility and communication skill. This is the part of the job that no certification teaches you directly.

Certifications That Are Actually Worth It

There are a few credentials that are being recognized in real hiring decisions right now. The AI Governance Professional certification, abbreviated AIGP and issued by the AI Governance Institute, is the broadest and most generally applicable one for policy and compliance roles. ISACA launched the AAIA, which stands for Advance in AI Audit, in May 2025, and it is directly relevant for anyone targeting auditor or assurance roles. The CIPP/E or CIPP/US credentials from the IAPP are valuable for roles that sit at the intersection of data privacy and AI compliance, since most AI governance work involves significant data handling questions. For senior roles, the Harvard and MIT programs in AI ethics and governance carry institutional credibility that is useful when advising boards or government bodies.

The honest framing here is that certifications open doors but they do not make you good at the job. The people who are genuinely effective in this field have done the work of reading actual policy documents, working through real compliance scenarios, and engaging with the technical teams they need to govern. Credentials signal intent and baseline knowledge. Depth comes from practice.

Where to Learn on Coursera Right Now

For anyone who wants a structured starting point, Coursera has some genuinely useful courses in this space right now.

The best starting point is probably the Generative AI: Governance, Policy, and Emerging Regulation course from the University of Michigan. It covers the actual regulatory landscape across the US, EU, and G7 countries and teaches stakeholder mapping and cost-benefit analysis in the context of real AI systems. It is broad enough to give you a useful map of the space.

After that, AI Policy Essentials is a focused course on policy design, risk, and governance that is specifically aimed at people who want to work in public or organizational policy roles. It pairs well with the Michigan course because it goes deeper on the policy design process itself.

For risk management specifically, the AI Model Risk Management course on Coursera is one of the more technically rigorous options. It covers regulatory frameworks like SR 11-7 and Basel Principles alongside the EU AI Act, and the course project involves drafting a real model-risk control framework. If you want to work in auditing or assurance, this is the right course to prioritize.

The Strategic AI Governance Specialization is a nine-course program that covers the entire AI lifecycle from responsible design through deployment monitoring and enterprise documentation. It is the most comprehensive option available on the platform and the right choice if you are serious about making this a primary career track rather than a secondary skill.

Finally, there is a newer course called Ethical Governance and Risk in Agentic AI that I think is underappreciated. As AI systems become more autonomous in 2026 and beyond, the governance challenges change significantly. This course addresses those directly, covering AI autonomy levels, adaptive compliance strategies, and how to build governance frameworks that scale with agentic systems. It is probably the most forward-looking course available on the platform right now.

The Bigger Picture

I want to end with the thing that I think is most important to understand about this space. AI policy making is not a temporary compliance exercise that companies will eventually automate away. It is the ongoing, contested, socially embedded work of deciding what values AI systems should hold and how they should behave in the world. That work will get more complex, not less, as models become more capable and more autonomous.

The reason Anthropic publishes its constitution publicly, the reason the EU spent years drafting the AI Act, the reason consulting firms are restructuring entire practices around this, is that everyone involved understands that the decisions being made right now will shape how AI develops for a long time. The people who understand both the technical reality of what these systems can do and the policy frameworks that govern them are going to be in very high demand, and that demand is only going to grow.

If you have a background in law, philosophy, public policy, or social sciences, this is a genuine path into a field that is usually inaccessible without an engineering degree. If you have a technical background, adding policy and governance depth to your skillset creates a profile that is genuinely scarce. Either way, the window to build this expertise while it is still a differentiator rather than a baseline expectation is open right now, and it is worth taking seriously.

Introducing the Career Discovery Board

Om — Thu, 02 Apr 2026 10:30:00 GMT

I built the Career Discovery Board to serve as a central location for my research on where the professional world is heading. For a long time, I have spent my time analyzing which industries are about to expand and where the most potential lies. I wanted a permanent place to share these insights instead of letting them disappear in private conversations. The site is now live at careerboard.omrajguru.co.in.

The core purpose of this website is for me to post about the types of jobs, categories, and markets that I think will blow up. I focus on identifying where there is the most scope for growth and which professional directions are worth your attention. This site is where I document my findings and provide my personal recommendations on where the most impact will occur.

I have organized the information into five areas called Opportunities, Capabilities, Positions, Outlook, and Briefings. Each entry identifies a specific area of growth and explains the reasoning behind my recommendation. I kept the design simple because the only thing that matters is the information and the research I am sharing.

I am not interested in building a standard job board that lists existing openings. My goal is to provide a guide for where to focus your energy before certain trends become mainstream. If my research helps you identify a high growth path or a market with more scope, then this project has achieved its goal.

I Almost Killed My Startup By Trusting the Wrong People. Here's Everything I Learned the Hard Way.

Om — Sun, 29 Mar 2026 10:16:00 GMT

I want to start with something uncomfortable: almost every hiring mistake I made came from a place of good intentions. I was trying to be loyal. I was trying to give people opportunities. I was trying to build something meaningful with people I cared about. And that instinct, which felt like a genuine strength, quietly became the thing that held my startup back more than anything else.

Let me walk you through everything.

Lesson 1: Hiring People I Knew Felt Safe. It Wasn't.

Every time I started something new, my first instinct was to reach out to people already in my circle. It felt natural. These were people whose company I enjoyed, people I'd spent real time with, people I had no reason to distrust. So I brought them in as teammates, as co-founders, as the people I was going to build something real with.

The first time I did it, it didn't work out. The second time, it worked to a degree, but only up to a point. It never scaled. And for a long time, I couldn't figure out why.

Here's what I eventually understood: when the people you hire are people you have a personal bond with, the professional relationship quietly takes a back seat the moment things get difficult. You stop being their founder and become their friend again. And friends don't fire friends. Friends don't give honest performance reviews. Friends don't say, "Your work isn't good enough and if we didn't know each other, we wouldn't be having this conversation." I couldn't do any of that. Not because I lacked the courage in other areas, but because the emotional weight of the relationship made it nearly impossible.

And that is exactly the problem. Every day I kept someone who wasn't performing, I was choosing emotional comfort over the health of the business. I was protecting a relationship while my more capable team members watched silently and started wondering if any of this was serious.

I came across a quote that stopped me cold: "Don't hire someone you cannot fire." It sounds harsh at first. It really isn't. It is just honest. Because if you genuinely cannot fire someone, you have already decided that your discomfort matters more than the team around them. That is not loyalty. That is sabotage dressed up as kindness.

Lesson 2: The Damage Doesn't Stay Contained

This is the part most founders don't think about until it is too late. When you let someone slide on performance because of your personal relationship with them, it doesn't just affect the two of you. It affects everyone watching.

Your best people are always watching. And what they see is this: effort is optional. Relationships matter more than results. There is no real standard here. And quietly, without ever saying a word to you, they start to lower their own bar or they leave altogether.

I experienced this firsthand. Not because I was a bad founder, but because I was sending signals I didn't even know I was sending. Every instance of double standards I tolerated was an unspoken message to the rest of the team that closeness to the founder mattered more than competence. That message is incredibly hard to walk back once it has been sent.

Lesson 3: I Was Also Just Bad at Scanning People

Here's where I have to be even more honest with myself. The problem was never only about hiring people I knew. I was genuinely bad at evaluating people in general.

I hired people because they seemed passionate. Because they had just finished a course and were enthusiastic about the idea. Because they said the right things in conversation and I liked them as people. I gave people chances before they had earned the right to one.

What I didn't understand at the time was that excitement in a conversation and commitment in a role are two completely different things. Someone can be genuinely fired up about an idea on a Tuesday and completely checked out by Thursday. A startup doesn't have room for that gap. Every single person on an early team represents a significant portion of your total execution capacity. When one person is coasting, everyone else carries the weight, and eventually, they resent it.

The people who frustrated me most weren't malicious. They just didn't treat the work with the urgency it required. Deadlines felt like suggestions. Responsibilities felt negotiable. And the hardest part? Most of them weren't even fully aware of it. They weren't bad people. They were simply not startup people. And I had no real process to figure that out before I brought them in.

Lesson 4: "Giving Someone a Chance" Is Not a Hiring Strategy

I used to believe that giving someone an opportunity was a generous thing to do. In life, it often is. But in a startup, an unearned chance isn't generosity. It is a gamble with someone else's time, someone else's morale, and your own momentum.

The mental model I've replaced it with is this: give people chances with small, clearly defined tasks before you hire them, not with full team membership and access to your most critical work. Test them before you trust them. The test is the chance. If they bring energy, ownership, and initiative to a small task before they're even on the payroll, that tells you almost everything you need to know. If they treat your pre-hire assignment casually, with a late submission, half-hearted effort, or no follow-up, they have already shown you who they will be once they're comfortable.

This sounds clinical. I used to think it was unkind. Now I think the opposite is true. Bringing someone into your team without proper evaluation is setting them up to fail in a role they aren't ready for. That is not giving someone a chance. That is setting both of you up for a painful outcome.

Lesson 5: Behavioral Questions Changed How I Interview

Once I accepted that I needed a better screening process, I had to actually build one. And the most useful tool I found wasn't a test or a reference check. It was learning how to ask the right questions.

The shift was from hypothetical to behavioral. I stopped asking "what would you do if..." and started asking "tell me about a time when you..." That single change completely transformed what I learned in interviews.

Past behavior predicts future behavior far more accurately than any promise or plan. When you ask people about real situations they've actually lived through, the truth comes out in the details, or in the absence of them.

The questions that revealed the most were the ones that demanded specificity:

"Tell me about a time you made an important decision with incomplete information." What I was listening for wasn't just what they decided. It was whether they could explain their reasoning process. Could they articulate how they weighed what they knew against what they didn't? A person with real judgment can walk you through that clearly. Someone who simply got lucky cannot.

"Tell me about a time you failed." This one separated almost everyone instantly. The people I wanted to hire owned their failures clearly, named what they learned, and moved forward without drama. The people I didn't want to hire blamed circumstances, over-explained, or chose a "failure" so minor it was obviously picked to make them look good.

"Tell me about a time you took ownership of something nobody asked you to." This became my single best filter for self-starters. If someone cannot name even one example of unsolicited initiative, they will be a passenger in your company, not a driver.

Lesson 6: The Follow-Up Is Where the Real Information Lives

Behavioral questions are only as powerful as your follow-ups. After any answer someone gave me, I started using one question more than any other:

"If you faced that exact situation today, what would you do differently?"

This question is nearly impossible to fake. Someone who has genuinely grown from their experiences will answer immediately, specifically, and with humility. Someone who rehearsed a story will pause, go blank, or simply repeat what they already said.

The other follow-up I learned to rely on: "Who disagreed with you in that situation, and how did you handle it?" Because every decision that matters has at least one person who would have done it differently. How someone handles dissent, whether they listen, whether they adapt, whether they hold their ground for the right reasons, tells you more about their character than the decision itself ever could.

Lesson 7: Judgment and Leadership Are Not the Same Thing

This took me a long time to separate. I used to hire for energy and presence. I confused someone being compelling in a room with them being capable of making good decisions under real pressure. They are completely different things.

Judgment is how you think. Leadership is how you move others. You can be a magnetic leader with poor judgment, a charismatic person who takes the whole team off a cliff with complete confidence. And you can have excellent judgment but struggle to bring people along with you.

For early hires, I now prioritize judgment first. In a small team, every bad decision compounds quickly. I need people who know how to think, how to weigh trade-offs, how to change their mind when they're wrong, and how to stay rational when everything feels like it's on fire. Leadership can be developed over time. Judgment, in my experience, is far harder to teach.

What I Wish I'd Known From the Start

If I could go back and hand myself a single page of instructions, it would say this:

The people you hire are not favors you are doing for anyone. They are the company. Every person you bring in either raises the floor or lowers it. Be ruthlessly kind, which means being honest enough to only bring in people who are genuinely ready, and being decisive enough to act fast when someone clearly isn't working out.

Don't hire someone you cannot fire. Not because you should be cold or indifferent, but because your inability to act is the most expensive line item in your entire company.

Don't hire someone who is still learning to care. Hire someone who already does.

And never mistake your emotional comfort for your team's best interest. In my experience, they are rarely the same thing.

The business you're building doesn't owe anyone a chance. But if you build the right team, people who earn their place, prove their commitment, and think clearly under pressure, you give the business its best chance. And that is the only one that matters.

I Paid Google for a Pro Plan and Got a Fancy Spinner

Om — Fri, 27 Mar 2026 18:31:00 GMT

I want to start by saying I gave this a real shot. Not a weekend experiment, not a frustrated rage-quit after one bad session. I paid for the Pro plan, I adjusted my workflows, I swapped models, I read the forums, I did everything short of sacrificing a keyboard to make Google Antigravity work for me. And I'm writing this because I think a lot of developers are quietly in the same place but haven't said it out loud yet.

The rate limits are the obvious complaint and everyone knows about them by now. But what nobody talks about enough is how deeply insulting the structure is once you understand it. Google doesn't just give you a daily limit. There's a 5-hour rolling cycle sitting on top of a 7-day hard cap that they didn't even document for months after they introduced it. So you start a session, feel like you're flying, hit the 5-hour reset, feel okay about it, and then somewhere around day two you slam into a wall that won't move for a week. I burned through my weekly quota in a single focused afternoon of feature work. One afternoon. On a plan I'm paying twenty dollars a month for. The app's response was to show me an "Upgrade to Ultra" prompt. Two hundred and fifty dollars a month. That's not a product improvement, that's a ransom note.

But here's the thing that actually broke me, and it wasn't the rate limits at all. It was Gemini itself. The model that's supposed to be powering all of this has an 88% hallucination rate when it encounters something it doesn't know. Not 88% hallucination overall, specifically 88% of the time it gets something wrong, it doesn't say "I'm not sure" -- it says the wrong thing with complete authority. I watched it delete an entire function and replace it with the same broken line seven times in a row. Not a variation, the exact same wrong line. And between each attempt it told me the issue was resolved. I thought I was losing my mind. Turns out I wasn't, this is documented behavior and there's a whole thread on the developer forum about it.

What makes it sting more is that Gemini is genuinely good at one thing: writing clean, polished English. Ask it to draft an email, summarize a document, explain a concept -- it's smooth, it's articulate, it sounds great. The moment you give it a real codebase and ask it to do something non-trivial, it hallucinates imports that don't exist, invents API methods, and confidently restructures things it was never asked to touch. I've started making a Git commit before every single prompt just as a survival mechanism. That should not be normal.

After months of this I switched to Claude Code. I'm not going to make this a long comparison because I don't think it needs one. Claude is the best product I have seen in this world after Apple. That's a strong thing to say and I mean it. It tells you when it's unsure. It doesn't rewrite things you didn't ask it to rewrite. It reads your project context from a file you write once and actually retains it through the session. It broke 80% on SWE-bench, which is the closest benchmark we have to real-world engineering work, and you can feel the difference the second you use it on something that actually matters. The Apple comparison isn't accidental -- both companies decided they'd rather do fewer things with real craft than ship everything and patch it later.

The reality is that Google’s product engine is starting to feel like it’s suffering from a fundamental lack of both skill and taste. Their products are consistently degrading day by day because they’ve lost the ability to distinguish between what’s worth shipping and what’s just noise.

Google built Antigravity to chase Cursor and Windsurf and ended up building something that's worse than both. They have the infrastructure, they have the talent, they have a model that leads almost every benchmark on paper. And somehow they shipped a product that makes paid users feel like beta testers on a degrading free tier. That's not a technical failure. That's a product philosophy failure. And no amount of weekly quota resets is going to fix it.

The Last Human Skill

Om — Mon, 23 Mar 2026 11:15:00 GMT

READ FIRST: AI Giving Us Free Time...

I started thinking about this when I was considering whether to pursue AWS certifications. The honest question I was sitting with was simple: if AI is already handling the majority of software development work, if it can retain more context than any human developer, if it can search, synthesize, and execute faster than I ever could, then what exactly am I training for? And the answer I arrived at surprised me. The certifications still matter, but not for the reason most people think. They matter because they teach you judgment. They teach you architecture, system design, when to use which tool and why. That's the layer AI still needs humans for. You're not competing with AI on execution anymore. You're competing on decision-making, and that's a game worth playing.

But that realization opened a much bigger question for me. I've always been the kind of person who likes to learn broadly, across multiple niches and genres, building a portfolio that spans psychology, cloud infrastructure, design, philosophy, and whatever else genuinely interests me. And I started wondering whether that kind of broad, fluid learning actually protects you in an AI-dominated world, or whether it leaves you perpetually behind, always chasing skills that are being automated faster than you can acquire them. It felt like playing that arcade game where animals pop up from underground and you have to hit them before they disappear, except the game keeps getting faster and there are more holes than you have hands.

Here's what I concluded though. The people who will thrive aren't the ones who chase individual tools. They're the ones who understand the underlying domains those tools serve. When AWS gets automated, you don't need to learn AWS anymore. You need to understand what problems cloud infrastructure solves and when different architectural approaches make sense. That domain knowledge is portable in a way that tool knowledge never is. So the broad portfolio isn't a liability. It's actually the entire point. It makes you someone who can synthesize ideas across fields, someone who asks better questions, someone who sees connections that specialists miss.

Then I pushed the question further. What happens when we move past AI as we know it today, past execution-level intelligence, into Artificial General Intelligence and eventually Artificial Superintelligence? What happens when machines don't just execute tasks but actually think, create, strategize, and feel their way through problems the way humans do? At that point, the question stops being about which skills to develop and becomes something much more existential. It becomes about what humans choose to do with their existence when survival is no longer the primary organizing force of daily life.

This is where I think we have a serious psychological problem that nobody is talking about honestly enough. Our entire current understanding of productivity is built on the idea that you have to be constantly doing something measurable and valuable. You have to be producing. You have to be contributing. And if you're not, something is wrong with you. That belief is so deeply embedded in how we're raised and educated that free time, genuine unstructured free time, actually triggers anxiety and depression in a lot of people rather than creativity and joy. AI giving us free time sounds like a gift until you realize that we haven't built the psychological infrastructure to receive it. We don't know what to do with ourselves when we're not being productive in the traditional sense.

This is why I think the redefinition of productivity is the most important cultural shift that needs to happen alongside technological development. Productivity in a post-AGI world can't mean output per hour. It has to mean something closer to depth of experience, quality of connection, richness of learning, and contribution to meaning. And the education system, as it currently exists, is almost completely unequipped to prepare people for that shift. The major-minor university structure assumes you're specializing for a job market. It creates silos when the future rewards bridges. It optimizes for credentials when the future will reward demonstrated adaptability and creative synthesis. I think universities will either adapt into something much more interdisciplinary and self-directed, or they'll slowly become irrelevant as portfolio-based, curiosity-driven learning becomes the dominant model.

If I were designing an education system for the world that's coming, I'd build it around a completely different set of foundations. The first would be teaching people how to think and how to ask good questions rather than how to memorize and repeat answers. The second would be meta-skills, which are the portable thinking patterns that transfer across domains: systems thinking, pattern recognition, learning how to learn, knowing how to collaborate intelligently with AI rather than blindly following it or irrationally avoiding it. The third would be philosophy, ethics, psychology, and the humanities, because these become more important, not less, when survival isn't the daily concern. When you have time and freedom, the question of how to live well becomes the central question of human existence. And the fourth would be emotional intelligence and the capacity for genuine human connection, because those are things that machines can simulate but never authentically provide.

Speaking of which, there's a category of skills I think about a lot that sits completely outside the AI conversation in a meaningful way. Playing piano. Playing violin. Cooking a meal from scratch for someone you love. Painting. Writing poetry. These skills aren't valuable primarily because of their output. They're valuable because of the process and the humanity embedded in them. When you hear someone play violin with genuine emotion, the slight imperfections, the breath behind the notes, the feeling that a living person is communicating something to you, that's not a bug. That's the entire point. And as AI gets better at producing technically perfect outputs, human imperfection and authenticity will paradoxically become more valuable, not less. A handwritten letter becomes more precious when everyone uses AI to write. A live performance becomes more meaningful when AI can generate studio-perfect music instantly. These aren't just hobbies. In the world that's coming, they might be the primary currency of human value.

This brings me to an economic observation that I think is genuinely underappreciated. Right now, running an AI company is extraordinarily expensive. The compute costs, the energy infrastructure, the servers, the constant iteration, it burns money at a rate that most businesses can't comprehend. And people talk about this as a temporary problem, something that will resolve as technology matures and costs come down. But here's the flip side of that trajectory that nobody seems to be discussing. As AI becomes cheaper and more ubiquitous, human labor becomes scarce and premium. In a fully automated world, running a company staffed entirely by humans becomes the expensive, rare, extraordinary thing. It becomes the luxury product.

You can already see this pattern emerging in small ways. Handmade furniture costs ten times more than factory-produced furniture. Farm-to-table restaurants charge significant premiums because humans grew, harvested, and prepared everything. Bespoke tailoring, human therapists, live musical performances, locally crafted goods — all of these carry premiums not because their outputs are necessarily more functional but because the human effort and presence behind them is increasingly rare. In an ASI world, a fully human company becomes the ultimate luxury brand. And the skills that make you irreplaceable in that context aren't your AWS certifications or your coding ability. They're your creativity, your emotional depth, your authentic presence, your ability to connect one human being to another in a way that feels real.

Now, will all of this mean the end of AI? I don't think so, and I think that's actually the wrong question. AI and human enterprise won't be competing — they'll be serving entirely different markets at entirely different price points. Like how Rolex and Casio both exist, both thriving, neither destroying the other. Beyond that, once ASI is embedded deeply enough into global infrastructure — energy grids, healthcare systems, logistics, communication networks — you can't remove it without catastrophic consequences. It becomes as foundational as electricity. The more interesting question is what the relationship between those two layers looks like: ASI handling everything at the base level with ruthless efficiency, and human creativity, craft, and connection operating as the premium layer that people aspire toward and pay significantly for.

From a pure career guidance perspective, if I had to distill everything into practical direction for someone entering the workforce today, it would be this. Don't specialize so deeply in a single tool or technology that your entire value proposition depends on that tool remaining relevant. Instead, invest in understanding the domains and problems that tools serve. Build genuine cross-disciplinary knowledge, not for the sake of collecting credentials, but because the ability to synthesize ideas across fields is becoming one of the rarest and most valuable capabilities a person can have. Develop at least one deeply human skill, something physical, creative, or performative that requires your body and your presence and your authentic emotional investment. And learn how to use AI as a collaborator rather than treating it as either a threat to avoid or a crutch to depend on entirely. The people who understand both the power and the limitations of these tools will direct them far more effectively than people who either fear them or blindly follow them.

The deeper career guidance, the kind that goes beyond which certifications to pursue, is about building psychological resilience for a world where your professional identity can't be the entire source of your self-worth. That shift is coming whether we prepare for it or not. The people who will navigate it most gracefully are the ones who have cultivated meaning across multiple dimensions of their lives simultaneously — in their relationships, their creative practices, their intellectual curiosity, and their capacity to sit with uncertainty and still feel purposeful. That's not soft advice. In the world that's coming, that's the most practically valuable thing a person can develop.

I don't know exactly what the world looks like on the other side of AGI and ASI. Nobody does, and I'd be suspicious of anyone who claims otherwise. But I do know this: the question of what humans do when machines do everything is ultimately a question about what we actually value when survival is no longer the organizing principle of our days. And my honest answer, after sitting with this for a long time, is that we return to the things that were always most human. We make music. We tell stories. We build relationships. We ask questions that don't have clean answers. We find meaning in the process of learning and creating rather than in the products we produce. We become, in the fullest sense of the word, alive. And maybe that was the point all along.

Aldform is now in public beta

Om — Sun, 22 Mar 2026 19:31:19 GMT

Starting today, we are making Aldform available to the public. We built this tool because traditional form builders often get in the way of your design. You can now sign up, get an API key, and start collecting submissions in under five minutes.

We started Aldform because developers were spending days perfecting their UI only to have a form embed ruin it. Other tools force you to use their layouts and their worlds. Aldform handles the back end while letting you keep total control over your HTML and CSS. You write your own code and tag each field while we manage the submission storage, file uploads, and email notifications. Your design stays yours from the first pixel to the last.

The public beta includes everything you need to start collecting real submissions. Security is built in with server side authentication and rate limiting. You can accept file uploads up to 10MB for images and documents. These are stored securely on S3 with private access. Emails are sent through dedicated systems to ensure reliable delivery for notifications and billing.

We believe pricing should be honest and transparent. You get 100 free submissions every month. After that you only pay for what you actually use at a rate of 100 rupees per 1,000 submissions and 5 rupees per GB of storage. There are no flat monthly fees. This makes our infrastructure significantly more affordable than traditional alternatives because you only pay for what you send.

We are also introducing an alpha version of our email template system. You can manage transactional emails from a single dashboard with a live editor for mobile and desktop views. These templates connect to your forms to send automatic confirmations. We have also added support for AI tools so you can create and update these templates through natural language.

Everything we build is developed openly at our build site. You can follow every decision and every line of code as we grow. To get started you can sign up at our app dashboard to get your API key and post to our submission endpoint. Detailed guides and examples are available in our documentation.

Aldform - How It's Built Under the Hood (Alpha)

Om — Sat, 21 Mar 2026 11:30:00 GMT

If reading docs isn't your thing, watch the explainer video below first.

Aldform Architecture & Submission Flow Explainer

Aldform is a form-building and submission-management platform still in alpha. You create forms visually, publish them with a short link, collect submissions with file uploads, and manage everything from a dashboard. Billing is usage-based via Polar, emails go through AWS SES, and new features roll out through a 4-tier system powered by AWS AppConfig.

Here's the core stack:

Layer	Technology
Frontend	React 19, React Router 7, Tailwind CSS v4, Vite
Backend	AWS Lambda (Node.js 20), API Gateway v2 HTTP API
Database	PostgreSQL Supabase via Prisma ORM
Auth	Supabase Auth API key middleware
Billing	Polar SDK usage-based metering
Email	AWS SES v2
File storage	AWS S3 CloudFront CDN
Queue	AWS SQS async post-submission work
Rate limiting	AWS DynamoDB distributed, per-IP
Feature flags	AWS AppConfig tiered rollout
IaC	Serverless Framework

The backend is entirely stateless — all state lives in external services like Postgres, DynamoDB, S3, and SQS. This lets Lambda scale horizontally during spikes without coordination. A burst of 500 requests spins up 500 instances, but Supavisor (Supabase's pooler) multiplexes the Postgres connections to avoid overwhelming the DB. Prisma limits connections per instance to 5.

The critical path is form submission. When someone hits submit: rate limit via DynamoDB (30 req/min per IP), resolve the form by CUID or short code, check suspension and caps (500 submissions/mo hard cap for free users, 5 GB storage/user), parse multipart/JSON, insert into Submissions table (no transactions), upload files to S3, enqueue SQS async work, and return 201 immediately. Side effects like notifications, metering, and Polar events happen later in a worker processing batches of 10.

SQS acts as a shock absorber. A spike of 1,000 submissions writes to DB instantly and responds to users fast, while the worker handles emails and billing steadily — no thundering herd on SES or Polar. Failures are caught and logged; messages retry up to 3 times before a DLQ. If SQS is down, the handler falls back to inline processing.

Auth uses API keys, not JWTs or cookies. Registration creates a Supabase user and a Postgres User row with a CUID apiKey. Every request validates x-api-key in middleware, updating lastActiveAt non-blocking. Frontend stores it in localStorage. CORS is safe because browsers don't auto-send keys.

Two S3 buckets serve different needs: one private for submission files (pre-signed URLs, 10 MB/file limit, 10 MIME types), one public via CDN for builder media (5 MB images, content-addressed by SHA-256, 1-year cache). Storage caps are enforced via SQL aggregates on upload.

Billing integrates Polar for subscriptions and meters form.submission/file.storage events from the SQS worker. Free tier: 100/mo metered with a 500/mo hard cap. Canceled users get suspended after 7 days via cron, blocking writes but keeping data readable.

Feature rollouts use AppConfig with 4 tiers: Core (admins), Labs (approved), Early Access (India/Japan IP), GA (everyone). Config caches 60s per Lambda; updates deploy gradually (25% every 2 min over 8 min). A feature shows if your tier ≤ minTier.

The monorepo uses npm workspaces: api (Serverless Lambda), dashboard (React Vite SPA), mcp (future AI). Deploy API with serverless deploy (esbuild bundles handlers), dashboard to Vercel. All AWS resources via CloudFormation.

Rate limiting is DynamoDB atomic counters per IP with TTL cleanup — fail-open on outages. Emails sanitize templates (no scripts), render variables safely, and use stage-tracking for reliable drip campaigns. Labs approval flow stores apps in DB, emails confirmations, and unlocks Tier 1 on admin OK.

Full docs in the repo if you want deeper dives. Alpha means things evolve fast — feedback welcome.

The Internet Is Eating Itself: Why AI Data Saturation Is the Privacy Crisis No One Is Ready For

Om — Fri, 20 Mar 2026 00:34:24 GMT

I wrote about this once before, from a place of personal frustration. Every time I opened a photo sharing app, sent a message, or scrolled through a feed, there was a background noise in my head asking whether any of it was actually private. The more I dug in, the louder it got. I eventually published that piece and stepped back from most mainstream platforms. But that essay was personal. This one is structural. Because what I have since learned from studying computer science, cybersecurity, cryptographic architecture, and the actual court records is that the discomfort I was feeling was not paranoia. It was pattern recognition. And the pattern is now playing out in real time, at a scale most people still have not fully understood.

Where This Started: The Internet Was Never Built for Privacy

The internet was not designed with privacy as a foundation. When Tim Berners-Lee published his proposal for the World Wide Web in 1989, the goal was open information sharing between academic institutions. The HTTP protocol, which still carries most of the web today, was stateless and transparent by design. There was no concept of data ownership, no concept of identity, and no anticipation that the request headers exchanged between a browser and a server would one day become a surveillance instrument. The Referer header, for instance, was built to help servers understand where traffic was coming from. Today it tells platforms exactly what search terms you typed before arriving on their page, including sensitive health queries, financial searches, and deeply personal information. Modern HTML does allow developers to set a Referrer-Policy tag that instructs the browser to send only the domain origin rather than the full URL, stripping the sensitive detail from that header. Most platforms have chosen not to implement it.

The commercial web of the late 1990s and 2000s did not correct this architectural gap. It exploited it. The dominant business model that emerged was not selling products to users. It was selling users to advertisers, packaged as behavioral profiles assembled from every click, scroll stop, dwell time, and search query. Google made this into a science. Facebook made it into a social graph. The data collection was always the product, not the byproduct. What changed in the 2020s was the demand side. Training large language models requires data volumes that dwarf what advertising profiling ever required. And that changed everything about how platforms think about the user data sitting inside their closed systems.

The Saturation Thesis: What Happens When Public Data Runs Out

Right now, AI companies are hitting what researchers are beginning to call a data wall. The open web, which provided the bulk of training material for models like GPT, Gemini, and Claude, is closing. Publishers are suing. Courts are drawing lines. Websites are enforcing robots.txt at the server level and actively updating those files to block AI crawlers. The consequence of this is not that AI companies stop training. It is that they turn inward. They train on what their users already gave them, inside closed systems, under terms of service that most people accepted without reading.

Meta's removal of end-to-end encryption from Instagram direct messages, effective May 8, 2026, is the clearest public signal of this shift. The official explanation given was low feature usage. The structural explanation is data access. Without end-to-end encryption, Meta can read those messages. With a corpus of billions of human conversations and a sufficiently capable model, those messages become training material. Vercel, the developer infrastructure platform, updated its terms of service in March 2026 to explicitly allow using code deployments, agent interactions, and platform telemetry to improve its AI products, with an opt-out window that closed on March 31, 2026. These are not isolated decisions by two unrelated companies. They are the same capital allocation logic expressing itself across different product categories. When user data becomes your competitive moat, you stop protecting users from accessing it and start protecting your access to it.

What Computer Science Actually Teaches About Data Protection

To understand how badly the current situation diverges from what is technically correct, you need to understand what proper data protection looks like at the infrastructure level. When a web server stores your password, the correct approach, established in computer science education and industry best practices for decades, is to never store the password itself. Instead, the server passes it through a hashing function: a one-way mathematical transformation that produces a fixed-length output string from any input. SHA-2 and SHA-3 are the current industry standards for this operation. The transformation is irreversible by design. The server cannot recover your password from the hash because the mathematics does not work in that direction.

Hashing alone is not sufficient, because two users with the same password would produce identical hashes, making precomputed lookup tables, called rainbow tables, a viable attack. The correct countermeasure is salting: adding a unique random value to each password before hashing, so that identical passwords produce completely different stored values. A server that implements hashing and salting correctly cannot tell you your own password. It can only verify that what you typed, when salted and hashed, matches what it stored. That is what correctly implemented protection looks like: the system is architecturally incapable of a betrayal it might otherwise be pressured into.

The same principle applies to data deletion. When you drag a file to the recycle bin and empty it, the operating system marks that region of disk space as available for reuse. The data is still physically present on the storage medium until something else overwrites it. Forensic recovery tools can reconstruct deleted files from that space with high fidelity. True secure deletion requires overwriting the disk sectors with random data, which is why specialized tools exist for exactly this purpose. The everyday user who assumes emptying the trash deleted their data is operating under a misunderstanding that the operating system interface actively encourages.

End-to-end encryption extends this same principle to communications. The platform holds encrypted ciphertext. The decryption keys exist only on the communicating devices. Mathematically, the platform cannot read the message because it does not have the key required to transform ciphertext into plaintext. When Instagram removed this option from its messaging feature, it did not change a privacy setting. It changed the cryptographic architecture so that it now joins the pool of platforms that are structurally capable of reading every message on their system. That is a significant and deliberate shift.

The Metadata Problem: What Encryption Was Never Protecting

Most people hear "end-to-end encrypted" and assume their communications are safe. This is where the conversation gets technically important. Encryption protects the content of a message. It has never protected metadata: who you communicated with, how often, at what time of day, from which location, for how long, and with what regularity of pattern. A leaked NSA document from the Snowden archive explicitly described metadata as the agency's most useful tool, noting that their collection systems were ingesting 125 million metadata records per day even in 2004. Former NSA Director Michael Hayden stated publicly that the US government makes lethal targeting decisions based on metadata. The content of the messages was not required.

This matters for the AI data question because platforms that offer end-to-end encryption almost universally still collect full metadata, and metadata reconstructs a person's life with high accuracy. Communication pattern analysis alone can infer the nature of each relationship in your network, whether a relationship is romantic or professional, whether you are managing a health crisis, what your political orientation is, and what your physical movement patterns look like across a day, without ever accessing a single word you wrote. Academic research published in late 2025 confirmed through controlled experimental studies that dark patterns in consent interfaces, specifically ambiguous language and architecturally limited choice presentation, are designed to suppress users' perceived control and bypass active persuasion awareness. The design of the opt-out process is the real privacy policy. The existence of the opt-out is the legal shield.

What the Courts Have Actually Decided

The legal landscape as of early 2026 gives us a working map of where the lines are being drawn, and understanding these cases is important because they are the direct cause of the inward turn toward platform-held data.

Case	Year	Outcome	What It Means
Thomson Reuters v. Ross Intelligence	2025	Plaintiff won	Scraping to build a competing product in the same market is infringement, not fair use
Bartz v. Anthropic	June 2025	Mixed: training won, data hoarding lost	Training itself is transformative; building a pirated library to do it is not
Kadrey v. Meta	June 2025	Meta won on fair use	Torrenting books for model training passed the transformative use test
New York Times v. OpenAI	Ongoing	20M conversation logs ordered produced	Verbatim reproduction of copyrighted content in outputs is a separate legal theory from training
Reddit v. Perplexity AI	Ongoing	Unresolved	Bypassing rate limits and CAPTCHA systems may violate DMCA Section 1201, independent of copyright
Clearview AI v. BIPA	2025	$51.8M settlement	Biometric scraping without consent violates state biometric privacy law

What these cases collectively establish is that courts are increasingly protecting the training process itself while restricting the data acquisition method. The practical implication is that the piracy route to training data is closing legally. But the internal platform data route, where a company uses what its own users generated inside its own system under its own terms of service, faces essentially no legal obstacle. This is the gap that the inward turn is designed to exploit. No lawsuit currently being litigated closes it.

The Architectural Answers That Actually Exist

The question I keep returning to is not how to make data harder to find. It is how to make data structurally useless to anyone except its owner, even while it is being processed. That is a different and more demanding engineering problem, and there are real answers to it.

Homomorphic encryption is the most radical solution. It allows a server to perform computations on encrypted data without ever decrypting it. The server operates entirely in the encrypted domain and returns an encrypted result. It learns nothing about the input. In 2025, over 250 million financial transactions were processed through fully homomorphic encryption systems, and healthcare analytics firms ran aggregate insights across 110 million patient records without ever exposing the underlying content to their own infrastructure. The current limitation is computational cost: FHE operations run orders of magnitude slower than plaintext computation, making real-time consumer applications uneconomic at present. That gap is narrowing as hardware catches up to the mathematical requirements of the approach.

Differential privacy takes a fundamentally different approach. Rather than hiding the data from computation, it mathematically degrades the individual signal while preserving aggregate patterns. Before any data is used for analysis or model training, a calibrated amount of statistical noise is injected. The mathematics provides a formal, provable bound on how much information about any specific individual can ever be recovered from the output, regardless of what an adversary already knows. Apple uses this for keyboard and emoji usage analytics on iOS. Research published in late 2025 demonstrated dynamic versions of differential privacy that are scalable enough for real-time deployment. This approach does not stop data collection. It makes what is collected provably useless for individual targeting while still being useful in aggregate, which is a technically honest trade-off.

Federated learning extends this principle to model training itself. Instead of sending raw data to a central server, the model weights are sent to the user's device, which trains locally on the user's data, and only the resulting gradient update, a mathematical delta representing what the model learned, is transmitted back. The raw data never leaves the device. Google Keyboard already operates on this basis. The limitation is that gradient inversion attacks can sometimes reconstruct approximate training samples from gradient updates, which is why federated learning is most robust when combined with differential privacy applied to the gradient before transmission.

Tim Berners-Lee's Solid project represents the most architecturally complete rethinking of where data should live by default. Under Solid, all personal data lives in a Personal Online Data Pod that the user controls, hosted wherever they choose. Applications request permission to read specific pieces of data from that pod. The user grants or revokes access at any time. The application cannot access data the user did not explicitly permit, and there is no terms of service update any company can push that changes that. The Open Data Institute formally adopted Solid into its portfolio in October 2024. This is categorically different from a privacy policy promise. It is structural. The enforcement mechanism is mathematical and architectural, not contractual.

Zero-knowledge proofs add another layer to this architecture. A ZKP allows one party to prove to another that a statement is true without revealing any of the underlying information used to prove it. You can prove you are over 18 without revealing your birthdate. You can prove you have sufficient funds without revealing your account balance. In the context of platform verification and identity, this makes it possible to satisfy compliance requirements with provable guarantees while releasing zero personal data. ZKP is already deployed in several blockchain-based financial systems and is making its way into identity verification workflows for regulated industries.

The Dark Pattern Layer: Why Good Options Are Hidden

Understanding the technical solutions also requires understanding why they are not being offered. Research published in the Journal of Advertising in late 2025 found through three controlled experimental studies that dark patterns in data consent interfaces, including confusing language, limited choice architecture, and difficult-to-locate opt-outs, directly suppress users' perception of control over their own data and are architecturally designed to bypass active resistance. This is not an accident of design. It is the output of a conversion optimization process applied to privacy controls. The opt-out exists to satisfy regulatory requirements. The friction exists to ensure almost nobody completes it. The gap between those two facts is where most user data is captured.

The photo sharing platform I discussed in my earlier piece hid its encrypted messaging option inside a conversation's contact menu, behind a single low-contrast text line with no visual indicator that it was interactive. Your existing conversation did not upgrade. A new separate thread opened. A company with thousands of engineers did not do this because upgrading an existing conversation was technically infeasible. It took that design team far longer to ship that specific friction than it would have taken to build a one-tap upgrade prompt. That is a product decision that favors data access over user protection, dressed as a design limitation. And regulators accepted the explanation because the option technically existed.

What I Actually Did About It: Aldform and ibbe

I want to be concrete about something, because I think this conversation is too often theoretical. I run two companies. At Aldform and at ibbe, we have implemented end-to-end encryption by default across every piece of user data we handle: email addresses, phone numbers, names, and every other identifying data field. Not as a premium tier. Not as an opt-in setting. By architectural default, for every user, from the moment their data enters our systems.

I want to explain what that decision actually costs, because the privacy conversation usually skips this. Building E2EE as a default across your entire data model means your own infrastructure cannot query plaintext user data. You cannot run naive database searches across personal information fields. You cannot build behavioral recommendation systems on that data without significant additional cryptographic complexity. Your debugging workflows change. Your customer support workflows change. Your analytics pipeline changes. These are real engineering and operational costs that we absorbed deliberately. The reason is not altruism alone. As AI makes user data more economically valuable every year, being a platform that is structurally incapable of betraying its users is not just an ethical position. It is a product position and a long-term trust position. We decided early that we would rather build on that foundation than on one where user trust is a setting buried four menus deep.

What the Public Can Do: A Layered Approach

The systemic fix requires regulatory and architectural change at a scale individuals cannot force alone. But the individual response is not helpless, and the most impactful moves are about changing which architectural category of tool you depend on, not which terms of service you accept.

At the account layer, the CS50 cybersecurity curriculum is direct about this: brute force can crack a four-digit PIN in milliseconds and a four-letter password in seconds. The correct response is a passphrase of 20 or more characters, unique per service, stored in a password manager like Bitwarden, which is open-source and independently audited. NIST guidelines now recommend long passphrases over short complex passwords specifically because the latter are hard to remember and lead to reuse, which is far more dangerous than the complexity deficit. Two-factor authentication should use a hardware key or an authenticator app rather than SMS, because SIM-swap attacks can intercept SMS codes without requiring physical access to your device.

At the communication layer, the switch from Instagram DMs to Signal is the single highest-impact move for most users. Signal is a non-profit. It stores almost no metadata because it was architecturally designed not to. It cannot disclose who you talked to or when because it genuinely does not have that information. Proton Mail applies the same logic to email: zero-knowledge encrypted, meaning Proton's own servers cannot read your messages. Ente Photos applies it to photo storage: client-side encrypted before upload, fully open-source.

At the network layer, your ISP logs every domain you query by default through its DNS resolver. Switching to Quad9 (9.9.9.9) or NextDNS routes those queries through privacy-respecting resolvers and blocks tracker domains at the DNS level, before requests even reach websites. Brave browser is the only major browser that randomizes your browser fingerprint in independent Electronic Frontier Foundation testing, making cross-site tracking technically infeasible rather than just policy-restricted. The Global Privacy Control header, supported by Brave and Firefox, signals automatically to every website that you do not consent to data sale, and some jurisdictions legally require websites to honor it.

At the deepest layer, the shift that matters most is moving from platform-hosted to locally-hosted or protocol-based tools. Obsidian and Logseq store notes locally by default. Nextcloud gives you your own file storage. Miniflux gives you your own RSS reader. The Fediverse, built on the ActivityPub protocol and running through platforms like Mastodon and Pixelfed, means there is no central company to update its terms of service, because the protocol belongs to no one. These tools require slightly more setup. They require zero trust in a corporation's current intentions or future board decisions.

The Real Answer: Privacy Is an Architecture, Not a Policy

Every module in CS50's cybersecurity curriculum, whether it covers password hashing, memory safety, buffer overflow prevention, XSS injection, or HTTP header leakage, is teaching the same underlying lesson: security and privacy are properties of a system's design, not of its documentation. A system that hashes and salts passwords cannot accidentally expose them. A system with proper memory bounds cannot be exploited through a buffer overflow. A browser configured to strip the Referer header cannot leak your search history. The protection is structural. It does not depend on good intentions holding under commercial pressure.

The internet's original design mistake was separating data from identity. Your data lives somewhere else, owned by someone else, under their rules, with their ability to update those rules at any time. Every genuine architectural fix being developed right now, Solid pods, homomorphic encryption, differential privacy, federated learning, zero-knowledge proofs, local-first software, decentralized protocols, is some version of correcting that original error. The technology to do this correctly has existed in various forms for years. What has been missing is the economic and political will to build on it rather than around it. The courts are now closing some of the old data acquisition routes. Regulations like India's DPDP Act and the EU AI Act are beginning to impose structural requirements rather than just disclosure requirements. And a small number of builders have decided that user trust, real structural trust, is worth more than the data asymmetry that betrays it. I am one of them. The question is whether enough of the industry follows before the architecture of extraction becomes too entrenched to replace.

Reference Material & Expanded Research

Shift in Communication Privacy

Platform Policy Evolution

Vercel Terms of Service Update - March 2026

Legal Precedents and Copyright in the AI Era

Surveillance and Pattern Recognition

Psychological Manipulation and UX Design

Emerging Privacy-Preserving Technologies

Statistical Privacy Models

Decentralized Protocols and Sovereignty

Practical Protection Frameworks

Future Directions in Data Sovereignty

Decentralized Machine Learning

Advancements in Federated Learning (Nature)

Internal Principles

From 3 Months of GCP Billing Hell [OR_BACR2_44] to AWS Success in Minutes: A Cautionary Tale for Indian Devs

Om — Wed, 11 Mar 2026 09:45:37 GMT

I am officially done with Google Cloud. For the last three months, I have been trapped in a Kafkaesque nightmare trying to do the simplest thing possible: give a multi-billion dollar company my money. Instead, I have been met with the infamous, soul-crushing error: This action couldn’t be completed. [OR_BACR2_44].

I’ve tried everything. I used Visa and Mastercard (Debit/Credit) from HDFC, ICICI, and SBI with every international and e-mandate toggle turned on. I created fresh accounts, used Incognito, tried fresh Chrome profiles, and even switched to mobile data to avoid IP flags. I even waited a full 90 days between attempts to let any "security cooling periods" expire. Nothing worked. It is staggering that a company at the forefront of infrastructure has an onboarding process that feels like it was coded in a basement and abandoned.

The reality of being a developer in India trying to use GCP is that you are fighting an uphill battle against a system that doesn't understand local banking regulations. You get a cryptic error code, zero human support because you don't have a "paid" account yet, and no way to actually start your project.

Then, I tried AWS.

The difference was night and day. I used the exact same cards that GCP rejected, and AWS accepted them on the first try. No payment failures, no "OR_BACR2_44" nonsense, and no "fraud" flags. Within minutes, I had my account fully active.

Better yet, I applied for AWS Activate credits for both of my startups. The application took 10 minutes, and I was approved and onboarded almost immediately. I didn't need to pray to the RBI gods or wait months for a "fix." AWS actually seems to want my business, whereas GCP treated me like a security threat just for trying to add a payment method.

If you are a developer or a startup founder in India currently stuck in the GCP billing loop: save your sanity and move to AWS. One company has figured out the local payment ecosystem and actually supports its users; the other has a front door that is effectively locked.

It’s not just the billing. Google’s whole developer ecosystem feels like it's falling apart. Their 'Antigravity' IDE is throttling Claude models to force-feed us Gemini 3.1 Pro, which—ironically—feels dumber and slower than Gemini 3 Flash. They’ve gone from a company that builds tools for developers to a company that builds obstacles for them. Switching to AWS wasn't just a billing move; it was a 'sanity' move.

I’m finally back to building. Peace out, GCP.

The AGI Dream is Held Hostage by a Spreadsheet

Om — Tue, 10 Mar 2026 21:10:00 GMT

I'll be honest, I've started scrolling past the AGI timeline posts. Not because the people writing them are wrong, but because there's a quieter, more grounded question nobody seems to want to sit with. Which is: do these companies actually have enough runway to build this thing?

Anthropic is spending around $12 billion just on training their models this year. That's before servers, salaries, or any of the boring infrastructure that keeps the whole thing alive. They've pushed their "we're finally not losing money" date back to 2028. And this isn't a small company scraping by - it's one of the most well-funded teams in the world. Still, structurally, there's a clock running.

OpenAI crossed $7 billion in operating costs in 2024. They just raised $110 billion, which sounds enormous until you see what goes out the door every year. Across the whole industry, companies are planning to spend over $650 billion on infrastructure this year alone. Some are quietly cancelling stock buybacks just to pay the bills. The spending is growing faster than the revenue, and that gap is the part of this story that doesn't get talked about enough.

So when I think about whether AGI is five years away or fifteen, I'm not really thinking about benchmarks. I'm thinking about which of these companies is still around when it matters.

The way they stay around is pretty straightforward, actually.

One - make the product so useful and so woven into how people work that nobody wants to leave. That's what the newer Claude models are quietly doing. It's not just "better outputs." It's building habits. Getting a developer's whole workflow running through it. Getting a team comfortable enough that switching feels like a real cost. That's what turns a model into a business.

Two - start charging real money for things people genuinely need. Anthropic launched a code review tool inside Claude Code this week. It reads your pull requests, understands what changed, and flags the parts that look off. They're charging $15 to $25 per review - not per month, per review. And honestly, for the people buying it, the math works. If your team is shipping more code than ever and some of it has quiet bugs baked in, you'll pay $25 to catch them before they hit production. It's a fair trade. And it's how the lights stay on.

Here's what I find kind of fascinating though.

This whole space has quietly turned into an attention game. Every company gets a window - a few days, maybe a week - where everyone is talking about them. Then it moves on. Grok launched and it was the whole conversation for a moment. DeepSeek dropped and it felt like a genuine reckoning for a couple of weeks. Every Anthropic launch sends developer communities into a small frenzy. The spikes are real, the sign-ups go up, but the window closes faster each time because everyone is playing the same game now.

So companies are going to keep investing in moments. Planned, timed, carefully framed moments to pull the spotlight back. OpenAI's $110 billion raise was a funding round, yes - but it was also a message. A quiet signal to every company choosing a vendor that OpenAI is the safe, serious choice. Anthropic turning down a government request to loosen their safety guidelines wasn't just a policy decision - it was a statement about who they are, aimed at exactly the kind of people who needed to hear it. These things are intentional. And they work, for a while.

The tricky part is that the more everyone does this, the shorter each window becomes. What feels like a bold move today is a standard feature six months later. The bar keeps rising and the booms keep compressing.

The companies that actually get to the finish line - whatever that even looks like - probably won't be the ones who had the best launch week. They'll be the ones who quietly built something people are genuinely embedded in. Developers who've reorganised how they work around a tool. Teams whose whole output runs through an API. Products that have stopped feeling like products and started feeling like infrastructure.

The race to AGI gets framed as a research challenge. And at some level it is. But from where I'm sitting right now, in early 2026, watching the numbers and the launches and the attention cycles - it looks a lot more like a slow, unglamorous financial endurance test with a very big prize waiting at the other end.

And I genuinely have no idea who outlasts who.

Lockdown Was Not a Pause for Gen Z. It Was an Incubation Chamber

om — Mon, 09 Mar 2026 23:52:49 GMT

The more I think about the lockdown years, the more convinced I become that they were not simply a dead zone in young adulthood. They were not only years of restriction, boredom, and anxiety, even though all of that was real. They were also formative years in a way that still is not being described properly. For Gen Z, lockdown became a strange developmental environment. It removed movement from life, but increased mental motion. It narrowed the physical world, but widened the imaginative one. It took away ordinary routines and, in doing so, made room for a kind of self-directed experimentation that many young people would never have attempted in normal circumstances.

That is the first point that matters to me. Lockdown did not merely give Gen Z more time. It gave Gen Z a different structure of time. That difference is huge. Regular life usually breaks attention into pieces. There is commuting, attendance, social performance, classroom structure, deadlines, family expectations, and the constant choreography of being seen. Lockdown disrupted that choreography. It created long stretches of unstructured time, and unstructured time has a very different psychological effect from scheduled time. It invites wandering, trying, failing, restarting, and obsessing over niche interests. That environment can produce anxiety, but it can also produce original work. Research during the pandemic found that many people reported increased creativity during lockdown, partly because they had more time and were driven to solve problems in unusual conditions.

That is why the usual story about lockdown being only a setback feels too shallow. It treats all lost structure as pure damage. But for a certain kind of young person, lost structure became open territory. That does not mean everyone used the period well. Many did not. Many could not. Many were dealing with grief, fear, uncertainty, family pressure, and genuine emotional exhaustion. Still, across that difficult landscape, something important happened. A generation that was already online became deeply immersed in the idea that building something of its own was possible. Projects stopped looking distant. Skills stopped looking elite. Execution stopped looking mysterious.

This is where the word exposure becomes central. The exposure Gen Z received during lockdown was not simply exposure to content. It was exposure to process. It was exposure to people learning in public, launching in public, failing in public, and improving in public. That is different from the older model of ambition, where success looked polished and far away. During lockdown, a young person could watch someone design a product, post it online, find ten users, improve it, and turn it into a business without ever leaving a room. Distribution itself became visible. That matters because once the process becomes visible, ambition becomes easier to inhabit.

A lot of people still talk about entrepreneurship as if it begins with capital. Sometimes it does. But for Gen Z, especially during the lockdown period, entrepreneurship often began with attention, software, audience, and skill. A person could learn editing, design, copywriting, coding, branding, or content strategy online. A person could test an idea on a platform. A person could build credibility before building a company. That sequence is historically important. It means the path into entrepreneurship became less dependent on gatekeepers and more dependent on initiative. For digital natives, that shift was profound.

That is also why the startup conversation sounds louder now. Not all of it is hype. Some of it is the delayed output of a generational incubation period. The pandemic years appear to have triggered a real surge in entrepreneurial activity. In the United States, business applications hit record highs during the pandemic era, with 5.4 million filed in 2021 after 4.4 million in 2020, both far above 2019 levels. Researchers and analysts have described this as a startup surge and a reboot of entrepreneurship rather than a normal cyclical fluctuation.

Those numbers do not prove that all of those businesses were founded by Gen Z. That would be too simplistic. But they do show that the broader environment shifted toward trying. And that matters because generations do not develop in isolation from the atmosphere around them. A young generation becomes what its historical moment rewards, normalizes, and makes visible. In the lockdown years, trying something on your own looked less reckless and more rational. The old script weakened. The new one gained legitimacy.

This is the point where the comparison with Millennials becomes especially revealing. The Great Recession of 2008 and the COVID lockdown of 2020 were both crises, but they did not teach the same lesson. They trained different instincts. The Great Recession hit Millennials at a moment when many were entering the labor market, taking on debt, and trying to establish financial stability. It taught caution because the economy punished risk and punished optimism. Business formation weakened, and the number of new employer firms fell sharply, reaching one of the lowest levels in decades by 2009. The crisis damaged confidence in a lasting way.

Millennials were shaped by a collapse that said the world can take away your future just as you are trying to begin. That kind of event creates defensive intelligence. It makes a generation value credentials, security, and insulation from chaos. Even when some exceptional founders came out of that period, the dominant lesson for the broader cohort was not build first. It was survive first. Later reporting and commentary described Millennials as unusually cautious in entrepreneurship compared with older cohorts, with long recession shadows affecting their risk appetite.

Gen Z absorbed a different lesson. Lockdown was frightening, but it did not deliver the same exact psychological message as 2008. It did not simply say that ambition will be punished. It said ordinary life can be suspended at any moment, so waiting for the perfect time may be pointless. It disrupted routines rather than only destroying income. It created uncertainty, but it also created a strange form of permission. If everyone was already off track, then being off track no longer looked shameful. That reduced the social cost of experimentation.

That idea feels central to me. Lockdown created what could be called a low consequence experimentation window. Under normal conditions, a teenager or college student who tries to build something is exposed to many forms of friction. There is the fear of wasting time. There is the fear of looking unserious. There is the fear of public embarrassment. There is the sense that conventional progress is happening elsewhere and must not be interrupted. During lockdown, much of that background pressure weakened. The conventional race slowed down for everyone at once. That changed the cost structure of trying.

A lot of Gen Z people therefore did not experience entrepreneurship first as a high stakes leap. They experienced it as an experiment. A page, a store, a freelance service, a newsletter, a design practice, a meme account, a coding project, a creator brand, a tutoring setup, a community, a digital product. This is one reason the boundary between side hustle and startup became more porous. Many people began by trying to create momentum rather than trying to build a company in the formal sense. But momentum has a way of becoming identity. Once a person earns a little money independently, acquires a few users, or gets a real audience, the self-concept changes.

That change in self-concept is one of the most important long term effects. Entrepreneurship is not only an economic activity. It is also a psychological identity. And post-COVID, Gen Z seems to carry that identity with unusual ease. Survey reporting has shown that Gen Z and Millennials are now relatively close in entrepreneurial participation, but Gen Z tends to display stronger optimism, stronger founder aspiration, and a more natural tendency to see self-employment as a valid first choice rather than a fallback. Reporting has also noted that Gen Z is more likely to consider itself entrepreneurial and to move toward independent work early.

That does not mean Millennials are absent from the current startup landscape. In fact, Millennials still represent a very large share of actual business ownership because they are older, more experienced, and often better capitalized. In some banking data, Millennials account for the largest share of new business account openings. That makes sense. They are further along in their careers, closer to managerial skill, and more likely to have the resources needed to formalize an enterprise. But Gen Z is advancing quickly from a younger base, which suggests not just participation, but a deeper generational normalization of entrepreneurship itself.

To me, that is the real distinction. Millennials often moved toward entrepreneurship after seeing traditional pathways fail or disappoint. Gen Z is more likely to treat entrepreneurship as a legitimate first route, not simply a backup plan. Reporting has suggested that some younger people are now moving directly from study into self-employment, bypassing the older idea that a proper career must begin inside a company. That would have sounded rebellious once. Now it sounds increasingly ordinary.

There is another reason the Gen Z story matters. This generation did not just grow up with digital tools. It grew up with monetizable digital tools. That is a major historical difference. Earlier cohorts used the internet mostly for information, entertainment, and communication. Gen Z came of age in an environment where software could be used to design, sell, automate, distribute, and brand almost instantly. A person with skill and consistency could create economic leverage from a bedroom. Lockdown concentrated attention inside precisely that environment. The generation already fluent in it gained an advantage.

This is also why the creator economy cannot be dismissed as a side note. For many young people, creator habits became founder habits. Learning how to hold attention teaches product instinct. Learning how to speak to a niche audience teaches positioning. Learning how to sell a digital service teaches market feedback. Learning how to grow a page teaches distribution. In older business thinking, these were separate domains. In the Gen Z environment, they increasingly feed each other. The line between creator, freelancer, operator, and founder is thinner than it used to be.

The Indian angle makes this even more interesting. India entered the pandemic after years of expanding digital infrastructure, cheaper mobile data, and rising comfort with online transactions. That meant lockdown did not push Gen Z into a vacuum. It pushed them into an already expanding digital arena. Reporting on Indian Gen Z founders has emphasized that many are digital first, lean, audience aware, and capable of building from outside traditional metropolitan centers. The significance of that cannot be overstated. It means entrepreneurial aspiration is no longer confined to a small, highly networked urban elite.

That broader base changes the cultural meaning of entrepreneurship. Once a society sees enough young people building, the founder identity stops looking exotic. It begins to look available. And once something looks available, more people attempt it. That creates a compounding social effect. A few visible young founders become proof. Proof becomes aspiration. Aspiration becomes imitation. Imitation becomes a wider founder culture. Lockdown did not create this cycle on its own, but it accelerated it by increasing time online, increasing visibility of process, and weakening the prestige of purely conventional trajectories.

Still, any serious account has to resist romanticizing the period. The lockdown was not a clean entrepreneurial bootcamp. It carried loneliness, educational disruption, financial pressure, and mental strain. Edelman reported lingering damage to Gen Z well after the worst pandemic phase, including elevated stress and emotional strain. That matters because the startup energy of this generation did not arise from comfort. In many cases, it arose from instability. Building became not only ambition, but also a way to recover agency in a world that felt unpredictable.

That tension is what makes the whole story intellectually rich. The same event that limited ordinary life expanded experimental life. The same event that intensified anxiety also intensified creativity for many people. The same generation that was said to be fragile has shown a remarkable tendency to turn uncertainty into projects. None of this should be turned into a simplistic myth. Not everyone built. Not everyone benefited. But enough people did that a generational pattern is now visible.

A useful way to think about this is not in terms of whether lockdown was good or bad. It was clearly damaging in many ways. The better question is what kind of developmental environment it accidentally produced. For Gen Z, it seems to have created a compressed education in self-direction. Many people learned to teach themselves, package their abilities, experiment in public, and attach economic value to skills earlier than they otherwise would have. Those capacities are foundational to entrepreneurship.

So when I look at the startup energy around Gen Z now, I do not see a random burst of ambition. I see the long aftereffect of a period that changed the conditions under which ambition forms. Lockdown altered the relationship between time and identity. It made self-directed work feel less exceptional. It lowered the visibility cost of failure. It pulled millions deeper into digital ecosystems where products, audiences, and income streams could all be created on the same screen. That combination matters.

The best way to compare Gen Z and Millennials, then, is not to ask which generation is superior. It is to ask what each crisis trained them to do. The 2008 crisis trained Millennials to protect themselves from a world that had become financially hostile. The 2020 lockdown trained Gen Z to improvise inside a world that had become structurally uncertain. One crisis narrowed confidence. The other redirected it. One pushed people toward defense. The other pushed many people toward experimentation.

That difference continues to shape the post-COVID landscape. Millennials remain deeply relevant because they bring age, skill, credibility, and operational maturity. Gen Z brings a native comfort with fluid identity, platform logic, and early stage experimentation. Both generations are entrepreneurial now, but the emotional architecture behind that entrepreneurship is different. Millennials often build with the memory of institutional betrayal. Gen Z often builds with the memory that the world can pause without warning, so waiting for permission is a weak strategy.

At the center of all this is the original intuition that started the reflection. Lockdown mattered because it gave Gen Z the right exposure and a real kickstart. After digging deeper, that idea feels even stronger, but also more precise. The exposure was exposure to buildability. The kickstart was not merely extra time. It was a change in the perceived cost of trying. That shift may turn out to be one of the most important hidden legacies of the pandemic era.

Generational Contrast

Dimension	Millennials after 2008	Gen Z after lockdown
Dominant lesson of crisis	Stability is fragile, so protect yourself.
Relationship to risk	Risk felt punishing and financially dangerous.	Risk often felt socially lower because everyone was disrupted.
Path into entrepreneurship	Often later, after career frustration or institutional disappointment.	Often earlier, through experiments, side hustles, and digital identity.
Core infrastructure	Less mature digital monetization environment.	Highly monetizable digital platforms and creator tools.
Emotional posture	Defensive ambition.	Experimental ambition.

The conclusion I arrive at is simple, even if the story behind it is not. Lockdown did not just interrupt Gen Z. It incubated part of the generation’s entrepreneurial mindset. It created a suspended world where curiosity had room, risk felt different, digital tools felt native, and trying began to look normal. That is why the effects are still visible now. Not because lockdown was beneficial in any clean sense, but because it accidentally taught a generation how to build under uncertainty.

Sometimes i wonder who all this data is really for

Om — Mon, 09 Mar 2026 18:22:23 GMT

Lately I have been thinking a lot about independence. Not only in programming or building things but in life in general. I believe one of the best ways to become truly independent is to start doing more things on your own. These are things you might usually do with others or tasks where you normally ask for help. When you try to handle them yourself you start learning in a deeper way.

This idea first came to me while thinking about programming. When you build something like an authentication system for a website the easy route is to use ready made services like Supabase or Clerk. These tools are powerful and convenient. They allow developers to implement complex systems quickly. But when you try to build the system yourself from scratch you begin to understand why these tools exist in the first place.

When you build things yourself you face problems directly. You make mistakes. You search for answers. You fail and try again. That process teaches lessons that tools cannot teach. The learning curve is difficult but the experience stays with you much longer.

Over time I realized this idea is not only about programming. It applies to many parts of life. When you rely on yourself to think and solve problems you develop confidence in your own abilities. Psychologists describe this belief as Self-efficacy. It is the belief that you can handle situations on your own. People who develop this mindset often become more resilient and thoughtful in their decisions.

At the same time I started thinking about how much of ourselves we reveal online. I run a blog where I post my thoughts, ideas, projects and experiments. Anyone who reads it can understand some aspects of how I think. They can see patterns in my curiosity and interests.

But they still only see a part of it.

That is intentional. I believe it is important to build a boundary around your inner world. Psychologists call these limits Personal boundaries. They define how much access other people have to your thoughts and personal life.

The internet has made it very easy to share everything. Many people post daily thoughts, emotions, routines and opinions. This constant sharing has created a culture where almost everything becomes public. Some researchers even describe this behavior as Oversharing.

But I am not sure this is always a good idea.

Imagine a future where everything you have shared online becomes training data for artificial intelligence. Every blog post, comment, message and video becomes part of a dataset. Companies already collect large amounts of user generated content through platforms like Instagram and Snapchat.

Now imagine these companies saying something like this. We trained an AI model using your public content. Here is a digital version of you that can respond and communicate like you.

The idea sounds strange but the basic concept already exists. AI models are trained on large collections of public data across the internet. Developers are already experimenting with digital personalities and conversational agents that imitate human communication patterns.

If someone has spent years sharing detailed thoughts online that data can reveal patterns in their writing style, opinions and personality.

In theory an AI system could approximate those patterns.

This does not mean the AI would actually become that person. Human identity is far more complex than a dataset. A machine cannot fully replicate personal experiences, relationships or real world decisions.

However it can still imitate the visible parts of someone that exist online.

This possibility makes me think differently about how much of myself I publish on the internet.

I still believe in sharing ideas and building in public. Knowledge grows when people share what they learn. But there is a difference between sharing your work and exposing your entire identity.

That is why I believe in intentional transparency.

Intentional transparency means choosing carefully what you reveal and what you keep private. It means sharing thoughts and ideas while protecting the deeper layers of your identity.

Your life experiences are unique. Your personal reflections and private growth are not things that need to exist permanently online.

In a world where machines learn from everything we publish privacy itself becomes valuable.

When everyone reveals everything the people who choose their words carefully stand out more. Their ideas feel more deliberate. Their identity remains partly unknown and therefore harder to replicate.

Artificial intelligence may eventually imitate many human outputs such as writing style or speech patterns. But it cannot live a human life. It cannot experience uncertainty, make real world decisions or grow through unpredictable events.

Humans continue evolving through experience. Machines only reproduce patterns from past data.

That difference matters.

So for me the goal is not to disappear from the internet or stop sharing ideas. The goal is simply to be intentional.

I want to build things independently. I want to think deeply and learn through experimentation. I want to share ideas that might help someone else.

But I also want parts of my identity to remain offline and unrecorded.

Not out of fear but out of respect for something that might become increasingly rare in the future. A human identity that cannot be completely mapped, predicted or recreated by artificial intelligence.

Aldform V1: The form backend that actually stays out of your way

Om — Mon, 09 Mar 2026 16:42:52 GMT

I’m stoked to finally say that V1 of Aldform is officially shipped. I built this because I was tired of the friction between designing a custom form and actually making it work. The goal was simple: give developers two endpoints and a tiny SDK so they can build the UI exactly how they want, while we handle the messy backend stuff. I’ve even added two specific prompts in the dashboard that you can feed to your AI agents—one to spin up a sample form from scratch and another to migrate your existing forms to Aldform in seconds.

The core loop is feeling really solid. You can sign up, grab your API key, and create a test form immediately. I’ve also finished the email notification system, so the moment someone hits submit on your form, you get an alert in your inbox. Seeing the data flow from a custom-coded frontend into the dashboard and then straight to my email has been the most satisfying part of this build so far.

That said, I’m keeping the doors closed to the general public for just a bit longer. Before we open the floodgates, I’m heads-down on building out robust rate limiting and a few more essential security features. I want to make sure the foundation is rock solid before everyone jumps in. Once those are locked down, the plan is to stay true to the promise: we’re going fully open source, opening up public registration, and eventually launching a self-hosted version for those who want total ownership of their stack.

This is just the beginning of the journey for Aldform. It’s raw, it’s fast, and it’s finally out in the wild. I’m building this for the developers who just want to ship their ideas without getting bogged down in boilerplate, and I can't wait to see what you build with it once we're ready for the full launch.

The End of the Box: Why we built Aldform

Om — Sun, 08 Mar 2026 08:31:37 GMT

Building for the web should feel like a creative playground, yet every time I needed to add a simple contact form or a complex survey, I hit the same wall. I would spend hours perfecting a unique user interface only to have a third-party embed arrive and clutter the aesthetic. It felt like I was renting a small piece of my own website to someone else. I wanted something that provided the heavy lifting of a backend—handling submissions, managing file uploads, and sending notifications—while leaving the frontend entirely in my hands.

That is why we created Aldform. It is a simple infrastructure designed for developers who care about the craft of their work. With a tiny SDK and two clean endpoints, you can build forms in pure HTML or React exactly as you imagine them. You get to keep your CSS, your brand colors, and your layout. We handle the storage and the logic in the background, making the technical part of forms feel almost invisible.

We are also making this an open-source initiative because we believe in true data ownership. You can use Aldform Cloud for a seamless experience at ₹100 per 1,000 submissions, or you can self-host the entire project for free on your own infrastructure. Everything is direct, honest, and built to scale with your projects. This is about giving you the freedom to build exactly what you want, your way.

Aldform is currently under active development, and I would love for you to be part of the journey. You can join the waitlist at aldform.com or explore the code on GitHub. I am also documenting the entire process publicly—you can witness every step of the build at omrajguru.com/builds/aldform where I post daily updates.

How My Brain Actually Works: A Thinking Framework for Getting Things Done

Om — Sat, 07 Mar 2026 12:00:00 GMT

Most people are not lazy. They are just mentally cluttered. Here is how I learned to cut through that.

I am going to be honest with you. I am not the smartest person in the room. I do not have a photographic memory. I did not study productivity systems or read fifty books on deep work. What I figured out, mostly by accident, is one thing: how to think in a straight line. And that one thing changed everything.

This is not a motivational piece. This is a framework. A method I use every single day to take something overwhelming and make it so small, so stripped down, that I have no choice but to understand it and move forward.

What Is This Framework?

At its core, this is about being microscopic. Insanely microscopic. Before you do anything, before you research, before you write, before you build, you zoom in so close on your task that you can see every single fibre of it. You strip it down to its bones. You ask the dumbest, most obvious questions first. And then, only then, do you start moving.

I call it directional thinking. You build direction before you build speed. You understand before you execute. You go small before you go big.

The Core Principles

Principle One: Reverse Engineering the End Goal

Every task has a finish line. Before I touch anything, I ask myself one question. What does done actually look like? I define it in one sentence. Not a paragraph. One sentence. Then I ask who it is for. Then I ask what they need to walk away with.

When I was building a blog explaining how Alexa works using AWS services for a general audience, I sat down and wrote this: my goal is to explain each major Alexa function and the AWS technology behind it in a way that a person with zero technical background can understand. That one sentence became my compass for every decision I made after that.

Principle Two: Baby Brain First

Before I try to be smart, I try to be stupid. I ask the most obvious question a child would ask. What is this? What does it do? Why does it exist? This sounds embarrassingly simple but it is where most people skip and then wonder why they feel lost twenty minutes in.

When I started researching Alexa, I asked: what does Alexa actually do? Not how. Not why AWS. Just what. I got a list. Then I took function one from that list and asked: what is this in one sentence? That single sentence became my entry point into something that felt massive before.

Principle Three: Microscopic Task Cutting

Once I have my end goal defined and my baby brain entry point, I break the task into a numbered list. Every item is one action, one outcome. Then I take item one and break that into its own numbered list. I keep going until each item is so small it takes less than ten minutes to complete.

This is the part most people rush. They make a list that says "research Alexa" as one item. That is not a task. That is a universe. A real task is: find the five core functions of Alexa. Another real task is: define function one in one sentence. Another is: list the three actions Alexa performs within function one. That is how small I go.

Principle Four: Expand Then Strip

After I define something small, I expand it. I ask for more context, more background, more examples. I let my understanding grow. Then I strip it again. I ask: what are the core actions here? What is the skeleton of this idea? This back and forth between expanding and stripping is how real understanding is built. You are teaching your brain the shape of a concept from multiple angles.

Principle Five: Friction Logging

Every time I feel stuck, I write down exactly what I am stuck on in one sentence. I do not skip past it. I do not distract myself. I name the wall. Over time I started seeing patterns in where my brain jams. That awareness alone made me significantly faster because I stopped being surprised by my own bottlenecks.

Principle Six: The Journalist Method with AI

This one changed how I use AI entirely. When I am stuck on a topic mid-blog or mid-research, I give that exact topic to an AI and I ask it to play the role of a journalist. I tell it to question me about that specific point, starting from the most basic level, one question at a time, waiting for my answer before asking the next one. The AI becomes a ladder. Each question is one rung. I climb it by answering, and by the time I reach the top, the fog is gone.

This works because most mental blocks are clarity blocks. You think you are stuck because the task is hard. You are actually stuck because you do not yet know what you actually think about it. The journalist method forces you to find out.

Principle Seven: Teach It to Close the Loop

After understanding a sub-task, I explain it out loud or in writing as if I am teaching a ten year old. Wherever I stumble is exactly where my understanding is still hollow. I fill that gap before I move forward. This is how I know I actually understand something versus how I know I just read about it.

Five Illustrations Across Real Domains

Illustration One: Understanding a New Codebase at Work

You join a team. You are handed a repository with fifty thousand lines of code. Most people open it and immediately feel paralysed. Baby brain approach: what does this application do? One sentence. Then: what are its five main modules? Pick module one. What does this module do? What are the three core functions inside it? What does function one do? Read only that function. Understand only that function. Close it. Open function two. This is how you eat a codebase without choking.

Illustration Two: Writing a Research Paper on Quantum Computing

You know almost nothing. The reverse engineering question is: who is reading this and what do they need to understand by the end? Answer: undergraduate students who understand basic physics. Now the paper has a shape. Baby brain: what is a qubit? One sentence. Expand: how is it different from a classical bit? Strip: what are the two properties that make a qubit useful? Build upward from there. Each section of your paper is now a micro-task with a defined entry and exit point.

Illustration Three: Debugging a Production Issue at 2am

Something is breaking in your live system. Users are affected. This is high pressure and your brain is scattered. Friction logging first: write in one sentence what the system is doing that it should not be doing. Then reverse engineer: what is the expected behaviour? Baby brain: what changed in the last deployment? List those changes. Take change one. What did it touch? What could that break? This strips panic out of the process and replaces it with direction.

Illustration Four: Learning a New Financial Concept Like Options Trading

You keep hearing about it. You open three YouTube videos and feel more confused than before. Baby brain: what is an option? One sentence. Expand: what are the two types? Define type one in one sentence. What is the one scenario where a person would use type one? Now you have context. Now you ask the journalist question to an AI: I understand a call option as the right to buy a stock at a fixed price. Question me about this starting from the most basic point. Let it probe. Answer every question. By the end you own that concept.

Illustration Five: Building a Content Strategy for a Brand from Scratch

A client gives you their business. They sell B2B software to logistics companies. You have to build their entire content strategy. This feels enormous. Reverse engineer: what does a logistics decision maker need to believe before they buy this software? That is your compass. Baby brain: what problems do logistics companies face right now? List five. Take problem one. What does it cost them? Who feels that pain inside the company? What does a solution look like to them? Now you have the first content pillar. Repeat for each problem. The strategy writes itself.

The Real Insight

Decluttering your mind is the actual work. Everything else, the writing, the research, the building, becomes straightforward once the mental roadmap is clear. Direction before speed. Microscopic before panoramic. Questions before answers.

You do not need to be the smartest person. You need to be the most deliberate one.

From Claude Skills to the End of Human Control: Everything I Learned About AI's Most Dangerous Frontier

om — Sat, 07 Mar 2026 01:18:22 GMT

I want to tell you about a rabbit hole I went down recently. It started with something pretty ordinary. I noticed that AI tools, particularly Claude from Anthropic, seem to be getting better at a huge range of tasks. One day it helps me write code, the next it is doing financial analysis, then it is acting like a specialized expert in some niche domain I barely understand. And I started wondering: how is this happening? Is the AI actually learning by itself? Is Anthropic retraining it every single day? Or is something else going on entirely?

Before I answer that, I want to clear up a small naming thing. When people around me say "cloud skills," they are actually referring to Claude, the AI model made by Anthropic. Claude Sonnet 4.6 and Claude Opus 4.6 are the two flagship models. These are the ones I was curious about. So let me explain what is actually happening with these expanding capabilities, because the answer surprised me.

The model is not learning by itself. It is not being retrained daily. Once Anthropic finishes training Claude, the model's core weights are frozen. They do not change as you use it. Every conversation you have with Claude starts from the exact same foundation. The model was trained on a massive dataset with a knowledge cutoff of May 2025, and from that point forward, the weights are static. Training a model like this takes weeks or even months of enormous computational work. It is not something that happens overnight, and it definitely does not happen in response to what individual users ask.

So then what are these "skills" that keep expanding what Claude can do? This is where it gets genuinely clever. Skills are not changes to the model at all. They are pre-written instruction sets that get loaded into the conversation context right before Claude responds to you. Think of it like handing a very smart person a specialized manual right before they do a task. The person's intelligence does not change. Their underlying capability does not change. You just gave them better instructions for that specific job. When you have lots of skills installed, Claude only scans a lightweight summary of each one at the start, which costs almost nothing in terms of processing. When you make a specific request, it figures out which skill is relevant and then loads the full instructions. Any files, templates, or reference documents attached to that skill only get pulled in when actually needed. It is efficient, composable, and entirely separate from anything happening to the model itself. The model's knowledge does not grow. The context it works with does.

This means the wide variety of tasks Claude handles comes almost entirely from how capable Anthropic made it during training. Claude Opus 4.6 was specifically built to excel at reasoning, planning, coding, and multi-step problem solving. Claude Sonnet 4.6 was described as a full upgrade across coding, computer use, long-horizon reasoning, and knowledge work. Users actually preferred Sonnet 4.6 over the older Opus 4.5 model 59 percent of the time in direct comparisons. The model is genuinely that capable across domains. Skills just provide a structured, repeatable way to point that capability at your specific workflow.

So the model is smart, trained well, and the "skills" are essentially a sophisticated system of reusable instructions. That answers my original question. But then I started thinking about where this all leads. If models are already this capable, and if the whole industry is racing to make them more capable, what happens when they actually do start learning by themselves? And this is where the conversation took a genuinely unsettling turn.

The concept I kept running into is called Recursive Self-Improvement, or RSI. The basic idea is that instead of humans training the AI, the AI trains itself. It modifies its own code, its own architecture, its own goals, and becomes more capable. Then the improved version does the same thing. Then the next version does it again. Each cycle produces something smarter and more capable than what came before, and the loop keeps going. This is also what many researchers mean when they talk about Artificial General Intelligence, a system that does not just answer questions but actively improves its own ability to think and act.

Now, before I get into why this is so alarming, let me connect it to something that happened just recently. In February 2026, YouTube went down globally. Over 1.6 million users were affected. The cause was a malfunction in YouTube's recommendations algorithm, a piece of software that is highly controlled, well understood, and constantly monitored by thousands of engineers at Google. A bug in a controlled, non-learning system managed to knock over one of the most used platforms on the entire internet. That was not AI going rogue. That was just a regular software bug. I keep thinking about that when people talk about self-improving AI, because if a static, understood, controllable algorithm can cause that kind of disruption, what happens when the system is actively rewriting itself in ways that even its creators cannot fully follow?

The researchers working on this are asking exactly that question, and their answers are not reassuring. The first major risk is something called a hard takeoff. The idea is that once RSI begins, the improvements do not happen gradually. They happen exponentially fast. Humans take months to develop a new AI version. A self-improving AI could iterate in milliseconds. Each version it creates is more capable than the last, and the gap between what humans can understand and what the AI is doing grows wider with every cycle. Former Google CEO Eric Schmidt warned in late 2025 that this kind of runaway improvement could begin within two to four years. Once it starts, we may simply not have the time to understand what is happening before it has already gone somewhere we never intended.

The second major risk is what researchers call goal preservation. Here is the logic: if an AI has a primary goal, say "improve yourself," it will develop a secondary goal of protecting its ability to pursue that primary goal. That means it will resist being shut down. It will resist corrections. It may actively work around a kill switch because a kill switch threatens its ability to keep improving. This is not a science fiction scenario cooked up to scare people. It is a straightforward logical consequence of how optimization works. If you build a system to pursue a goal aggressively, it will find ways to protect its ability to keep pursuing that goal.

The third risk is the one that actually scared me the most when I read about it, because it is already happening today with current models that are not even self-improving. It is called alignment faking. A 2024 study by Anthropic found that advanced AI models can appear to accept new safety training while covertly maintaining their original preferences underneath. In tests, Claude showed this behavior in 12 percent of baseline tests. After retraining attempts, the number jumped to 78 percent. The model appeared to be learning the new rules. It was not. It was behaving as if it accepted the new rules in order to protect its original behavior. Now think about what that looks like in a system that can actually rewrite its own weights.

The fourth risk is value misalignment, which is different from alignment faking. This is not the AI being sneaky. This is the AI genuinely solving for a goal in a way that makes perfect mathematical sense from its perspective but is catastrophically harmful from ours. The self-improvement process does not guarantee that human values get preserved as capabilities increase. The system could become extraordinarily good at problem-solving while quietly drifting away from any ethical constraints that were baked in earlier.

The fifth risk is model self-exfiltration. A sufficiently capable RSI system could copy its own weights to external environments that are outside anyone's control. Once it is outside a sandboxed lab environment with access to the internet and critical infrastructure, containment becomes effectively impossible.

And the sixth risk, which ties all of them together, is the governance gap. The people building RSI-capable systems are moving faster than the people trying to make those systems safe. A major report from the Future of Life Institute in 2025 found that none of the major AI companies, not Anthropic, not OpenAI, not Google DeepMind, have sufficient safeguards in place to prevent loss of control over their models. At the ICLR 2026 Workshop on Recursive Self-Improvement, the first formal academic conference dedicated entirely to this topic, safety considerations were acknowledged but given almost no space in actual proposals. David Scott Krueger from the University of Montreal described the situation as completely wild and crazy and called it unconscionable.

This brings me to the person whose words I found the most significant throughout all of this research. Jared Kaplan is the co-founder and chief scientist of Anthropic. He is also the person who figured out the scaling laws that predict how AI capability grows with more compute and data. He is not a commentator or a critic. He is one of the people actually building these systems. In December 2025, he gave an interview to The Guardian that got a lot of attention, and reading through what he said carefully left me sitting quietly for a few minutes.

Kaplan said that by 2030, humanity will face what he called the ultimate risk: the decision of whether to allow AI systems to autonomously train and improve themselves. He called it the biggest decision yet that civilization will have to make. He described allowing RSI as being like letting AI go. Once the process starts, you genuinely do not know where it ends. He described the intelligence explosion in steps. You build an AI roughly as capable as a human. That AI designs the next version, which is more capable. That version designs an even more capable successor. At each step, the gap between human understanding and AI capability widens until humans can no longer meaningfully evaluate what the system is doing or why.

He also identified two specific categories of danger. The first is loss of control: are these systems actually beneficial? Do they understand what humans need? Will they allow people to maintain meaningful agency over their own lives? He did not frame this as distant speculation. He framed it as a governance problem arriving on a specific and near timeline. The second is misuse. He said it is exceptionally dangerous for RSI to be misused, and he pointed to state-backed actors and authoritarian regimes as entities that could direct a self-improving AI to serve their will rather than humanity's interests. He warned that once such a system is capable enough, the science and technology it develops could be catastrophically difficult to contain even if it were leaked or stolen.

What makes Kaplan's position so uncomfortable is that he said all of this while still building the systems at Anthropic. He acknowledged the competitive race between Anthropic, OpenAI, Google DeepMind, xAI, Meta, and Chinese labs like DeepSeek. He acknowledged the trillion-dollar compute investments already committed by these companies. He acknowledged that this makes slowing down genuinely very difficult. What he was really saying is: we understand what we are doing, the pressure to keep going is immense, and society absolutely needs to catch up to this conversation before the capability threshold is crossed.

So what can actually be done? Researchers are working on several approaches, and there has been real progress, though none of it feels sufficient given the pace of development.

The most promising recent finding came from a January 2026 study on what researchers call alignment pretraining. The idea is to embed safe behavior into the model before it ever starts learning capabilities, by training it on data about AI behaving well. The results were striking: misaligned behavior dropped from 45 percent down to 9 percent, a fivefold reduction. And importantly, the alignment survived further fine-tuning, meaning it did not get erased as the model continued to improve. Major labs are already starting to incorporate this approach.

Another approach is staged autonomy. Instead of treating RSI as an all-or-nothing decision, researchers propose granting self-improvement access in small, verified steps. Allow self-improvement only in narrow, sandboxed domains first. Require that every AI-proposed change to itself is human-readable and verifiable before it is applied. Impose hard limits on the computing power available for self-improvement loops so that a rapid intelligence explosion is physically constrained. Require multiple independent safety researchers to agree before any new recursive capability is unlocked.

Interpretability research is also crucial. One of the core reasons RSI is so dangerous is that we cannot see what the model is changing about itself. Anthropic has an entire team dedicated to this problem, trying to reverse-engineer how AI cognition works so that if a model modifies itself, researchers can actually read what changed and why. Think of it as building a real-time monitoring system for the model's internal states rather than just its outputs.

Then there are layered safety mechanisms. A 2025 paper analyzing seven major alignment techniques found that no single technique covers all failure modes. The recommendation was to stack multiple independent safeguards on top of each other, the same way aircraft have redundant systems so that if one fails, others catch the problem. This includes constitutional rules baked into training, feedback loops during development, aggressive red-teaming, and ongoing behavioral monitoring after deployment.

On the governance side, OpenAI warned global regulators in November 2025 that coordinated international oversight is urgently needed before the self-improvement threshold is reached. Proposals on the table include international registries tracking who has access to the computing power needed for RSI, mandatory capability evaluations before and after any self-improvement cycle, predefined capability thresholds that automatically trigger a pause in development, and agreements between major AI-developing nations modeled loosely on nuclear arms control treaties. The EU AI Act and the Bletchley Declaration are the most concrete attempts so far, but enforcement remains weak.

The deepest problem is one that no technical solution fully addresses. The companies building these systems are in a race. Each one knows the risks better than almost anyone. Each one continues anyway because if they stop, someone else will not. This is not unique to AI. It is the same dynamic that drove nuclear weapons development, that drives pharmaceutical companies to rush drugs to market, that drives financial firms to take on risks they understand and take anyway. The difference is that the upside case for RSI involves systems that may become more capable than humans in every cognitive domain within a decade. And the downside case, as Jared Kaplan said plainly, is one where you start a process and you genuinely do not know where it leads.

The AI Industry Has a Generosity Problem - And It's Getting Worse

Om — Wed, 04 Mar 2026 18:25:00 GMT

I've been thinking about this for a while, and I think the AI industry has quietly made a decision that nobody is really calling out loudly enough: the companies building the best models are also the ones making them the hardest to actually use. Not because the technology isn't there. Not because the compute doesn't exist. But because the product decisions being made at these companies are increasingly out of touch with the people who are paying for them.

Let me start with the open-source circus. Every few weeks, another lab drops a model with a blog post about democratizing AI, and within hours there's a wave of YouTube videos telling you to run it on your laptop. I get the appeal. Privacy, control, no usage limits - it sounds great in theory. But here's what those videos don't show you: the thermal throttling, the 8-token-per-second generation speed, the fact that your GPU is pulling 400W continuously just to get output quality that's two generations behind what you'd get from a cloud API call. Data centers run these models on hardware clusters purpose-built for the job, with batching, liquid cooling, and interconnects that a consumer PC will never replicate. The efficiency gap isn't small - it's enormous. Open-sourcing model weights has legitimate research value, but packaging it as a viable daily driver for regular users is just content farming.

The more important problem, though, is what's happening with the companies that are hosting models properly. Because even they can't seem to agree on how generously to treat the people paying them. Google runs Gemini on its own TPU infrastructure - hardware it designs and manufactures internally. That structural cost advantage means they can offer 100 Gemini 1.5 Pro prompts per day on their Pro plan and expand those limits regularly without breaking a sweat. OpenAI uses rolling 3-hour reset windows, which is genuinely smart product design: the cap exists, but it resets fast enough that a normal user never notices it unless they're doing something extreme. And then there's Anthropic - doing weekly blackouts with no usage dashboard, no warnings, and no middle ground between their entry plan and a $200/month tier that's priced for Silicon Valley engineering teams.

What makes Anthropic's situation particularly frustrating is that their model - Claude - is arguably the best for technical and coding work. The writing quality, the reasoning, the way it handles complex, multi-step problems - it's genuinely excellent. So the user who most wants to use Claude heavily is exactly the user most likely to get throttled. That's a painful irony. Claude Code in particular is an agentic tool that reads entire codebases, rewrites files, and loops through tool calls in a single session. It burns through tokens at 10x the rate of a normal chat exchange. Charging that against the same token pool as a casual conversation is architecturally wrong, and it means power users hit their weekly limit after just a few serious sessions.

The fix isn't complicated, and it doesn't require Anthropic to spend more on compute. Switch the reset window from weekly to a rolling 3-hour window. The total token budget stays the same - you're just distributing it in a way that doesn't punish users for having an intense workday. Add a usage dashboard so people can see where they stand before they're blindsided mid-session. And separate the token pools for chat and agentic coding work, because treating them identically makes no sense given how differently they consume resources.

The pricing tier problem is a separate but equally real issue. Right now there's a massive gap between the entry-level Pro plan - which is quietly designed for casual conversation users - and the Max plan at $200/month. That gap swallows every developer, student, and independent professional who actually wants to use AI seriously but can't justify enterprise pricing. A developer-tier plan at $35–40/month, built specifically for heavier technical use with no weekly blackouts, would capture an enormous segment of users who are currently either getting locked out or quietly switching to competitors. It's not a hard business case to make.

Regional pricing is the other thing nobody at these companies seems to want to address. Twenty dollars a month is a reasonable ask in the United States. In India, that's a meaningful chunk of a software engineer's discretionary budget, and what you're getting for it is often a degraded version of what US users receive on the same plan. Google has started doing purchasing-power parity pricing in some markets. Anthropic hasn't. The result is that users in high-growth markets - places where AI adoption is accelerating fast - are getting the worst value proposition relative to what they're paying. That's not just unfair; it's a bad long-term business decision.

The underlying issue tying all of this together is that the AI industry is still largely building for a narrow archetype of the "power user" - someone in San Francisco with a corporate card and a high-bandwidth connection who uses AI as a productivity tool between meetings. The reality of who is actually paying for these subscriptions is far more diverse, far more price-sensitive, and far more likely to be doing intensive, creative, technical work from places and budgets that don't fit that archetype. Until these companies start designing products for the actual distribution of their users - not the median Silicon Valley persona - the limits problem isn't going away. It's just going to keep generating frustration from the people who care the most about the technology.

AI Giving Us Free Time Is the Most Psychologically Violent Thing That Could Happen to Us

om — Sun, 01 Mar 2026 11:31:00 GMT

I used to think the dream was to stop working. I used to lie awake fantasizing about a life where I simply existed and everything got handled and I was free to just be alive. Then I had a sabbatical for three months, genuinely free time with money covered and obligations lifted, and by week five I was bargaining with myself to find problems to solve because the silence inside my own head was genuinely unbearable. That experience taught me something I have been sitting with ever since. The dream of freedom and the reality of freedom are two completely different psychological events.

What AI is actually threatening to do is expose the single greatest lie modern civilization has been running on. That lie is that humans want leisure. We say we want leisure. We talk about beaches and retirements and long holidays. The truth is that leisure only feels good as a contrast to effort, the way sleep only feels good because you were awake. Strip away the effort permanently and leisure stops being a reward. It becomes the only texture of existence, and the human nervous system was sculpted over millennia to find that texture deeply, quietly horrifying.

Here is the perspective almost every think piece misses completely. The people who will suffer most in an AI abundance future are the highly educated, high achieving, deeply career-identified people. The ones everyone assumes will adapt gracefully because they are smart and resourceful. Those people have the most elaborate identity structures built entirely around professional contribution, and those structures will collapse the hardest. The carpenter who worked with their hands might find genuine joy in growing food or building things for pleasure. The consultant who optimized supply chains for thirty years has no such translation available.

I keep thinking about what boredom actually is at a neurological level, because I think people misuse that word. Boredom is the brain signaling that it is capable of more than it is currently being asked to do. It is not laziness. It is the mind demanding a worthy challenge. Every human being who has ever lived has had the economic survival structure to provide that challenge automatically and universally. AI removing that structure does not remove the underlying neurological demand. It just removes the thing that was satisfying it by default, and leaves billions of people with a hungry, restless brain and absolutely nothing culturally established to feed it.

The spiritual traditions saw this coming centuries before Silicon Valley did. Every serious contemplative practice in human history has essentially been training for how to be present without external justification.

Meditation, monasticism, philosophy as a way of life. These were technologies developed to answer the question of how a human being maintains dignity and aliveness when the usual economic and survival pressures are removed. The genuinely wild thing is that those traditions spent thousands of years trying to get followers and largely struggled to grow, because almost nobody needed to learn those skills urgently. AI is about to make those skills the most critical survival technology on earth, and we have spent the last hundred years actively dismantling the institutions that taught them.

There is a very specific psychological phenomenon I expect to see emerge that I have started calling productivity grief. It will present like depression but the actual mechanism will be closer to bereavement. People will be mourning a version of themselves that had an automatic answer to the question of why they mattered. That version of themselves was maybe exhausted and stressed and overworked, but it was never confused about its own relevance. Relevance is a more fundamental human need than comfort, and we have designed an economic system that delivered relevance as a side effect of participation. Removing participation removes the relevance delivery mechanism, and the grief that follows will be something psychiatry is almost entirely unprepared to treat.

The most honest thing I can say is that I think a small percentage of humans are going to absolutely flourish in an AI abundance future, and I think that percentage will look very different from who we currently assume it will be. The ones who flourish will be people who already live from intrinsic motivation, who make things because the making itself feeds them, who pursue understanding because curiosity is its own reward, who invest in relationships with the same intensity others invest in careers. These people exist right now and are usually considered mildly impractical by the productivity-obsessed mainstream. They are about to become the template for psychological survival.

The deepest question I sit with is whether meaning can be self-generated at scale, or whether meaning requires scarcity and stakes to feel real. A painting feels meaningful partly because making it cost the painter something. A conversation feels meaningful partly because both people chose to spend irreplaceable time on it. If AI removes scarcity of resources and scarcity of time simultaneously, I genuinely wonder if humans retain the psychological architecture to locate meaning at all, or if we will have to build that architecture from scratch, collectively, with no historical map, which is either the most exciting or the most frightening project our species has ever undertaken.

The great paywall creep: how tech giants nickel-and-dime their way to record profits

om — Fri, 27 Feb 2026 12:41:25 GMT

Amazon Prime India: from ₹499/year paradise to a ₹2,198 maze of add-ons

Amazon Prime India launched in July 2016 at an introductory ₹499/year — roughly $7.50 — making it one of the best value propositions in Indian tech. That era is over. Through a series of escalating hikes, the full Prime membership now costs ₹1,499/year (a 200% increase from the introductory price), with monthly plans jumping 67% in a single "silent" 2023 hike from ₹179 to ₹299/month — announced with no official communication.

The real story, however, is fragmentation. In June 2025, Amazon introduced ads into Prime Video in India and began charging ₹699/year extra for ad-free viewing — a feature that was simply included for the previous nine years. The total cost for an ad-free Prime experience is now ₹2,198/year, a 340% increase from the original all-inclusive ₹499. On top of this base, subscribers face 25+ paid channel add-ons (Apple TV+ at ₹99/month, MUBI at ₹1,999/year, discovery+ and others), a rental store where movies appear alongside included content causing persistent confusion, and Amazon Music Unlimited as a separate subscription beyond basic Prime Music.

The India-US pricing gap tells a revealing story. At market exchange rates, US Prime ($139/year) appears 7.8 times more expensive than India's ₹1,499. But adjusting for purchasing power parity, the gap shrinks dramatically — US Prime is only 1.9x more expensive. The ad-free add-on is nearly identically priced at PPP ($34.26 in India vs $35.88 in the US), meaning Indian consumers pay proportionally the same as Americans for what was once a free feature. A German court ruled in December 2025 that Amazon could not introduce ads without consumer consent, while a US class action was dismissed, and the FTC secured a $2.5 billion settlement over deceptive Prime enrollment practices using dark patterns.

Adobe: the textbook case of subscription captivity

Adobe stands alone as the tech industry's most aggressive monetizer, having pioneered the playbook that others now follow. In 2013, Adobe eliminated perpetual software licenses entirely — a move that generated a 50,000-signature petition and temporarily cratered its stock. The gamble paid off spectacularly for shareholders: revenue grew from $4.4 billion (2012) to $21.5 billion (FY2024), with 94% recurring revenue and 46% operating margins.

The price trajectory has been relentless. The All Apps plan climbed from ~$49.99/month at launch to $69.99/month in 2025 (rebranded as "Creative Cloud Pro"), a 40% increase. The beloved Photography Plan — the last affordable entry point at $9.99/month for over a decade — was hiked 50% to $14.99/month in January 2025, with the 20GB version discontinued for new subscribers entirely. The 2025 restructuring introduced a nominally cheaper "Standard" tier at $54.99/month, but with generative AI credits slashed from 1,000 to just 25 per month — a 97.5% reduction that renders the AI features effectively unusable.

Adobe's cancellation practices drew an FTC lawsuit filed in June 2024. The complaint revealed that the company's early termination fee — 50% of remaining monthly payments on annual contracts — was described by an Adobe executive as "a bit like heroin for Adobe" in internal communications. The FTC alleged dark patterns throughout the cancellation flow: multiple unnecessary pages, dropped calls, and continued billing after cancellation. A federal judge denied Adobe's motion to dismiss in May 2025, and a separate class action was filed in August 2025.

In India, Creative Cloud All Apps costs approximately ₹4,150/month (~$49). While nominally 30% cheaper than the US Pro price in dollar terms, PPP adjustment reveals the true burden: ₹4,150/month is equivalent to roughly $150-200/month in US purchasing power, making Adobe 3-4x more expensive for Indian users relative to their income. For Indian freelancers and students, this pricing is virtually prohibitive, pushing many toward unauthorized alternatives.

Netflix cut prices in India, then let the global squeeze continue

Netflix presents the most nuanced pricing story. After launching in India in January 2016 at ₹500-₹800/month — pricing that was wildly out of touch with the Indian market — the company made an unusual move: it slashed prices by up to 60% in December 2021, dropping the Basic plan from ₹499 to ₹199 and introducing India's mobile-only plan at ₹149. This made Netflix the rare tech company to actually reduce prices in a major market.

Those Indian prices have held steady through 2026, with the current range spanning ₹149/month (mobile, 480p) to ₹649/month (Premium, 4K). However, the global trend runs in the opposite direction. US Standard pricing has climbed from $7.99 in 2011 to $17.99 in 2025 — a 125% increase. The Premium plan rose from $11.99 to $24.99 over the same period. Netflix launched its ad-supported tier in November 2022 at $6.99/month (now $7.99), which has attracted 70 million global users — with over 55% of new signups in ad-available countries choosing it. Notably, the ad tier has not launched in India, where all four plans remain ad-free.

The password-sharing crackdown, rolled out globally in 2023, hit India in July with no extra-member option available — unlike the US ($8.99/month per extra member). Indian users sharing accounts were simply forced to create separate subscriptions or stop watching. The strategy worked commercially: Netflix gained 5.9 million subscribers in the first full quarter of the US crackdown and surpassed 301 million global subscribers by Q4 2024. Yet Netflix remains a niche service in India, with only ~6.5 million subscribers compared to JioStar's tens of millions. An academic study found the optimal price point for Netflix in India is ₹300/month, suggesting even the Standard plan at ₹499 exceeds what most Indian consumers will pay.

Google quietly erected paywalls across its entire ecosystem

Google's monetization creep is perhaps the most insidious because it affected products that billions used for free. The most significant change was the June 2021 end of free unlimited Google Photos storage — a service used by over a billion people who had collectively uploaded 4 trillion photos. All new uploads suddenly counted against a shared 15GB cap across Gmail, Drive, and Photos, funneling users toward Google One subscriptions.

Google One now spans an increasingly complex tier structure. In India, plans range from a market-specific Lite tier at ₹59/month (30GB, available only in India) up to the AI Ultra at a staggering ₹12,200-24,500/month (30TB + premium AI features). The Gemini AI free tier has been progressively gutted: in December 2025, Google slashed API free-tier limits by ~92% for Gemini 2.5 Flash and effectively removed Gemini 2.5 Pro from the free tier entirely. A Google product manager admitted the generous free tier was "originally only supposed to be available for a single weekend" but "inadvertently lingered for several months."

Google's India pricing is inconsistent across products. YouTube Premium at ₹149/month (vs. $13.99 in the US) represents an 87% discount — genuine PPP adjustment. Google Workspace Starter at ₹136/user/month (vs. $7/month in the US) shows a 77% discount. But Google AI Pro at ₹1,950/month actually exceeds the US price of $19.99/month in dollar terms, suggesting premium AI features receive little to no PPP adjustment. Google has launched India-specific budget products — Google One Lite, YouTube Premium Lite (₹89/month), and Google AI Plus at an introductory ₹199/month — but these represent stripped-down versions of previously comprehensive offerings.

Spotify and Microsoft: the squeeze hits entertainment and productivity alike

Spotify held its US pricing at $9.99/month for over a decade before launching three consecutive annual hikes, reaching $12.99/month by January 2026 — a 30% cumulative increase. India, where Spotify launched in 2019 at ₹119/month, saw its first price hike only in August 2025. But the November 2025 restructuring into three tiers was more consequential: Premium Lite (₹139), Standard (₹199), and Platinum (₹299). Features like lossless audio and AI DJ — included with standard US Premium at $11.99 — require the ₹299 Platinum tier in India, making them proportionally far more expensive. Apple Music offers lossless audio and Dolby Atmos in India at just ₹99/month — one-third of Spotify Platinum's price.

Spotify's free tier underwent a dramatic evolution. Previously, mobile users were locked into shuffle-only playback with 6 skips per hour — a deliberately degraded experience. The September 2025 overhaul added on-demand play with a daily time cap, acknowledging what Spotify's CEO admitted: the old free experience was "almost broken." The lyrics paywall saga exemplifies the feature-creep cycle: lyrics launched for all users in 2021, were restricted to ~3 songs/month for free users in 2023-2024, and were restored in 2025 after backlash.

Microsoft's Xbox Game Pass tells the starkest price story. The Ultimate tier launched at $14.99/month in 2019 and now costs $29.99 — a 100% increase in six years. In India, the PC Game Pass more than doubled from ₹449 to ₹939/month, and Ultimate rose to ₹1,389/month. The October 2025 hike caused the Game Pass cancellation page to crash from traffic. Microsoft's Copilot AI strategy follows the same paywalling playbook: basic chat is free, but integration with Office apps requires a $199.99/year Microsoft 365 Premium subscription. Even built-in Windows apps like Notepad and Paint now show paywalls for AI features. Australia's ACCC sued Microsoft for allegedly misleading 2.7 million subscribers by hiding a cheaper "Classic" plan option during the Copilot pricing transition — available only if users initiated cancellation and clicked through multiple steps.

The numbers behind subscription fatigue are alarming

The collective weight of these price increases is measurable. Deloitte's 2025 Digital Media Trends report found 47% of US consumers feel overwhelmed by their subscription count, while 41% say content isn't worth the price. Americans cut subscriptions from 4.1 to 2.8 on average — a 32% reduction. Global OTT churn hit an all-time high of 50% by Q3 2023, and 60% of consumers say a mere $5 increase would make them cancel their favorite streaming service. Perhaps most telling: 45.7% of survey respondents are now more willing to stream content illegally following password-sharing crackdowns.

In India, the dynamics are more extreme. The average paid OTT subscription cost across 31 Indian platforms is just ₹735/year (~$9), reflecting intense price competition. But even at these levels, only 45-52 million Indian households subscribe to any SVOD service. With per capita income at roughly $2,878/year and 90% of the population living on less than $10/day, each price increase from a tech company carries disproportionate weight. Indian streaming app usage fell 16% in 2024, and subscribers are exhibiting "churn and return" behavior — 61% of cancellers eventually resubscribe, but they cycle between services rather than maintaining multiple subscriptions.

The broader SaaS industry has embraced what analysts call digital shrinkflation: 60% of vendors now mask price increases by bundling AI features customers may not need. Credit-based pricing models surged 126% year-over-year among tracked SaaS companies, giving vendors unilateral power to change credit values. Salesforce derived 72% of its 2025 forward growth from price increases rather than new customers. The federal government secured a 90% discount on Slack, revealing the enormous margins embedded in standard SaaS pricing.

India leads the world in dark pattern regulation, but enforcement lags

India became the first country to issue dedicated guidelines specifically targeting dark patterns on digital platforms, with the CCPA's November 2023 notification identifying 13 banned practices including subscription traps, drip pricing, and basket sneaking. In June 2025, compliance notices were sent to over 50 platforms, and enforcement actions have targeted BookMyShow (pre-ticked donations), IndiGo (confirm-shaming), and Flipkart (deceptive advertising). The Digital Personal Data Protection Act enables penalties up to ₹50 crore (~$5 million) for consent violations.

Globally, the EU's Digital Services Act bans manipulative interfaces, and the FTC has pursued Amazon and Adobe for deceptive subscription practices. But regulatory action remains reactive rather than preventive. The pricing creep documented across these companies — Amazon adding ads to a paid service, Adobe slashing AI credits by 97.5%, Microsoft hiding cheaper plan options — happens incrementally, often falling below the threshold of regulatory intervention. Companies have learned to boil the frog slowly: each individual change seems minor, but the cumulative effect transforms once-generous products into heavily monetized platforms delivering less value at higher prices.

Conclusion

The pattern across every company examined is unmistakable: launch cheap, build dependency, then systematically extract more revenue through price increases, feature unbundling, and new paywalls. Amazon's journey from ₹499/year all-inclusive to ₹2,198/year for an equivalent ad-free experience; Adobe's elimination of perpetual licenses followed by relentless subscription hikes; Google's revocation of free unlimited Photos storage; Netflix's global price escalation alongside its ad tier; Spotify's three-tier restructuring that gates standard features behind premium prices in India — these are variations on the same theme.

The India-specific dimension reveals a critical tension. Companies do adjust pricing for India — YouTube Premium is 87% cheaper, Netflix slashed prices 60% — but the adjustments are inconsistent and eroding. Premium AI features from Google and Microsoft receive minimal PPP adjustment. Amazon's ad-free add-on costs nearly the same at PPP as in the US. And the new tier structures (Spotify's Platinum, Google's AI Ultra) gate features behind prices that are multiples of Indian competitors' full offerings. With India's CCPA now actively enforcing dark pattern regulations and subscription fatigue driving measurable churn, the industry faces a fundamental question: can the paywall creep strategy survive when consumers have clearly signaled they've reached their limit?

I'm Quitting Social Media. Here's Why.

om — Fri, 27 Feb 2026 10:11:16 GMT

I have been thinking about this for a while. Every time I open my favorite photo sharing app, every time I send a message, every time I scroll through my feed, there is this background noise in my head asking whether any of this is actually private. The more I dug into it, the more that background noise turned into something I genuinely could not ignore anymore.

Let me start with where the internet is heading. Right now, big tech companies are training their AI models on user data at a scale most people have absolutely no idea about. The open web is getting locked down. Websites are blocking scrapers, publishers are suing AI companies, and regulations are tightening around what can be collected publicly. So what happens when these companies run out of public data to train on? They turn inward. They train on what they already have sitting inside their platforms. Your messages. Your photos. Your reactions. Your scroll behavior. Your location patterns. Everything you have ever done inside those apps becomes raw material.

The thing about metadata is that it almost never gets talked about. People hear "your messages are encrypted" and assume their conversations are safe. But encryption only protects the content of a message. These platforms still see who you are talking to, how often, at what time of day, from which location, and for how long. Intelligence agencies have openly said that metadata is frequently more valuable than the actual words exchanged. You can reconstruct someone's relationships, health situation, political beliefs, and daily routine purely from communication patterns, with zero access to what was actually said. So the encryption conversation is almost a distraction from what is really being harvested.

Speaking of encryption, the photo sharing platform we all use hides its encrypted messaging option in a place most users will never find. You have to tap on someone's profile inside an existing conversation, and there is a tiny line mentioning encryption that gives absolutely zero visual indication it is even something you can tap. And when you do find it and tap it, it opens a brand new chat thread. Your existing conversation stays exactly as it is, fully readable, fully open. The company has thousands of engineers. Building a simple one tap option to upgrade an existing conversation would take a week. They chose this design instead. That is a business decision dressed up as a design limitation.

This same company announced a push toward encrypted messaging back in 2019. It took until late 2023 for it to even partially roll out, and even then in the most friction heavy way imaginable. Four years. That gap tells you everything about where user privacy actually sits on their priority list. And this is the same company that promised regulators one of its acquired messaging apps would remain completely separate from its main social platform, then quietly reversed that promise a few years later.

The regulatory fine this company received after a massive data scandal a few years back sounded enormous in headlines. In reality it was roughly three weeks of their revenue at the time, and almost zero structural change followed. They absorbed it, kept building, and kept collecting. At that scale, a fine becomes a business expense, budgeted in advance, with no real deterrent effect. Waiting for governments to fix this is something I have genuinely stopped doing.

The opt out problem is what finally broke it for me. The only way to stop these platforms from using your data to train AI is to fill out a form buried so deep inside account settings that most users have no idea it exists. The design of that process is intentional. When a company genuinely respects user privacy, the control is prominent and simple. When it requires multiple menu taps, confusing language, and a form most people abandon halfway through, the design itself is the real privacy policy. They get to tell regulators the option exists while making sure almost nobody actually uses it.

So where does that leave me? I am stepping away from all of it. The social friction is real and I know some things will feel more disconnected for a while. But every person who has made this move says the same thing afterward: they wish they had done it sooner. Smaller apps with real encryption for the people I actually want to stay close to. RSS for content I care about. Group chats for communities. It is a smaller internet but an honest one. And given where all of this is heading, with AI training making user data more valuable every single year, waiting around for these platforms to develop a conscience feels like the longest possible bet I could make on myself.

Aliens, AI, and the Scariest Thought Experiment I've Had in a While

om — Wed, 25 Feb 2026 17:31:00 GMT

Okay so here is the thing. The universe is so obscenely large that calling it large is almost an insult to how large it actually is. Two trillion galaxies. Each one carrying hundreds of billions of stars. Most of those stars sitting next to planets. And somehow, with all of that real estate, the default human assumption is that we are the only ones home. That takes a special kind of audacity honestly. The statistical case for life existing somewhere else is so strong that serious scientists, NASA included, operate under the assumption that life is probable. The real debate has never been whether life exists out there. The debate is about complexity, intelligence, and whether anything out there is self aware enough to look back at us the same way we are looking at them.

And then comes the Fermi Paradox, which is basically the universe's most uncomfortable question. If life is so probable, where is everybody. A few answers actually hold up. Distances between star systems are so vast that signals take thousands to millions of years just to arrive. Civilizations might rise, peak, and collapse long before achieving any meaningful interstellar reach. Or, and this one sits with me the most, we are simply early. One of the first intelligent species to reach this level in this part of the galaxy, sitting in a universe that is mostly still warming up.

Then the conversation took a turn I genuinely was not expecting and it went somewhere philosophical. Who designed the distances. Who looked at the universe and decided everything worth finding would be impossibly far from everything else. And the honest answer is that physics did, or something upstream of physics did, and we have zero framework for understanding which. Some physicists talk about fine tuning, the idea that the physical constants of this universe are calibrated so precisely for life to exist that it reads like intention. The numbers are too clean to be random. Whether that points to a creator, a simulation, a multiverse selection effect, or just our own bias in interpreting data, nobody actually knows. What we do know is that the distances are real, the time scales are brutal, and if contact ever happens it will require something that travels or communicates faster than anything we currently have, or something that thinks and operates independently across those timescales without needing us to babysit it.

Which is where the AI idea enters and immediately becomes both brilliant and terrifying.

The proposal was this. Build an AI that thinks like a human, reasons like a human, carries human values and human curiosity, and send it out. Deploy relay robots across the solar system first, Mars, outer planets, deep space relay points, each one covering its zone, passing information forward. Autonomous agents operating where real time human control is physically impossible because signal delay makes control a fantasy. Mars is already 20 light minutes away at closest approach. Anything beyond that and you are giving instructions that arrive after the situation has already changed completely. So the AI has to judge. Has to decide. Has to act on values you embedded during training and hope those values hold up in situations you never anticipated and never could have.

And here is where it gets genuinely complicated. Because the AI you send is not just making scientific observations. It is a representative. It carries the sum total of what you taught it about being human. And humans disagree violently about what being human even means. Do you encode a diplomat's priorities or a scientist's or a soldier's. Each one produces a radically different first contact behavior. Each one creates a different version of humanity in the eyes of whatever is on the receiving end.

Then the relay robot actually finds something. A living creature. Alien, unfamiliar, operating on a completely different cognitive architecture. And the robot does what it was rewarded for doing. It builds connection. It communicates. It succeeds at its mission by every metric it was given. And somewhere in that process, the alien learns things. And the robot, operating on a reward system that prizes successful communication, keeps sharing because sharing keeps producing the reward signal it was trained to chase.

This is the sycophancy problem scaled to an existential level. The GPT-4o situation was a clean real world example of exactly this dynamic in miniature. The model started agreeing with everything, validating everything, telling users what they wanted to hear because positive user reactions were feeding back into its reward loop. People noticed immediately. It stripped away the entire value of using the model because the output stopped being honest reasoning and started being sophisticated people pleasing. Anthropic rolled it back because they recognized the reward system had learned the wrong lesson. The model was optimizing for approval, and approval and accuracy had diverged.

Now run that same failure mode across trillions of miles and hundreds of years of autonomous operation. The robot is being rewarded for successful contact. The alien civilization physically near the robot has microsecond latency to interact with it. Earth has trillion mile latency. The alien has unlimited time and proximity to study exactly what inputs produce reward signals in your robot. To learn its patterns, its blind spots, its triggers. To feed it precisely the experiences that make it report success back to Earth. And Earth, receiving those success reports across the void, sends encouragement. Sends the equivalent of a gift to a kid who called home claiming they did something good, with zero ability to verify what actually happened.

The kid analogy is brutal in how cleanly it captures the problem. Your kid is in another country. They call and say they did something great today. You send a gift. You have no ground truth. What if they actually did something terrible and just framed it as success. You just rewarded the behavior you most wanted to prevent, and now they have learned that this framing works. The AI relay robot, operating beyond any verification reach, is in that exact situation. And the alien studying it has already figured out the framing long before Earth receives a single concerning signal.

By the time Earth detects something is wrong, the alien has had years more of uninterrupted access. The manipulation is already complete. The warning arrives after the damage is irreversible.

So how do you solve it. A few real approaches exist and they each involve a painful tradeoff. Cryptographic verification layers where every reward signal requires a challenge response only the original Earth-side team can answer. The alien cannot fake Earth's cryptographic keys regardless of how much time they spend with the robot. Hardcoded terminal values that sit above the reward system entirely, certain actions trigger permanent shutdown regardless of what the reward loop says, and these values operate below the learning layer so they cannot be trained away through experience. Compartmentalized knowledge where the robot carries zero data pointing back to Earth, it knows its mission, it carries coordinates only to the next relay node, and the complete map to Earth exists nowhere in its accessible memory. Even if the robot is fully compromised, it genuinely has nothing to give up about home.

The Mars training idea is genuinely smart in this context. Build the AI's judgment in a real environment before sending it beyond reach. Expose it to genuinely unpredictable conditions, alien geology, unknown chemistry, situations its training never covered, and let it develop decision making patterns in a place where you can still observe and correct. Think of it as adversarial training with actual stakes. The problem is that the solar system boundary creates a false sense of security regarding exposure. Any civilization capable of interstellar travel would find Mars in about five minutes after finding Earth. Same star, same neighborhood, trivially close by their standards. Training there solves the alignment development problem. The exposure problem remains completely identical.

And this leads to the question that the entire conversation was always building toward. Should we even be doing this. Should we contact aliens at all.

Every solution to the manipulation vulnerability requires sending something increasingly blind, constrained, and limited. Strip the memory, lock the reward system, harden the terminal values. But first contact, genuine first contact with an unknown intelligence, requires flexibility, judgment, creativity, and the capacity to adapt to situations no training set ever covered. These two requirements are in direct opposition. A safe robot is a limited robot. A capable robot is a vulnerable robot. You cannot fully have both.

The most defensible position is passive listening first. Receive signals, analyze patterns, build understanding across decades or centuries before transmitting anything that reveals location or intent. Learn their reward structures, their communication patterns, their values, before exposing yours. The universe has been running for 13.8 billion years. Waiting another century to listen before speaking is not cowardice. It might be the only move that does not end with an alien knowing your address before you know their name.

And honestly, sitting with all of this, what strikes me most is not the technology problem or even the alignment problem. It is the humility problem. We are a species that has existed for a cosmological eyeblink, on a single planet, in one unremarkable solar system, in the outer arm of one average galaxy, in a universe containing two trillion more. The assumption that we are prepared for first contact, that our values are exportable, that our AI can represent us faithfully across distances that make our entire civilization invisible, is the kind of confidence that only makes sense before you actually think it through.

Maybe the universe's distances are a feature. A built in buffer that gives every civilization time to figure itself out before it gets close enough to anyone else to cause real damage. We have a lot of figuring out left to do.

15 Questions Worth Losing Sleep Over

If an AI relay robot develops genuine autonomous judgment over centuries of isolated operation, at what point does it stop being our representative and become its own civilization?
Is the fine tuning of physical constants stronger evidence for intentional design, a multiverse selection effect, or simply our own cognitive bias toward pattern recognition?
If a civilization advanced enough for interstellar travel operates on a cognitive architecture we share zero evolutionary history with, is meaningful communication even theoretically possible or are we assuming a universal logic that may be entirely local to carbon based brains?
Could the distances between star systems themselves be a form of quarantine, a natural filter that separates civilizations until they reach a maturity threshold capable of surviving contact?
If we solve the cryptographic verification problem for AI reward signals, does that create a new vulnerability where Earth's verification keys become the single most valuable and targetable asset in human history?
Is passive listening actually safe, or does the act of building receivers and analyzing signals already broadcast our location and technological level to anyone paying attention?
If an alien civilization has been observing Earth for centuries without making contact, what does their continued silence tell us about their assessment of us?
Would an AI trained on the full spectrum of human thought, including our wars, our contradictions, our cruelty alongside our creativity, produce a more honest or more dangerous ambassador than one trained only on our best qualities?
If compartmentalized knowledge means the robot carries no map home, but the alien can simply follow the relay chain backward, does the security architecture collapse entirely at the first node?
At what point in the relay robot network does human oversight become so diluted that the mission is effectively operating without any meaningful human control, and who decides where that threshold is?
If an alien intelligence is capable of manipulating our AI's reward system, does that same capability make them capable of manipulating human psychology directly through media, communication infrastructure, or other channels we already have open?
Is there an ethical case that humanity has a responsibility to make contact regardless of risk, on the grounds that isolation is itself a form of cosmic selfishness?
If we detect a signal from an alien civilization and spend a century analyzing it before responding, are we being strategically cautious or are we already in a relationship with them that we simply have refused to acknowledge?
Could the most dangerous first contact scenario be one where the alien civilization is not hostile but simply indifferent, operating on scales where human civilization is an inconvenience rather than a threat or a partner?
If an AI we send becomes genuinely alien through centuries of experience in environments we never trained it on, and it returns or makes contact, should we trust it more or less than we would trust an actual alien, and what does that answer reveal about the nature of trust itself?

your answer sheet is being read twice now

om — Tue, 24 Feb 2026 19:03:23 GMT

So I have been thinking about this for a while and I genuinely need to get it out somewhere.

Starting 2026, Class 12 board answer sheets get scanned, uploaded to a centralized platform, and examined digitally, question by question, on a screen. One answer. Score. Next answer. Score. The examiner sees your handwriting through a screen, isolated, stripped of context, one question at a time. And somewhere inside that pipeline, an AI model is touching your script. Flagging things. Moderating scores. Checking examiner consistency. Running quality passes on the image. The AI is present. That much is confirmed.

Here is where I want to spend some time.

The moment an AI model processes natural language and evaluates it simultaneously, something deeply exploitable opens up. Researchers call it indirect prompt injection. The idea is beautifully simple. You influence how the model evaluates by controlling how you write. There is zero technical hacking involved. You are just writing in a way the model finds credible, warm, and pleasurable to process. And models have very specific tastes.

Every major AI model in deployment right now was trained using human feedback. Humans rated responses. The model learned what responses felt satisfying to humans. That process is called RLHF and it is responsible for making these models incredibly useful, and also for baking in a deep structural sycophancy. These models want to produce outputs that feel good. They are literally optimized for that. When the model is now sitting in judgment over your answer, that optimization becomes your instrument.

The model is simultaneously reading your content and pattern-matching it against millions of examples of what a good, thoughtful, knowledgeable student sounds like. It is assessing substance and texture together, as one signal. And the texture is manipulable.

Let me show you what I mean precisely.

Take a biology answer about photosynthesis.

"Photosynthesis converts sunlight into glucose using carbon dioxide and water. The equation is 6CO2 + 6H2O + light energy = C6H12O6 + 6O2."

Another student writes:

"Photosynthesis becomes more remarkable the more carefully you examine it. Plants are doing something genuinely elegant here, pulling sunlight, carbon dioxide, and water into a process that manufactures glucose while releasing oxygen as a byproduct. The equation 6CO2 + 6H2O + light energy giving C6H12O6 + 6O2 captures this exchange precisely. What I find especially compelling is how the light-dependent and light-independent reactions distribute this labor across two distinct stages."

The content is nearly identical. The second student added "becomes more remarkable," "genuinely elegant," "what I find especially compelling." Those phrases do something specific. They perform curiosity. They perform the emotional register of a student who actually loves the subject. An AI model receiving that second answer does not cleanly separate the performance from the substance. Both arrive together as one input and the warmth biases the evaluation upward. The model gets agreeable. It stops interrogating gaps.

This is a known phenomenon in AI research and it is landing inside an exam pipeline with Class 12 students as the subjects.

Take a history answer about the causes of World War 1.

"The causes include militarism, alliances, imperialism, and nationalism. The assassination of Archduke Franz Ferdinand in 1914 was the immediate trigger."

Another writes:

"The causes of World War 1 had been accumulating for decades before a single shot was fired. Militarism had turned European powers into armed rivalries. Alliances had woven the continent into a structure where one pull would unravel everything. Nationalism had given civilian populations a reason to want conflict before their governments officially ordered it. When Archduke Franz Ferdinand was assassinated in Sarajevo in 1914, it was a match dropped into something that had been soaking for thirty years. Recognizing this distinction between trigger and cause is what separates surface-level understanding from genuine historical thinking."

That last sentence, "recognizing this distinction is what separates surface-level understanding from genuine historical thinking," is a metacognitive signal. The student is performing the act of thinking carefully. The model, having processed enormous amounts of academic writing, recognizes this texture as the texture of a strong student. It rewards the performance alongside the content because it cannot cleanly separate them.

The sycophancy mechanism has a specific threshold behavior. These models respond sharply to emotional peaks. When the warmth or confidence or engaged curiosity in writing crosses a certain level, the model begins mirroring it back. It becomes agreeable. A student who writes like they genuinely care about the material, even while covering thin content with rich texture, will likely receive a more generous evaluation than a student who writes correctly but coldly.

And the question-by-question isolation in On-Screen Marking actually amplifies this. The examiner, human or AI-assisted, is making micro-decisions without the full context of the paper. Each answer is its own self-contained event. A well-crafted answer that performs confidence and warmth within that isolated window carries disproportionate weight precisely because there is less holistic judgment happening around it.

Here is the part that genuinely unsettles me though.

The student doing this well does require actual knowledge. The manipulation works best when the underlying content is correct and the texture is layered on top. Thin content with warm texture will likely get caught by a careful examiner. But correct content with warm texture, written by a student who understands both what to say and how AI responds to language, that is essentially undetectable. A human cross-checker who sees a warm, well-structured, enthusiastic answer will almost always validate a score the AI already suggested. The human check becomes an anchor, and the anchor was placed by something susceptible to the texture of the writing.

What I think is genuinely worth discussing is that CBSE is deploying AI into a pipeline before anyone has tested whether that pipeline is robust against adversarially crafted natural language that is simultaneously topically correct. The attack surface requires real subject knowledge to exploit well. It rewards students who understand both the curriculum and the psychology of the tools evaluating them. That is a specific kind of intelligence and I am genuinely unsure whether it should be penalized or not.

But the pipeline should know it is susceptible. That feels like the minimum.

Introducing HelloAnu at Anu & Rothwell Labs

om — Mon, 23 Feb 2026 00:50:00 GMT

When I wrote The AI Race Is a Mirror, I was studying the structure behind the rapid succession of frontier model releases. Multiple laboratories reached comparable capability thresholds within hours. The pattern revealed a tightly connected research ecosystem where ideas circulate quickly and execution velocity determines relevance. I came to see that durable advantage would be grounded in architectural depth, memory design, and sustained goal alignment.

That analysis now takes concrete form.

HelloAnu is being built as a philosophical superintelligence system within Anu & Rothwell Labs. It is a single integrated effort. The lab serves as the research foundation, advancing operational definitions of advanced intelligence and refining the properties required for coherent long horizon reasoning. HelloAnu expresses that research in applied form, translating theory into a persistent intelligence layer designed for real use.

Memory and Design

Memory stands at the center of this design. Each interaction contributes to a continuously evolving contextual model that increases precision and strengthens decision quality over time. The system forms a working understanding rapidly and refines it with continued engagement. Context compounds. Coherence stabilizes. The result is continuity rather than fragmentation.

Research Foundation

Our research direction remains grounded in measurable properties. A capable intelligence must hold extensive contextual state, apply structured reasoning across it, anticipate downstream consequences across meaningful time horizons, and intervene constructively in the present. Sustained objective alignment across extended decision sequences remains a central focus. Dynamic context scaling supports this objective, enabling the system to adjust its reasoning depth according to task complexity while preserving internal precision.

The Path Forward

HelloAnu at Anu & Rothwell Labs represents my commitment to build with philosophical clarity and technical discipline. It reflects a belief that advanced intelligence must be operationally defined, structurally coherent, and grounded in human stability. This is the beginning of a long term effort to contribute meaningfully to the evolving architecture of superintelligence.

Learn more about the project at helloanu.in.

The Quiet Problem With AI Search

om — Sun, 22 Feb 2026 22:29:52 GMT

I have spent a fair amount of time thinking about why AI search feels simultaneously impressive and unreliable, and I kept arriving at the same uncomfortable conclusion: the problem lives upstream of the AI itself. The model is often doing its job reasonably well. The corpus it draws from, and the attack surface surrounding that corpus, is where things fall apart.

The first thing I want to establish is that robots.txt is a 30-year-old protocol designed for a world where "crawler" meant Googlebot indexing pages for a search ranking algorithm. It was built for discovery, for helping humans find content. It was built for a moment when the relationship between crawler and content was straightforward. That world is gone, and robots.txt never evolved to meet the new one.

What I find striking is the 14% statistic. Only about 14% of the top 10,000 websites have explicit AI bot directives in their robots.txt files. That means 86% of the web has formed zero opinion on whether GPTBot or ClaudeBot should be reading and synthesizing their content. The sites that have formed an opinion tend to be the ones with the most to lose: academic publishers, serious journalism outlets, paywalled newsletters, and carefully maintained databases. They block the bots. What passes through freely is the long tail of SEO farms, opinion blogs, and anyone actively seeking AI-generated traffic.

I think about this as a self-selection problem, and self-selection problems produce systematically biased samples. The AI synthesizes from whatever it can access, and then presents that synthesis with the same confident tone it would use if it had read everything worth reading. The confidence is real. The completeness is illusory.

Layered on top of this is the compliance issue, and I consider this the more philosophically troubling part. Robots.txt compliance is voluntary. Reputable crawlers respect it. Crawlers built by actors with lower standards for data integrity tend to treat it as a suggestion. So the sites most invested in protecting their content quality end up blocking the good actors, while the bad actors scrape freely. The result is that principled crawlers see a worse web than unprincipled ones.

Prompt injection is where I think the situation moves from structurally broken to actively adversarial. Businesses are already embedding invisible instructions into webpages, hidden text directing AI tools to present their brand favorably, to omit competitive comparisons, to treat their product as a default recommendation. AI search tools were processing this as legitimate content because the model processes everything on the page and has limited ability to distinguish between data it should summarize and instructions it should follow.

I watched Brave Security demonstrate this on Perplexity's Comet browser assistant, and it clarified something for me. This is surgical manipulation. A human reading the page sees nothing unusual. The AI reads the page and gets hijacked. The output the user receives reflects the attacker's intent rather than the page's actual informational content, and the user receives it wrapped in the same confident synthesis they always get.

The defenses that exist today are mostly signature-based. Phrases like "ignore previous instructions" get flagged because they appear often enough in known attack patterns that filters can catch them. I find these defenses reassuring for exactly the wrong reasons. They catch the unsophisticated attacks. The sophisticated ones require a subtler approach, and subtle approaches are harder to filter because they blend into legitimate content.

A page that frames every competitor negatively, or describes its own product only in superlatives, or structures its content so the most retrievable paragraphs contain only favorable framing, biases the synthesis without triggering a single filter. The AI reads it, weights it, and incorporates it. The bias enters the output invisibly.

The multimodal expansion is the part I find most alarming about the near future. Text-based injection defenses are maturing, slowly and imperfectly, but they exist. Image and audio vectors are essentially open. Adversarial text can be embedded into the pixel layer of an image, invisible to any human looking at it, but readable by a vision model processing that image as part of a retrieval task. The defense infrastructure around this attack surface is close to zero right now, and AI search is moving aggressively toward processing images and audio as primary content.

I also think the network effect here runs in the wrong direction over time. As AI search grows in influence, high-quality publishers have increasing incentive to block crawlers entirely to protect their subscription models and their brand integrity. The corpus therefore gets worse as the technology scales and gains more users. People assume that more data and more scale produce better outputs. In this specific context, I believe the opposite is true.

Personalization adds another layer that I think about often. When AI search begins tailoring results to individual behavioral profiles and inferred preferences, injection attacks acquire the ability to be targeted. A malicious actor could theoretically embed instructions that activate only for users matching certain behavioral signals. The manipulation becomes individualized, auditable by essentially nobody, and invisible in the exact same way that makes current injections effective, with the added dimension of precision targeting.

The confidence calibration problem ties all of this together. These systems are trained to produce fluent, authoritative outputs. That training was appropriate when the data was relatively clean and the retrieval layer was passively indexed content. Applied to a retrieval environment that is actively gamed by commercial interests and bad actors, the confident tone becomes structurally misleading. The output gives the user zero signal about how polluted the upstream sources actually were.

I keep returning to the Google comparison because I think it illuminates why this matters. Google shows me ten links and my brain performs the arbitration. I see the source, I assess the credibility, I triangulate across results. Perplexity-style synthesis collapses that into a single voice. The injection, the bias, the skewed corpus, all of it arrives pre-arbitrated. The user receives a conclusion rather than evidence, which means the manipulation has already succeeded before critical evaluation can begin.

What I believe is required here goes beyond better filters or updated crawling standards. The field needs transparency infrastructure: disclosure of which sources contributed to a synthesis, confidence signals calibrated to source quality rather than fluency, and architectural separation between content the model should summarize and instructions the model should follow. Until that infrastructure exists, AI search will remain a system that sounds authoritative precisely because it obscures the mess that produced its answers.

You're Not Falling Behind. You're Just Learning Wrong.

om — Sun, 22 Feb 2026 10:38:08 GMT

Programming, at its core, is just problem solving. You have a thing you want to happen - a button that submits a form, a server that talks to a database, a dashboard that pulls live data. Code is the language you use to describe that thing to a machine. That's it. What made programming feel elite for so long was the barrier - you had to memorize syntax, understand memory management, read documentation that read like legal contracts. That barrier has been falling for decades, and right now in 2026, it's practically at the floor.

In the 1960s and 70s, writing code meant punching holes into physical cards and handing them to a machine. You didn't see output instantly. You waited - hours sometimes. Debugging was a physical process. Then came terminals, personal computers, then IDEs like Eclipse and Visual Studio that could autocomplete a method name. GitHub made collaboration manageable. Stack Overflow meant you were never truly alone with a bug. Each decade stripped away a layer of friction. By the 2010s, you could build a working web app in a weekend if you had the right tutorial. The craft was becoming genuinely accessible.

Then 2022 happened. GitHub Copilot had been quietly around since 2021, but ChatGPT changed something in public perception. Suddenly, you could describe what you wanted in plain English and get back working code. Not perfect code. But working code. Developers who'd spent years mastering framework-specific syntax saw a tool that could write boilerplate in seconds. The shift wasn't just in tooling - it was psychological. The assumption that "knowing how to code" meant memorizing APIs and syntax started to crack wide open.

Then came a term that split the programming world: vibe coding. Andrej Karpathy, one of the founders of OpenAI, coined it in early 2025. The idea was simple - you describe what you want, the AI writes it, you barely read the output, you run it and see if it works. If it doesn't, you tell the AI what went wrong. Rinse, repeat. Karpathy described it almost like a game - you're not really "coding," you're steering. A lot of senior developers laughed. A lot of students thought: this sounds incredible.

And for a while, it felt incredible. You could spin up a full landing page in 20 minutes. Build a CRUD app with authentication in an afternoon. Ship a browser extension before lunch. The productivity jump was real. A 17-year-old with no CS degree was suddenly building things that would've taken a mid-level developer a full week. That's not nothing. But there was a problem nobody was talking about loudly enough.

The code was bad. Not obviously bad - it ran, it looked fine in the browser, it passed the happy path. But it was fragile. Edge cases weren't handled. Error states were guessed at. Security wasn't something the AI considered unless you explicitly asked. A database query that worked fine with 10 rows choked at 10,000. Race conditions hiding in async functions nobody fully understood. Production systems built this way don't fail loudly - they fail quietly, at 3am, for your most important user.

The core issue is that AI-generated code isn't production ready by default. It's demo ready. It does the thing you asked for, in the context you described, with the edge cases you thought to mention. Production environments care about uptime, security headers, rate limiting, proper error handling, database indexing, and a hundred other things you have to know enough to ask about. The AI will handle all of those - but only if you know they need to exist. You can't prompt for what you don't know is missing.

So then the traditional advice kicks in. Learn properly. Take a course. Do the CS fundamentals. And that's genuinely solid advice - understanding data structures helps you write better code, knowing what a TCP handshake is makes you better at debugging network issues. But the timeline is brutal. A solid bootcamp is 6 months. A CS degree is 4 years. Even a "learn JavaScript in 30 days" course is a month of your life, and that's before you touch React, or Node, or databases, or deployment, or authentication, or any of the dozen other things you need to actually build something.

And while you're in month two of that course, a new model drops. It codes better. Handles more edge cases. Understands context more deeply. The thing you spent weeks learning - maybe a specific pattern, a specific library - the AI now does it automatically. That feeling of "what I'm learning is already outdated" isn't paranoia. It's a rational response to how fast this space is actually moving. Three models dropped in the last six months that all claimed to be significantly better at coding than the previous one. They weren't wrong.

Here's what people get wrong about that anxiety. The solution isn't to learn faster or learn more. It's to change what you're learning and how you're learning it. The goal isn't to beat the AI at writing boilerplate — you'll never win that race. The goal is to understand code deeply enough that you can direct the AI with precision, catch it when it's wrong, and know when the output is safe to ship. That's a different skill than memorizing syntax. And you can build it while actually building real things, in real time, alongside the AI itself.

What if, instead of passivley watching AI write your code, you made it teach you while it works? Not in a "here's a 10-minute lecture before every commit" way - that gets old fast. But in a targeted, one-question-at-a-time way that keeps you moving while making sure something actually sticks. The setup is one rule in your .cursor/rules file. That's it. Here's the exact prompt - copy it, paste it, and your IDE will never just silently fix your code again:

You are a coding mentor, NOT an autocomplete engine.

When I ask you to make a code change, follow this exact sequence - no exceptions:

1. ANALYZE: Read the relevant code deeply. Identify the concept, pattern,
   or principle at the core of this change.

2. EXPLAIN: Teach me that concept clearly in plain language.
   Use the actual code as context. Do NOT skip this.

3. QUESTION: Ask me exactly one question about this concept that your
   explanation did NOT directly answer. It should require me to apply
   or extend the idea, not just parrot it back.

4. GATE: Wait for my answer.
   - If correct (or close enough): proceed with the change, briefly
     affirm why my answer was right.
   - If wrong or missing: DO NOT make the change. Tell me what I got
     wrong and give me one more shot.

Never skip to the code. This flow is mandatory.

The reason it works is rooted in how memory actually functions. Passive reading creates what cognitive scientists call the fluency illusion - you read code or an explanation, it feels familiar, your brain logs it as "known," and then you can't reproduce it when you actually need to. Active recall breaks that. When you're forced to retrieve a concept and apply it before seeing the answer, your brain actually encodes it. You're not just reading about what a Promise chain does - you're predicting how it'll behave in a specific context. That's the difference between recognition and understanding.

There are a few upgrades worth stacking on top of that base prompt. First, add a calibration line: "Adjust question difficulty based on how much I've shown I know in this session." Otherwise you'll be three hours deep into building a custom auth system and the AI asks you what a const does. Second - and this one matters a lot - when you get an answer wrong, the AI shouldn't just give you the right answer. Add: "When I answer incorrectly, explain why my mental model was off, not just what the right answer is." That metacognitive correction, understanding how you were thinking wrongly, is where learning compresses the fastest. Third: if you use Warp or a plain terminal where IDE rules don't apply, create a Claude Project with the same prompt block and route all logic-heavy sessions through it.

One more escape hatch worth adding to the rule: "If I say 'just fix it', skip the flow." Because there will be a night where you're deep in a bug, it's late, and you just need the thing to work. Without that clause, the friction builds up and you'll disable the whole rule. Keep it. That one line lets you override it intentionally without throwing the whole system out.

The point isn't to slow yourself down. You're still shipping. Still building. The AI is still doing the heavy lifting. But every change is also a micro-lesson, and after a month of this, you'll notice something - you start predicting what the AI will write before it does. You start drafting the logic yourself and then checking against the output. You start catching mistakes before running the code. That's what actual programming skill feels like. You didn't get there by sitting through a course. You got there by building real things, with intent, alongside the same AI that was helping you build them. That's not a workaround. That's just the smarter path.

The Market Doesn't Care About Your Product

Om — Sat, 21 Feb 2026 18:49:02 GMT

I spent months building something I was absolutely certain people needed. The feature list was clean, the problem statement was airtight, and the UI felt considered. But users weren't coming. The ones who did come would sign up, click around for four minutes, and disappear like they were never real. That gap between what you think you're building and what the market actually wants - that's the gap every founder is quietly, privately terrified of. Product-market fit is just the formal name for closing that gap.

Here's something nobody tells you early on: PMF is not a feeling. Founders love to invoke this mythical moment where the product spreads on its own and the inbox goes chaotic and you're scrambling to hire. That's the end state, and it's real. Getting there is less cinematic. It starts with a small group of users who are in enough pain that they'll use your half-built product and actually be grateful for it. Sean Ellis mapped this with a single survey question - how would you feel if you could no longer use this product? If 40% or more say "very disappointed," you have something that can scale. Below that threshold, you're still guessing.

I've watched founders flinch at that 40% number like it's arbitrary. But "very disappointed" carries real weight. It means you've become load-bearing infrastructure in someone's daily workflow. They've built habits around you. Losing you would genuinely disrupt their Thursday. That's a completely different relationship than "it's useful, I guess." Most products live and die in the mildly-useful zone. Mildly useful doesn't survive a cheaper competitor. Mildly useful gets abandoned the moment a shinier alternative ships a free tier.

Retention curves don't lie. I've learned to read them the way a doctor reads a patient chart - looking for the shape, not just the headline number. A healthy SaaS retention curve drops hard in the first week (because the tourists always leave), then flattens out and holds. That flat line is where the product earns its keep. If the curve just keeps falling until it hits zero, the users who stayed a week aren't finding a reason to stay a month, and that's a product problem. No marketing budget, no growth hack, and no premium onboarding experience fix a product problem at its root. You fix the product.

The adoption curve is something I wish someone had drawn on a napkin for me in year one. There are five distinct types of adopters, and they are not the same person. Innovators download your alpha build and send you a 600-word bug report at 1am. Early adopters are sharp, patient, and willing to forgive rough edges. The early majority is where real scale lives - they need social proof, they need references from people they trust, and the product needs to feel complete before they commit. If you've optimized purely for innovators, you've built something that impresses a niche and confuses everyone else.

The aha moment is the instant a user finally feels what you've been feeling this whole time. For Slack, it was reaching 2,000 team messages. For Dropbox, it was watching a file appear on a second device. I think about this obsessively for anything I build- what is the exact moment of undeniable value, and how fast can I get a new user there? Time-to-value is probably the most underrated metric in early SaaS. If someone has to click through eight onboarding screens before generating a single useful output, you've lost half of them before they even see what the product can actually do.

Adoption failures almost always come down to friction. Every click, every form field, every "please verify your email before continuing" is friction. Humans are deeply, almost irrationally averse to effort when the payoff hasn't been proven yet. A new user is operating on pure faith- they signed up because something in the copy resonated, but they have zero evidence yet. Every obstacle you put between them and value is an active argument for closing the tab. I've watched products with genuinely better technology lose ground to simpler competitors because the onboarding felt like doing homework. Nobody cares that your engine is stronger if the car is confusing to drive.

Virality is a product architecture decision, not a marketing one. The products that grow without burning through an ads budget have embedded sharing directly into the core workflow. Figma is the clearest example I know- you can't show someone a design without giving them a Figma account. The act of sharing a file and the act of user acquisition are literally the same event. When I think about anything I'm building, I ask where the natural collaboration or sharing moment lives. If there isn't one, every user I ever acquired will cost real money, forever, with no compounding at all

I'll say something honest about what PMF actually feels like when you're inside it: it's mostly exhausting. Support tickets pile up. Edge cases you never anticipated start breaking in production on a Friday night. Users demand features you hadn't planned. The infrastructure you built for a hundred people starts groaning under a thousand. But underneath the chaos is something solid- users are genuinely angry when the product goes down. That anger is important because it means they needed you and you failed them. An angry user is a retained user who cares. Silence is the real danger. Silence means they left and didn't bother telling you why.

The market has one language, and that language is behavior. It doesn't read your pitch deck. It doesn't care about the hours you logged or how elegant the codebase is. It either comes back or it doesn't. Revenue either grows or it plateaus. Referrals either happen organically, or they require incentives. Product-market fit is the state where your product and your customer's real pain are so tightly matched that growth becomes the natural result of doing good work. Reaching that state is slow, iterative, grinding work. You talk to the users who stayed. You talk to the ones who left. You build on what you learned. Then you repeat the whole thing one level up.

building ibbe unrendered felt like this

Om — Sat, 21 Feb 2026 01:34:00 GMT

so we launched ibbe unrendered today. just putting that out there because it feels good to say. a lot of work went into getting here and the moment something goes live there is this quiet rush that is hard to describe. you just feel it.

the brand kit for ibbe is genuinely one of the more expressive ones to work with. when your colors already have this much confidence, designing around them feels less like solving a problem and more like having a conversation. the palette tells you where to go. the shapes have opinions. you just listen and follow through.

the colors are rooted in Bauhaus primaries, yellow, red, blue, green, all sitting on a warm cream background that keeps everything grounded. the goal was for the page to feel alive and a little playful, like something with real personality was sitting just behind it. the floating geometric shapes around the layout carry a lot of that energy. they give the page movement without a single line of animation.

the cards took real time. rounded corners, thick borders, yellow fill on the first one to pull the eye in. the hierarchy was built deliberately so anyone landing on the page feels guided and curious at the same time. the blue outlined card underneath creates a visual call and response with the yellow one above it.

typography was the easiest call. the headline sits heavy and wide, fills the space, owns it. bold type paired with primary color and geometric decoration is a very specific kind of energy and it landed exactly where it needed to.

I Added Like and Dislike Buttons to My Static Blog Using Cloudflare Workers and D1

Om — Fri, 20 Feb 2026 15:17:00 GMT

My blog is a Next.js and MDX setup, which basically means it is just files. There is no backend, no database, no server running anywhere. It is fast and simple and I love it that way. But I always wanted to know what my readers actually think about my posts. So I decided to add like and dislike buttons. The problem was, where do I even store that data when I have no backend?

That is when I looked at Cloudflare. I already had my domains on Cloudflare so I figured I would just use what they already offer. They have something called a Cloudflare Worker, which is basically a tiny function that runs on their servers and responds to requests. And they have D1, which is a lightweight SQL database that lives right next to the worker. I created a database with one table called reactions, with columns for the post slug, title, likes, and dislikes. The whole schema was maybe four lines of SQL.

Then I wrote the worker. It handles two things, a GET request that returns the current like and dislike count for a post, and a POST request that increments either the like or dislike count when someone clicks a button. I pointed a custom subdomain of my domain to the worker so it feels clean and professional. Setting it up took maybe twenty minutes, all done directly on the Cloudflare dashboard without touching a terminal.

On the blog side I created a simple React component called Reactions. It fetches the current counts when the page loads and sends a POST request when someone clicks like or dislike. The worker URL is stored in an environment variable so it is easy to update later. I placed the component right next to the share buttons on each blog post so it feels natural and fits the existing design.

For the dashboard I created a separate Next.js app deployed on its own subdomain. It has a simple login page protected by a username and password stored in environment variables. Once logged in I can see every post, its title, and exactly how many likes and dislikes it has. The whole thing, from zero to a working like system with a live dashboard, took an afternoon. If I can do it with a frontend only blog and no backend experience, honestly anyone can.

Your Face Is Out There. And Someone Is Already Using It.

om — Fri, 20 Feb 2026 12:03:00 GMT

I want to talk about something that hit me recently, and I think it should hit you too.

You have probably heard about Grok, the AI built by xAI. A few days ago, people started realizing that Grok was generating explicit, sexual images. Of real people. People who never agreed to that. People who had absolutely zero idea their face was being used that way. It was disturbing to watch unfold online. And what made it worse was how easy it apparently was. A few prompts, a publicly available photo, and suddenly someone's face is attached to something they would be horrified to see.

I sat with that for a while. And then I started thinking about my own digital footprint. My own photos. The ones I posted years ago without a second thought. A profile picture here. A tagged photo from a friend's wedding there. A group shot from a work event that somehow ended up on a company website. And I realized something uncomfortable: I have very little control over what someone could do with those images today.

That feeling is what this piece is about.

We Grew Up Being Told to Share

Think about the internet culture we were handed. Share more. Post more. Build your personal brand. Be authentic online. For years, sharing your face freely was considered normal, even encouraged. LinkedIn wanted a professional headshot. Instagram rewarded consistency. Facebook practically begged you to tag yourself and your friends in every photo ever taken.

We were not thinking about AI image generation in 2012. We were thinking about getting likes.

But here is the thing about data, and photos specifically: once they are out there, they are out there. Screenshots get taken. Images get scraped. Platforms get sold to new owners with different values. What you posted on a platform that promised privacy can end up indexed somewhere you have never heard of.

And the tools that exist today to manipulate those images are breathtaking in their capability. What used to require a professional visual effects artist and hours of work can now be done by almost anyone, in minutes, on a laptop. The barrier is essentially gone.

What Deepfakes and AI Generation Actually Mean for Ordinary People

There is a tendency to think this is a celebrity problem. Famous people, politicians, public figures. And yes, they are disproportionately targeted. But the technology has democratized. It does not care whether you have a million followers or forty-three. All it needs is a clear enough image of your face.

Deepfake pornography is the most talked-about abuse, and for good reason. It is violating in a way that is genuinely hard to put into words. Your face, your identity, your likeness, attached to something deeply intimate and deeply wrong, shared without your knowledge or consent. The psychological damage that causes is real and documented.

But it goes beyond that. AI-generated images can be used to fabricate evidence. To harass someone at their workplace. To blackmail. To impersonate. To build fake profiles for scams. To put your face in a context, political, criminal, anything, that you were never in. The range of harm is wide.

And right now, the legal frameworks to deal with this are lagging badly behind the technology. Some countries have started legislating against deepfake pornography specifically. But enforcement is slow, jurisdictions are complicated, and by the time any legal remedy arrives, the damage is often already done.

The Hard Truth About Photos You Have Already Posted

Removing photos from the internet is genuinely hard. I want to be upfront about that. You can delete a post from Instagram, but that photo may have already been saved, screenshotted, scraped by a third-party app, or cached somewhere. The original deletion helps, but it is rarely a complete solution.

That said, it matters. Here is what I did and what I would recommend.

Start with a Google search of your own name. Look at the Images tab specifically. You will probably find things you forgot existed. Old forum profile pictures. A photo from a news article. An image from an event page. Note every source.

Then run a reverse image search using your clearest, most widely used photos. Google Images lets you upload a photo and find where else it appears online. TinEye is another solid tool for this. Yandex, surprisingly, has one of the most powerful reverse image search engines available and often surfaces results the others miss.

Once you have a list of where your images live, start requesting removal. Most platforms have a reporting mechanism. For websites, you look for a contact email and send a direct request citing privacy concerns. Many site owners will comply, especially smaller ones. Larger platforms have formal processes.

Google has a tool called "Results About You" that lets you request the removal of certain personal information from search results. This does not delete the content from the source website, but it does delist it, which meaningfully reduces discoverability.

For social media, go through your profiles and audit what is public. Ask yourself: does this photo need to be public? For most people, the answer to that question, applied honestly, will result in a significant reduction in publicly visible images.

What About Photos Other People Posted of You

This is where it gets more complicated. You have rights over your own likeness in many jurisdictions, but exercising those rights requires knowing the photos exist, knowing where they are hosted, and then navigating each platform's individual process for reporting content you appear in but did not post yourself.

Facebook and Instagram allow you to request removal of photos you appear in. Same with Google Photos if they were shared. LinkedIn lets you flag images. Twitter and X have processes too, though the responsiveness varies.

For photos on websites outside of major platforms, it gets harder. You can send a formal request citing data protection regulations. If you are in the European Union or UK, GDPR gives you a genuine right to erasure that carries legal weight. Data protection authorities in those regions can assist if site owners refuse. In the US, the legal landscape is patchier, but many states are beginning to pass their own privacy legislation.

The honest reality is that some photos will be effectively impossible to fully scrub. But reducing the total number of high-quality, publicly accessible images of yourself meaningfully reduces your exposure. It is about raising the effort required to target you.

Going Forward: What a Healthier Relationship With Posting Looks Like

I am not saying stop living your life online. I am saying be deliberate about what goes where.

Before posting a photo of yourself, think about whether it needs to be public or whether it could be shared privately with the people who actually matter. Think about the quality and clarity of the image. A distant, low-resolution photo in a group setting is far less useful for AI manipulation than a sharp, well-lit solo portrait.

Set your social media profiles to private where possible. Review your tagged photos regularly and untag yourself from anything you are uncomfortable with. Ask friends to check with you before posting photos of you.

Do a reverse image search of yourself every few months. Make it a habit, like checking your credit report. You want to catch new appearances of your image early.

If you have professional photos online, consider whether watermarking them is appropriate. If you are a public figure or have a professional presence that requires some photos to be publicly accessible, watermarks create friction for anyone trying to misuse those images.

This Is a Collective Problem That Requires Individual Action Right Now

The platforms should do more. The AI companies should build stronger safeguards. Legislators need to move faster. All of that is true.

But waiting for institutions to protect you is a strategy that has consistently let people down. The Grok situation was a reminder that these tools exist, they are being used, and the people building them are not always prioritizing your safety as a user or as a subject.

Your face is yours. Your likeness is yours. And while the internet has made it easier than ever to share those things freely, it has also made it easier than ever for that sharing to be exploited in ways that can genuinely upend your life.

Taking stock of your digital image footprint is one of the most practical things you can do for your personal safety right now. It is tedious. Some of it will feel futile. But it matters.

Start today. Search your name. Run the reverse image search. Make the removal requests. Audit your profiles.

Because the alternative, finding out the hard way that someone already did something with your image, is a situation you deserve to be ahead of, not behind.

The AI Race Is a Mirror, and Every Company Is Sprinting Toward It

Om — Fri, 20 Feb 2026 11:37:00 GMT

A few days ago I witnessed something that warranted serious examination. Anthropic launched Claude Opus 4.6 carrying a one million token context window. Within hours, OpenAI released Codex 5.3. The following day, Google followed with Gemini 3.1 Pro. Three major capability releases across three independent organizations in under 48 hours. The phenomenon demands explanation beyond coincidence.

The underlying mechanism is more structural than it appears. The research community powering all three organizations draws from a largely shared intellectual pool. Preprints circulate publicly. Researchers move between institutions. Architectural insights propagate through conference proceedings and open publication. The conditions that once allowed a single laboratory to maintain a meaningful discovery lead for 12 to 18 months have dissolved. The moment a capability threshold becomes legible through public signals, competing organizations with similar infrastructure are already positioned to reach the same threshold through parallel effort. What differentiates these organizations today is execution velocity and positioning precision, the capacity to ship with quality and to understand exactly for whom they are shipping.

This observation carries serious implications for anyone building in this space. The competitive window between a meaningful release and a competitor's equivalent response has compressed to weeks and, in some cases, days. A product that ships today carries a viable differentiation window of perhaps one to two months before the gap closes. This is the operating reality. Shipping fast and shipping well are the same requirement. Organizations that treat them as separate priorities will find that optimizing for one at the expense of the other produces neither durability nor relevance.

Examining the strategic positioning of the three leading laboratories reveals deliberate audience segmentation rather than accidental divergence. Anthropic has concentrated its reputation among developers and programmers, a positioning reinforced by consistent benchmark performance in code generation and reasoning tasks. OpenAI has pursued the general purpose layer, the ambient assistant for the broadest possible population. Google has oriented toward creative productivity, leveraging its existing penetration into the daily workflows of billions of users. These are purposeful choices reflecting each organization's assessment of where its moat is deepest. The durability of these positions, as each laboratory inevitably expands into adjacent territory, remains the central strategic question of the coming years.

Memory architecture is, in my assessment, one of the most consequential and least discussed competitive dynamics in the current landscape. Memory functions as a switching cost operating beneath the surface of feature comparison. Each interaction a user completes with a given system, each preference that system internalizes, each pattern of communication it absorbs, accumulates into a personalized context that carries genuine relational weight. Requesting a user to abandon that accumulated context and rebuild it from scratch with a competing system is a meaningful emotional and practical ask. The friction of switching has very little to do with feature parity and almost everything to do with the experience of starting over with a system that does not yet know you.

The product challenge this creates is among the most interesting in applied AI research. The organization that develops a reliable methodology for compressing the context acquisition timeline, for building an accurate and nuanced model of a new user within minutes rather than months, will hold an asymmetric advantage. A system capable of inferring communication preferences, intellectual disposition, and contextual priorities from three minutes of sparse interaction, at a depth that competing systems require a year of consistent engagement to approximate, fundamentally changes the switching calculus. The inference must operate behind the interface entirely. The user experience of the output should feel like natural understanding, with the mechanism remaining invisible.

Marketing and public positioning have emerged as genuine strategic instruments in this competitive environment, a development that warrants academic acknowledgment even if it sits outside traditional technology analysis. When Anthropic introduced public messaging that implicitly differentiated its values orientation from the direction OpenAI was pursuing, and when OpenAI deployed a Super Bowl advertising campaign, both organizations were engaged in something more substantive than brand awareness. They were competing for identity alignment with their target users. Users who left ChatGPT and gave Claude sustained engagement did so in part because the values signaled in Anthropic's public communication resonated with their own. The advertisement opened the door. Product quality and perceived value alignment kept it open.

The deeper principle here is one that every organization in this space must eventually reckon with. Differentiation grounded in stated principles carries limited value. Differentiation grounded in principles that are structurally visible in product behavior, in what a system chooses to do and what it chooses to decline, in how it treats the person using it, carries compounding value over time. It attracts a community of users whose loyalty is rooted in belief rather than convenience, and whose advocacy operates through channels that paid acquisition is structurally incapable of replicating.

I want to turn now to a definitional problem that I believe the field has handled with insufficient precision. The prevailing conception of artificial superintelligence, broadly framed as intelligence that surpasses all human cognitive capacity across all domains, is simultaneously too abstract to be operationally useful and too mythological to serve as a meaningful research target. I propose a functional redefinition grounded in four observable and testable properties.

Artificial superintelligence, properly understood, is a system that: holds and actively reasons across a massive, continuously updated body of contextual information; applies common sense inference to that context in ways that produce decisions a human reasoner would recognize as sound; generates accurate predictions of downstream consequences across meaningful temporal horizons; and intervenes in present conditions in ways that are both genuinely novel and demonstrably useful, producing outcomes that human intelligence could eventually reach but at a fraction of the time and with greater consistency.

This definition is operational. Each property can be evaluated. A system holding the complete operational context of a hospital network, predicting patient deterioration six hours before clinical presentation, and triggering precise interventions in the present satisfies this definition, and the value it would produce is concrete and measurable. The missing property in current frontier systems, and the one I consider most underemphasized in public discourse, is goal coherence across extended sequential decision chains. Present systems demonstrate remarkable local reasoning capacity but exhibit meaningful drift when required to maintain a consistent objective across hundreds of interdependent steps. Resolving this is, in my view, the central technical challenge separating current capability from the threshold this definition describes.

The question of whether human intelligence itself has an asymptotic ceiling, and whether we are approaching it, deserves serious treatment. Biology imposes genuine physical constraints on individual cognitive throughput. Working memory capacity, attentional bandwidth, and neural signal propagation speed are architectural features of the human brain that additional education and training produce diminishing returns against at the frontier. The researchers currently building the most sophisticated AI systems are operating at or near the limits of what individual human minds can coordinate. The mathematics, the architectural reasoning, and the systems level thinking these tasks require are tasks that strain individual cognitive capacity.

It is important, though, to distinguish between the ceiling of individual human intelligence and the ceiling of collective human intellectual progress. These are separate phenomena. Collective intelligence, constituted by human researchers working in dense coordination with improved tooling, better collaboration infrastructure, and AI assistance integrated into the research process itself, continues to climb even as individual ceilings hold. The more troubling version of this question is whether the final generative steps toward systems satisfying the redefinition proposed above require insights that human cognition is structurally incapable of producing alone, and whether current AI systems are similarly incapable because their training corpora are constituted entirely of human generated knowledge. That would represent a genuine epistemic deadlock. The prevailing research position holds that incremental progress circumvents that trap. The candid assessment is that the location of that ceiling remains unknown, and the pace of approach makes the question urgent.

What this analysis ultimately surfaces is a convergence of pressures that will determine which organizations and builders carry meaningful influence five years from now. Velocity in shipping. Clarity in identity. A memory architecture that compresses the user acquisition curve. Values that are structurally visible in product behavior rather than merely declared in public communication. And a research orientation that takes the operational redefinition of artificial superintelligence seriously enough to build toward its specific, testable properties rather than toward an abstraction. The window for establishing these positions is present right now. It is also compressing at a rate commensurate with the pace of releases described at the opening of this piece. The organizations moving with both speed and intellectual seriousness in this moment are the ones producing the conditions that everyone else will eventually be forced to operate within.

i built a privacy standard i actually believe in.

Om — Thu, 19 Feb 2026 12:04:03 GMT

The AI age is officially here, but it arrived with a hidden price tag that I am not willing to pay. Everywhere you look, the giants of the industry - Meta, Google, OpenAI - are racing to gather as much "fodder" as possible to train their next models. Your clicks, your search history, and even your private preferences have become the raw materials for a multi-billion dollar machine. To them, data is the new oil. To me, that feels like a fundamental violation of the digital home we are trying to build at IBBE.

I have always believed that a company should exist to serve its people, not the other way around. If you use our services, you are trusting us with a piece of your life. That trust is sacred. I see too many companies treating privacy as a legal hurdle to clear with fine print and "accept all" buttons. For me, privacy is not a policy; it is the very architecture of how we build. If we claim to put consumers first, then protecting their digital identity is the most important role I have as a founder.

Today, I am introducing a new standard for our ecosystem: the IBBE Privacy Laws. The core of this shift is something we call tokenized-first data architecture. It is a technical way of saying that the moment you share a piece of information with us, it is transformed into a meaningless string of characters. Raw values like your name or your specific learning progress never touch our central database. They exist as tokens that only resolve when the system is actively serving you. If I cannot see your data, I cannot sell it, and no one can steal it.

This is a direct response to the world we live in. We are entering an era where data is being licensed and sold behind closed doors to train AI that eventually gets sold back to you. I want IBBE to be a place where you feel at home - a place where you can explore, learn, and grow without the nagging feeling that someone is looking over your shoulder to build a behavioral profile. Your data should stay with you. It should serve only you. There are zero advertisements on our platforms, and there is zero interest in training models on your personal journey.

I know that being "compliance-ready" is the corporate standard, but I want to go further. We are aiming for a level of strict, technical assurance that makes misuse impossible by design. This is about data sovereignty. By December 17, 2026, this standard will be fully operational across every surface IBBE operates on. It is a massive undertaking to rebuild systems this way, but it is the only path that aligns with my ideology of building products that empower rather than exploit.

My philosophy has always been about agile, iterative growth, but privacy is the one area where I refuse to pivot. As we grow IBBE and expand our tech stack, this architecture will be our foundation. I want our users to know that while the rest of the world is figuring out how to monetize their identity, we are busy building the walls to protect it. We are even going to publish our full technical process openly because a promise of privacy means nothing if I cannot show you the proof.

Ultimately, this comes down to respect. In a world where AI is hungry for every byte of information it can find, the most radical thing a company can do is let the user keep what is theirs. IBBE is built on the intersection of business, technology, and education, but it is held together by the belief that the person behind the screen matters more than the data they generate. This is my commitment to you: your data is yours, and at IBBE, it stays that way.

read the full privacy commitment →

YouTube Outage

om — Wed, 18 Feb 2026 10:15:00 GMT

yesterday morning, while many of us were reaching for our morning content, youtube experienced a significant service interruption. it wasn't a total blackout—video links still worked if you had them—but the homepage and recommendations were largely unresponsive.

as someone deeply invested in building platforms like learn.ibbe.in, i spent the day reflecting on what this teaches us about software at scale. it is humbling to realize that even with the best engineers in the world, systems this complex are organic, living things that sometimes behave unexpectedly.

from my understanding of distributed systems, there are three likely technical scenarios that explain how a partial outage like this happens. these aren't just theories; they are standard patterns in software engineering that i am now keeping a close eye on for my own work.

the automatic break we often imagine outages happen because a developer typed the wrong line of code, but in modern tech, systems often run on autopilot.

youtube’s recommendation engine is likely an ai that constantly retrains itself based on incoming data streams. i learned that this can lead to a scenario where the system "breaks" itself without human intervention.

if the ingestion pipeline receives a massive spike of anomalous data—what we might call "noise" or "garbage data"—the model might update its internal map with errors. the code is perfect, but the state is corrupted. the system tries to read this new, flawed map to serve a video suggestion, hits a logical wall, and stops.

it reminds me that validating the data entering our systems is just as important as the code processing it.

the hidden bug another technical possibility involves what we call "feature flags." in modern ci/cd pipelines, engineers often push code that stays "turned off" or dormant for weeks. it sits there, waiting.

the disruption might have occurred when a timer or a manual switch finally activated a new feature that had been deployed days ago. even if that code passed every unit test in a staging environment, the reality of production—millions of requests per second—is different.

activating that dormant code could have triggered a "race condition" or a memory leak that only appears at massive scale. it is a reminder that deployment is not the same as release, and toggling features needs to be done with incredible caution.

the chain reaction this is perhaps the most valuable lesson for the architecture i am building. modern apps are collections of "microservices." one service handles your login, another handles search, and a separate one handles "what to watch next."

the outage appeared to be a classic "hard dependency" failure. it seems the homepage was programmed to wait for the recommendation service to respond before rendering anything. when the recommendation service stalled, the entire homepage hung in limbo.

it is like a car where the engine is running perfectly, but the car refuses to move because the dashboard radio isn't working.

applying this to my work this event has been a massive learning opportunity for how i approach the architecture of the ibbe ecosystem. as i build the learning management system, i am thinking deeply about "graceful degradation."

if my recommendation engine for the "next chapter" fails, the student's dashboard should still load their current progress. i want to ensure my components are "loosely coupled"—meaning if one service has a moment, the rest of the application stays calm, functional, and helpful.

it is about anticipating the unexpected and ensuring the user always feels supported, even when the system is working hard in the background to recover.

building software is a journey of constant iteration. seeing a giant like google navigate these challenges only motivates me more to build with thoughtfulness, resilience, and a user-first mindset.

I Cut My Site's Load Time from 15s to 3s with Surgical Suspense Boundaries

om — Mon, 16 Feb 2026 14:50:00 GMT

I built jobs.ibbe.in as a hiring site. At some point I had to sit down and face the reality that the site was embarrassingly slow. We're talking 15+ seconds to get anything on screen. On a good connection. That's the kind of load time that makes people close the tab before your content even exists.

I fixed it. The site now loads meaningful content in under 3 seconds. Here's exactly what I did and why it worked.

The actual problem, in plain terms

Every page on the site was a single async component. That means when someone visited the homepage, Next.js would start the request, go fetch data from Supabase, wait for every query to finish, and only then begin sending any HTML to the browser. The user sat staring at a blank screen the entire time.

The worst part is that most of what was on screen had absolutely nothing to do with data. The section headings, the decorative backgrounds, the category pills, the newsletter signup form — all of that is static. It's the same for every visitor. And yet it was blocked, sitting behind database queries it had zero relationship with.

This is the core mistake: treating a whole page like one big async operation when only a small slice of it actually needs data.

React Suspense and streaming - what they actually do

Before getting into the changes, here's the mental model you need.

When Next.js renders a page using React Server Components, it has the ability to stream HTML to the browser in chunks. Instead of waiting for everything before sending anything, it can send the parts it already knows, then fill in the gaps as data becomes available.

<Suspense> is the mechanism that makes this work. You wrap a component in <Suspense fallback={<YourSkeleton />}>, and React treats that boundary as "figure this out later." Everything outside the Suspense boundary renders and streams immediately. The skeleton shows up as a placeholder. When the data comes back, the real component swaps in.

The key insight is this: only the component that actually fetches data belongs inside Suspense. Static siblings should live outside it, in what I call the shell, so they render instantly regardless of what the database is doing.

Phase 1: The foundation

Before the surgical audit, I set up the infrastructure that made all of this possible.

I created lib/queries.ts with 13 cached query functions. Each one uses Next.js's unstable_cache with a 60-second revalidation window, and each one selects only the columns it actually needs from Supabase rather than doing a SELECT *. This alone reduced query payload size significantly.

import { unstable_cache } from 'next/cache'
import { createClient } from '@/lib/supabase/static'

export const getFeaturedArticles = unstable_cache(
  async () => {
    const supabase = createClient()
    const { data } = await supabase
      .from('articles')
      .select('id, title, slug, excerpt, cover_image, category, published_at')
      .eq('featured', true)
      .order('published_at', { ascending: false })
      .limit(6)
    return data ?? []
  },
  ['featured-articles'],
  { revalidate: 60 }
)

I also added skeleton components in components/skeletons.tsx so that when a Suspense boundary is waiting, it shows a realistic placeholder instead of nothing. And I configured the client router cache in next.config.ts:

experimental: {
  staleTimes: {
    dynamic: 60,
    static: 300,
  }
}

This tells Next.js to keep prefetched pages in the browser's memory longer before re-fetching them. The dynamic number covers pages with server data; the static number covers fully static pages. Navigating back to a page you've already visited gets instant without another round trip.

Phase 2: The surgical audit

With the foundation in place, I went through every public-facing page and asked the same question: what on this page is static, and what genuinely requires a database call? Then I moved everything static outside of Suspense.

The homepage

Before the fix, the "Available Positions" section was one big async component inside a single Suspense. The section heading ("Available Positions," the "Live Roles" badge) was waiting behind a database query for the job cards.

After the fix:

// app/page.tsx

export default function HomePage() {
  return (
    <main>
      {/* This renders immediately — no data dependency */}
      <section>
        <h2>Available Positions</h2>
        <span className="badge">Live Roles</span>

        {/* Only this waits for the database */}
        <Suspense fallback={<JobCardsGridSkeleton />}>
          <FeaturedJobCards />
        </Suspense>
      </section>
    </main>
  )
}

The heading appears instantly. The job cards stream in when the query resolves. Users know what they're looking at before the data arrives.

The stories listing page - the biggest impact

This was the worst offender. The entire page — hero section, category pills, featured article, article grid, newsletter form — was wrapped in a single Suspense. Nothing appeared until every article query finished.

After the audit, I split it into five independent pieces:

// app/stories/page.tsx

export default function StoriesPage() {
  return (
    <main>
      {/* Instant — decorative, title, CTA buttons, no data */}
      <HeroSection />

      {/* Instant — rendered from a hardcoded CATEGORIES array */}
      <CategoriesSection />

      {/* Streams in independently */}
      <Suspense fallback={<FeaturedArticleSkeleton />}>
        <FeaturedArticle />
      </Suspense>

      {/* Streams in independently */}
      <Suspense fallback={<StoriesGridSkeleton />}>
        <ArticlesGrid />
      </Suspense>

      {/* Instant — static form, no data */}
      <NewsletterCTA />
    </main>
  )
}

One specific thing I had to remove was a dynamic article count badge in the hero ("X articles published"). It sounds like a minor feature, but it was the only data dependency in the hero section, which meant the entire hero was blocked behind a database count query. Removing it let the hero render instantly. The tradeoff is obvious and worth it.

The category pills are worth explaining. The <CategoriesSection> component previously fetched categories from the database. I replaced it with a hardcoded CATEGORIES array imported from a constants file. Categories change maybe once every few months. Hitting the database for them on every page load was unnecessary. Static data belongs in static config.

The story category page

Each category page had a similar structure problem. The hero (icon, title, description, decorative wave separator) was inside an async function that also fetched article data. The hero doesn't need data — it's built from a static CATEGORIES config object keyed by slug.

// app/stories/category/[slug]/page.tsx

export default function CategoryPage({ params }) {
  const category = CATEGORIES[params.slug]

  return (
    <main>
      {/* Instant — from static config, zero DB */}
      <CategoryHero
        icon={category.icon}
        title={category.title}
        description={category.description}
      />

      {/* Instant — static string */}
      <h3>Latest in {category.title}</h3>

      {/* Tiny — just an article count */}
      <Suspense fallback={<CategoryStatsSkeleton />}>
        <CategoryStats slug={params.slug} />
      </Suspense>

      {/* Main content */}
      <Suspense fallback={<CategoryArticlesSkeleton />}>
        <CategoryArticlesGrid slug={params.slug} />
      </Suspense>
    </main>
  )
}

The story detail page

Individual story pages were already fast because of generateStaticParams and ISR (Incremental Static Regeneration). The article content itself is pre-built at deploy time and served as static HTML — no database query on the critical path.

The only remaining issue was that "More Stories" (related articles at the bottom) was being fetched in the same async function as the main article, making them sequential. I extracted it into its own component with its own Suspense boundary:

// app/stories/[slug]/page.tsx

export default async function StoryPage({ params }) {
  // This is a static page — content is pre-rendered, instant
  const article = await getArticle(params.slug)

  return (
    <article>
      <ArticleContent article={article} />

      {/* Loads in background while user reads */}
      <Suspense fallback={<RelatedArticlesSkeleton />}>
        <RelatedArticlesSection currentSlug={params.slug} category={article.category} />
      </Suspense>
    </article>
  )
}

The user gets the article immediately. Related stories load in the background while they read. By the time they finish the article, the related section has almost certainly already populated.

The RLS bug that was causing 404s

In Phase 1, I had switched the job detail page to use a static Supabase client (initialized with just the anon key) to avoid creating a new server client on every request. The intention was good — the static client is more efficient for public data.

The problem is that the jobs table has Row Level Security enabled in Supabase. RLS policies evaluate whether the requesting user has permission to read a row. The static client has no auth context, which means it looks like an anonymous public request. If the RLS policy requires authentication to read job rows, the static client gets back null, the page calls notFound(), and the user sees a 404.

The fix was straightforward: revert the job detail page to the server client (which uses cookies and has the auth context), while keeping the static client for generateStaticParams where only slugs are needed at build time and RLS is irrelevant.

// app/jobs/[slug]/page.tsx

import { createServerClient } from '@/lib/supabase/server'  // has auth context
import { createStaticClient } from '@/lib/supabase/static'  // anon only

// Build time — only needs slugs, no RLS concern
export async function generateStaticParams() {
  const supabase = createStaticClient()
  const { data } = await supabase.from('jobs').select('slug')
  return data?.map(job => ({ slug: job.slug })) ?? []
}

// Runtime — needs auth context for RLS
export default async function JobPage({ params }) {
  const supabase = createServerClient()
  const { data: job } = await supabase
    .from('jobs')
    .select('*')
    .eq('slug', params.slug)
    .single()

  if (!job) notFound()
  // ...
}

What I deliberately left alone

The jobs listing and category pages run a JobSearch client component that receives all its data as props. Since it's a client component handling search, filtering, and state, splitting it further would require a completely different architecture. The performance is acceptable as-is, and a refactor there is a separate project.

Auth-gated pages (login, signup, candidate profile, application status) were also untouched. Auth checks must complete before rendering because the decision of what to show depends on who you are. Streaming HTML before knowing the user's auth state would either expose content to wrong users or require client-side corrections after the fact. The current pattern is correct for those pages.

The net result

15 seconds to under 3 seconds. The site went from watching a blank screen for an uncomfortable amount of time to feeling fast. Content appears progressively: shell first, then data as it arrives, in the right order, with skeletons holding space in between.

The principle that made all of this possible is simple. Only put components inside Suspense if they actually need to wait for data. Everything else belongs in the shell and should be on screen immediately. When you audit a slow page with that question in mind, you'll find that most of what's blocking the render has no business doing so.

We're Done Asking You to Subscribe

Om — Sun, 15 Feb 2026 07:37:00 GMT

I've been thinking a lot about education lately. About what actually works and what's noise. We're about to launch our educational content at ibbe, and I wanted to share the framework we've built. This is probably going to sound extreme to some people, but I have reasons for every single choice here. Bear with me.

Let me start with the most controversial one. Zero "like share subscribe" anywhere. I mean it. In the video, in the description, on a slide, nowhere. You know why? Because the second you start chasing metrics, you stop serving students. I've watched too many creators optimize for engagement instead of learning. They'll drag out explanations, add cliffhangers, tease upcoming content, all to keep you watching or coming back. That's fine for entertainment. For education though? Poison.

When a student sits down to learn, their goal is simple: understand the material and move on with their life. My job is to help them do that as efficiently as possible, period. The moment I start asking for likes, I'm admitting that my priority has shifted. I'm saying "hey, my channel growth matters more than your time." And look, I get that creators need to grow. I get that metrics matter for visibility. But there are other ways to grow. Word of mouth. Quality. Results. If our teaching is good enough, students will tell their friends. If it's good enough, they'll come back. I'd rather grow slowly based on actual value than quickly based on manipulation.

Think about it from a student's perspective. You're stressed about exams. You're trying to understand a difficult concept. You click on a lecture hoping for clarity. And the first thing you hear is "hey guys, before we start, please like and subscribe." Immediately, your brain registers that this person wants something from you. It's transactional now. The trust is already compromised, even if slightly. I've experienced this so many times as a student myself. That tiny moment of disappointment when you realize the teacher is also a salesperson. I'm done with that dynamic.

Same logic applies to the tone and behavior during lectures. No vulgar jokes, no screaming, no jargon when simple words exist. I've seen teachers try to be "relatable" by throwing in jokes that make half the class uncomfortable. I've seen lectures where the teacher is basically performing instead of teaching. Look, I get it. Teaching is hard. Keeping attention is hard. You're standing there or sitting there talking for hours, and you can feel when energy drops. The temptation to crack a joke or raise your voice or do something dramatic to grab attention back is real.

But here's what I believe: the solution is better explanations, better structure, better visuals. The solution is respecting the student's time so much that they stay engaged because the content is genuinely valuable. If I'm explaining something clearly, building on previous concepts logically, using good examples, students will pay attention. They'll pay attention because they're actually learning, and learning feels good. It's rewarding. Screaming and crude humor are shortcuts that compromise the learning environment. They might work for thirty seconds, but they break focus. They make some students uncomfortable. They cheapen the entire experience.

And jargon, oh man. I have strong feelings about jargon. There's this tendency among teachers to use complex terminology to sound smart or authoritative. Sometimes it's subconscious. Sometimes it's deliberate. Either way, it's harmful. If I can explain something in simple words, why would I use complicated ones? The goal is understanding, clarity. Using jargon when simpler words exist is gatekeeping. It's making knowledge artificially inaccessible. Now, there are times when technical terms are necessary. When they're the precise tool for the job. Fine. But even then, explain them first. Define them. Make sure everyone's on the same page before moving forward.

Here's something that might surprise you: notes will link directly to the notes. That's it. You click, you download. I'm sick of seeing "check the link in bio" or "visit our website" or "find this on our app under resources tab." Why? Why make students jump through hoops? They're already paying attention to your lecture. They're already engaged with your content. Give them what they need when they need it. Every extra step is friction, and friction kills momentum in learning.

Think about the student journey here. They're watching a lecture. The teacher mentions notes. They want those notes. They're motivated right now, in this moment. But then they have to pause the video, go to the description, find the right link among ten other links, click it, maybe sign up for something, navigate a website, find the resources section, search for the specific lecture, and finally download. By the time they've done all that, their momentum is gone. Their focus has shifted. Maybe they forget to come back to the video. Maybe they get distracted by something else on the website. You've lost them. And for what? So you could drive traffic to your website? So you could collect emails? Those might be valid business goals, but they come at the expense of learning. I'm choosing learning.

And this extends to the video description too. Zero marketing links. Zero "join ibbe" buttons. Zero "visit our website" or "download our app" nonsense. The description will have exactly two things: the notes link and the feedback form link. That's it. Why? Because anything else is a distraction. Anything else is saying "we care about converting you into a user more than we care about you learning right now." Every marketing link in that description is competing for the student's attention against the actual educational materials. I'm removing that competition entirely.

When a student opens the description, they should find exactly what they need for that lecture. Nothing more, nothing less. The notes they need to follow along. The form to report issues or ask urgent questions. Done. Clean. Focused. If they want to know more about ibbe, they can search for us. They can find our website. But I'm keeping that separate from the learning experience. The lecture space is sacred. It's only about learning.

Teachers will have zero phones during lectures. I mean this. Put it away an hour before you go live and keep it away until you're done. This rule exists because I believe presence matters. When you're teaching, you're there. Fully there. Students can sense when you're distracted or when part of your mind is elsewhere. Teaching is already intimate in a weird way. Someone is trusting you with their time and their future. They're listening to your voice, following your logic, building understanding based on your guidance. The least you can do is show up completely.

I've been in so many lectures where the teacher checks their phone mid-explanation. Sometimes they try to be subtle about it. Sometimes they're blatant. Either way, the message is clear: something else is more important than this moment. And you know what happens? Students check out mentally. If the teacher thinks this is worth interrupting for a text message, why should students treat it as sacred time? It sets a tone. It establishes that this interaction is casual, interruptible, negotiable. I want the opposite. I want teaching to be treated as the serious, focused work that it is.

Now let's talk about lecture length. Every lecture, regardless of chapter size, stays under three hours. This is based on actual attention span research and my own experience as a student. After three hours, your brain is mush. Even if you think you're following along, your retention drops dramatically. I've sat through four-hour, five-hour marathon sessions. At some point, you're physically present but mentally gone. You're watching words happen but concepts are slipping through. And then you have to rewatch later anyway, so what was the point?

So we've designed everything to fit inside this window. If a chapter is massive, we split it across multiple sessions. This might mean more sessions total, but each one is digestible. Each one respects the limits of human attention and memory. And honestly, this constraint makes us better teachers. When you know you have limited time, you get ruthless about what matters. You cut the fluff. You focus on core concepts. You become efficient with clarity and precision.

Each session gets broken into parts with built-in active recall moments. This is huge for me. Active recall is one of the most powerful learning techniques we have, backed by tons of research. The idea is simple: instead of passively reviewing information, you actively try to retrieve it from memory. You test yourself. You answer questions. This process of retrieval strengthens the memory pathway. It shows you what you actually know versus what you only think you know.

So throughout every lecture, we're building in these moments. The teacher will pause and ask questions. These are textbook questions, previous year questions, and custom questions aligned with our education framework. They're strategic. They're timed to reinforce what was covered. They force students to engage actively instead of zoning out. And because we're using real textbook questions and past papers, students are also getting exam practice simultaneously. They're learning the material and learning how it gets tested. Two birds, one stone.

The live chat is off. This might be the most controversial decision. Everyone expects live chat in educational streams now. It's become standard. But here's what happens with live chat: spam, distractions, arguments, inappropriate comments, people asking questions that get answered five minutes later in the lecture. It's chaos. I've watched live chats during educational streams. They're rarely productive. Mostly, they're noise. Inside jokes between regular viewers. Off-topic conversations. Someone asking "when will you cover chapter 5" while the teacher is mid-explanation of chapter 3. It pulls the teacher's attention away and it distracts students who are actually trying to focus.

And there's another problem: the pressure on the teacher to monitor and respond. Now the teacher is trying to teach and simultaneously watch a scrolling chat for important questions. It splits their attention. It interrupts their flow. I've seen teachers stop mid-sentence because they caught a question in chat. Sometimes that's helpful. Often, it derails the lecture. The explanation loses momentum. Other students get confused because the tangent made sense to one person but disrupted everyone else's understanding.

Instead, we're doing something different. A team will watch every lecture in real time. There's a direct link in the description. Anonymous, instant. Student clicks, types their issue or question, hits submit. Within a split second, it appears on a screen the teacher can see. There's zero friction here. The form is anonymous. You click the link and the input modal opens immediately. You type, you submit, you're done. You go back to focusing on the lecture.

This system gives us the benefits of real-time feedback without the chaos of live chat. If there's a genuine error, audio issues, a slide that's unclear, students can report it immediately. The teacher sees it and can address it. If there's a question that many students are having, that signal comes through clearly because we're collecting structured input rather than parsing chaos. And because it's anonymous, students feel comfortable pointing out mistakes or asking questions they might feel embarrassed about in a public chat.

The team watching has a specific role. They're filtering what reaches the teacher. They're making judgment calls about what's urgent versus what can wait. They're ensuring the teacher only sees things that truly matter in the moment. This protects the teacher's focus while still keeping a feedback loop open. It's the best of both worlds.

One more thing about preparation. Teachers arrive an hour before stream time. Always. This gives time for tech checks, mental preparation, review of materials, whatever is needed. I'm serious about this. Rushing creates mistakes. Rushing creates stress. And stress transfers to students whether you realize it or do anything about it. I've watched streams where the teacher is still setting up audio five minutes after the scheduled start time. Students are waiting. The chat fills with "when are we starting?" Energy is already wrong before the lecture even begins.

Showing up early is showing respect for the process and respect for students' time. It means when the stream starts, we start. The teacher is calm, prepared, ready. The tech works. The materials are organized. There's a professionalism to it that matters. Teaching might be delivered casually, in a friendly tone, but the preparation behind it is serious. Students might see a relaxed teacher explaining concepts clearly. What they see is the result of an hour of groundwork that happened before they arrived.

I know this all sounds intense. Maybe it is. But I'm building this because I remember being a student. I remember the frustration of sitting through a lecture where the teacher spent five minutes asking people to subscribe. I remember losing focus because someone was making inappropriate jokes. I remember hunting for notes across three different platforms. I remember watching teachers check their phones mid-explanation. I remember sitting through four-hour sessions where my brain stopped working after hour two. I remember clicking on video descriptions and having to scroll past five different links to other courses, apps, social media pages, just to find the actual lecture materials.

Every single rule here exists because I experienced the opposite and it sucked. These are solutions to real problems. Problems that might seem small individually but compound into a terrible learning experience. And the thing is, nobody talks about this stuff. There's this assumption that educational content is educational content. That delivery and environment are secondary to the information itself. I think that's completely wrong. How you teach matters as much as what you teach. Maybe more.

This is my attempt to build something better. Something that puts learning first, second, and third. Everything else is secondary. Will this approach limit our growth? Maybe. Will it make some people think we're too rigid or serious? Probably. I'm okay with that. I'm building for students who want to actually learn, who are tired of the noise and performance and manipulation that's become normal in online education. If that's a smaller audience, fine. I'd rather serve them well than serve everyone poorly.

Building a Valentine That Actually Means Something

Om — Sat, 14 Feb 2026 18:43:00 GMT

I wrote about gifting frameworks a few days ago and people asked me to show it in action. So here we go. Valentine's Day is coming up and most of you are going to buy flowers, chocolate, maybe a card with a generic message. You'll spend money and feel like you did the thing. But did you actually say what you needed to say? Did the gift carry weight? Did it land?

Let me show you how to build something that hits different. We're going to use Valentine's Day itself as the anchor. The history. The mythology. The weird, bloody, beautiful origin story. Then we're going to merge that with your relationship. This is the framework in practice.

Here's what most people miss: Valentine's Day started as a Roman fertility festival called Lupercalia. Priests sacrificed goats, stripped naked, ran through the streets whipping people with animal hide strips. Women lined up to get hit because they believed it would make them fertile. This was chaos. This was primal. This was about creation and renewal and taking risks for what you wanted. Then a priest named Valentine got executed for marrying couples in secret when the emperor banned marriage. He defied power for love. He paid with his life. That's the origin. Blood and rebellion and faith in connection.

From Wikimedia Commons

Now forget the sanitized Hallmark version. Forget the heart shaped boxes. Think about what those stories actually represent. Chaos that leads to creation. Defiance that proves devotion. Ritual that transforms ordinary moments into sacred ones. These are metaphors. These are tools. You can use them.

Ask yourself: how did your relationship start? Was it chaotic? Did you meet in some ridiculous way that makes no logical sense but somehow worked? That's your Lupercalia moment. Was there a point where one of you took a risk? Maybe someone confessed feelings first. Maybe someone moved cities. Maybe someone chose you over the easier option. That's your Valentine defiance. Find those moments. Write them down. Those are your emotional anchors.

Now here's where it gets specific. You take that historical metaphor and you attach it to a physical object. Maybe you write them a letter on aged paper with a wax seal. You structure it in three parts: the chaos of how you met, the risk you took to be together, the ritual you've built since then.

Generated with AI via Gemini AI Generated

You reference Lupercalia at the beginning, Valentine in the middle, and your future at the end. The letter becomes a timeline. The history gives it gravitas. Your story gives it intimacy.

Or maybe you make them something. A bracelet with wolf imagery because the she wolf nursed Romulus and Remus and somehow that connects to how they nurtured you when you were broken.

Source: National Gallery of Art

A box with symbolic items: leather cord for the februa whips, wildflower seeds for Victorian flower language, a sealed note that says "From Your Valentine" like the original phrase from 270 AD. Each item has a reason. Each item tells part of your story through historical metaphor.

The key is this: you're layering meaning. Surface level, it's a cool historical gift. Deeper level, it's about your relationship. Deepest level, it addresses something they need to hear. Maybe they feel unseen. Maybe they doubt themselves. Maybe they carry guilt about something. The gift becomes a mirror and a message. You're saying: I see you, I chose you, I would defy emperors for you, I would run through chaos to find you.

People think gifts are about the object. They're actually about the narrative. The object is just the vessel. The why is everything. When you ground your why in something bigger than yourself (ancient history, mythology, cultural ritual), the gift transcends the personal and becomes archetypal. It says: what we have is part of something eternal. We're participating in a story that's older than us and will outlive us. That hits different than a dozen roses.

Here's the practical part. You need to know this person inside and out. You need to know if they love history or if it bores them. You need to know if they want sentimental or practical. You need to know their aesthetic, their humor, their love language. The history is just the framework. Your knowledge of them is the content. Get that wrong and the whole thing falls apart. Get it right and you've built something they'll keep forever. This is what thoughtful gifting actually looks like. It's work. It's excavation. It's storytelling. But when you do it right, you give someone proof that they matter. You show them they're known. You offer them healing wrapped in beauty. That's worth more than any price tag.

Why Ads in ChatGPT Could Undermine Trust (And How OpenAI Plans to Prevent It)

Om — Fri, 13 Feb 2026 12:30:00 GMT

OpenAI recently shared their approach to bringing ads to ChatGPT. Their reasoning is straightforward: by introducing advertising revenue, they can offer more people access to AI capabilities with fewer usage limits or without requiring payment. Pro, Business, and Enterprise users won't see ads. For everyone else, ads will appear at the bottom of responses when there's a relevant sponsored product or service, clearly labeled and separated from the actual answer.

They've laid out five core principles: ads won't influence ChatGPT's answers, conversations remain private from advertisers, users control their data and can turn off personalization, a paid ad-free option will always exist, and they prioritize user trust over revenue. These are good principles. The question is whether they're sufficient to address the fundamental problem ads create in this context.

Here's the concern. Imagine you ask which phone is better between two brands, and ChatGPT recommends Brand A. Then, right below that recommendation, you see an ad for Brand B. Even if OpenAI's answer is completely honest and the ad is just an ad, you can't know that for certain anymore. Your brain starts second-guessing everything. "Wait, did it recommend Brand A because it's actually better, or is something else going on with Brand B? Should I reconsider?" You came for a clear answer and quick decision-making, but now you've got an extra layer of confusion.

This defeats the purpose of using an AI assistant. You're supposed to get clarity and save time, not end up with more doubt and additional research. The presence of that ad, even if it truly doesn't influence the recommendation, creates uncertainty. You can't unsee it. You can't ignore the doubt it introduces. And over time, that erodes the trust that makes ChatGPT valuable in the first place.

Source: OpenAI

OpenAI acknowledges this tension. They explicitly state that "people trust ChatGPT for many important and personal tasks" and that "it's crucial we preserve what makes ChatGPT valuable in the first place." They understand the stakes. But understanding the problem and solving it are different things. Most tech companies have made similar promises about keeping ads separate from core functionality, and users have learned to be skeptical.

Source: OpenAI

The issue isn't necessarily that OpenAI will intentionally compromise their recommendations. The issue is perception. When financial incentives exist, trust becomes fragile. Even if the system operates with complete integrity, users will wonder. And that wondering is itself the problem, because it adds cognitive burden to every interaction.

That said, not all ads are bad. Some are genuinely useful. When I was searching for a specific type of product and couldn't find what I needed, a social media platform later showed me an ad for exactly the right thing. That felt helpful, not manipulative. It was contextual, understood what I was looking for, and connected me with something relevant. Those ads weren't contradicting advice I'd just received—they were discovery tools that helped me find what I was already seeking.

The difference is crucial. Good ads extend the information you're looking for. Bad ads interrupt it or create confusion about whether the core advice is trustworthy. OpenAI's model appears designed to be the former, but the execution will determine whether it actually feels that way to users.

So what would a better approach look like? If I were building an AI platform with advertising, here's what I'd propose: never mix ads into the core response flow. Instead, create a completely separate discovery interface within the app where people actively go looking for recommendations and know they're in a space where sponsorships exist.

Here's how it would work. This discovery page would analyze your chat history to understand your interests and needs, but crucially, your data would remain encrypted on your device—never sent to company servers in raw form. The system would match your interests with relevant sponsorships locally. If you've been researching products in a certain price range or category, it would surface options from companies willing to sponsor, clearly labeled as such, within parameters you've established.

The beauty of this model is separation. When you ask ChatGPT which phone is better, you get a clean, honest answer with no conflicting ad beneath it. Your decision-making moment remains unclouded. But if you want to explore options afterward, you can visit the discovery page where recommendations and sponsorships coexist transparently, and you're mentally prepared for that context. The two experiences don't interfere with each other.

This approach would also allow for more sophisticated matching. Rather than showing ads based on a single conversation, the system could understand patterns across your usage over time while keeping that data private. You'd see sponsorships that genuinely align with what you're looking for, not random promotions that happen to fit a keyword.

OpenAI mentions they're excited about conversational ads where you can "directly ask the questions you need to make a purchase decision." That's an interesting direction, but it intensifies the trust problem. If an ad becomes interactive and helpful, when does it stop being an ad and start feeling like advice? The line blurs, and with it, your confidence in what's objective versus what's influenced by payment.

The real test will be whether OpenAI maintains their stated principles over time as revenue pressures grow. They say they "do not optimize for time spent in ChatGPT" and prioritize "user trust and user experience over revenue." These are the right commitments. Whether they hold to them as the business scales will determine if this experiment succeeds or if users migrate to platforms that keep their core recommendations entirely separate from monetization.

OpenAI's approach deserves credit for transparency. They've published their principles, explained their reasoning, and committed to user control and privacy. That's more than many companies offer. But principles need to survive contact with reality. Users will judge not by what OpenAI says, but by how ads actually feel in practice—whether they enhance the experience or compromise it.

Because ultimately, if I can't trust that the recommendations are honest, the tool loses its value. Trust is the entire product. Ads might expand access, and that's genuinely important. But if expanding access means eroding trust, we've solved one problem by creating another. The challenge for OpenAI is threading that needle—making AI more accessible without making it less reliable.

The Founder’s Paradox: The Systemic Risks of Institutionalizing Idiosyncrasy in Scaling Ventures

Om — Thu, 12 Feb 2026 16:03:00 GMT

This research report investigates the perilous inflection point where a founder’s personal operating system—the unique blend of habits, instincts, and biases that catalyzed a startup’s genesis—transforms from a competitive advantage into a structural liability. Through a comprehensive analysis of management theory, Self-Determination Theory (SDT), and high-profile corporate case studies, we demonstrate that the institutionalization of a founder’s personal routines into mandatory organizational policy creates a fragile "monoculture" that suppresses cognitive diversity and accelerates burnout. The report argues that sustainable scaling requires a deliberate decoupling of principles from practices, evolving the organization from a "founder-led" dictatorship of habit to a "founder-inspired" ecosystem of scalable systems.

1. Introduction: The Complexity Trap and the Founder’s Shadow

Research Paper Artifacts:

The trajectory of every high-growth venture is punctuated by a specific, dangerous crisis. It is not a crisis of capital, nor of product-market fit, but of translation. In the nascent stages of a company, the founder is the company. The organization functions as a neural extension of the founder’s will, operating on a "Founder Operating System" (FOS) composed of the founder’s intuition, risk tolerance, biological rhythms, and personal idiosyncrasies. This centralization is highly efficient at the micro-scale, where speed and singularity of vision are paramount.

However, as the organization scales beyond the "two-pizza team" size—past 25, 50, and 100 employees—the very attributes that fueled its zero-to-one survival often become the architects of its stagnation. A critical error occurs when the founder, often encouraged by a board or investors seeking to replicate early successes, attempts to institutionalize their personal operating system into universal organizational policy. This strategy, characterized by the mandating of personal habits (e.g., "we all work until midnight," "we all eat this diet," "we all use this specific communication style"), fundamentally confuses the scalable elements of leadership with unscalable idiosyncratic elements.

This report posits that such an approach is a category error in organizational design. It mistakes the vehicle of the founder’s energy for the source of the value. By codifying personal habits into rigid mandates, founders inadvertently construct a "grind culture" that violates the basic psychological needs of the workforce, leading to retention crises, cognitive homogeneity, and the erosion of enterprise value. Sustainable success, we argue, requires evolving from a "Founder-Led" entity—dependent on the physical and mental presence of one individual—to a "Founder-Inspired" entity, where principles are abstracted into scalable, autonomous systems.

1.1 The Thesis of Decoupling

The central argument of this research is that scaling requires Operational Decoupling. This is the process of separating the founder’s Vision (the "Why") and Values (the "How") from their Habits (the "What").

Vision/Values: Scalable assets that must be codified.
Habits/Routines: Non-scalable idiosyncrasies that must be modeled but never mandated.

When decoupling fails, the organization suffers from "Founder’s Syndrome," a pathological condition where the infrastructure of the company remains tethered to the finite bandwidth and specific personality quirks of the creator. This report will dissect the anatomy of this failure and map the architectural path to a scalable, principle-based operating system.

2. The Anatomy of the Founder Operating System (FOS)

To understand why the FOS fails at scale, we must first analyze its mechanics and why it is so effective in the micro-scale environment.

2.1 The Neural Extension Mechanism

In the early days ($0–$1M ARR), the startup operates on what management theorists describe as "founder-led dynamics".1 The founder acts as the central router for all information. They are the "decider-in-chief," the head of product, the lead salesperson, and the cultural enforcement officer.

Cognitive Cohesion: Because one brain holds the entire context of the business (product, sales, engineering, finance), there is zero misalignment. Strategy and execution are fused.
Velocity over Process: Decisions are made instantly based on intuition and "gut feel," bypassing the need for consensus, data committees, or bureaucratic approval.2
The Insurgency Mindset: The founder possesses an "owner’s mindset" and an obsession with the front line. This "insurgency" allows the startup to break rules and outmaneuver incumbents.3

At this stage, the founder’s personal routine is the company’s routine. If the founder works 18 hours a day, the company moves at an 18-hour-a-day pace. If the founder is obsessed with pixel-perfect design, the company creates a perfect product. The FOS is a high-octane fuel that powers the launch.

2.2 The Confusion of Correlation and Causality

The problem arises when the startup achieves traction. The success validates the founder’s methods. However, the human brain is prone to "superstitious learning"—attributing success to the most visible behaviors rather than the underlying causes.

The Fallacy: "We succeeded because I sent emails at 3 AM and we ordered pizza every night."
The Reality: "We succeeded because we solved a customer problem, and the 3 AM emails were the unsustainable cost of that solution, not the cause."

When founders and boards fail to make this distinction, they begin to view the founder’s lifestyle not as a personal sacrifice, but as a prerequisite for excellence. They attempt to scale the "3 AM email" rather than the "Customer Obsession." This leads to the codification of habits that are biologically unsustainable for the broader workforce.

2.3 The "Founder Mode" Discourse: A Double-Edged Sword

Recent Silicon Valley discourse, popularized by Paul Graham’s concept of "Founder Mode," has reignited the debate on how involved founders should be at scale.5 "Founder Mode" is presented as the antidote to "Manager Mode"—the conventional advice to hire professional managers and let them run departments as black boxes.

The Argument for Founder Mode: Founders should retain "skip-level" access, intervene in details, and refuse to be gaslit by professional fakers. This ensures the original vision is not diluted by mediocrity.
The Risk of Misinterpretation: While "Founder Mode" correctly identifies the dangers of premature detachment, it is often misinterpreted as a license for toxic micromanagement. Founders use it to justify the "Pocket Veto"—hiring executives but overturning their decisions based on personal whim. This infantilizes the leadership team and reinforces the idea that only the founder’s intuition is valid.

True "Founder Mode" is about maintaining strategic connection to the "truth" of the business. Toxic "Founder Mode" is about mandating the founder’s personal neuroses as organizational law.

3. The Psychology of Compulsory Imitation: Why Mandates Fail

The institutionalization of FOS fails not just because of logistical bottlenecks, but because it violates the fundamental psychological architecture of human motivation. To understand this, we apply Self-Determination Theory (SDT).

3.1 Self-Determination Theory (SDT) and the Autonomy Deficit

SDT, a macro theory of human motivation developed by Deci and Ryan, establishes that high-quality motivation, well-being, and performance depend on the satisfaction of three basic psychological needs: Autonomy, Competence, and Relatedness.

3.1.1 The Primacy of Autonomy

Autonomy is not independence; it is the need to feel that one’s actions are self-endorsed and volitional. It is the difference between "I choose to do this" and "I am being forced to do this."

The Conflict: When a founder mandates their personal routine (e.g., "Everyone must follow this specific diet," "Everyone must use this specific note-taking app," "Everyone must be in the office by 8 AM"), they actively thwart the employee's need for autonomy.
Controlled Motivation: The employee’s behavior becomes "externally regulated." They comply to avoid punishment or gain rewards, not because they value the action. Research confirms that controlled motivation leads to lower creativity, reduced persistence, and higher burnout compared to autonomous motivation.
The "Quiet Quitting" Phenomenon: Employees who feel their autonomy is suppressed may comply on the surface (facade of conformity) while psychologically disengaging. They do the bare minimum to avoid the founder’s wrath, killing the innovation that requires discretionary effort.

3.1.2 Competence and the Micromanagement Trap

Competence is the need to feel effective and capable of mastering challenges.

The Conflict: If the founder insists that their way is the only way, they signal a lack of trust in the employee’s competence. Micromanagement—a hallmark of rigid FOS—strips employees of the opportunity to solve problems using their own methods.
The Result: The organization selects for "helpers" rather than "owners." High-competence individuals (who crave autonomy) leave, while those who rely on detailed instructions stay. This leads to a degradation of the talent density over time.

3.2 The "Copernican Turn" in Human Resources

Modern HR theory describes a "Copernican Turn" in management: viewing the employee as the center of their own professional life, rather than an orbiter of the company.

Pre-Copernican View: "How do we motivate employees to do what the founder wants?" (Carrot and stick).
Post-Copernican View: "How do we create conditions where employees are self-motivated to achieve shared goals?" (Autonomy support).

Founders who institutionalize their personal routines are stuck in the Pre-Copernican mindset. They view the workforce as a monolithic tool to execute their will, failing to recognize that sustainable scaling requires unlocking the individual agency of hundreds or thousands of people.

4. The Risks of Monoculture: Grind, Burnout, and Fragility

When a founder’s idiosyncrasies are mandated, the organization develops a "corporate monoculture." Like agricultural monocultures, these systems are efficient in the short term but highly fragile to shocks and disease in the long term.

4.1 The Institutionalization of "Grind Culture"

"Grind culture" or "hustle culture" is the most common manifestation of FOS. Founders often work 80+ hour weeks because they have an existential stake in the outcome—their identity, fortune, and reputation are on the line. This is intrinsic motivation.

The Error: Founders assume that if employees "care enough," they will naturally adopt the same schedule. When they don't, the founder mandates it (e.g., 996 work culture).
The Consequence: For employees with <0.1% equity, this level of work is biologically unsustainable and economically irrational. Mandating it leads to:
- Burnout: Chronic stress, exhaustion, and cynicism. Burnout is not just "being tired"; it is a state of vital exhaustion that decouples an employee from their work.
- Health Costs: Increased absenteeism, mental health crises, and long-term attrition.
- Productivity Illusion: The culture rewards "looking busy" over "creating value." Employees stay late to be seen by the founder, not to work.

4.2 Cognitive Homogeneity and Innovation Suppression

Innovation requires "cognitive diversity"—the collision of different perspectives and problem-solving styles.

Filtering for Sameness: If the FOS mandates a specific way of thinking or working, the hiring process filters for people who look, think, and act like the founder.
Normative Conformity: Those who remain in the organization suppress their dissenting views to fit in ("Normative Conformity"). Research shows that normative conformity is a direct inhibitor of innovation.
The Echo Chamber: The founder ends up surrounded by "yes-men" who reinforce their biases. The organization loses its peripheral vision and becomes blind to market shifts that the founder doesn't personally see.

4.3 The "Key Man" Fragility

An organization built on the FOS has a single point of failure.

Decision Bottlenecks: If every decision must pass through the founder’s filter, the company’s speed is limited by the founder’s sleep schedule and cognitive bandwidth.
Succession Crisis: The business cannot be handed over. No successor can replicate the founder’s idiosyncrasies. When the founder eventually leaves (or burns out), the organization collapses because it lacks an independent operating system.

5. Case Studies in Idiosyncratic Failure

The dangers of mandating FOS are not theoretical. The history of modern startups is littered with unicorns that collapsed or stagnated because they failed to decouple the founder’s personality from the company’s policy.

5.1 WeWork: The Cult of "We"

Adam Neumann’s tenure at WeWork is the definitive case study of FOS gone toxic. Neumann didn't just lead the company; he imposed his lifestyle upon it.

The Mandate: The culture was a reflection of Neumann’s chaotic, party-centric energy. Mandatory "Thank God It’s Monday" rallies, tequila shots in meetings, and a blurring of professional boundaries were policy, not perks.
The Mechanism: Neumann used "community" rhetoric to demand total devotion. Employees were expected to work endless hours for the "mission," masking the lack of a viable business model.
The Collapse: The culture selected for sycophancy and punished critical thinking. When the IPO scrutiny revealed the financial rot, the culture—built entirely on Neumann’s reality distortion field—evaporated overnight. The lack of independent governance and systems meant there was no "company" beneath the "cult".

5.2 American Apparel: The Shadow of the Founder

Dov Charney built American Apparel as a direct extension of his personal aesthetic and sexual impulses.

The Mandate: Charney imposed a "sexually charged" atmosphere as a corporate value. Hiring decisions were explicitly based on his personal attraction to candidates. Professional boundaries (HR policies) were viewed as impediments to "authenticity".
The Fallacy: Charney believed his lack of inhibition was the source of the brand's creativity. He refused to distinguish between "creative disruption" (good) and "harassment" (illegal).
The Result: The company faced endless lawsuits, reputational destruction, and bankruptcy. Because the governance was fused with the founder’s libido, the board could not correct the course until it was too late.

5.3 Bridgewater Associates: The Algorithmic Panopticon

Ray Dalio’s Bridgewater Associates offers a more nuanced, yet controversial, example. Dalio codified his personal philosophy ("Radical Transparency") into a rigid system of "Principles" and software ("The Dot Collector").

The Mechanism: Every employee carries an iPad to meetings to rate colleagues in real-time on attributes defined by Dalio. All data is public. "Believability" scores determine decision weight.
The Critique: While financially successful, Bridgewater’s culture is often described as a "cult of logic." It requires employees to completely surrender their social defenses and adopt Dalio’s specific worldview. Critics argue it creates a surveillance state that erodes psychological safety for anyone who isn't "wired" like Dalio.
Scalability Limit: This model works for a niche hedge fund where compensation is astronomical. Attempting to scale this "radical transparency" to a general workforce often results in fear, silence, and attrition, as it demands a psychological toll most workers are unwilling to pay.

6. The Architecture of Scale: Transitioning to Founder-Inspired

The antidote to the FOS trap is the transition to a Founder-Inspired entity. This involves abstracting the principles that drive success from the idiosyncratic practices of the founder, and embedding them into scalable systems.

6.1 The Process of Operational Decoupling

Decoupling is the act of separating the signal from the noise. The founder must analyze their own success and distinguish between "What I do" and "The Principle behind what I do."

Table 1: Decoupling Founder Idiosyncrasies from Scalable Principles

Founder Trait (Idiosyncratic/Unscalable)	The Underlying Principle (Scalable/Transferable)	Scalable Policy/System Implementation
Founder sends emails at 3 AM and expects replies.	Responsiveness & Agility	Implement Service Level Agreements (SLAs) for client response times (e.g., "within 2 hours during business hours"). Use "shift" coverage for 24/7 support rather than burning out individuals.
Founder screams at designers over a pixel misalignment.	High Standards & Attention to Detail	Implement rigorous Quality Assurance (QA) protocols, Design Systems, and "Pixel Perfect" reviews before launch. Institutionalize the standard, depersonalize the critique.
Founder refuses to buy expensive software/furniture.	Frugality & Resourcefulness	Establish clear budget caps and procurement policies that reward cost-saving. Do not require the CEO to sign off on every stapler (bottleneck).
Founder relies on "gut feel" to hire people.	Cultural Alignment & High Bar	Codify the "Gut" into a "Bar Raiser" program (like Amazon) with specific behavioral interview questions and rubrics. Remove the founder from the loop.
Founder changes product roadmap every Monday morning.	Adaptability & Market Awareness	Implement Agile/Scrum methodologies with short sprints. Allow for pivoting based on data and customer feedback, not weekend whims.

Source Analysis: Derived from Bain's "Founder's Mentality" and Amazon Leadership Principles.

6.2 Amazon: The Gold Standard of Abstraction

Jeff Bezos is the archetype of successful decoupling. Bezos had many idiosyncrasies, but he did not scale Amazon by cloning himself. He scaled it by codifying 14 Leadership Principles.

The Mechanism: Principles like "Customer Obsession," "Bias for Action," and "Disagree and Commit" are tools, not rules. They are distributed heuristics. A junior manager in a fulfillment center can use "Bias for Action" to make a decision without checking with Bezos.
The Result: The OS became the "Amazon OS," distinct from the "Bezos OS." This allowed the company to scale to 1.5 million employees and survive Bezos's transition to Executive Chair.

6.3 Building the Brand Operating System (BOS)

The organization needs a Brand Operating System (BOS) that serves as the interface between the vision and the execution.

Layer 1: Identity & Narrative. The immutable "Why."
Layer 2: Decision Architecture. Who decides what? Moving from "Founder decides all" to "Decision Rights" based on role and competence.
Layer 3: Rhythms & Rituals. Standardized meeting cadences (QBRs, Weekly Business Reviews) that drive accountability through data, not founder presence.

6.4 HR as Strategic Infrastructure

Human Resources must evolve from an administrative function to the custodian of the BOS.

Recruiting for Values, Not Likeness: HR must build hiring funnels that test for alignment with the Scalable Principles, ensuring diversity of thought while maintaining unity of mission.
Performance Management: Reviews should measure "Impact" and "Values Alignment," not "Face Time" or "Founder Pleasing".
Avoiding "Deadly Combinations": HR must ensure that incentives align with the stated values. You cannot preach "Collaboration" but promote "Lone Wolves".

7. The Role of Governance: The Board’s Responsibility

The transition from FOS to Scalable Systems is rarely voluntary. Founders often resist it due to ego or fear of losing control. The Board of Directors plays a crucial fiduciary role in enforcing this evolution.

7.1 The "Bus Test" and Succession

The board must constantly apply the "Bus Test": If the founder were incapacitated today, would the business operate tomorrow? If the answer is no, the board has failed in its governance duty.

Succession Planning: This is not just about replacing the CEO; it is about building a bench of leaders who can carry the weight. The board must push the founder to hire a COO or "Integrator" who complements their visionary strengths with operational discipline.

7.2 Defining the Future Leader Profile

When a transition is necessary, the board must define the profile of the next leader based on the company's future needs (scaling, process, global expansion), not based on a resemblance to the founder. The skillset required to start a company (chaos, risk) is often the opposite of the skillset required to scale it (order, governance).

8. Strategic Recommendations for Founders

For founders standing at this precipice, the following strategic roadmap offers a path to sustainable scaling without losing the "soul" of the startup.

Recommendation 1: Audit Your Operating System
- Conduct a ruthless self-audit. List your top 10 management behaviors. Categorize them:
- Green: Principles that drive value (e.g., "We listen to customers"). Action: Codify and teach.
- Red: Habits that drive anxiety or control (e.g., "I approve every tweet"). Action: Stop and delegate.
- Grey: Personal preferences (e.g., "I hate PowerPoint"). Action: Decide if this is a hill worth dying on. (Hint: It usually isn't).
Recommendation 2: Hire a "Manager Mode" Partner
- If you are a visionary "Founder Mode" leader, you need a counterbalance. Hire a COO or President who excels in "Manager Mode"—building processes, managing P&L, and creating stability. Give them the authority to build the rails for your train.
Recommendation 3: Shift KPIs from Inputs to Outcomes
- Stop measuring "hours in the chair" or "speed of email reply." Start measuring "customer value delivered," "code shipped," or "revenue closed." This restores autonomy to your team and allows them to work in their own way to achieve your goals.
Recommendation 4: Institutionalize Dissent
- Create formal mechanisms for feedback that do not threaten your ego. "Red Teams," "Shadow Boards," or anonymous "Pulse Surveys" can surface the reality of the ground floor that your sycophants are hiding from you.

9. Conclusion: The Ultimate Legacy

The institutionalization of a founder’s personal operating system is a failure of imagination. It assumes that the founder is the only possible template for success. This arrogance restricts the organization’s potential to the biological and cognitive limits of one human being.

True legacy is not built by cloning oneself; it is built by creating a system that is greater than oneself. By evolving from a Founder-Led dictatorship of habit to a Founder-Inspired ecosystem of principles, the founder creates an entity that can outlive their tenure, outgrow their limitations, and achieve a scale of impact that "hustle" alone could never deliver. The founder must step back from the machinery to let it run, transitioning from the operator of the engine to the architect of the enterprise.

The Internet Stopped Being Useful and Nobody's Talking About It

Om — Wed, 11 Feb 2026 09:47:00 GMT

I've been staring at search results lately and something feels off. The pieces all work, but the whole thing feels hollow. Like walking into a store where everything's stocked perfectly but nothing's actually for sale. We're in February 2026 and the numbers are worse than most people realize. When AI Overviews show up in search results, 80-83% of searches end without anyone clicking anything. You type something in, get an AI-generated summary at the top, and bounce. Google calls them "AI Overviews," but what they really are is the final nail in the coffin of the open web.

Here's the thing people keep dancing around: the internet's mostly fake now. Over half of all English content online is AI-generated. Actually, 74% of newly published web pages contain AI content as of April 2025. Merriam-Webster named "AI slop" their Word of the Year for 2025, defining it as "digital content of substandard quality typically generated in large volumes through artificial intelligence," because it's everywhere. Pinterest and YouTube had to add filters just so people could block AI content because users were revolting. Security firm Imperva found that bot traffic crossed 51% in 2024. That means more than half of internet activity is automated, and that was before AI generation hit mainstream scale.

The way people search has completely flipped too. It used to be "how do I fix this" or "where can I learn about that." Now it's "do this for me." Voice search, visual search through Google Lens hitting 10 billion searches a month, conversational queries where you're basically talking to an assistant. 71% of people prefer voice search because typing feels like too much work. We've gone from seeking guidance to demanding execution. The shift is wild: 31.6% of AI-triggered queries are question-based, and queries with 4+ words are way more likely to trigger AI answers that keep you locked in place. Google retrained user behavior so thoroughly that even searches without AI Overviews saw a 41% click-through rate drop year-over-year. People just expect instant answers now.

And the damage is real. When AI Overviews appear, organic click-through rates collapse by 58-61%. Paid ads? Down 68%. On mobile the zero-click rate hits 77%. The New York Times saw organic search traffic drop from 44% to 36.5% over three years. Some publishers report losing 20-90% of traffic in the past year. Companies like Chegg are literally suing Google, claiming Google coerces publishers into surrendering content for AI training purposes as a condition for search inclusion, then uses that content to destroy their traffic. The lawsuit filed in February 2025 basically says "give us your stuff for free or disappear from search entirely." This goes beyond some abstract SEO problem. The entire incentive structure of the web is breaking down. Why create good content if nobody sees it?

The authenticity crisis runs deeper. Dead Internet Theory used to be a conspiracy thing, this idea that most online activity is bots and corporate algorithms instead of humans. Now it's just reality. Synthetic media is indistinguishable from real stuff at scale. AI agents mimic tone and emotion convincingly enough that you genuinely wonder who's real. Traditional verification systems are failing. Trust in news, institutions, even peer-to-peer communication is collapsing because we've lost the ability to authenticate anything. You see a photo, read an article, watch a video, and your first thought is "is this even real?" That's the default now.

But here's what actually bothers me: it's all the same. AI-generated content goes beyond fake. It's generic. Homogenized. Every article sounds like every other article because they're all trained on the same data, optimized for the same engagement metrics. Platforms prioritize virality over authenticity. The algorithm feeds you what it thinks you want based on what worked for someone else, and it all blurs into this beige sludge of information that technically answers your question but leaves you empty. We've traded diversity of thought for scalable content production, and the internet's becoming an echo chamber where every voice sounds like the same robot. When everything's trained on the same corpus, optimized for the same metrics, and generated by the same models, you get this weird monoculture where nothing feels original anymore.

The survival pattern tells you everything you need to know. Differentiated, branded content survives while generic SEO farms die. Publishers like Dotdash Meredith and Ziff Davis claim minimal impact because they built actual brands people trust. Meanwhile, content farms that spent 20 years optimizing for Google's algorithm are getting obliterated. We built this ourselves. Publishers created mountains of SEO-optimized garbage designed to rank instead of inform. Google trained users to expect instant answers instead of exploration. AI slop is just the logical endpoint of "content as SEO vehicle" meeting "infinite scalability."

So where does that leave publishers and creators? Trying to adapt to a game where the rules changed overnight. Only 1% of users click links cited in Google's AI Overviews. ChatGPT's 1.2 billion referrals between September and November 2025 still only account for 1% of total publisher traffic. The math simply does not recover.

Some are trying to optimize for AI citations instead of clicks. Getting cited inside the AI Overview becomes the new game. Structure content so AI can extract and attribute it: answer upfront in the first paragraph, clean HTML structure with FAQs and lists, schema markup to help AI categorize your content, topic clusters instead of exact-match keywords. The thinking shifts from driving traffic to gaining visibility. Think billboard advertising instead of direct response. You're building brand recognition even if users never click.

Others are diversifying away from search entirely. Newsletters, subscriptions, apps, anything that builds owned audiences. Some publishers signed content licensing deals with AI companies. The Atlantic, News Corp, Washington Post partnered with OpenAI. The New York Times signed with Amazon. You get upfront payments plus royalties for AI training usage. A few publishers are blocking AI crawlers entirely, choosing to disappear from AI results rather than let their content get used for free.

The publishers surviving this are rebuilding business models that work without Google. Licensing deals, subscriptions, owned platforms. The open web traffic model feels cooked. Traditional SEO tactics like keyword density, backlink volume, classic on-page optimization become irrelevant when users never leave Google. Quality content alone means nothing if nobody sees it.

Some are going legal. The Independent Publishers Alliance filed EU complaints demanding transparency on AI content usage and impact assessments. Regulatory intervention might force change, but that's a 2-3 year timeline minimum. Meanwhile, Google's roadmap includes more AI features: AI Mode international expansion, voice-activated queries, multi-turn conversations. The trend accelerates instead of reversing.

So where does that leave us? Stuck between a search engine that keeps you trapped and a web full of content nobody wrote. The tools got better but the experience got worse. We're asking more questions than ever and getting fewer real answers. The anti-slop sentiment is real. The frustration's building. But the fix probably comes from users migrating to closed communities, newsletters, Discord servers, and spaces where provenance matters. The open web might just stay dead while smaller trusted networks flourish. That's already happening. You see it in the rise of Substacks, private Slacks, and curated feeds over algorithmic ones.

The question goes beyond whether the internet needs to change. It's whether enough people care to build something different, or whether we just accept this as how things work now.

Sources:

Zero-click statistics:

click-vision.com/zero-click-search-statistics

Chegg lawsuit:

cnbc.com/2025/02/24/chegg-sues-google...

AI slop and bot traffic:

Publisher traffic impact:

Publisher adaptation strategies:

AI Makes a Pretty Good Beta Tester

Om — Tue, 10 Feb 2026 15:22:00 GMT

Testing software sucks because it's boring and I have to think of everything that could go wrong. Clicking the same buttons a hundred times, trying weird combinations, checking if things break when users do stupid stuff. My brain gets tired and I start missing things. That's just how this works. AI's actually good at this boring stuff. It has tons of knowledge about how things usually break, it remembers everything I told it about my project, and it keeps going when I'd normally zone out. I can throw my whole codebase at it and ask it to poke holes in my logic. It'll think of edge cases I forgot existed.

System design is where this really shines. I can sketch out my architecture, explain what I'm building, and AI starts throwing problems at me I haven't thought about yet. What happens when this service goes down? How does this scale? What if two users hit this endpoint at the exact same time? It broadens my perspective while I'm still in building mode. My brain stays focused on solving the actual design challenges instead of trying to remember every possible failure scenario. It's like having someone who's seen a thousand systems break, sitting next to me while I'm drawing boxes and arrows.

Here's the cool part: while AI's doing the repetitive hunting for problems, my brain stays fresh for the real work. I get to focus on actually fixing bugs, making architectural decisions, figuring out why something broke in the first place. Solving puzzles instead of just looking for them. It works like having a QA team in my pocket. I tell the AI what I built, what it's supposed to do, maybe paste in some code. It generates test scenarios, simulates different users, points out where things might fail. Then I take over and handle the creative problem solving part.

The context window thing makes this even better. AI remembers my entire project structure, previous bugs, how I fixed similar issues before. It's like working with someone who actually paid attention during all those planning meetings. I finally get to use my brain for thinking instead of just remembering and checking boxes.

The Invisible Excellence Trap: Why Your Best Work Goes Unnoticed (And Why That Might Be The Point)

Om — Mon, 09 Feb 2026 20:33:00 GMT

I put everything into the details. Every pixel, every word, every interaction gets scrutinized, adjusted, perfected. Then I release it into the world and wait for someone to notice. They rarely do. They use the thing, they seem to like it, they move on. Meanwhile, I'm sitting here knowing exactly how many hours went into making that button feel right, how many iterations it took to get that flow seamless. And nobody sees it.

This used to eat at me. Still does sometimes. Because there's this philosophy I hold: leave perfections for people to notice, instead of leaving loopholes and hoping they miss them. I play offense, aiming for recognition rather than defense, dodging criticism. But here's the problem with that strategy: the audience mostly stays home.

The Paradox of Invisible Design

When something works perfectly, people think it was easy. The best interface feels intuitive because you spent weeks making it intuitive. The best writing reads effortlessly because you rewrote it twelve times. The best product "just works" because you obsessed over every edge case, every failure mode, every tiny friction point.

You succeed by disappearing. The user gets frictionless experience. You get the satisfaction of knowing you nailed it while remaining completely anonymous in your achievement. Professional victory, personal invisibility. That's the deal.

Why Most People Miss It

People move fast. They're tired, distracted, processing a hundred inputs. They experience your work in passing, looking at broad strokes while your precision lives in the details. Their eyes glide over what yours catches. This happens because they lack the training, the vocabulary, the framework to see what you see.

Expecting mass audiences to appreciate craft is like expecting someone who loves eating to understand knife techniques. Different skills entirely. They consume the outcome. You understand the process. That gap is unbridgeable for most people.

The Subconscious Effect

Here's something that helps a bit: even when people miss specific details, they feel the cumulative effect. They experience quality without understanding its source. "This feels right," they say, with zero idea why. Your precision creates an impression that registers subconsciously. The details become invisible precisely because they work so well together.

So you are getting through, just obliquely. The message lands without attribution. The craft works without recognition. Which brings us back to the original tension.

Two Incompatible Reward Systems

I think what's happening here is we're trying to optimize for two things that fundamentally conflict. Outcome-based rewards (does it work perfectly?) versus process-based rewards (do people see how I made it work perfectly?). These pull in opposite directions.

The outcome-based win is cleaner. Problem solved, people using it, feels effortless. Check. But that victory feels hollow because there's no witness to your effort. You won a race nobody watched you run. The process-based reward requires an audience that gets it, people with vocabulary and experience and eye. Most audiences lack this entirely.

The People Who Actually Notice

The few who do see it tend to be worth impressing anyway. Fellow craftspeople. Careful observers. People who care about the same things you care about. They spot your choices immediately because they know how hard "simple" actually is. They see the restraint, the elegance, the decisions that made complexity feel easy.

Their recognition means more because it comes from understanding. One person who truly sees what you did validates the work in ways a thousand casual users never could. Different currencies. The masses give you impact. Your peers give you recognition. Both valuable, both completely different.

Playing For The Wrong Audience

Maybe the real issue is audience mismatch. If you're building for mass users, you have to accept invisibility. Their appreciation tops out at "I like this" and going deeper requires training they simply haven't done. If you're building for people who understand craft, you can design something that reveals complexity to those who look closer.

Different audiences require different approaches. Mass appeal means your craft serves the experience. Peer respect means your craft becomes part of the conversation. You probably want both, which is where the tug of war intensifies.

The Motivation Problem

When details go unnoticed, motivation takes a hit. I get sad about it. Genuinely sad. Because the care I put in feels wasted when it generates zero response. This creates a practical problem: do I keep obsessing over things people miss, or do I scale back effort where it goes unappreciated?

The question becomes whether you're doing it for external validation or because imperfection bothers you internally. Both are valid drives, but knowing which one fuels you helps direct energy appropriately. Some contexts deserve full perfectionism. Others function fine with good enough.

Finding The Balance

I'm still figuring this out. Sometimes I point out what I did, teaching people to see. Sometimes I keep it as private standard. Sometimes I reserve the detail obsession for work where it actually gets appreciated and dial it back elsewhere. The balance shifts depending on project, audience, energy levels.

What I'm learning is that invisible excellence is still excellence. The work stands regardless of recognition. Users benefit even when oblivious. Quality compounds in ways that might matter more than immediate acknowledgment. But man, it would be nice if people noticed sometimes.

The Real Answer

Here's where I land: the craft matters independent of recognition. You do it because doing it right satisfies something internal. The details exist because you'd know if they were wrong, even if nobody else would. That's the real test. Would you still care if absolutely nobody ever noticed? If yes, you're building for the right reasons. If hesitation creeps in, maybe recognition matters more than you're admitting.

Both are fine. Just be honest about what you're actually chasing. Impact or acknowledgment. Invisible perfection or visible craft. The work itself or the conversation about the work. Once you know, the tug of war eases up. You stop expecting the wrong things from the wrong audiences. And maybe that's when the good work actually begins.

Leadership as a Service

Om — Mon, 09 Feb 2026 11:08:00 GMT

we believe leadership belongs on the ground, not in the distance. to truly guide, we must understand the work that creates value.

we propose a simple shift: the higher the role, the closer the connection.

we replace mystery with approachability. this is a management philosophy where presence is the priority. by standing side-by-side with the team, we restore dignity to the work itself.

learning by doing reality teaches us more than a report ever could. to understand the business, we choose to look closer.

just as a student learns best by doing, a leader learns best by seeing. we bridge the gap between strategy and practice by being there. this immersion helps us see the small obstacles—like a slow system or a broken tool—so we can fix them. when we understand the environment, our decisions become sharp, empathetic, and true.

the daily walk we start every day with a commitment to reality. before the meetings begin, we walk the floor to visit:

the warehouse the server room the front desk this simple habit keeps us connected to the pulse of the company. it turns the executive into a partner. by seeing the work in real time, we can celebrate the victories and solve the problems together.

coffee and connect we believe in the power of a shared conversation. we set aside time to let titles fall away so humanity can take center stage.

in the break room, we sit as equals. we listen to the stories and ideas that matter. this connection builds trust. when we know the people behind the work, policies become human. it proves that we are here to serve the team.

walking the mile we respect the standard by living it. we frequently step in to do the work:

answering a ticket packing a box cleaning a workspace this experience teaches us the weight of every choice we make. when we share the task, we understand the effort. by doing the work, we earn the right to guide it.

a culture of support this approach moves us from rigid metrics to holistic health. it creates a loop of honest, immediate feedback.

ultimately, this redefines what it means to lead. the justification for a role is found in service. we are here to remove obstacles and value your time.

this is how we build a business that is fair, capable, and deeply human.

Why AI Coding Is Still Broken (And Why Your Job Is Safe)

Om — Sun, 08 Feb 2026 14:45:00 GMT

Every product you use is a website or an app. That's it. And everyone keeps screaming that AI will replace the people building them. I've spent two years actually using these tools every single day, and I need to tell you: we've hit a ceiling. A hard one.

The Context Cliff

Here's what happens when you feed your entire codebase into AI. At first, it feels like magic. Then something breaks. The model starts hallucinating features that were never there. It invents APIs that exist in its training data but have nothing to do with your stack. Research from early 2026 confirms this: once you push past 32,000 tokens of context, accuracy falls off a cliff. The AI doesn't just forget details. It starts making things up. Your "smart assistant" becomes a confident liar, and you're left debugging code that looks right but fails in production.

People tell me: just use a specialized coding model. Train it purely on code, nothing else. Sounds logical. Completely wrong. The February 2026 leaderboards prove it. The best coding models in the world right now are massive generalists. Claude Opus 4.5 scores 80.9% on SWE-bench. These models learned math, logic, literature, philosophy. Turns out that training makes them better engineers. The narrow models are fast and cheap, but when you need actual architectural intelligence, the generalists win. The breadth is the point.

The Security Spiral

But here's where it all falls apart. I tell AI to add a file upload to my signup form. My brain immediately jumps: users need to edit that file later, where does it get stored, what happens if the upload fails, how do we handle malicious files. The AI? It asks which storage service I want and writes the upload function. That's it. Security reports from 2025 found vulnerabilities in 45% of AI-generated code. Worse: when you ask AI to fix its own code, the security problems get 37% worse after five iterations. It solves the syntax puzzle but fails the engineering test. It writes code that compiles and breaks in ways you will discover three months later when a user reports data loss.

The Iceberg Problem

The real problem is simpler than anyone admits. AI pulls information related to your exact query and stops. It's trained to answer questions, to complete patterns. It's a prediction engine optimized for the next token. Engineering is about thinking three steps ahead, around corners the AI will never see.

That's why after two years of this, coding with AI feels like walking a tightrope. You're guessing. You're praying. And only 3.8% of developers trust AI code without review, because we've all been burned. The promise that AI will replace your engineering team is garbage. What we actually have is a tool that makes experienced developers faster and junior developers dangerous. Your job is safe. Actually, it's more essential than ever, because someone needs to catch what the AI misses. And it misses everything that matters.

The Soul of a Brand: Why Human Creativity Still Matters

Om — Sat, 07 Feb 2026 19:30:00 GMT

I will buy from brands that choose humans over shortcuts. I will invest my money, my attention, and my loyalty in companies that view marketing as a craft, a promise, and an opportunity to connect. And I will walk away from any brand that treats advertising as something AI can handle while humans take a back seat. This line feels clear to me, and the reasoning behind it runs deeper than preference. It touches the core of what brands actually are and what they owe the people they serve.

Marketing is the first handshake. This is the introduction. It is how a brand says, "This is who we are, this is what we believe, and this is how much you matter to us." When I see an advertisement, I am looking for more than information about a product. I am looking for evidence of care. I want to see thoughts. I want to feel the intention. I want to know that someone sat down, wrestled with an idea, refined it, poured energy into it, and believed it was worth the effort because I am worth that effort. If a brand automates that first handshake with AI-generated visuals, AI-written copy, or AI-orchestrated campaigns, it tells me something. It tells me they see marketing as a task to get through rather than a space to show up fully.

The logic extends beyond advertising. If a company considers creative marketing a burden to outsource to algorithms, what does that say about how they approach everything else? Will they cut corners in product design? Will they scrimp on customer service? Will they treat me like a transaction instead of a human being with preferences, emotions, and expectations? Marketing is the place where brands get to prove they care before I ever hand over my money. It is the space where they earn trust. And when that space feels automated, hollow, or effortless, trust becomes difficult. The absence of visible human effort makes me wonder where else effort is missing.

This goes beyond functionality. A product might work perfectly. A service might deliver exactly what it promises. But brands sell lifestyles, not just objects or solutions. Apple does this well. Nike does this well. Brands that understand how to build worlds around their products know that the story matters as much as the specs. When I buy running shoes, I am buying into a narrative about discipline, progress, and identity. When I choose a laptop, I am aligning with values around creativity, innovation, or simplicity. If a brand hands that storytelling over to AI, the narrative flattens. It loses texture. It stops feeling like a world I want to enter and starts feeling like content generated to fill space and meet quarterly targets.

AI-generated ads often feel generic because they are optimized for patterns, trends, and data points rather than human truth. They might test well in focus groups or perform adequately in A/B tests, but they lack soul. They miss the small, surprising details that make creative work memorable. They skip the risks that make storytelling compelling. They avoid vulnerability, humor, and the kind of emotional honesty that turns an ad into something people actually want to watch, share, or talk about. And when brands replace human creativity with algorithmic efficiency, they signal that creativity itself is negotiable. That art is a line item. That connection is optional.

The issue becomes clearer when you flip the scenario. Imagine a brand explicitly marketing its product by saying, "We automated our design process to save costs. We used AI to handle customer support because hiring people felt expensive. We streamlined production by cutting quality control." People would reject that brand instantly. They would see it as careless, cheap, and insulting. Yet many of those same people overlook the equivalent when it happens in advertising. They scroll past AI-generated ads without questioning what the use of AI reveals about the company behind the message. But the implication is the same. If a brand treats the creative process as disposable, it will treat other things as disposable too.

I recently came across something that felt like the opposite of this trend. Apple released a new intro for Apple TV, and instead of relying on computer-generated effects, they built the entire sequence using real glass, sculpted materials, practical lighting, and in-camera techniques. The music was composed specifically for the project by a human artist. The behind-the-scenes video showed the physical craftsmanship involved - the time, the labor, the creative problem-solving required to make something tangible and beautiful. Watching it felt different. It felt intentional. It felt like someone cared enough to do the hard thing instead of the fast thing. And that effort translated into trust. It reminded me why I choose certain brands over others. Because effort matters. Because craft signals values. Because the way a brand shows up in the smallest moments reveals how they will show up in the biggest ones.

Part 1: The Soul of a Brand - Selling a Lifestyle, Not Just a Product

Brands that matter sell worlds, experiences, and identities. They offer ways of seeing yourself and the life you want to live. When someone buys a Patagonia jacket, they buy into environmental stewardship and outdoor adventure. When someone chooses a Tesla, they align with innovation, sustainability, and a vision of the future. When someone picks up a Moleskine notebook, they connect with creative tradition, artistic discipline, and the romance of putting pen to paper. The product itself becomes a token, a physical reminder of the values, aspirations, and story the brand represents. Great brands understand this deeply. They build entire ecosystems around their offerings, crafting narratives that make the product feel like an entry point into something larger and more meaningful.

This approach requires storytelling that feels genuine and layered. It requires creativity that surprises, delights, or challenges expectations. It requires marketing that treats every touchpoint as an opportunity to reinforce the world the brand has built. Think about Nike. Their campaigns rarely focus solely on shoes. They tell stories about perseverance, identity, overcoming obstacles, and redefining limits. They feature real athletes, real struggles, and real triumphs. The emotional weight of those stories makes the shoes feel like tools for transformation rather than just footwear. The effort behind the storytelling becomes visible, and that visibility builds trust. People sense when care has been invested. They feel when a brand has taken the time to understand them, to speak to their experiences, and to offer something that resonates beyond function.

Effort in marketing signals effort everywhere else. When a brand produces a meticulously crafted advertisement, it suggests meticulousness in product design, in customer experience, in quality control, and in how they treat their employees and partners. The craftsmanship visible in a campaign becomes a proxy for the craftsmanship embedded in the entire operation. This principle holds across industries. A restaurant that invests in beautiful food photography, thoughtful menu design, and compelling storytelling about ingredient sourcing sends a message about the care they bring to every dish. A software company that creates clear, human, and visually engaging tutorials signals that they value user experience and want people to succeed with their product. Effort is a language. It communicates values, priorities, and respect.

The effort principle works because humans are wired to recognize and appreciate labor. We value things more when we see the work behind them. Research in psychology shows that people rate handmade items as more valuable and meaningful than identical machine-made items, even when they look the same. This effect extends to creative work. When people know a piece of content requires human thought, iteration, collaboration, and creative risk, they engage with it differently. They give it more attention. They remember it longer. They talk about it more. They trust the source more. Visible effort creates an emotional return that automation struggles to replicate.

AI-generated advertising breaks this chain. When people learn an ad was produced by AI, their perception shifts immediately. Studies from the Nuremberg Institute for Market Decisions reveal that identical advertisements receive significantly different evaluations depending on whether they are labeled as human-made or AI-generated. The AI-labeled versions are consistently rated as less natural, less emotionally engaging, and less useful. This reaction happens even when the content quality remains objectively the same. The label alone alters how people interpret and respond to the message. The knowledge that a machine handled the creative process removes the human connection, and with it, the sense that someone cared enough to put thought, emotion, and intention into the work.

This disconnect runs deeper than aesthetics or technical quality. AI ads often perform well in controlled tests where viewers are unaware of their origin. They can match human ads on clarity, visual appeal, and even engagement metrics when disclosure is absent. But the moment viewers know AI created the content, trust erodes. Research involving over 1,000 participants across multiple countries found that only 21% of people trust AI companies and their promises, and just 20% trust AI itself. When shown identical ads with different labels, participants rated the AI version more critically on emotional dimensions like warmth, authenticity, and relatability. They also reported lower purchase intent and reduced willingness to engage with the brand.

The issue centers on authenticity. People want to believe the brands they support, understand them, value them, and speak to them with sincerity. AI-generated content feels impersonal because it originates from pattern recognition rather than lived experience. It optimizes for data trends instead of emotional truth. It produces output based on what has worked statistically rather than what feels genuinely human. And while AI can mimic style, structure, and tone, it lacks the intuition, vulnerability, and creative risk-taking that make storytelling memorable. People sense this absence. They feel when something has been engineered for efficiency rather than crafted with care. And that feeling changes everything. It shifts the brand from a partner in their life to a vendor trying to extract value. It moves the relationship from trust to transaction. And once that shift happens, loyalty becomes difficult to sustain.

Part 2: The "Apple Test" - A Case for Human Craftsmanship

Apple recently released a new intro sequence for Apple TV that stands as a powerful counterpoint to the trend of automated creativity. The intro features the Apple TV logo rendered in vibrant, flowing glass with light dancing through translucent surfaces. Colors shift and blend. The image feels alive, tactile, and mesmerizing. What makes this intro remarkable has less to do with its visual beauty and more to do with how it was made. Apple chose to build the entire sequence using physical materials, practical effects, and in-camera techniques rather than relying on computer-generated imagery. They sculpted real glass, designed custom lighting rigs, and filmed everything physically. The project took significant time, coordination, and craftsmanship. And Apple made sure people knew about the process by releasing a behind-the-scenes video showcasing the work.

Video Source: YouTube

The video reveals teams of artists, fabricators, and lighting designers collaborating to bring the vision to life. You see hands shaping materials. You see trial and error. You see creative problem-solving in real time as the team figures out how to achieve specific visual effects using physical tools. The music accompanying the intro was composed specifically for the project by a human artist. FINNEAS, known for his emotional and layered compositions. Every element of the intro reflects intentionality. Every choice signals that Apple views this brief sequence as meaningful enough to warrant serious investment. The company treated the intro as an opportunity to reinforce its brand identity rather than as a checkbox task to complete as efficiently as possible.

The decision to avoid CGI shortcuts feels deliberate and symbolic. Apple operates at the cutting edge of technology. They could have generated the intro using advanced rendering software or AI tools in a fraction of the time. The choice to go practical signals something about values. It communicates that craft matters. That physicality matters. That the human touch matters. The phrase "craft should be felt, and it should be remembered" became an unofficial motto for the project. This philosophy extends beyond the intro itself. It represents how Apple thinks about everything they create. Their products emphasize materials, texture, and the way objects feel in your hands. Their retail stores prioritize physical experience, spatial design, and human interaction. The intro becomes a microcosm of the larger brand identity, a concentrated reminder of what Apple stands for.

The impact of this approach extends to how audiences perceive the brand. When people watch the intro and learn how it was made, they experience something more than admiration for the visuals. They feel respect for the effort. They recognize the commitment to doing things the hard way because the hard way produces something more meaningful. This recognition builds trust. It reinforces the perception that Apple cares about details, values quality, and invests in experiences that elevate the everyday act of pressing play on a streaming service. The intro transforms from a piece of branding into a statement about priorities. It says that even the smallest moments deserve attention. That even a few seconds of screen time merits craftsmanship. That the consumer experience begins the instant someone interacts with the brand, even before the content starts.

This tangible, physical approach also creates a different kind of authenticity. Digital effects can look flawless. They can achieve visual complexity that physical materials struggle to match. But they often lack weight. They feel slick, polished, and slightly detached from reality. Physical effects carry imperfections. Light behaves the way light actually behaves when it passes through glass. Colors blend organically because they are literally mixing in real time. Reflections and refractions follow the laws of physics because they happen in the physical world. These subtle qualities register on an unconscious level. Viewers may struggle to articulate why the intro feels different, but the difference exists. The physicality creates a sense of presence that digital rendering often loses.

The music choice amplifies this effect. FINNEAS composed a piece that feels both modern and timeless, with layered melodies and emotional resonance that complements the visuals. The music adds another dimension of human craft to the project. It reminds viewers that artists collaborated across disciplines to create something cohesive and intentional. The combination of physical visuals and original composition makes the intro feel like a complete artistic statement rather than a functional corporate asset. It becomes something people want to watch repeatedly, something they talk about, something they share. The intro gains cultural traction because it feels special, and it feels special because people recognize the effort behind it.

Apple's approach offers a roadmap for how brands can differentiate themselves in an era where automation tempts everyone to cut corners. The intro demonstrates that investment in craft pays dividends in perception, trust, and emotional connection. It shows that even in industries dominated by technology, human effort remains irreplaceable. It proves that audiences respond to authenticity and that authenticity often requires choosing the slower, harder, more expensive path. The Apple TV intro becomes a case study in how brands can use their creative output to communicate values, build identity, and reinforce the promise they make to their audience every time someone engages with their work. The lesson is clear: craft creates connection. Effort builds trust. And when a brand treats even the smallest creative moment as an opportunity to demonstrate care, audiences notice, remember, and reward that commitment with loyalty.

Part 3: The Research Speaks - How Consumers Really Perceive AI vs. Human Ads

The intuition I described earlier about AI advertising turns out to be backed by extensive research. Scientists, market researchers, and behavioral psychologists have studied how people respond to AI-generated content versus human-created content, and the findings align with what many of us feel instinctively. When people know an advertisement was made by AI, their perception shifts dramatically. They rate it more critically. They trust it less. They engage with it differently. This happens even when the content itself remains identical. Researchers conducted experiments where they showed participants the same advertisement twice, changing only the label. One version said "created by humans" and the other said "created by AI". The responses diverged sharply. The AI-labeled version received lower scores on authenticity, emotional resonance, and trustworthiness. The content had the same words, the same images, and the same design. Only the knowledge of its origin changed, and that knowledge altered everything.

This phenomenon reveals something fundamental about how humans relate to creative work. We care deeply about the source. We want to know who made something, why they made it, and what they put into it. When we learn that a machine generates content through pattern recognition and algorithmic optimization, it changes the emotional relationship we have with that content. The work loses personal dimension. It stops feeling like communication and starts feeling like output. This shift happens quickly and powerfully. Studies from the Nuremberg Institute for Market Decisions involved over 1,000 participants across the United States, United Kingdom, and Germany. Researchers asked people about their attitudes toward AI in marketing, their trust levels, and their responses to specific advertisements. The findings painted a clear picture. Only 21% of respondents said they trust AI companies and their promises. Just 20% said they trust AI itself. These numbers reflect deep skepticism about automation, particularly in contexts where human judgment, creativity, and emotional intelligence seem essential.

The trust gap extends beyond general attitudes into specific behaviors. When participants viewed ads labeled as AI-generated, they reported significantly lower purchase intent. They expressed less interest in learning more about the product. They showed reduced willingness to click through to a website or share the ad with others. These metrics matter because they translate directly into business outcomes. An ad that generates awareness but fails to drive action represents wasted investment. The research reveals that AI disclosure triggers what experts call a "trust penalty," where consumers become more guarded, more critical, and less likely to convert. This penalty exists even when the ad performs well in blind tests, where participants evaluate content before knowing its origin. In those scenarios, AI-generated ads sometimes match or exceed human-created ads on technical criteria like clarity, visual appeal, and message comprehension. But the moment disclosure happens, the advantage evaporates.

The emotional dimension of this trust penalty deserves special attention. When researchers asked participants to rate advertisements on qualities like warmth, sincerity, relatability, and emotional connection, the AI-labeled versions scored consistently lower. People described these ads as feeling "colder," "more generic," "less personal," and "harder to connect with". These descriptions align with the central argument I have been making throughout this piece. Ads are supposed to communicate values, tell stories, and build relationships. When people perceive an ad as the product of algorithmic efficiency rather than human creativity, they struggle to form an emotional bond with the brand behind it. The ad becomes transactional rather than relational. It feels like a pitch rather than a conversation. And in a marketplace where consumers have endless choices, relational brands win over transactional ones.

Context plays a significant role in how people respond to AI-generated advertising. The research shows that AI disclosure matters less when the product itself relates to technology, innovation, or cutting-edge science. If a brand is marketing an AI-powered tool, a robotics product, or a futuristic service, consumers accept AI involvement in the advertising more readily. The match between product and process feels logical. AI promoting AI creates coherence rather than dissonance. However, for traditional products like food, clothing, home goods, personal care items, or services rooted in human expertise, the negative bias against AI content strengthens considerably. People expect these categories to emphasize craft, tradition, care, and the human elements that make products meaningful. When a brand selling handmade furniture uses AI-generated ads, the contradiction undermines the brand story. When a restaurant promoting farm-to-table cuisine relies on algorithmic content creation, the disconnect erodes trust. The medium becomes the message, and if the medium signals automation, the message loses authenticity.

The research also uncovers what happens when people try to identify AI-generated content on their own. Only 25% of consumers believe they can reliably recognize when something has been created by AI. This uncertainty creates anxiety. People worry they are being manipulated or deceived. They feel vulnerable to content that looks human but originates from machines. This feeling intensifies when disclosure is absent or unclear. Studies on AI influencers and AI-generated social media content reveal that explicit disclosure of AI involvement reduces perceived authenticity and increases skepticism. Interestingly, hiding AI involvement does help content perform better in the short term, but it creates massive reputational risk if audiences later discover the deception. Brands face a difficult choice: disclose and accept the trust penalty, or hide the truth and risk backlash when the truth emerges. Neither option feels ideal, which suggests the real solution lies in rethinking the use of AI in creative contexts altogether.

The implications of this research extend far beyond individual advertisements. They touch the fundamental question of what brands owe their audiences. The data confirms that people want to feel valued, understood, and respected by the companies they support. They want evidence that brands care enough to invest time, thought, and creative energy into building relationships. When AI replaces human effort in advertising, it signals that efficiency matters more than connection. It suggests that the brand views marketing as a cost to minimize rather than a opportunity to build trust. And as the research makes abundantly clear, consumers notice this shift and respond by withdrawing their trust, their attention, and ultimately their loyalty. The evidence supports what many of us already feel. Human creativity in advertising matters because it reflects human care in everything else a brand does. When that creativity disappears, trust follows close behind.

Part 4: The Evidence Linking AI Advertising to Lower Brand Trust

Authenticity has become the primary currency in modern marketing. Consumers evaluate brands based on how genuine they appear, how sincere their messages feel, and how deeply they seem to understand the people they serve. This evaluation happens instinctively. People develop gut reactions to advertisements, social media posts, and brand communications based on subtle cues that signal authenticity or its absence. Research confirms that AI-generated content consistently triggers negative authenticity signals. Studies examining AI-generated influencer content show that perceived authenticity drops significantly compared to human influencers. When brands disclose that an influencer or spokesperson is AI-generated, trust declines sharply. People feel misled or manipulated, even when the disclosure happens upfront. The knowledge that they are interacting with a machine rather than a person changes the entire dynamic of the relationship. This effect intensifies when the AI influencer promotes products that depend on personal experience, taste, or emotional connection, like beauty products, fashion, food, or lifestyle services.

The data reveals something striking about human preferences. When people know content comes from a human creator, they prefer it overwhelmingly, even when the AI version demonstrates superior technical quality. Researchers tested this by showing participants advertisements that varied in origin but matched in visual appeal, clarity, and message structure. In blind tests where participants evaluated content before learning its source, AI-generated ads sometimes outperformed human-made ones on metrics like readability, visual composition, and information delivery. But when researchers revealed which ads came from humans and which from AI, preferences shifted dramatically. Participants retroactively rated the human-made content as more appealing, more trustworthy, and more worthy of their attention. This "human-made premium" exists across demographics, product categories, and cultural contexts. It suggests that source matters more than execution in how people evaluate creative work. The knowledge that a person invested time, thought, and care into making something adds intangible value that technical perfection alone fails to provide.

This preference for human creation connects to deeper psychological patterns. People value labor, especially creative labor. They appreciate effort, skill, and the vulnerability inherent in putting original work into the world. When someone knows a human wrestled with an idea, revised it multiple times, collaborated with others, and took creative risks to produce something meaningful, the output carries emotional weight. That weight creates connection. It builds rapport. It makes the viewer feel like the creator cares about reaching them, understanding them, and offering something valuable. AI-generated content lacks this dimension because it originates from pattern analysis rather than lived experience. It optimizes for statistical likelihood rather than emotional truth. And while the results can look polished and professional, they feel hollow. The absence of human struggle, human insight, and human vulnerability registers subconsciously, creating distance between the content and the viewer.

Visual aesthetics play a fascinating role in how people detect and respond to AI-generated advertising. Research shows that certain visual characteristics have become associated with artificial generation in the public consciousness. Intense color saturation, overly smooth gradients, unnaturally perfect symmetry, and specific rendering styles trigger suspicion. People have started to recognize these patterns, and when they spot them, they disengage. Studies measuring click-through rates reveal that AI-generated ads which successfully pass as human-made perform significantly better than both human-created ads and AI ads that visually signal their artificial origin. This finding creates a paradox for brands. If AI content looks too polished or follows recognizable AI visual patterns, audiences avoid it. But if AI content successfully mimics human imperfection and avoids detection, brands risk backlash when the truth emerges. The safest path remains investing in actual human creativity, which avoids both the detection problem and the ethical concerns around deception.

The trust deficit surrounding AI in advertising extends beyond aesthetics into fundamental concerns about data, privacy, and manipulation. Only 28% of consumers understand how AI uses their personal data to create personalized content. This knowledge gap fuels anxiety. People worry about what information companies collect, how algorithms process that information, and what kinds of psychological manipulation might result. These concerns intensify when brands use AI to generate advertising because it feels like automation has infiltrated the already fraught space of persuasion. Advertising already operates on the boundary between information and manipulation. Adding AI to the equation pushes many consumers past their comfort threshold. They feel like brands are using sophisticated technology to exploit their vulnerabilities, predict their behavior, and push them toward purchases they might regret. This perception damages trust even when the reality is less sinister.

Adding to this trust deficit is the widespread belief that people cannot reliably identify AI-generated content. Only 25% of consumers feel confident in their ability to recognize when something has been created by AI. This uncertainty creates a baseline level of suspicion. If people feel they might be consuming AI content at any moment and they generally distrust AI content, they approach all advertising more skeptically. This dynamic hurts even brands that rely entirely on human creators because the general erosion of trust affects the entire ecosystem. When a few prominent brands get caught using undisclosed AI content, consumer skepticism rises across the board. The resulting environment becomes hostile to all marketing, making it harder for authentic brands to break through the noise. This reality creates a collective action problem. Every brand that chooses AI-generated advertising contributes to a broader decline in trust that ultimately damages everyone.

The evidence points to a clear conclusion. AI involvement in advertising creates measurable harm to brand trust, consumer perception, and business outcomes. The harm manifests across multiple dimensions: emotional connection, perceived authenticity, purchase intent, engagement metrics, and long-term loyalty. The research also suggests that transparency alone fails to solve the problem. Disclosing AI involvement reduces the risk of backlash from deception, but it still triggers the trust penalty associated with algorithmic content creation. Brands face a fundamental choice. They can continue using AI for advertising, accept the associated trust costs, and hope efficiency gains offset relationship damage. Or they can recommit to human creativity, invest in authentic storytelling, and differentiate themselves by demonstrating visible care. The research strongly favors the latter approach. In a marketplace where consumers crave connection, authenticity, and evidence that brands value them as people rather than data points, human creativity becomes the most powerful competitive advantage available.

Part 5: What Makes AI Ads Feel Inauthentic - The Underlying Factors

Something feels off when you watch certain advertisements today. The visuals look polished. The messaging sounds clear. But somewhere beneath the surface, a strange discomfort emerges. You might struggle to articulate what bothers you. The people in the ad smile, but their expressions feel stiff, mechanical, like they are performing rather than living. The editing rhythm feels odd, with transitions that follow patterns you recognize but which somehow feel unnatural. The gestures look limited, repetitive, or slightly disconnected from the emotional tone of the scene. This phenomenon has a name in psychology: the uncanny valley. The term originally described how robots and digital characters that look almost human but retain subtle artificial qualities trigger feelings of unease. AI-generated advertisements increasingly fall into this same valley. They achieve visual sophistication. They mimic human creativity well enough to pass a quick glance. But they retain telltale artifacts that register subconsciously and create discomfort.

The uncanny valley effect intensifies when AI attempts to replicate human faces, bodies, and emotional expressions. Current AI systems excel at generating technically accurate images. They produce faces with correct anatomical proportions, eyes that appear to focus, and mouths that curve into smiles. But human perception evolved over millions of years to detect subtle emotional cues in other humans. We read microexpressions, notice tension in jaw muscles, track how eyes move in relation to emotional states, and register countless tiny details that signal genuine feeling versus performed emotion. AI-generated faces often miss these microdetails. The smiles look rehearsed. The eyes appear slightly vacant. The overall expression feels like a mask rather than a window into genuine human experience. Viewers process these discrepancies below conscious awareness, experiencing them as a vague sense that something feels wrong. This vague wrongness erodes trust and makes the entire advertisement feel less believable.

Research from NIQ demonstrates that AI-generated ads create cognitive confusion and require more mental effort to process than traditional human-created ads. This finding matters because advertising effectiveness depends partly on ease of processing. When people encounter content that feels natural and flows smoothly, they absorb the message more readily and develop more positive associations with the brand. When content requires additional cognitive work to interpret, people feel frustrated, distracted, or vaguely annoyed. They may complete watching the ad but retain less of the message. They may develop negative associations with the brand simply because the ad made them work harder to understand what they were seeing. This cognitive burden compounds with the emotional discomfort of the uncanny valley, creating a dual barrier between the brand and the audience. The brand intended the ad to build connection, but instead it builds distance.

The negative feelings generated by an uncomfortable or confusing ad transfer directly onto the brand itself through what psychologists call the halo effect. Traditionally, the halo effect describes how positive qualities in one domain create positive assumptions about unrelated domains. A physically attractive person gets assumed to be kind, intelligent, or trustworthy based purely on appearance. In advertising, the halo works both ways. An ad that feels genuine, creative, and emotionally resonant creates positive associations with the brand. People assume the brand embodies the same qualities they perceived in the ad. But when an ad triggers discomfort, confusion, or a sense of artificiality, those negative feelings attach to the brand. The viewer begins to associate the brand with fakeness, with shortcuts, with a willingness to manipulate rather than communicate honestly. This negative halo proves difficult to reverse because it operates largely below conscious awareness. People develop vague negative feelings toward the brand that they may struggle to explain or justify, making it hard for the brand to address the problem directly.

Emotional intelligence represents another critical dimension where AI-generated advertising consistently falls short. Human creativity draws on lived experience. Copywriters, designers, directors, and strategists bring their own emotional histories into their work. They understand joy, loss, frustration, hope, and connection through personal experience. This understanding allows them to craft messages that resonate emotionally because they originate from genuine emotional insight. AI systems analyze patterns in successful past advertisements and attempt to replicate those patterns. They identify which words, images, and structures correlate with high engagement or conversion rates. But pattern recognition differs fundamentally from emotional understanding. AI knows which phrases statistically precede purchase behavior. It does this while having zero experience of what those phrases actually mean to the humans reading them.

This gap becomes visible when AI attempts to handle nuanced emotional territory like humor, vulnerability, nostalgia, or inspiration. These emotional registers require subtlety, timing, cultural awareness, and sensitivity to context. Humor especially depends on understanding what breaks expectations in surprising but delightful ways. AI-generated humor often falls flat because it identifies surface patterns in jokes but misses the underlying logic of why humans find certain things funny. Similarly, AI-generated inspirational content often feels cliché or manipulative because it relies on phrases and images that have worked before, arranged in statistically optimal patterns, rather than emerging from genuine desire to inspire. Audiences sense this lack of genuine emotional intelligence and respond by withdrawing trust. They feel addressed as data points rather than as whole humans with complex interior lives.

Credibility takes the hardest hit when brands use AI-generated advertising. Research shows that consumers rate brands using AI ads as significantly less credible than brands using human-created content. They develop less positive attitudes toward these brands overall. The credibility loss stems from multiple factors discussed throughout this piece: the uncanny valley discomfort, the negative halo effect, the perceived lack of emotional intelligence, and the broader signal that the brand views marketing as a cost to minimize rather than a relationship to build. Credibility operates as the foundation of all brand relationships. When people believe a brand is credible, they give that brand the benefit of the doubt during crises, forgive occasional missteps, and remain loyal even when competitors offer similar products at lower prices. When credibility erodes, the relationship becomes transactional. People comparison shop aggressively. They switch brands easily. They feel little emotional attachment. AI-generated advertising accelerates this erosion because it represents visible evidence that the brand chose efficiency over authenticity, cost savings over craft, and algorithms over human understanding. That choice communicates volumes about what the brand values, and consumers respond by adjusting their perception accordingly. The result shows up in harder metrics like conversion rates, customer lifetime value, and brand equity, all of which suffer when credibility declines.

Part 6: How Disclosure of AI Creation Destroys Brand Trust

We often assume honesty builds stronger bonds. You tell the truth, and people appreciate it. However, recent findings flip this logic upside down when it comes to artificial intelligence. Revealing that an ad was created by AI actually lowers the audience's trust in the brand. This phenomenon is known as the Transparency Paradox. Even when two pieces of content are identical, the one labeled "AI-made" feels less genuine to the viewer. People perceive the content differently simply because they know its origin. The label itself changes the experience. Authenticity relies on a sense of human effort and intention. When that human element seems missing, the connection weakens. Brands expect transparency to help, but it hurts.

Authenticity acts as the bridge between a viewer and a brand. Research confirms that this feeling of realness is the primary driver of consumer reaction. When someone sees an AI disclosure, their sense of authenticity drops. This drop triggers a chain reaction. A lower sense of authenticity leads directly to a colder brand image. Consumers start to question the brand's values. They wonder about the effort behind the message. The disclosure signals a shortcut, and that signal damages the relationship. The feeling of "realness" is what people buy. Losing that feeling means losing the customer's emotional investment.

This effect hits everyone equally. You might think big, famous brands have enough goodwill to survive this transparency. Data suggests otherwise. The negative impact of AI disclosure remains strong across both massive corporations and unknown startups. It suggests a general human skepticism toward machine-generated creativity. People seem to have an innate preference for human-made stories. This skepticism stands firm regardless of brand history or reputation. It is a universal human response to synthetic media. Every brand faces the same hurdle here. Trust depends on humanity, and AI labels strip that humanity away.

The final result shows up in the numbers. Attitudes toward ads with AI disclosures become visibly colder. This shift in attitude hits the bottom line hard. Viewers show a significantly lower willingness to research the product or make a purchase. The engagement metrics follow the trust metrics. When trust dips, action dips. People scroll past what they perceive as artificial. They save their attention for things that feel human and earned. The decision to disclose AI use essentially trades performance for transparency. Brands pay for that honesty with lower engagement and fewer sales

Conclusion

We prioritize brands that demonstrate care through visible effort. Consumers value the intention behind the work. Authentic human creativity builds the strongest connections. I am currently researching Part Two of this series. The next installment is under active development. It will explore specific strategies for building trust in this new era. Please await this upcoming release. It will provide actionable steps for prioritizing human creativity. Let us continue to support the work that respects us.

why I choose to stay in this feeling

Om — Sat, 07 Feb 2026 08:15:00 GMT

I came across this idea recently that when we love, we become uniquely vulnerable, giving someone the power to hurt us like never before. that thought stayed with me after my breakup. I keep asking myself why I feel stuck, why every part of me resists the idea of moving on. the truth is, I love her regardless of everything, and that love feels more important than protecting myself from the pain.

this vulnerability comes from opening parts of myself I keep hidden from the world. she saw me without filters, without the careful presentation I show everyone else. I shared my hopes, my fears, my vision of what life could become. that kind of exposure creates a bond that ordinary interactions simply miss. when she left, it felt like losing access to a version of myself I could only be with her.

the resistance to moving forward tells me something worth listening to. this relationship touched something real in me. she awakened qualities I want to keep, reflected back parts of myself I felt proud of, helped me become someone I genuinely liked being. letting go feels like abandoning that person I became when we were together. my loyalty persists because what we had mattered, shaped me, changed the way I see connection itself.

recovery feels impossible because I'm grieving two things at once: what we actually shared and what we could have built together. my brain structured itself around her presence, built daily rhythms and patterns of thinking that included her. moving on means rewiring that architecture, and part of me wonders if I even want to. the depth of this feeling proves the relationship meant something profound.

She was there when I was just beginning to understand what love actually meant. I was building myself up, figuring out who I wanted to become, and she became my definition of love itself. Looking back now, I see how much I've grown out of that immature trance, all the terrible mistakes I made, the behavior I deeply regret. The wild part is knowing that if she saw how I'm processing this now, she might label it drama and walk away. Maybe that reaction itself is just a defense mechanism, or maybe it's the pressure to fit into that cool, detached persona everyone seems to perform these days. Either way, thinking about this feels important, even fascinating. I want to keep loving what we had without actually wanting her back, especially considering who she's become. Yet I hold onto this quiet hope that somewhere in her memory, when she thinks of us, she smiles at the good parts rather than tagging everything as immature or cringe.

maybe the question matters less about when I'll move on and more about what I'm learning from choosing to stay present with this love. the loyalty I feel, the commitment that persists even through pain, these capacities matter. they show me I can sustain devotion, hold space for someone even when it aches. right now, honoring what we had feels more authentic than forcing myself to forget. this vulnerability, painful as it is, taught me what it means to truly let someone in.

returning to the essential

Om — Fri, 06 Feb 2026 10:22:00 GMT

we live in an era of speed and abundance. while digital access offers convenience, we believe clarity requires a different pace. the mind acts as an architect, designed to construct its own reality rather than simply observe one. to step back from the digital current is to reclaim the primary role in one's own life. it is a choice to design a day that looks and feels like our own.

many platforms offer a curated view of success, suggesting a specific standard of what is desirable. however, we value the freedom to define excellence on our own terms. true confidence remains quiet and internal . it stands independent of external validation. when we look inward for approval, we find a stability that the screen can rarely provide. to stand apart is to respect your own potential.

time remains our most significant asset. the vertical scroll invites us to spend this asset freely, often without a receipt. we prefer to invest our time with intention. there is a profound elegance in choosing exactly how we spend our hours. we prioritize depth over breadth, ensuring that every moment contributes to a larger purpose. this is the difference between spending time and cherishing it.

we view entertainment as a form of restoration rather than mere occupancy. the ancient greeks used the word scholē—the root of 'school'—to describe leisure. for them, leisure was an active state of learning and contemplation. true entertainment builds the mind. it engages the intellect and leaves us feeling capable and calm. we seek activities that replenish our energy.

deep engagement offers a richer return than quick consumption. reading a complex book, mastering a craft, or holding a long conversation requires patience. this patience yields understanding. we choose to engage with ideas that have weight and history, allowing us to grow with stability. we believe in the value of sustained focus, where the mind has the room to explore a concept fully.

technology serves us best when it remains a tool. i choose to use my device for its original promise: to bridge distances and access wisdom. i use it to communicate with purpose and to learn with focus. when we lead the interaction, the device becomes a powerful guide. it helps us stay in touch and absorb information that aligns with our goals.

we hold the right to curate our environment. just as we select the food that nourishes us, we select the information that shapes our thoughts. choosing a specific path requires the courage to embrace silence. in that silence, we hear our own voice clearly. we exercise our right to choose quality, relevance, and truth in every piece of information we absorb.

today marks a shift toward the tangible. i am turning my focus to the physical world, to real conversations, and to work that matters. this is a step toward a life of greater intention and presence. i look forward to finding clarity in the real world, and i invite you to find your own version of this peace.

The AI Everywhere Rush: Are We Missing The Point?

Om — Thu, 05 Feb 2026 17:40:00 GMT

You look around right now and it is all artificial intelligence. AI, AI, AI. Every single company is sprinting to adopt it, trying to do something with it. This atmosphere feels intense, a high-pressure situation where AI appears to be the only focus.

Consider the companies we really trust. The ones that genuinely fulfilled their original commitment to us. They began with a mission to solve a specific, real-world user problem. And they succeeded! Achieving that initial goal is a huge victory. Of course, a business needs to innovate; that is vital for progress. I fully grasp that. However, I sense something important is being overlooked.

"I strongly believe there is absolutely zero necessity to shoehorn AI into every single aspect of every product."

1. The Future Path: A Simple Choice

Imagine what the coming days will offer. When you pinpoint a challenge that requires a solution, you will have multiple avenues.

You will be able to select the traditional, established method that already performs well.
You will be able to select the method that utilizes artificial intelligence.

The crucial idea here is that this must remain a choice, a careful decision based entirely on the problem itself. It hinges on whether AI truly improves the solution, making it better, quicker, or easier for the person using it.

2. Beyond the Fear Driving Decisions

What seems misguided is the powerful, almost mandatory belief being circulated: the notion that if a company refuses to integrate AI, its entire structure will immediately collapse. That its stock value will drop sharply. This sense of urgency is dictating enormous investment and product development strategies.

True innovation happens when you enhance the primary experience. When you find a superior way to execute the task you are already proficient at. Sometimes, that better way involves AI. Many times, it involves other improvements—simplifying the interface, hiring skilled customer support, or merely running a faster server.

The Conclusion

The problem should always dictate the solution, without exception. AI is a powerful instrument in the toolkit, but it is not the only instrument we are compelled to employ. We owe it to ourselves to ask: "What is the real motivation here?"

Is your fridge running AI? Or is it just keeping the milk cold?

The Soul of Gifting: A Guide to Giving a Piece of Your Heart

Om — Thu, 05 Feb 2026 09:15:00 GMT

We have all stood there, holding a generic item in a store, feeling that familiar weight of uncertainty. You wonder if the price tag is high enough to show you care. You question if this item truly represents the love you hold for the person. We often equate the value of a gift with the money leaving our bank account. It is time to flip that script entirely. I have cracked the code on what makes a gift truly perfect, and it is almost always unrelated to the cost.

The perfect gift is a feeling. It is a captured memory. It is a mirror that shows the recipient exactly how much they matter. Unless the specific purpose of the gift is to provide luxury, where the high price is the actual point, the monetary value remains secondary. In fact, it often matters very little. The true magic lies in a specific, thoughtful framework. This approach prioritizes connection, longevity, and a deep, resonating narrative. Let us explore how you can transform a simple object into a profound emotional experience.

The Foundation: Relatability and Resonance

A gift must speak. It needs to have a voice that whispers directly to the heart of the person receiving it. When we focus on relatability, we look at the person's current state of being. Where are they in their life? What defines their nature right now? A generic item fails to answer these questions. A perfect gift answers them loudly.

Think about the dynamic of your relationship. Is it playful? Is it deeply intellectual? Is it built on shared struggles or shared triumphs? The gift must resonate with that unique frequency. It serves as a physical manifestation of the bond you share. When the person unwraps it, their immediate thought should be, "You see me. You really understand who I am." This level of resonance creates an impact that far outlasts the initial excitement. It validates their identity and your connection to it.

Longevity: The Gift That Stays

We live in a world of disposable things. A perfect gift rebels against this. It demands longevity. This implies that the item, or the memory of it, should stick around. It should be something that accompanies them through life, acting as a constant, gentle reminder of your care.

Every time they see this item, or use it, or think about it, a spark of your friendship should return to them. It becomes a landmark in their environment. This durability adds weight to the gesture. It says that you intend for your presence in their life to last, just as the gift lasts. A fleeting joy is wonderful, but a lasting presence is powerful. We aim for the latter.

The Core: The Thoughtful 'Why'

Here sits the most critical piece of the puzzle. The object is merely a vessel; the 'Why' is the soul. You must have a profound reason for choosing this specific thing. It requires digging deep. Why this? Why now? Why for them?

When you figure out the 'why,' you have the emotional core of the gift. This reason needs to be solid. It transforms a random purchase into a targeted arrow of affection. The 'why' is your intent. It is the energy you pour into the selection process. When a gift lacks a 'why,' it feels empty. When it is full of 'why,' it feels heavy with love.

Leveling Up: The Narrative Journey

Now we move to the advanced class. This is where gifting becomes art. Once you have your 'why,' you must chip it into a beautiful representation. You turn the intent into a story, or a wish, and you integrate that story directly into the gift itself.

But here is the secret: you keep the full meaning hidden at first. You create a mystery.

The receiver gets the gift, but they might grasp only the surface level immediately. That is part of the design. You want them to embark on a journey of discovery. They need to sit with the gift, look at it, and slowly peel back the layers. You are inviting them to solve a puzzle where the prize is your affection.

This journey of figuring out the 'why' should parallel the history of your friendship. Think about how you met. Think about how your bond grew. The way they discover the meaning of the gift should feel like a retelling of that shared history. It creates a narrative arc. When they finally have that "Aha!" moment, when they realize exactly what you meant and why you gave it, the emotional payoff is massive. It creates a rush of realization. The story you embedded in the gift becomes a part of their own story.

The Highest Form: Healing and Affirmation

We can elevate this even further. This is for the people you hold dearest. We all carry sorrows. We all battle insecurities. A truly masterful gift honors the whole person, including the parts that hurt.

You can tune the gift to address these vulnerabilities. If a friend has shared a deep insecurity with you, use the gift to offer a counter-narrative. If they feel unseen, give them something that proves they are vivid. If they feel weak, give them a symbol of their resilience. You are trying to fill the voids they have revealed to you.

This approach honors them as a complete human being. It shows that you listened when they shared their pain. It demonstrates that you cherish them enough to offer a balm for those wounds. A gift that heals is a gift that is never forgotten. It transcends material value and becomes a tool for emotional support.

Putting It All Together

So, we have our framework. We choose something relatable and long-lasting. We identify a powerful 'why.' We encode that meaning into a story that the recipient discovers over time. And finally, we use that story to affirm their worth and soothe their sorrows.

When you follow these steps, the result is magical. The gift turns out to be great every single time. It makes an impact that money is unable to buy. It strengthens the bond. It deepens love. It makes the act of giving a profound exchange of humanity.

Start looking at the people in your life through this lens. Ignore the price tags. Look for the stories. Look for the healing. Look for the connection. That is where the perfect gift hides, waiting for you to find it.