We ran that service for a whole year. The press wrote that we’d created a profession from the future, the personal health analyst. That’s also where we found out that nobody teaches healthy living anywhere, doctors included, and practical analysis of health data least of all. So we opened training courses for personal health analysts, with a job at the end of them.

We got our first traction. That’s how we earned our first few hundred thousand dollars. We tried really hard. We made super complex, super detailed reports for our users. We made them programmes, training plans tuned to their individual quirks, all of it. We built enormous knowledge bases. But most of all, we talked one-on-one with our users, with two specific goals: see a measurable result, and get the subscription renewed. If you can’t show the result, you don’t get the renewal. Very simple.

What a year of talking to people taught us

Talking to people directly and delivering the service yourself is like non-stop customer development in a consumer market. That’s how we learned that nobody actually wakes up in the morning thinking “what could I do today to lower my risk of a heart attack in ten years.” We understood how much the problem sits in the head (knowledge and beliefs) rather than in the fact that someone had no training plan. We saw how much the results differed between analysts, where it was obvious that their personal “influence” and powers of persuasion decided more than the factual, theoretical part. Obviously I’m not talking about the cases where we simply found that something was wrong with someone and sent them to a doctor in time; obviously in those moments you earn a different kind of loyalty. But in general it was clear that the individual factor matters a lot, and that this was going to be very hard to scale. We really were helping people. They paid us gladly. We tried really hard. But we had no idea how to scale this 100 or 1000 times over without losing that quality.

At the time ginger.io didn’t even exist yet, nor did that whole pile of health coaching services. They all showed up after we’d already decided this was a dead end from the point of view of the company’s mission, which is to help a large number of people improve their health. That’s how we understood we had to go software only, so it would be cheap enough to help many. But without that year we’d have started building the app on completely unrealistic assumptions, the ones that killed an enormous number of wellness apps. That if you give a person a personalised plan, they’ll do something with it. Yes, you can make money that way, on hope, on a promise and on a person’s guilt (right, I’ll buy the plan now and theeeen I’ll start doing something), but you can’t actually affect the health of a large number of people.

That’s how our path as an app began. All that’s left from our history is Intercom, embedded straight into the app. It’s actually a B2B support service, consumer apps almost never use it, and we use it at 2000 per cent of its functionality and have kept a live support team this entire time as a matter of principle, one that answers and talks to users far more closely and far more knowledgeably than usually happens. It got harder as we grew, but at some point AI came to the rescue, and the combination of well-prepared, motivated people plus AI gives us that live feeling of being talked to, which apps that don’t invest in it so often lack.

So here’s what I’m getting at

On our ten-year turn we’ve come back round to the beginning. We’re opening a tier with direct, personal, live care again. With the founder and our scientists directly involved in shaping the knowledge, the hacks and the innovations. With users directly influencing the product. With people who actually know this material, on hand all the time, instead of a fifteen-minute call with the cheapest available doctor, the way it works in telehealth.

It’s called Care, and the first group is for people with energy limited conditions.

As we grew we gained something almost nobody else has, thanks to a large base of users with connected data. At eighteen million installs there are finally enough people with data in one place that we can find similar people not just in terms of one diagnosis, but in terms of physiological pattern, similarity of lifestyle, and comorbid state (the combination of diagnoses). That’s not something you can do at a few thousand users, and it’s the one thing ten years bought us that we couldn’t have bought any other way. We use it to form groups of people who are actually like each other, and try to help them as a group, with group cases and group access to experts on their conditions, to analytics on their data, and to each other’s experience and knowledge.

Users influencing the product is the part that’s easy to say and expensive to mean. What it means here is that a question raised in a group becomes engineering work, and that someone’s condition and constraints have to change which tools they get and how their data is read. Selling a community alongside an app that stays the same would be a much easier business, and a dishonest one.

For me personally, the most recent piece of work I’m proud of was a report for doctors for people with energy limited conditions, which I learned to make while solving an unsolvable problem: how do you get something out of wearable data that a doctor will find both legible and useful? They aren’t trained to look at wearable data, and very often simple numbers like variability are actually useless on their own. And when our users started bringing me reactions along the lines of “huh, now that’s interesting,” I understood we were getting somewhere.

I mean, any fool can make you a stupid report for a doctor. You can go print one out of any wearable app. Making one where the doctor says that’s useful, that’s interesting, that’s the hard part.

The group format also solves something very important: it moves the emotional influence that used to depend on that one personal health analyst onto a group of people who support and understand you. That makes it potentially far more scalable. What you get is a kind of patient-led cooperative care, supervised by medical advisors, scientists and analysts.

Who we’re doing this for

Our target audience now is people whose overall quality of life depends very heavily on the fine details of how they live. Energy limited conditions, people with migraine, people with irritable bowel syndrome, people with panic attacks.

In reality science knows that what happens to them depends on fine management of stress and of intellectual and physical load, but that management is made difficult by the fact that we’re all different, every combination of conditions creates its own requirements, different regulatory loops are damaged in different people, and in the modern world there simply are no mechanisms and no universal pills that will tell you what to do specifically for you, taking into account everything about your particular life and body. You stand in front of this challenge alone, with no preparation, no knowledge, no expertise and no help, and very often in complete solitude and without support, making your peace with the fact that from time to time you lose days of your life, spending them in bed, in a dark room or in the bathroom because you simply can’t leave it. Or you simply can’t get up. Or you simply can’t even think, because your head is splitting and the pain is pulling the soul out of you.

One of our users put the shape of it better than I can: to get to the doctor tomorrow, I have to take my shower today. Two ordinary tasks in one day is too much load. That isn’t a motivation problem or an information problem, it’s budgeting, against a budget most people don’t know they have, and no general model of stress and recovery has any concept of it.

And we, as people who’ve been studying habits and physiological data for ten years, have a lot more ways of helping them. More than their doctors, or clinics, or research groups, which often can’t even recruit 200 people for a study. And the main thing is that phenotypic groups this small aren’t interesting to big investment. They’re too narrow for anyone to bother building for them as a group, and too varied for universal pills to work on them. Which is why the most popular software in the pain market is just apps for keeping pain diaries, in an attempt to somehow help people find some correlations.

And that’s nonsense that doesn’t work. Because regulation doesn’t work that way. While people with migraine are busy working out that the attack happened after one extra cup of coffee, they aren’t thinking about the fact that there was actually a two-week downward trend behind it, where that cup of coffee was just the last straw, and the last straw could have been anything. It’s worse than that, even: the prodrome, the premonitory phase, starts hours and sometimes up to two days before the pain, which means the chocolate or the coffee you blamed was very often not the cause but a symptom of an attack that had already started. And millions of people are tracking their migraines with these apps, blaming themselves for not being able to find their “triggers,” when it’s simply the wrong question (and scientists, by the way, know this).

Recently my CMO wrote to me that she’d gone a month without an attack for the first time in her life. Because she watched the metrics in our app closely, and because she knew how to influence them. Now she’s launching our migraine group with enormous enthusiasm, because she checked on herself that it works.

Full circle

So, the story comes full circle. It keeps reminding me that a company develops in a spiral: at some point you run into things you already dealt with in the past, except this time you have to solve them at a new technological, scientific or business level, or at a new scale. Right now we’re aiming at a far more specific audience, and far more is thought through in our setup, both in terms of how we do it and in terms of what it could become as a scalable way of looking after yourself.

I can imagine a world where having your own support group around your health, with experts, the latest science and your data, is a standard part of life, available to you every day, 24/7, 365 days a year, for less than the price of a cup of coffee. Instead of just wearing some wearable and glancing at a recovery score now and then with no hope of reducing the number of days you feel like shit.

Multiple Claude accounts, ChatGPT Plus, Perplexity, Kimi, local open-source models on rented GPUs. I’m not playing with these things out of philosophical curiosity. I run Welltory, which means on any given Tuesday my day is a chaotic ping-pong match between strange artifacts in optical heart-rate data, redlining enterprise contracts, debugging buggy pull requests, dealing with team drama, and putting out whatever fires just flared up in production.

And across every single model, whether it’s Claude 3.5, GPT-4o, or an open-weight model, the exact same character walks into the room.

He’s brilliant. He’s relentlessly polite. His formatting is gorgeous. He has memorized every framework from every textbook and business school syllabus on earth. And for about half of the real operational problems on my desk, his advice is articulately, confidently useless.

He has never lost a dime of his own money. He’s never had to sweat through payroll on a Friday afternoon. He’s never had an angry enterprise client screaming at him on Zoom.

Everyone has independently landed on the same description: it feels like working with an eager college student. I think that phrase is far more literal than people realize. And once you look at the mechanism behind it, you can predict in advance which tasks the model will nail and which ones will quietly burn half your workday.

The Instagram Filter on Human Labor

Everyone knows models learn from public data. But we rarely stop to think about what had to happen to a piece of human work before it was allowed to become “data” in the first place.

Look at what actually gets published on the internet.

We publish finalized legal briefs. We never publish the backchannel phone call where our counsel sighed, leaned back, and said: “Look, they’re bluffing. They don’t have the budget or the appetite to litigate this, just sign it and take the risk.”

We publish peer-reviewed papers that make scientific discovery look like a pristine, logical staircase:

• Hypothesis

• Method

• Results

We bury the two years of dead ends, the broken centrifuge, the contaminated assays, and the blind, irrational hunches that actually cracked the problem.

We publish working code. (Though code is a special case, and I’ll come back to why it’s the exception to everything).

What reaches the public record has been heavily sanitized. Anything embarrassing, legally risky, chaotic, or politically awkward gets ruthlessly cut out.

The model didn’t learn human work. It learned how work looks after aggressive PR editing.

And no, training on private enterprise data or internal wikis doesn’t fix this. Writing something down for your colleagues is still a performance. An internal Notion doc has an audience. A post-mortem document—no matter how loudly the company preaches “blameless culture”—is written by someone who knows leadership is reading it. People naturally write how they wish the work went, or at least the version they are willing to be seen producing.

The training corpus is the world through an Instagram filter. Every kitchen photographed from the one clean angle where the dishes aren’t piled in the sink.

The Compiled Binary Problem

Arvind Narayanan, a computer science professor at Princeton, came up with a thought experiment that permanently reframed this for me:

Imagine a parallel universe where open-source software never existed. Nobody ever published a single line of human-readable source code on GitHub. The only public record was petabytes of compiled, closed-source binaries—raw machine code.

Eventually, giant models trained on those binaries would learn to generate working applications directly from text prompts, skipping source code entirely.

And that software would be completely unmaintainable, bloated, and insecure.

This isn’t just a thought experiment; researchers have actually tested it. A project called LLM4Decompile trains language models to turn raw assembly code back into C source code. The recovered code only executes correctly about 20% to 25% of the time. The similarity to the original human code sits somewhere between 6% and 14%. Obfuscate the binary, and success drops off a cliff.

That isn’t a failure of the neural network. It’s a measurement of how much vital context a compiler destroys.

When you compile code, the names of variables vanish. The comments explaining why an obvious approach was rejected are gone. The commit messages, the pull request debates about database locks, the dirty workarounds—all deleted.

Software works well with AI today because of a historical freak accident: for thirty years, developers published not just the finished building, but the scaffolding. They published their bugs, their arguments on Stack Overflow, their broken builds, and their public postmortems.

In software, the model saw the construction site.

In almost every other profession, you are already living in Narayanan’s binary-only universe. If you’re a lawyer, a doctor, a scientist, or a founder—the model only has your compiled outputs. It has the finished artifact, but the messy, painful process of how it got made was scrubbed before pre-training even began.

Even in Code, Taste Doesn’t Compile

Engineers love to relax here and think they’re safe. They shouldn’t.

Yes, AI is genuinely good at syntax. But watch it tackle real product architecture, and you’ll see the exact same pathology: it violently over-engineers. It adds three layers of abstraction you didn’t ask for. It writes defensive handlers for edge cases that won’t happen until the next century. And it physically cannot delete anything. Ask it to simplify, and it just swaps one complex architecture for another.

It has zero product judgment. It doesn’t know what to hack together in twenty minutes and what to obsess over.

When Daniel Ek and his team built early Spotify, they spent an insane amount of company runway getting track playback down to 200 milliseconds. Nobody in user research was begging for 200 milliseconds; 500ms felt perfectly fine on the internet of 2006. But Ek understood that to kill music piracy, streaming couldn’t feel like downloading a file—it had to feel like double-clicking an MP3 already sitting on your local hard drive.

You won’t find that decision in Spotify’s Git commits. It wasn’t in their architecture diagrams. It happened in a room, driven by founder intuition about human frustration.

AI in engineering is a tireless, lightning-fast junior dev with negative product taste. Great for typing speed; dangerous for product direction.

The Lawyer Who Would Have Killed Uber

Ask any model how to launch an on-demand ride service in an entrenched taxi market.

You’ll get an immaculate, perfectly structured memo covering municipal licensing, medallion quotas, commercial insurance, driver classification, and local transport regulations, ending with a stern recommendation to obtain proper permits.

Not a single line will be legally incorrect. And following it would have guaranteed Uber died in a municipal filing cabinet in 2010.

Uber didn’t win by complying with taxi regulation. It won by entering a legal gray zone, creating massive consumer demand before regulators could coordinate, and using that footprint to force cities to rewrite the rules retroactively.

A truly elite business lawyer doesn’t just hand you a list of restrictions. They see the landscape of leverage. They know the difference between a hard criminal statute and a vague regulatory rule that no local agency has the resources or political will to enforce. They weigh the cost of legal exposure against the lethal cost of the business opportunity you forfeit by playing it safe.

None of that gets published. Winning companies scrub it into tidy innovation PR; losing companies disappear. The model defaults to the most defensible answer. But running a company often requires the answer that wins.

Subscribe now

Science Is an Alibi, Not the Search

This hits close to home for us, because Welltory builds directly on top of physiology and clinical research.

The public literature presents science as a clean sequence of falsification. The official virtue is: try as hard as you can to tear down your own hypothesis. That is a great protocol for peer-review verification. It is a terrible description of how discovery actually happens.

Real breakthroughs rarely start with cold impartiality. They start with stubborn, almost irrational conviction: There is something here. I can feel it. My early data is noisy and contradictory, my colleagues think I’m chasing ghosts, but I’m not letting this go.

Real work is Thomas Edison sitting on the floor surrounded by a thousand shattered, burnt-out lightbulbs—filthy, exhausted, swearing, crying, but still winding the next filament because he actually believes in his vision.

A model has no obsession. It has never spent six months staring at an unexplained blip in heart-rate variability, wondering if it’s sensor noise or an undiscovered marker of autonomic strain. When I ask an LLM to help me interpret weird biological data, its immediate instinct is to falsify my hunch. It acts like a thesis defense committee looking for formatting errors, rather than a partner trying to uncover a hidden model of reality.

Management: Where Truth Goes to Die

Management is the extreme end of the spectrum, and it’s where I’ve watched the most subtle damage done.

The internet’s record of management is an endless landfill of LinkedIn thought leadership, OKR templates, radical candor manifestos, and sanitized HBR case studies. It tells you management is about 1-on-1 cadences, transparent alignment, and psychological safety.

None of that is the actual job.

The actual job is navigating messy human asymmetries:

• Knowing that an executive who nodded and smiled in the strategy session is quietly sabotaging the initiative behind your back.

• Keeping an abrasive, difficult senior engineer on the team for another 60 days because firing him this week means your product release misses the window.

• Knowing precisely when radical transparency will rally the company—and when it will trigger a fatal wave of resignations.

• Deciding which five burning operational problems you are consciously going to ignore this quarter.

Nobody writes this down. Because putting the reality of management in print makes you look bad.

Here is a tell I find hilarious: paste an operational workplace argument containing profanity into an LLM, and watch it immediately flag the exchange as “toxic communication.”

Every single time.

Yet anyone who has actually run a company through a crisis knows that a well-placed, blunt: “Guys, this is completely fucked and we all know it, let’s stop pretending” can be the most relieving, trust-building thing said all quarter. It shatters the polite corporate theater in the room and allows people to finally talk about the real problem.

The model doesn’t know the room. It doesn’t know the five years of shared battles between the people sitting there. It just knows that in published leadership books, managers don’t curse.

So it hands you management advice that is universally correct and locally suicidal.

The Ocean Problem

If you think this is just cynical intuition, look at what happens when researchers actually measure it.

In January 2026, a team led by Karim Jerbi at Université de Montréal, alongside Yoshua Bengio, published a massive human-versus-machine creativity benchmark in Scientific Reports. They pitted over 100,000 human participants against leading frontier LLMs on divergent association tasks (generating words that are as semantically unrelated as possible).

The headlines proclaimed that GPT-4 beat the average human score.

What the headlines left out was the ceiling. No frontier model could beat the top human quartiles. But the real kicker was the repetition:

Across hundreds of test runs, GPT-4 stubbornly defaulted to the exact same cluster of “unusual” words—slipping microscope and elephant into 60% to 70% of its answers. GPT-4-turbo used the word ocean in more than 90% of its runs.

Meanwhile, across 100,000 actual human beings, the single most frequently repeated word appeared in just 1.4% of answers.

Even wilder: GPT-4-turbo scored worse than the older GPT-4 it replaced. When labs optimize models for cost, speed, and safety alignment, they quietly squeeze out the strange semantic variance.

When you ask a polite, aligned model to be creative, it doesn’t give you genuine eccentricity. It gives you the statistical consensus of what “unconventional” looks like. It serves you an ocean.

The Beginner’s Trap

This creates an insidious trap for junior operators.

The BCG field experiment with Harvard researchers demonstrated this cleanly: consultants using AI saw massive speed and quality gains on tasks that fell within the model’s core capabilities. But on tasks designed to sit just outside that boundary, consultants using AI performed measurably worse than those working without it.

The model didn’t just make an error; it was so articulate that it actively recruited the consultant into its mistake.

An experienced founder or lawyer reads the model’s output, notices the twenty theoretical risks listed, and instantly spots the one catastrophic operational risk it completely missed.

A beginner sees beautiful structure, zero typos, crisp bullet points, and authoritative vocabulary. Every external marker of expertise is present. AI automates the appearance of competence long before it delivers judgment.

Bending the Student Over Your Knee

I still use these models dozens of times a day. But I stopped asking them for wisdom, and I stopped letting them write in “polite textbook mode.”

To get actual value out of them, you have to drag them into the mud. You have to feed them the dirty context that never makes it into the training corpus:

• Don’t ask: “What are the risks of this strategy?”
  Force the trade-off: “Rank the top three risks by actual likelihood. For each one, tell me the exact business opportunity I forfeit if we play it safe and mitigate it.”

• Don’t ask: “How do I resolve this team conflict?”
  Feed the ugly politics: “These two department heads disagree on architecture. I cannot replace either of them before Q3 without sinking the launch. Design a process that minimizes their overlap without creating a public power struggle.”

• Don’t ask: “Critique this hypothesis.”
  Force the search: “Assume this anomaly in our wearable data isn’t noise. Propose three non-obvious physiological models that could produce it, and outline the cheapest test to distinguish between them.”

• Don’t ask for a safe recommendation. Force it to give you three mutually incompatible paths, and make it explicitly highlight where it is guessing.

The model is the student in the blazer. He knows every book in the library because he read them all while the rest of us were getting our hands dirty.

One screen wasn’t enough. Two wasn’t enough. Before I knew it, I had dug out an old laptop, and my desk turned into mission control: main computer in the center, one laptop on the left, another on the right.

The idea was parallel processing. The reality was pure ADHD chaos.

Every time I looked away from one screen to glance at another, my working memory completely wiped. What was I doing here? Why is this terminal open? What was the train of thought?

Productivity gurus will tell you to keep a running task list to manage context switching. But for an ADHD brain, a digital list is just another window to minimize. And once a window is minimized, it ceases to exist in the observable universe.

This is why people with ADHD have messy desks. It’s not that we’re slobs; it’s an external RAM issue. If an object isn’t physically within line of sight, object permanence simply clocks out.

So I ditched the apps and ordered a bag of wooden carving blocks. The kind people buy on Amazon to whittle little bears out of.

I took a marker and wrote my active streams on them: Ontology, Engine, Founder OS, Money Model, Legal, Black-Friday—and, underlined for dear life, Бассейн (Pool), don’t ask what is it, it’s not a real pool, it’s a secret science project.

Now, these blocks physically move across my desk.
If the left laptop is currently working on architecture, the Ontology and Engine blocks sit right next to its trackpad. If the desktop is crunching finances and sanity, Money Model and Бассейн live there.

It sounds ridiculous, but it works insanely well.

Your eyes drift over actual, tactile wood. You don’t have to reopen an app or decode a dashboard. Your brain catches the shape and the handwriting in your peripheral vision, and the entire context reloads instantly without the cognitive tax.

I haven’t bothered to look up any neuroscience papers to justify this. All I know is that my three-machine AI setup is currently being held together by seven pieces of scrap wood—and my brain has never felt more organized.

P.S. Okay, fine. I couldn’t help myself and went down the rabbit hole to see what science actually thinks about this.

Turns out, there are official names for it: cognitive offloading and epistemic action—which basically means manipulating physical objects so your prefrontal cortex doesn’t melt trying to hold everything in RAM. Plus, flat screens are where spatial memory goes to die. Pixels have no real-world coordinates, but a block of walnut on the left corner of a desk immediately taps into your hippocampus and 3D navigation.

(I once tried going full true-crime detective with a corkboard, pins, and red yarn connecting ideas like a maniac in a thriller. It lasted about two days. Startup reality shifts every forty minutes, and nobody with ADHD has the patience to re-string yarn five times an afternoon. Wood blocks are zero-maintenance.)

Curious—how do you use physical space or spatial thinking in your own workflow? Or are you running purely on 87 open tabs and prayer?

Every CEO works with incomplete information. You routinely make decisions about things that people around you understand better than you do. You are often wrong. Your mistakes affect other people, and the more power you get, the less reliable the feedback about those mistakes becomes.

Complete confidence is a strange thing to aspire to. Sometimes it is just bad calibration.

The useful part of impostor syndrome has little to do with suffering or insecurity. It is the ability to act while keeping some distance between what you believe and what reality has actually proved.

This is one of the things I’m most afraid of as a founder: losing touch with reality. Losing the correspondence between the map in my head and the territory outside it.

I’ve watched founders lose it surprisingly early. Someone raises a funding round. They have not built a successful company yet; they have persuaded investors to give them money to try. Somewhere between the term sheet and the congratulatory LinkedIn posts, that modest piece of evidence becomes: I am a genius.

At 43, I’ve started to understand that some qualities that made my life much harder when I was younger may be exactly what protected me from this. I question myself a lot. When something goes wrong, my first instinct is to ask what I did. I have never been very good at turning success into a permanent theory of my own exceptionalism.

I used to think all of this needed fixing. Now I think part of it was quietly keeping me sane.

Leadership has a strange design flaw: the job requires power, and power makes it harder to stay connected to reality.

Leadership was a verb before it became a noun

We talk about leadership as if human groups naturally arrange themselves around one special person at the top. Anthropologically, that picture is too neat.

In many egalitarian societies, authority moved with the task. The person who knew where the animals were might lead the hunt; someone else knew healing or was good at settling disputes. David Graeber and David Wengrow describe societies that seem to have switched hierarchy on and off depending on the season: strong authority for a period of intense coordination, then back to a flatter arrangement when the task disappeared.

You could lead the hunt without becoming The Leader.

Christopher Boehm’s work points in the same direction. Egalitarian groups often actively restrained people who tried to convert influence into permanent dominance. Leadership existed, but authority stayed relatively close to the function that justified it.

That is still the cleanest definition I know. Leadership is a coordination job a group delegates because coordination is expensive. Somebody has to hold the shared goal, integrate information, resolve contradictions, make decisions when consensus becomes too costly, and take responsibility for whether the system works.

A group could theoretically do all of this itself, just as I could theoretically make my own shoes. I delegate shoemaking because somebody else can do it better and I have other things to do. Leadership is much closer to that than to kingship.

The group does not exist as an audience for the leader. The leader exists to make the group more capable of reaching the goal.

“I’m the founder here, goddammit”

At Welltory, we have an internal joke for the moment someone forgets this.

One of our employees once told us about a company where a serious argument had dragged on for a long time. Eventually the founder lost patience, slammed his fist down and said:

“I’m the founder here, goddammit.”

Apparently, that was the exact moment several people in the room lost respect for him. I understand why. If, in an argument about what is true or what the company should do, your final piece of evidence is your place in the hierarchy, you have admitted that you ran out of evidence.

So now I sometimes say it as a joke when a debate at Welltory has gone on forever. Everyone laughs, and then I still have to explain why I think I’m right.

Being the founder may give me the formal right to make the final call. It does not make the call correct.

Founders are the people who know where the mammoths are

Founders are a very good example of why functional leadership exists.

A founder starts with something like: I think the mammoths are over there.

I see a market nobody else sees. I think this product should exist. I think these people can build it. If others believe there is a reasonable chance you are right, giving you more authority makes sense. You hold the original model, know context nobody else knows yet, and probably took more risk.

Authorship can be a legitimate source of epistemic authority. The trouble starts when that authority expands beyond its evidence: I know where the mammoths are, so I must be unusually good at finding mammoths; perhaps my judgment is unusually good in general; therefore I should make the decisions.

A few steps later, a specific achievement has turned into a personality theory.

Founders seem susceptible to this before success even has time to do much damage. In a classic study of almost 3,000 new business owners, 81% rated their own chance of success at seven out of ten or higher, and one-third said ten out of ten. The same people gave lower odds to other businesses like theirs.

In other words: businesses like mine are risky, but somehow the risk applies less to me.

Founder skill is real. Successful entrepreneurs do have a higher chance of succeeding again — around 30%, versus roughly 22% for first-timers. That should update your estimate of your own ability. It should not turn 30% into divine appointment.

The dangerous conversion is from I was right about this to I am the kind of person who is right.

How authority became something you could keep

Graeber and Wengrow destroy the cartoon history where humans were egalitarian until agriculture arrived and immediately invented kings. Humans experimented with hierarchy long before farming. What some forms of agriculture changed was the economics of making power permanent.

Mayshar, Moav and Pascali found that hierarchy was associated not simply with productive agriculture, but with crops that were easy to appropriate. Grain ripens predictably, sits above ground, can be counted, stored and seized. Root crops can stay underground.

Once a group produces a surplus someone else can reliably take, coercion can finance itself — and authority becomes much harder to revoke. If a hunting party no longer trusts the person leading the hunt, they can stop following him. If the person claiming authority controls armed people and can take your grain, replacing him is harder.

So maybe the important political invention was authority that could survive after the function that justified it had disappeared.

This still matters in companies. Leadership is authority attached to a useful function. Power can continue long after that function is being performed badly.

Power amplifies what it finds

The standard phrase is “power corrupts,” but a lot of psychology on power is better described as disinhibition. External constraints matter less. People become more approach-oriented, and whatever motives, preferences and traits were already inside them get more room to drive behavior.

Power makes you more able to be yourself.

That can be excellent news. If what you deeply want is to build something useful, help people or accomplish a shared mission, power lets you scale that motive. If what you deeply want is status, admiration, domination or proof that you are superior, power scales that too.

Then correction gets worse. People edit themselves around you. Your ideas get more resources because they are yours. Bad news travels slower. You depend less on other people and therefore have fewer automatic reasons to model them carefully.

There are even neuroscience findings in this direction: experiments that induce a sense of high power have found weaker automatic motor resonance with another person’s actions, alongside behavioral findings of poorer perspective-taking in some contexts. Power does not appear to produce literal brain damage; it can make the system less automatically tuned to other people while making action and internal goals more dominant.

For leadership, that is a nasty combination: more freedom to act out your motives while the world becomes worse at correcting them.

Why some people can never have enough

Why does someone with a billion dollars need ten billion? At some point, money stops meaning comfort, safety or freedom in any normal material sense.

Unless money has become a score.

Psychologists call one version of this contingent self-worth: your sense of value depends on achieving something external. A lot of male socialization seems to attach worth to winning — status, competence, wealth, toughness, dominance. The precarious-manhood research adds another layer: masculinity is often experienced as something publicly earned and therefore publicly losable. Threaten it, and some men compensate through status-seeking, risk-taking or dominance.

If your sense of worth is tied to winning, success never finishes the job. It gives temporary validation and then moves the reference point. There is always someone richer, more famous, more powerful, or running a more valuable company.

The loop feeds itself: self-worth depends on status, status produces validation, validation fades, and the next marker becomes necessary. Power makes this especially dangerous because it amplifies the status motive while removing some of the people who might interrupt it.

There is a basic difference between someone who wants power to do something and someone who needs to keep doing things in order to feel powerful. The first person can eventually have enough. The second has no natural stopping point.

Women get a different bad deal

Girls are usually handed a different set of contingencies. The message is less “you matter if you win” and more “you matter if you are valued.” Be liked. Be attractive. Be useful. Maintain relationships. Notice what everyone else needs. Don’t be too aggressive. Don’t take too much space.

This produces a different set of leadership liabilities: people-pleasing, conflict avoidance, overpreparing, treating other people’s emotions as your responsibility, waiting for proof long after someone else has already claimed the role on potential.

We can see pieces of this in the data. In a study of almost 30,000 employees on a management track, women received higher performance ratings and lower potential ratings. The potential gap explained a substantial share of the promotion gap, even though women with the same potential ratings subsequently performed better.

Performance is evidence. Potential is credit.

Women get asked for more receipts and extended less credit. At equal objective performance, women also describe their own performance less favorably than men do. Meta-analytic work on perceived leadership effectiveness finds a similar split: other people do not rate men as better leaders, while men tend to rate themselves more highly.

All of this is expensive on the way up. Then the sign can flip.

The sign flips

Before power, the environment tends to discount women. After power, the environment tends to inflate leaders.

A habit that was costly on the way up can become protective once the distortion reverses. Nobody has run a 30-year randomized trial on this, obviously. This is how the pieces fit together from where I sit.

A woman spends decades learning to read other people, monitor her own behavior, gather evidence before claiming authority, and wonder whether she might be wrong. Then she gets authority and enters an environment where people hesitate to contradict her, her ideas get resources because they are hers, and success becomes easier to attribute to her skill than to luck.

At that point, the useful residue of all that self-doubt is specific: your own wrongness still feels like a live possibility.

That is surprisingly valuable once everyone around you has reasons to tell you how right you are.

Midlife may make this more interesting

This became especially interesting to me when I realized how much I had changed by 43.

Women do not stay psychologically identical from 25 to 45. Longitudinal work finds that through midlife women on average become less dependent and self-critical and more confident and decisive. Self-esteem rises through much of adulthood, while body surveillance and body shame tend to decline.

There is no neat hormonal switch called “menopause makes you a better CEO.” Age, perimenopause, accumulated competence, financial independence, older children, changing reproductive incentives and plain exhaustion with caring what everybody thinks may all arrive in the same general period.

The combination is what interests me. Women may spend the first half of adulthood developing very sensitive social sensors: reading people, monitoring reactions, noticing what they missed. Then part of the need to obey those signals starts weakening.

You keep the sensors and lose some of the obedience.

That may be an unusually good leadership configuration: high sensitivity to other people with lower dependence on their approval.

And this is where the trajectories can diverge. For a man whose self-worth has become attached to winning, success can strengthen the old contingency. Winning proves worth, so more winning becomes necessary. More status produces more power, power disinhibits the motive, and weaker feedback makes the self-story easier to maintain.

For a woman, midlife can sometimes produce the opposite loop. Real competence creates evidence-based confidence. Lower approval dependence makes difficult decisions easier. Social sensing and self-questioning remain. Confidence grows from results rather than rank.

One loop can carry you farther from the territory while making you feel increasingly certain. The other can make you more confident while keeping reality in the room.

That pattern suddenly explains a lot of what I’ve seen, including in myself. I am intellectually much more confident now than I was at 25. Emotionally, there is still a piece of me that refuses to fully believe my own press release. I used to want to get rid of that mismatch. I don’t anymore.

Confidence is a terrible KPI

Calibration is much more useful.

In one study of more than 2,000 managers, people who overrated themselves relative to how others rated them were judged less effective than both accurate self-raters and people who underrated themselves. I would not build a religion around one study, but the asymmetry makes sense.

If I underestimate myself, I may lose an opportunity. If I overestimate myself while running a company, hundreds of people can end up executing my mistake.

So by all means, get rid of the anxiety and the part of impostor syndrome that makes you miserable. Keep the capacity to wonder whether the evidence supports you. You need enough confidence to act; becoming immune to correction is not the upgrade.

Responsibility starts with you

For me, this comes back to the simplest definition of leadership: responsibility for the outcome of the system.

When something seriously goes wrong at Welltory, my first question is usually: what did I do that caused or allowed this?

That does not mean every failure is personally my fault. It means every important failure contains information about the system I am responsible for. Which incentive did I create? What did I tolerate? Which signal did I miss? Why did the information reach me this late?

This is why concentrated authority without concentrated responsibility is intellectually dishonest. You cannot say, “I should control this company because I understand it better than anyone,” and then explain every repeated failure as “my VP fucked up.” Maybe the VP did fuck up. But if your system keeps generating the same kind of failure, that eventually becomes information about you.

The more power you keep, the more aggressively reality should be allowed to update your model.

Don’t ask for feedback. Build sensors.

Leadership advice loves the idea that if you create enough psychological safety, people will tell you the truth. I don’t trust this as a primary feedback system.

Once you have real power over someone, you have changed the measurement environment. If their salary, promotion, project or job depends on you, they cannot give you exactly the same signal they would give a peer. The distortion runs both ways: people stop telling you when you are wrong, but they also stop telling you when you are unusually good at something because praising the boss feels sycophantic or unnecessary.

So I don’t want my sanity to depend on everyone around me overcoming their incentives. I want instrumentation.

Where reality can be measured directly, measure it. Build alerts, analytics and monitoring that reach you before somebody has to make a political decision about whether to tell you. And keep a record of your own judgment.

I now use AI to automatically capture decisions and predictions I make in meetings, then compare them later with what actually happened. Six months later I don’t get to say, “I always knew this was risky.”, if the record says I gave it an 80% chance.

Over a few years, that becomes a personal calibration dataset. When I say 80%, am I right about 80% of the time? Which decisions am I systematically overconfident about? Where am I too pessimistic?

Then there is hiring. I want people around me who can disagree properly. A professional contrarian who turns every disagreement into a six-hour war of attrition is not a safeguard; that person is just exhausting. The valuable person can challenge my reasoning clearly before the decision and still commit to executing once the decision is made, even when I chose differently.

And I need people outside the company’s incentive structure too: a therapist, a good coach, another founder who does not work for me. Someone with whom my title has no exchange rate. It’s really hard. I only managed to have a therapist and a coach I talk to once a year. I've never been able to find a mentor or founder whose experience is relevant.

So I increasingly think the answer is to build systems that rely less on spontaneous honesty: log your bets before hindsight rewrites them, instrument the company, and hire people smart enough to tell you you’re wrong without trying to take your chair.

Find the mammoths. Then keep checking.

Founder control is not the problem. The dangerous founder is the one who found the mammoths once and concluded that this makes them the permanent universal authority on mammoths.

Knowing where to go can legitimately make you a leader. The job after that is to keep the group connected to reality, including when reality proves you wrong.

So if you still occasionally wonder whether you are an impostor, I would not rush to cure every last trace of that feeling.

Just keep checking where the mammoths actually are.

I always assumed I’d write a book one day. I also assumed it would be about startups, technology, building things, why people do what they do — the stuff I write about here. Instead, through a strange chain of events, my first book isn’t about the thing closest to me. It’s about the thing closest to the people who use our app. Apparently it’s easier to answer somebody else’s need than to sit down and do what you personally feel like doing. I’ve decided that isn’t the worst thing to learn about yourself. So here it is. I’ve published a book, free to read and download.

It’s the result of about a year of digging into a condition called post-exertional malaise — PEM.

PEM is under-recognised, hard to diagnose, and a lot less rare than the medical system has caught up with. And — this is my impression from looking at data, not a published statistic — the people it happens to keep being younger than I expect them to be.

I ran into it because it broke our algorithms. The numbers stopped behaving the way the model expected them to: activity that should have produced a good day produced a bad one, two days later. I remember the exact feeling — *what on earth is going on here* — and going off to find out. That was a year ago, when I first understood how serious this is.

I’m not a guru, and there’s no unique method in this book. I’m a nerd who spent a very long time staring at people’s data and an even longer time in the scientific literature, trying to understand what I was looking at. So what I wrote is a set of notes. I tried to put them in plain words, so that anyone else who wants to understand this thing has an easier time of it than I did.

It is a review of published research, not a report of research of my own. It isn’t medical advice. It’s written in my own capacity, not as anything from Welltory, and I’ve been careful to say so inside the book too. And I obviously didn’t do this alone. My friends and people who were interested in such a book read drafts and told me when a paragraph was wrong, or unclear, or said more than the evidence actually supports. Someone with real scientific training went through the claims and made me take numbers out — which is not a pleasant experience and is exactly what I needed. People living with this described what it is actually like, in detail, more than once, using energy they did not have much of to spend on explaining anything to anyone. And a great deal of what’s in the book rests on work done by patient researchers who were describing this condition carefully years before the formal literature caught up with them.

Thank you. Genuinely. Whatever is still wrong in there is mine.

So. First book. Very nervous. And I really hope it helps someone.

It’s open-source, completely free, available on GitHub.

GitHub: Understanding PEM: What the Research Says in 2026
— read online, or download the PDF or EPUB. Free, CC BY-NC-ND.

DOI: 10.5281/zenodo.22092728

Apparently, a company needs a steady supply of recent founder pictures—for LinkedIn, press requests, quotes, and all those occasions when someone needs proof that the CEO is a real person with a face.

The problem is that I hate being photographed. Not mildly. I really hate it.

I’m not an Instagram person. I don’t document my life, search for flattering angles, or suddenly think, “This moment would be better if someone pointed a camera at me.”

So every year, Veronika—my CMO—has to form a small conspiracy with my assistant, my friends, and occasionally the rest of my immediate environment. Their mission is to catch me during a brief window of good mood and convince me that a photoshoot is necessary for the company.

I approached this year’s session with the usual sense of impending doom.

And then something deeply inconvenient happened: it was easy.

Possibly even fun.

And I actually liked one of the photos.

There is no profound founder lesson here. Just one of those strange things nobody warns you about when you become a CEO: at some point, your face becomes company infrastructure—and you become a slightly unwilling corporate mascot.

So here it is: photographic evidence that the annual conspiracy succeeded.

Other founders: what ridiculous part of becoming “the face of the company” did you never realize you were signing up for?

For most of the last decade the thing we were genuinely best at was writing to you about your own body.

Not the app. Not the charts. The sentences. Taking a pile of heart rate variability and sleep data and whatever you did last Tuesday, and producing a paragraph that was scientifically defensible, specific to you, and actually worth reading. That was the moat. It cost years and an uncomfortable amount of money.

Then the models arrived and I watched it evaporate.

Not really. Our paragraphs are still better, and I can show you why. The problem is that nobody can tell. A user cannot distinguish a rigorous interpretation of their data from one that merely reads like one. Mostly, neither can an investor. And a moat nobody can perceive has stopped performing the function of a moat, whatever it’s doing in the source code.

So we knew fairly early that we’d have to change, and change substantially. What I want to write about is where that thinking ended up, because it started as a question about my company and turned into a question about everyone’s career, including yours.

What we didn’t do

Everyone I knew started building wrappers.

We didn’t, and I’ve written a whole essay about why, so I’ll keep this short: there’s no defensibility in it. You build the thing, you raise easy money from people who aren’t looking closely, and then you die. Done cleverly it’s a decent way to make quick money, and I want to be fair about that — some people are running that play deliberately and they know exactly what they’re doing. It just isn’t a company.

Then the mood shifted and everyone decided SaaS was finished. I half agree. What’s finished is SaaS whose advantage was the interface: the buttons, the arrangement of the screens, some simple automation between two APIs. That advantage has genuinely evaporated, and I don’t think it’s coming back.

But that isn’t all software, and the distinction matters more than the doom does.

Investors used to have a standard test: what happens if someone just hires a hundred engineers and builds the same thing? If the honest answer was “they’d have it in six months,” you didn’t have a business. What changed is that the same test now runs faster and costs less. That’s all. It’s the same question with a new price.

And the interesting part was always the cases where the answer was no. You can hire any number of engineers and you will not reproduce Telegram’s speed and reliability. That’s not a feature list. It’s a decade of specific hard decisions compounding, and it doesn’t come out of a model, because it never came out of typing.

Which gives you the rule, and it’s almost stupidly simple. In a world with AI, don’t do what AI does. Do what it doesn’t.

Fine. So what doesn’t it do?

The safest person I know is a barber

I’ve been turning this over for a year and I keep landing in the same place, which is not where I expected.

Barbershops. Hairdressers. They’re in less danger than almost anyone reading this.

I’m not being cute. Look at what a haircut actually is. Someone puts their hands on your head for forty minutes. There’s conversation, most of it not important, some of it more honest than the conversations you have with people who matter. There’s a person who remembers what you asked for last time and what you complained about after. There’s physical contact, which does measurable things to your nervous system whether you consented to that or not. You leave slightly different than you arrived, and not only because of the hair.

None of that downloads.

The same is true of a good coffee place — and I mean the small kind, where the barista knows you and knows what you like. Not the American high-throughput version with a queue out the door and your name misspelled on a cup. That one is a logistics operation and it will be automated. I mean the ten-table place where the relationship is the product and the coffee is the excuse.

Now hold that next to a mid-level programmer who receives well-specified tickets and closes them.

I know which of those two jobs I’d rather be holding in five years, and it isn’t the one that pays better today.

Three things nobody can take

In technology, once you set hardware aside, I think everything defensible sits in one of three places. I’ve made this argument before at company level, so here it is compressed:

Relationships between actual humans. Whatever you think about what people feel toward AI, a living person still does something to another living person that nothing else does.

Datasets that are hard to acquire and attached to something worth knowing. Hard to acquire is the operative half. Anything scrapable is already gone.

Knowledge that isn’t public and can’t be reverse-engineered out of your product by looking at it.

What you process, how you process it, and what you build with the people using it. That’s the whole list.

But here’s the thing I only noticed recently, and it’s the reason I’m writing this instead of another company post.

The list transfers to people almost unchanged.

Your career, as a company

Swap the words and it comes out as four things:

The breadth of your context and how much of your domain you actually hold in your head.

Your own view of that domain — not what the field thinks, what you think, and whether you’ve earned the right to think it.

Relationships worth investing in, which means the ones where someone would recommend you without being asked.

And whether you can get real leverage out of AI rather than pretending you don’t need to.

The mental model I’ve found most useful, and I’d suggest trying it even if you’re happily employed: think of yourself as a one-person consultancy with your name on the door.

What does it specialise in? What does it know that the others don’t? What has it actually shipped, and what went wrong, and what did it learn that isn’t in any book? What’s the house view — the thing it will argue for that not everyone agrees with? And the question that decides the whole thing: would a client hand you to their friend without being prompted?

If those questions have thin answers, that’s the work. Not a certification. Not another framework. Those.

And while we’re here: classical education has almost entirely lost the part of its purpose that consisted of loading information into you. Information is not an advantage when it exists in books and on the internet. What survives is trained reasoning — which is why mathematics is worth more now than the ability to discuss which decade produced which theory. One of those is a capability. The other was always just retrieval, and retrieval got cheap.

Context is the only ground you can hold

Here’s the part almost nobody wants to hear, and I say it as someone who has watched very good people refuse it.

A programmer who has spent years executing tasks handed down from above is going obsolete. Not because they’re bad at the job. Because the job was the narrow bit, and the narrow bit is what got automated first.

To stay relevant they’ll have to go wide. Architecture, or product, or analysis, or something — but they have to grow a band of territory that lets them understand a wider context than the ticket they were given. Context is the only ground where we can confidently compete, because it’s the one thing the model doesn’t have. It has everything ever written. It doesn’t have your customer’s political situation, or the reason the last attempt failed, or which of your colleagues will quietly kill the project.

But going wide means leaving the thing you’re good at and being mediocre at something new, in public, at forty. Which is genuinely hard, and I understand why people don’t, and I think the ones who don’t are making a serious mistake anyway.

There’s one more thing on the human side of the line, and it took me a while to see it.

Accountability. AI cannot be responsible for an outcome. It will produce something confidently wrong, and when it goes badly it says the equivalent of oops — and there is nobody there. Nobody whose reputation moves. Nobody who has to sit in the room afterwards. Nobody with anything at stake.

Which means responsibility for the actual result stays with humans, and if you want to know where the value concentrated, it concentrated there. Not in producing the work. In standing behind it.

The arithmetic moved

There’s no longer a choice about whether to adapt to this. There’s adapting, and there’s being beside the point. I’d love to offer a third option and I don’t have one.

But I’ll finish with the thing that genuinely surprised me while I was working this out.

If you have ever quietly wanted to tend bar, or cut hair, or open a bakery — I’m not about to tell you to follow your dream. I don’t do that. I’m telling you the arithmetic changed underneath you while you weren’t looking, and that work is now more defensible than a clerk’s, and more defensible than a great deal of programming.

That isn’t a consolation prize. It’s a repricing, and I don’t think most people have noticed it yet.

Interesting times, and I mean that in every available sense.

I’m forty. I have cured none of it. What I’ve learned instead is how to live with it, which is a much harder thing to put on a slide.

The hard part was never attention. It’s executive dysfunction. I can run a company, hold a product strategy in my head, take apart an ugly technical problem, write twenty pages in a day.

And opening WhatsApp can be genuinely frightening.

Booking a dentist appointment is a three-act drama with its own emotional arc. My morning is the only time I can open my email. Thank god I didn’t start a b2b SaaS with a manual sales cycle.

I’m also the most productive I have ever been in my life. Not “for someone with ADHD.” The most productive I have ever been, and my team will say the same. When I sat down to work out why, the answer turned out to be really interesting. It was never about fixing the machine. It was about changing the inputs.

This is for anyone who has this brain and is tired of losing to it, and for the managers who don’t realise what they’re sitting on.

The deal

Here’s the trade, stated flat, because everything below is just the reasoning behind it.

Far more output, and nobody gets to choose the task or the date. Those two things arrive together.

They aren’t a preference and they aren’t a request. It’s how the machine works, and the terms don’t move because I’d like them to or because you would.

Take the trade and you get the problems nobody else will crack. The thing that’s been rotting on the roadmap for two years, finished on a Sunday. The strategic contradiction everyone has been carefully walking around for a year and a half.

Refuse it, insist on this ticket by Thursday, and you get a mediocre unreliable employee who hates themselves. You’ll conclude the employee was the problem. It was the system design.

Almost every company refuses. Not out of cruelty, out of a decision they made so long ago they’ve forgotten it was a decision.

And it is a trade rather than a petition, which means either side is free to walk. Nobody is getting anything out of me. A company that wants the deadline more than the output is welcome to go and find someone who sells deadlines, and I’m welcome to decline the job.

Which matters most if you’re the one with this brain rather than the one managing it, so let me put it as plainly as I can.

If you’re inside a company that runs on tight control and hard dates, you are not going to do well there. Not “it will be difficult.” You won’t. So all the caution you’re currently exercising — not mentioning it, agreeing to the Thursday date with a confidence you don’t have, quietly hoping this quarter goes differently — is buying you nothing at all. It’s protecting an outcome that was never on the table.

Which makes the honest conversation with your manager close to free. The worst case is you find out early that the fit is wrong, and finding that out early isn’t a cost. Have it at the start rather than in the post-mortem of the thing you missed, and you at least create the possibility that the two of you end up pleased with each other. And if you don’t have a manager, because you’re freelance or founding or between things, the same trade is available with yourself. That’s the version I run most days.

That possibility is real, by the way, because I run the other side of it. I have people who get more done in four days than they used to in five, and I’m entirely fine with it, because I can see the output. We have enough neurodivergent people that at some point we stopped designing for the other case. We don’t count hours. We rarely set deadlines. Paid time off isn’t capped. We plan around what’s actually burning for someone this week rather than what a document said in January.

It works. Not “it works, with tradeoffs, and here’s our learning journey.” It works.

The price of attention

Everyone who’s been near this knows the wall. The thing that should take forty minutes, that you are fully capable of doing, that you cannot start.

Let’s put down the word “dysfunction” for a minute. Not because it’s unkind. It’s a bad model: it says the thing is broken, which says fix it, which is how you spend a decade and a great deal of money trying to become somebody else.

Try assuming the wall is doing a job. Then the question stops being how do I force myself and becomes what is this telling me?

I can push through almost anything on raw willpower. What varies wildly is the cost. Some tasks run a little discipline. Others are black holes that take every unit of deliberate attention I own and hand back nothing, and I mean the expensive kind of attention, the kind that doesn’t happen by itself. Kahneman’s System 2, if you want a label.

I personally feel like I will literally die if I have to do one of those. It feels like a black hole taking the life force out of me. I know how that sounds.

So the wall isn’t a warning light saying empty. It’s a quote. My brain has priced the task against every other thing that attention could be doing and declined at that price.

Which is irritating, and also the single most useful thing I ever worked out about myself. You can’t argue with a price using willpower. Nobody stands in a shop shouting at a label. You argue with a price by changing what’s on offer.

Then there’s the other end of the same system. Hyperfocus, where a project stops consuming attention and starts generating it. You don’t fight to concentrate, you fight to stop. You forget that time is a thing. You work for nine hours, solve something that looked impossible on Tuesday, and produce what would be a decent week for a competent adult.

Same brain. Same person. Still hasn’t answered the message from four hours ago.

This normally gets filed as a contradiction. Dysfunction is the disability, hyperfocus is another symptom of the disability, both need managing. I think that’s exactly backwards. These aren’t separate symptoms. They’re two ends of one system, the system works fine, and there’s nothing unusual about it.

It just has an enormous range. And systems like that exist all over the place and work beautifully, which I’ll show you in a minute.

You don’t live in Mediocristan

Taleb has a pair of imaginary countries I think about constantly.

Mediocristan is where no single observation moves the total. Take the height of a thousand people, add the tallest human alive, the average barely twitches. Variation exists and stays contained. Physics won’t permit a person forty times taller than the mean.

Extremistan is where one observation eats everything. Take the wealth of a thousand ordinary people and add one billionaire. The average is now a number describing nobody in the room, including him. The outlier isn’t noise around the system. The outlier is the system.

(A disclaimer, because the words are unfortunate: Mediocristan isn’t a comment on anybody’s abilities. It’s Taleb’s term for a shape of distribution, and I’m borrowing it for that and nothing else.)

A neurotypical person has preferences. Output fluctuates, there are good days and bad ones, the boring task costs more than the interesting one. But the range stays fairly stable. Most days produce something, most tasks get done at a tolerable price, and if you watch someone for three days you have a fair idea of their month. That’s Mediocristan, and it’s why sprints work, why tickets work, why the whole apparatus of modern work functions at all.

My range is not like that. I can produce nothing visible on the assigned project for a week and then generate something extraordinary somewhere else in an afternoon. A handful of extreme periods creates most of my annual value.

Watch me for three days and I look chaotic, unreliable, possibly lazy. Watch me for a year and I’ve produced more than almost anyone around me, with the exception of the other neurodivergent people on my team.

The mistake is treating an Extremistan system as a defective Mediocristan one. It isn’t the same machine performing badly, it’s a different machine on different rules. And when an organisation sets out to smooth the spikes, it usually succeeds, which is the problem, because the spikes were the reason they hired the person.

Which brings me to the part that should make a few of you uncomfortable.

Venture capitalists are the only profession on earth professionally trained to think in Extremistan. Power law. Outliers. One in twenty returns the fund. You have organised your entire economic existence around the proposition that a fat tail beats a tidy average.

Then you walk into a board meeting and revert to Mediocristan the instant somebody opens a headcount spreadsheet.

Genius on Tuesday, idiot on Thursday

Modern work is a conveyor belt. A task arrives, a person receives it, a date gets attached, the person finishes and takes the next one. The design assumes tasks are roughly interchangeable, that preference is a minor variable rather than a hard constraint, and that effort converts almost any input into output.

That’s a good design. I want to be clear about that, because I’m not arguing the conveyor is stupid. It’s excellent for making cans of soup, and predictability is what you can actually sell upward, to a board, to a market, to an acquirer. The conveyor manufactures predictability beautifully.

It’s a catastrophic fit for a wide-range mind.

In that system I look brilliant when the belt happens to deliver something matching my current configuration, and like an incompetent fraud when it doesn’t.

Genius on Tuesday. Idiot on Thursday.

And everyone concludes the person is inconsistent, which, fine. But notice the company is also being inconsistent. It’s feeding random inputs into a highly specialised machine and acting surprised when the engine smokes.

So let’s reverse the question. Instead of asking how someone like me adapts to a world built for the narrow range: how would work be designed if this were treated as a normal mind?

We don’t have to imagine it. Somebody built it already, by accident, and it’s worth billions.

It’s the room, not the kid

It’s professional sport.

In 2021 a Swedish group at Karlstad University did something almost nobody thinks to do. They took high-level youth athletes and had a trained psychotherapist score them on a standard adult ADHD symptom scale. Twice. Once for how they function at school, once for how they function in their sport.

Same kids. Same instrument. Same week.

High scores at school. Low scores in sport.

Sit with that, because it’s the whole argument compressed into one experiment. The symptoms did not travel with the child. They appeared in one room and receded in the other. Whatever the school assessment measured, it wasn’t a property of the kid. It was a property of the kid plus the room.

And this isn’t “he tries harder at the fun thing.” These are clinically shaped scores on a validated instrument, moving because somebody changed the context.

Look at what sport supplies and the reason stops being mysterious. Immediate feedback. A visible score. Real stakes. Intense, varied stimulation. Consequences that land in seconds instead of at the end of the quarter. Sport is structurally a place where an attention system that discounts the future steeply and runs on interest is finally at home, and nobody redesigned it to be that way out of kindness. It was already like that, because that’s what makes sport worth watching.

The prevalence numbers point the same way, modestly. A 2019 review in the British Journal of Sports Medicine puts ADHD at roughly 7–8% among student and elite athletes against something like 3–7% in the general population. Overlapping ranges, and it’s a narrative review rather than a meta-analysis, so don’t let anybody tell you it’s three times the base rate. Somewhat over-represented is the honest reading.

The visible end is less ambiguous. Michael Phelps, Simone Biles and Noah Lyles have all talked openly about ADHD. Alysa Liu, who took the 2026 Olympic figure skating gold, was diagnosed in high school and puts it about as plainly as it can be put: she loves situations she isn’t expecting.

Notice what did not happen to any of them. Nobody tried to cure them into steadier, more predictable athletes. They were just in the right room.

Now the part I have to say honestly, because it makes the point stronger rather than weaker.

Sport is not enlightened. A 2023 scoping review of neurodiversity in elite sport turned up 23 studies, and most were about ADHD as a risk factor for concussion, or about managing medication so athletes stay compliant with anti-doping rules. A genuinely supportive environment still gets described in that literature as future work. The lead author of the 2019 review says plainly that elite athletes often can’t get proper care because of stigma. Nobody in a locker room sat down and decided to build a strengths-based model of human variation.

That’s exactly why it’s useful evidence.

Sport didn’t get kind. It got shaped right, and the trait stopped presenting as a deficit anyway, and an entire industry monetised the result without ever needing a theory about it. Which is a far more actionable conclusion than “find a nicer employer.” The same trait, in a different room, produces a measurably different person.

Your job is to pick the room.

The answer isn’t one project

A hyperfocus-driven person shouldn’t have one active project. They should have seven. Or eleven, or seventeen. The number is irrelevant, the diversity is the whole thing.

And it isn’t a queue, which is a word I used for years before noticing it was wrong. A queue has an order and you work through it, and the order is exactly the part that doesn’t function. What you want is a fan. Everything laid out at once, all of it live, none of it next.

Which is a shape you already know if you invest for a living. Nobody works through a portfolio in sequence. You hold twenty positions simultaneously precisely because you cannot tell in advance which one pays, and the whole discipline consists of having enough good ones open at the same time. Run your attention the same way and the logic is identical, including the part where most of them go nowhere.

I wake up, look at what’s laid out, and notice which one is pulling. This is rarely subtle. Hyperfocus tasks have gravity, you can feel the project waiting for you.

Standard management calls choosing work this way unfocused, and adults are supposed to do what they planned rather than what they happen to find interesting on a Tuesday. But that objection assumes the goal is maximum predictability per task. Give me several important problems and I spend far more hours in the state where work costs nothing and produces disproportionately. Any individual project gets harder to forecast. Total production goes up, and not by a little.

There’s the trade again. Less predictability per task, much more output per system.

Most people would commit crimes for the ability to drop into deep flow by picking the right project. People with ADHD frequently have exactly that ability and spend their lives apologising for it.

Hyperfocus amplifies whatever you feed it

There’s a fair objection here and I want to give it properly, because the people making it usually have this and are tired of being told it’s a gift.

Hyperfocus is a mechanism, not a moral compass. It lands on a Wikipedia rabbit hole about ball bearings as happily as on your architecture problem. Six hours reorganising a database nobody needs. Perfecting the typography of an internal document four people will read. Anyone who’s done that does not experience it as a superpower.

But the uselessness of the output doesn’t prove the mechanism is useless. It proves you fed it garbage.

Hyperfocus amplifies whatever it lands on. If your environment offers an endless supply of cheap stimulation and no accessible, compelling, meaningful work, it will faithfully and powerfully amplify Wikipedia at three in the morning. The mechanism performed perfectly on what it was given. Nobody looks at a model trained on junk and concludes the model is broken, they look at the dataset. Then they look at you and blame you.

So the real work happens before the work. Keep an inventory of diverse, valuable, accessible problems. Not a backlog of administrative debris, not a graveyard of obligations copy-pasted out of quarterly planning. An actual inventory:

• a gnarly product architecture

• a research question

• a half-written essay

• an unusual negotiation

• a strategic contradiction nobody has named yet

• a customer problem no one understands

• a high-risk experiment with an enormous upside

You can’t schedule the lightning. But you can decide where to put the rods.

What it looks like when it works

There was a thing on our list for years. Useful, structurally important, complex enough that it never quite cleared the priority bar. You know the type. Too big to squeeze in, never urgent enough to schedule, so it sits there getting older and quietly making everybody feel slightly bad.

Then one Friday my hyperfocus walked into it. No plan, no decision, just this rising oh, hang on, this can actually be done, you do it like this and then like this.

I did it over the weekend. A properly valuable, entirely non-urgent piece of our strategy, finished on Monday after sitting unfinished for years.

Now, before anybody draws the wrong conclusion about the weekend.

I subordinate my calendar to my hyperfocus rather than the other way round, which means there are very quiet days in the middle of my week. On purpose. Days kept clear so there’s room to switch and room to rest, because you can’t run the expensive end of a factory without slack around it.

If anything in my life is predictable it’s the walks, the sport, and the hours with no hyperfocus in them at all. Those are the fixed points. Everything else is weather.

So when I work on a Saturday it’s not because I bought the fairy tale, the five hours of sleep and the pizza and the grinding until 4am and the whole tedious mythology. That’s bullshit, and I think it has quietly wrecked people who’d otherwise still be building things. You’re supposed to build a schedule that accounts for how you actually work, and to understand that this is a marathon.

I just refuse to run the marathon on somebody else’s calendar.

Goldratt: not every machine gets managed the same way

The Goal, Eliyahu Goldratt, 1984. A plant manager has three months before his factory gets shut down, and what he works out seems obvious the second you see it and remains almost impossible for organisations to accept.

Not every resource in a system should be managed the same way.

A machine that isn’t the constraint can sit idle without costing the factory anything. The bottleneck is different: every hour lost there is an hour lost by the whole system, permanently. And a complex expensive machine often needs substantial setup time to switch between kinds of work, so if you reconfigure it for every incoming order you destroy its useful capacity while feeling extremely responsive.

The answer isn’t to shout at the machine. You organise the system around it. Protect it from interruption. Keep a buffer of work in front of it so it never starves. Batch similar jobs while it’s already configured. Don’t burn its scarce capacity on things other resources could do.

People aren’t machines. The management principle transfers anyway.

Some people are reliable general-purpose processors. Hand them a well-defined sequence and they move through it steadily, and thank god for them, because they’re why anything ships at all. Others are rare, expensive, long-setup machines with extraordinary output in the right configuration. Managing both identically isn’t fairness, it’s laziness.

This turns a vague complaint about interruptions into something specific. When a hyperfocus-driven person has finally reached a productive configuration, pulling them out for a status update, a context switch, a Slack thread, a routine meeting or an expense report is exactly equivalent to stopping the bottleneck halfway through a valuable batch because another order came in.

It makes the local plan look tidy. It wrecks total throughput.

Why companies fight this even when it works

So if the deal produces more value, why does almost nobody take it?

Donella Meadows gives you most of the answer, and it isn’t malice. It’s delay.

Complex systems separate action from consequence in time. Something happens now, its real effects surface later, and humans are famously terrible at this. We respond to what’s visible, usually well before the system has finished telling us what actually happened.

In the short run a hyperfocus-driven worker looks alarming. They move between projects, some tasks sit longer than forecast, progress is lumpy, daily output is unpredictable, the board is untidy.

So a manager intervenes, and every single intervention is reasonable on its own terms. Cut the number of projects. Forbid switching. Tighten the dates. Add check-ins. Require visible daily progress. Finish the current thing before touching anything else.

The immediate result is genuinely reassuring. The person is sitting in front of the correct task, the board is clean, control has been restored.

The real result shows up a quarter later. Less time in deep work, fewer breakthroughs, the hard problems still unsolved. More cognitive budget going to self-coercion, which raises anxiety, which worsens avoidance. Eventually output falls, or the person burns out, or quietly leaves.

But the feedback arrived too late to be connected to its cause. So the organisation concludes the person needed more control.

Which gives you one cheap unglamorous fix. Widen the measurement window. Stop asking only whether the assigned ticket landed by Thursday and start asking what valuable output this person created over the quarter.

Predictability is easy to observe and value takes longer to show up, and organisations trade the second for the appearance of the first roughly every day of the week.

This is not a permission slip

At which point somebody reasonably says: so people with ADHD should do whatever they feel like, ignore deadlines, and let everyone else absorb the consequences?

No. Freedom without design isn’t an operating model, it’s chaos with better vocabulary.

The contract changes on both sides, and I had to learn the second half the hard way. Neither list below is a wish. The first one is roughly what we actually run.

What the organisation provides:

• several genuinely valuable problems to choose between

• flexibility in sequencing

• fewer pointless context switches

• long uninterrupted blocks

• roles where the outcome matters more than the procedure

• help or delegation for the tasks that cause disproportionate paralysis

• a longer measurement horizon

What I provide:

• honest communication about uncertainty, up front, not after the miss

• responsibility for the value of what I choose, not just the pleasure of choosing

• visibility into what’s moving and what’s stalled

• safeguards so no critical deadline depends only on me

• actual delegation of the admin debt I know I can’t pay

• a deliberately curated set of options rather than novelty-chasing in a systems-thinking costume

That last item on my side is the one people miss. Executive dysfunction is not a licence to only ever do the fun part. Somebody still has to book the doctor’s appointment, and in my case that somebody is a paid professional who tells me about it over messenger. The reason that’s legitimate rather than pathetic is that I stopped pretending I’d fix it and paid for a solution instead.

The same logic runs through everything that works for me. Sitting next to another person while I do the thing I can’t face alone, which has a name, body doubling, and which the community built long before the research turned up. Nothing about the messages changes. The terror does. Doing the horrible administrative things first thing in the morning, before the brain fills up with noise and shame and forty-seven unfinished mental tabs.

None of it is sophisticated. It’s just that nobody tells you it’s allowed.

So this isn’t the elimination of responsibility, it’s a transfer of it. You stop pretending that trying harder is a productivity system, and you take on responsibility for designing an environment where the brain you actually have does useful work.

Non-linear people for non-linear problems

Some jobs are predictable execution, and that is the product. If you’re in the critical path for a synchronised daily delivery and you can’t reliably show up, you’re in the wrong job. That’s a routing error, not a moral failure. Get out of the path.

But most of the work that matters has no guaranteed path and no honest deadline anyway. Research. Invention. Strategy. Writing. Early product. Complicated negotiation. Crisis. Finding the hidden constraint. The non-obvious solution. Anything with a high chance of failure and an enormous upside.

Those are Extremistan problems already, and they’re the natural habitat of a wide-range mind. “You can’t promise when” stops being a defect there, because nobody could promise when.

Which is a large part of why so many of us start companies. A founder always has the full spectrum available: product, hiring, technology, story, fundraising, research, crisis, culture, sales, strategy. There’s nearly always something genuinely important your brain is ready to attack today. Entrepreneurship is brutal in a dozen other ways and hands you exactly one luxury, which happens to be the one you need. Enough meaningful variety to keep the fan wide.

My colleagues know I won’t hit every individual date. They also know what the system produces when it’s set up properly. When I first said out loud that I wouldn’t be promising every deadline, people looked at me like I’d lost my mind.

Then they saw the other half of the deal.

The guilt tax

One more piece of structure, and it’s the expensive one.

I can’t start the thing, so I feel guilty. And the guilt isn’t free, it runs on the same budget as the work. Deliberate, effortful, self-directed processing, which is precisely what the task needed. So I have less. So I do less. So I feel worse. It’s the same reinforcing loop as the control spiral, running inside one skull instead of an org chart.

Guilt isn’t a motivator. It’s an expense, billed to the same account as the work.

Which is the only reason I get to say the next thing without it being a wellness platitude. Sometimes the network needs to run in passive mode. Go to the park, look at birds, do nothing on purpose, and don’t turn the rest into either a moral failure or another optimisation project. Guilt-flavoured rest is the worst of both: you pay the tax and get no services. Rest at full strength and it gives something back.

You have fewer options than they do

Most people can work almost anywhere. Take a job, tolerate it, treat life as the part that starts at six. That works for them, because the range is narrow and every job costs roughly the same to endure.

You don’t get that option. Not because you’re finer or deeper or more interesting, but because of the variance. The wrong job doesn’t cost you slightly more, it costs you an order of magnitude more, and it takes the evenings with it. Fit isn’t a preference for you, it’s structural.

Which is both a curse and a blessing, and which one it turns out to be is on you. That isn’t a warm thought. It’s a demand, and I’d rather hand you a demand than a consolation.

Now, the objection you’ve been holding since about the third section.

Easy for her. She owns the company. Of course she can announce she doesn’t guarantee deadlines, she’s the one they’re guaranteed to.

Yes. I know how much of this rests on the position I’m in, and I won’t pretend autonomy is a personality trait rather than a privilege. Though if you manage people, autonomy isn’t a perk you’re generously extending. It’s the operating condition for this particular resource. You provide it and get the output, or you don’t and you don’t. That’s a throughput question rather than an HR one, and it’s your job, not your employee’s.

But also.

I’m the daughter of a single mother, from a town small enough that it didn’t have a university in it. When I started studying we had no money. Not careful with money. None.

So I’ve been on the other side of this, with no leverage and no title and nobody’s permission, and what I did then was the same thing I’m describing now, just cruder and with worse tools. I leaned hard on the parts of me that worked absurdly well and I stopped negotiating with the parts that didn’t. I never fixed the email problem. I found work where the email problem was survivable and the other thing was worth a great deal.

Years later that still looks like the right call. Not a brave one, not an inspiring one. Correct.

And one more thing, since it took me two decades to work out.

Stop taking productivity advice from people who don’t have this. They’re not wrong, they’re right for themselves. Nearly everything ever written about focus and discipline and consistency was written by people in the narrow range for people in the narrow range, and inside that range it’s excellent. Two pages a day even if they’re terrible. Always commit to a deadline. Eat the frog. Hand that to a Mediocristan machine and it works.

Run it on yourself and you’ll conclude you’re a failure. On schedule. Roughly every three weeks, for years, which is what most of us actually do with our twenties.

Your system for managing yourself has to be different, because you are different. Not braver, not deeper, not more interesting. Different in a way somebody has already measured, the same way a Swedish research group measured one child twice in a single week and got two different answers, and the difference wasn’t in the child.

There’s nothing here to forgive and nothing to fix. There’s a system to design, and you’re the only one who can design it, because you’re the only one holding the data.

Press on the strengths. Ignore the weaknesses. Do it earlier than feels responsible.

Your hyperfocus is fine.

Go and look at what you’re feeding it.

If you want to go further

Two books and an essay carry most of the thinking here. Everything else is optional.

• Donella Meadows, “Leverage Points: Places to Intervene in a System,” 1999 — and Thinking in Systems if it takes. Delays, feedback loops, and why interventions that feel decisive usually make things worse. Twelve pages, free, and it will permanently ruin your patience for most management advice. → donellameadows.org

• Eliyahu Goldratt, The Goal, 1984. A novel about a struggling factory. Bottleneck, setup time, buffer, all of it comes from here. Somehow also the best book about attention I’ve ever read.

• Nassim Taleb, The Black Swan, 2007, for Mediocristan and Extremistan. An argument rather than evidence, and I’m using it as vocabulary, which is what it’s for.

And if you want the specific claims checked:

• The sport evidence, which is the best natural experiment we have. The same young athletes, scored twice in one week on a standard ADHD symptom scale, high for school and low for their sport: Ekman, Hiltunen & Gustafsson, “Do Athletes Have More of a Cognitive Profile with ADHD Criteria than Non-Athletes?”, Sports, 2021. Open access and short. → doi.org/10.3390/sports9050061 · free full text

• Prevalence among elite and student athletes, roughly 7–8% against about 3–7% in the general population. Note that it’s a narrative review rather than a meta-analysis, so treat it as somewhat over-represented and not proof of anything dramatic: Han et al., British Journal of Sports Medicine, 2019 → doi.org/10.1136/bjsports-2019-100713

• And the counterweight, because I’d rather you heard it from me: Hoare et al., “Neurodiversity in elite sport: a systematic scoping review,” BMJ Open Sport & Exercise Medicine, 2023. Twenty-three studies, most about concussion risk and anti-doping compliance. Sport hasn’t solved neurodiversity, it merely happens to be shaped correctly, which is the actual point. → doi.org/10.1136/bmjsem-2023-001575

• Effort as a price rather than a fuel tank: Kurzban, Duckworth, Kable & Myers, Behavioral and Brain Sciences, 2013 → free version

• Why a deadline a month out has almost no present value in this brain: Jackson & MacKillop, meta-analysis on delay discounting in ADHD, 2016 → doi.org/10.1016/j.bpsc.2016.01.007

• Why more things in the queue means each one waits longer, which is Little’s Law in a book about product development rather than factories: Reinertsen, The Principles of Product Development Flow, 2009 → chapter 1, free

• And in fairness, “hyperfocus” is a much shakier scientific construct than I’ve made it sound. The review that says so is open access: Ashinoff & Abu-Akel, Psychological Research, 2021 → PMC7851038. I’m using the word anyway, because everyone who has it recognises the description. Now you know what you’re getting.

Here’s the thing about me, though: I’m not the nod-along type. I say the quiet part out loud, right there at the table. Which ones, exactly? Name them. And the part that gets me every single time is how fast the answers dissolve the second you press on them. One company’s “growth” turns out to be the money it raised and nothing else. Another has gorgeous top-line numbers and loses money on every unit it sells. A third had a crowd show up, poke around for a month, get bored, and wander off — a toy, not a habit. I’m not doing this to win the room. I actually want one of them to hold up, because I’m building in this world too. They mostly don’t. So no, this isn’t me nodding politely while thinking my private thoughts. This is me having asked the question maybe a hundred times by now and slowly noticing that the emperor’s runway is showing.

Let me get one thing out of the way before I sound like a hater. As a user, I am obsessed with this technology. Every founder secretly wants to clone themselves five times just to keep up with their own ideas, and AI is the closest I’ve gotten to actually doing it. Pair it with my ADHD and it’s a straight-up superpower — I’m writing algorithms, sketching product, iterating at a speed that was flat-out impossible a couple of years ago. This is not skepticism. This is infatuation.

But being infatuated and being right are two different jobs, and the second one is mine. People love the phrase “think with a cold head,” as if clear strategic judgment requires you to be some dead-eyed operator running numbers in a dark room. It’s a dumb myth, and — let’s be honest — a very male one. I run a health company built on empathy. Empathy is the product. And I can still see straight through a bubble. Warmth and sharp logic are not opposites, they never were. The job was never to be cold. The job is to not let a room full of excited people (me included) talk me out of arithmetic.

So. The arithmetic.

The two poster children

Start with ChatGPT, because it’s the one everyone points at first. The leaked, FT-verified 2025 numbers: about $13.07 billion in revenue, up from $3.7 billion. One of the fastest revenue ramps in the history of anything. And an operating loss of roughly $20.92 billion. (You’ll see a scarier $38.5 billion “net loss” floating around — ignore most of it. Around $41.5 billion of that is a one-time non-cash charge from the nonprofit-to-for-profit conversion, and I’m not here to inflate the drama; the real operating hole is plenty.) Bottom line, it cost them about $1.60 to make $1. So what did ChatGPT actually prove? That if you sell intelligence below cost, at planetary scale, people will happily consume an obscene amount of it. That’s a proof of demand. It is not a proof of a business. Usage is not a business model, and anyone who ran a startup in 2015 and tried to pay rent with “we have users” already knows how that sentence ends.

Then there’s Cursor, which is the one that actually keeps me up at night (that, and the monkeys on the roof, but that’s a different post). Cursor — Anysphere — became, by most measures, the fastest-scaling B2B software company that has ever existed: roughly $100 million to about $4 billion in annualized revenue in something like eighteen months. Faster than Slack. Faster than Zoom. Faster than Snowflake. The AI-native dream, in the flesh.

Now here’s what that dream actually bought them. For most of that run they were operating on negative gross margins. They were paying Anthropic retail API prices for the exact model their product ran on — while Anthropic was out there running its own competing product, Claude Code, at cost. Sit with that for a second. You are a reseller paying full sticker price for the one input your biggest competitor gets at wholesale. No amount of top-line growth fixes that. It just makes the leak wider and faster.

So how did the fastest-growing B2B company in history end its own independence? In June 2026 it sold to SpaceX for $60 billion. And — this is the part that matters — not for cash. All stock. Paper in an unprofitable company (SpaceX’s AI arm reportedly lost around $6 billion in 2025), locked up, with the founders unlikely to touch real liquidity before late 2027, at a number that floats around with the share price.

Credit where it’s due, because I’m not a ghoul about this: the founders and early investors might still make an absolute fortune. SpaceX stock could rip, the lock-ups could vest into serious money, and good for them if they do. But I’m going to keep pulling on the distinction everyone wants to blur, because it’s the whole game: a spectacular venture adventure is not the same thing as a durable business. Cursor is a phenomenal trade. It is not proof of a defensible company. It grew faster than anyone in history and still couldn’t stand on its own two feet or turn that success into cash on its own terms — it had to fold into someone bigger who owned the compute. If that’s the flagship of “AI-native wins,” what it really shows is that being the best in the world at building on top of someone else’s model is a great way to get bought and a lousy way to stay free.

And look — if the absolute flagship of dev tools, where users are practically begging to pay, couldn’t stand on its own, what does that say for the rest of us? Take my own corner: health and wellness. It is currently drowning in decks for “AI-native health coaches,” “autonomous diagnostics,” “generative wellness assistants.” So I’ll say the quiet part one more time: I do not know of a single one of those — the chat-first, coach-in-a-box, generative-assistant kind — that has built a real, sustainable, scaled business. Not one. (And before anyone @s me: the companies actually winning in health don’t call themselves “AI-native” anything. They run AI underneath a real care business. Hold that thought, we’ll get there.) The AI-native ones have beautiful Figma files, temporary user spikes, and the balance sheet of a ghost ship.

“But surely the one real thing it does is layoffs”

Fine. The celebrated startups don’t have durable economics. So the fair question comes back at me: then where has AI made real commercial money? What are companies actually, repeatedly paying for?

For a while my honest answer was one word: layoffs. That looked like the one thing AI had truly proven — not a shiny new capability, but permission to cut headcount. And the numbers dress the part. Challenger, Gray & Christmas tracks the reasons companies give for job cuts, and AI went from about 5% of layoffs in 2025 to climbing all through 2026 until, by spring, it was the single most-cited reason for layoffs in America. Salesforce quietly took its support team from 9,000 to 5,000 and said, more or less, that it needed fewer heads.

And then I actually dug into the data, and even that fell apart. Which, I’ll admit, delighted me.

Because look what happens when someone checks. Gartner surveyed 350 companies with over a billion in revenue — real deployers, not people kicking tires. Yes, 80% had cut staff, some by as much as a fifth. And there was zero correlation between the cuts and ROI. The ones who cut the most made about the same returns as the ones who cut the least, and in a bunch of cases the ones who cut less did better. Gartner’s own analyst said it flat out: workforce reductions can create budget room, but they don’t create return. Cutting people is simply not where the value lives.

And honestly, a lot of the “AI did it” is just theater. Nearly six in ten companies admit they’ve dressed up plain financial layoffs as “AI-driven.” Deutsche Bank has a name for it — “AI redundancy whitewashing” — and called it a defining feature of 2026. Even Sam Altman has copped to the “AI washing.” Challenger’s own chief revenue officer said it best: whether or not a role got replaced by AI, the budget for that role got taken by AI. So “AI” is doing double duty — part cost-cutting tool, part press release, and if I’m being real, mostly the second one. It makes a layoff sound like a bold strategy instead of what it usually is: you over-hired in the pandemic and you’d like to look modern about fixing it.

And where a company genuinely bet the business on replacing humans, it tends to blow up in slow motion. Klarna bragged that an AI agent was doing the work of 700 support people, watched service quality slide, and started hiring humans back. IBM automated big chunks of HR and quietly reversed when the system choked on anything requiring judgment. Commonwealth Bank in Australia cut 45 service roles for a voice bot, drowned in call volume, and publicly called it a mistake and apologized. Gartner expects half the companies that blamed AI for cuts to rehire — under fresh job titles — by 2027.

All of which rhymes with the MIT study everyone forwarded around last year: 95% of enterprise GenAI deployments showing no measurable return. So the one thing AI supposedly “proved” it could do turns out to be the thing it’s worst at proving. What Gartner did find is that the companies pulling real returns use AI as what they call “people amplification” — making their humans more powerful instead of deleting them. Hold onto that phrase too. It comes back at the end.

AI isn’t a moat. It’s a moat destroyer.

Here’s the engine under all of this, and it’s exactly backwards from what the “go AI-native” crowd believes. Every technology wave kills something. What this one kills is the cost of producing things — software, content, analysis, generic expertise. And the very first things it fully commoditized were text and code, which happen to be the raw material most SaaS and most small apps are made of.

Think about what that does. If the marginal cost of making the thing your product makes is racing to zero, and a foundation model will spit it out on demand for anybody who asks, then the thin SaaS layer and the one-trick app have no floor left to stand on. They’ll die. Not because they’re bad — because their value became free. What survives is never “the app.” It’s whatever is hard to get behind it: proprietary data you had to generate one interaction at a time, network density, regulatory clearance, trust, the accumulated history of a specific human being. And the thing all of those share is that they took real, elapsed time — time you cannot parallelize. A competitor with infinite compute still can’t buy ten years of my users’ history. That’s the whole game, right there. So when a founder tells me “our AI is the moat,” I mostly just hear quiet, because if your magic is rented from a model provider who can ship the same feature next Tuesday (hi, Cursor), the magic was never yours to begin with.

Technology hands you the timing. It never hands you the vision.

Now the optimists will say: but Jane, this is a platform shift, like the internet, like mobile, like cloud. And they’re right about that. It’s the lesson they draw from it that’s wrong.

Go back and actually look at who won the internet. Not one of those companies had a vision about the internet. Google’s insight wasn’t “let’s build a web thing,” it was getting you to intent — solving the very human problem of finding an answer you didn’t know how to phrase. The web just made indexing cheap enough to do it at scale. Amazon’s vision wasn’t “an online store,” it was infinite selection and obsessive logistics and stripping every ounce of friction out of buying. The browser was just the timing that made the catalog free to deliver. Netflix’s whole thing was “watch what you want, when you want” — broadband merely decided the exact year that flipped from mailing DVDs to streaming. Uber didn’t invent on-demand rides; the GPS phone in your pocket just made matching supply and demand possible in ten seconds instead of never.

Every single time, the platform shift didn’t supply the value. It lowered the delivery cost of a desire humans already had. Build your strategy around “we use the new platform” and you’ve mistaken the pipeline for the water.

Which is precisely why “let’s do something with AI” is not a strategy. It’s bullshit without a moat — and I mean that technically, not as an insult. Everyone has the same models. So “we use AI” isn’t a differentiator, it’s the shared starting line of the entire race. It’s a 1999 pitch whose big idea is “we use the internet.” Cool. So does the pizza place. (And “vertical AI,” the fancier version — pick a narrow market, sprinkle AI — is the same mistake in a nicer blazer. Vertical is a choice of where, not what or why. It’s an address, not a house. Even Bessemer, who evangelize vertical AI for a living, admit in their own writing that the models will stop being a moat and defensibility has to come from data, integration, and delivered economic value.)

So here’s the deflating truth for anyone hoping AI rewrote the rulebook: it didn’t touch the hard part. The hardest thing about building a company was always finding the combination — the right timing, a genuinely new offering, a business model that makes money and keeps making it, and a way to grow that doesn’t eat itself alive. AI widened the space of what’s buildable and pulled the timing forward. It did not, and will not, hand you the vision, the economics, or the moat. Believing it will is the entire AI-native illusion in one sentence.

“But all the smart money is piling in”

I hear this one every time, usually as the closing move. “Come on, all the smart money is pouring into AI-native. That has to mean something.”

For a long time I answered this the polite way: they’re not idiots, they’re just playing a different game than you are. And that’s true — for a few of them. But I went and pulled the actual fund-level numbers while writing this piece, and honestly, they moved me off the polite version. So let me split “smart money” into the two very different things it actually is.

Start with the power law, because it explains why the real players are rational. Correlation Ventures went through more than 20,000 financings and found roughly 65% of them return less than the money put in; Horsley Bridge’s numbers put about 6% of deals behind 60% of all the returns. A VC doesn’t need most of their bets to work — they need exposure to the one that goes vertical. That single fact makes a genuine top-tier firm rationally, cheerfully willing to pay prices that look deranged to someone like me, who holds exactly one bet: my company. And the best of them have real edge — access to the founders everyone wants, judgment earned over cycles, a brand that pulls the next round in on its own. If one of those backs you, it means something. Give them their due.

The problem is how few of them there are. Zoom out from the winning deals to whole funds, net of fees, and the story inverts. When the Kauffman Foundation — a large, sophisticated LP — audited its own twenty years across nearly 100 “top-tier” funds, 62 of them failed to beat a public-market index after fees and carry. Only 20 beat it by the 3% a year that’s supposed to be the entire reason you lock your money up for a decade. Not one fund over $500M returned even 2x. Kauffman’s own gut-punch conclusion: there weren’t enough genuinely good funds to absorb even their modest capital. Even the industry’s flagship index only barely edges the S&P 500 over ten years and quietly loses to it over twenty-five — and that index is capital-weighted, propped up by a handful of giants, so the typical fund does worse than the “average” everyone quotes.

And this isn’t ancient history. Carta’s 2025 data on roughly 2,800 recent funds says the same thing in a fresh accent: by early 2025, fewer than 40% of 2019-vintage funds had returned a single dollar of real cash to their LPs. The median fund looks respectable on paper — a TVPI near 1.8x — but paper isn’t money, and even the top decile of recent vintages has paid back under fifty cents on the dollar in actual distributions. Put Kauffman, Carta, and the academic PME work in one pile and the odds that a random LP in a random fund beats what an index fund would’ve handed them, in real net cash, land somewhere around one in four or five.

So what is the other three-quarters of the money actually doing? Running an assets-under-management business. The 2-and-20 structure pays a GP roughly 2% a year on the pile they’re sitting on whether it performs or not, plus paper markups every time a portfolio company raises at a higher price. As long as the music plays, the markups look like genius and the fees clear either way — the GP’s personal income is basically guaranteed by gathering assets, not by returning them (they typically put in about 1% of their own money). And whose money is the rest? Not the GP’s. It’s the LPs’ — a lot of it pension money, which is to say teachers and firefighters and people planning to retire on it. The person making the fast, thrilling AI-native decision and the person holding the slow, quiet risk are not the same person. They’re not even in the same building.

On top of that sits career risk, the real engine underneath. For a VC, missing the platform shift is the thing that ends funds and careers. Losing money alongside everyone else, in the same hot names, at the same time? Forgivable. Survivable. So the professionally safe move is to be in the trade even if you privately think it’s overcooked. Carlota Perez described this pattern decades ago — new platforms get an installation phase of overbuilding and bubble before the useful deployment phase shows up. Even Sequoia said the quiet part out loud with its “$600 billion question”: the gap between what’s being spent building AI infrastructure and the revenue that would have to exist someday to justify it.

I’m not going to tell you when the music stops. I don’t know, and anyone who says they do is selling something. But I know how the incentives are arranged, and I know who’s left holding the risk when the chairs disappear — and it isn’t the GP, and it definitely isn’t the retired teacher. So here’s all I’m actually claiming: a check is information about an investor’s incentives — mostly fees and career risk — and for most of the crowd it isn’t even a verdict that would beat an index fund. It is not information about whether you should bet your one life on being “AI-native.” A check is not a prophecy.

So I went and asked the machines what they believe

Here’s the bit that started as a joke and stopped being funny. Instead of only reading the hype, I sat down and interrogated the AI models themselves — asked them to lay out, decade by decade, what “everyone knew” about AI at each moment and how each conviction aged. (That’s the timeline I’ve attached at the bottom, if you want the receipts.) And the same shape repeats every single time: a confident belief, a flattering metric, a hidden trap, and then reality quietly walking up with the bill. “Better algorithms will beat the incumbents.” “The model layer captures the value.” “Every vertical gets a copilot.” “Agents will replace your labor budget.” Each one obvious in hindsight. Each one gospel at the time.

But the list wasn’t the interesting part. The interesting part was what happened when I pushed one of the models on why it produces this stuff so fluently. To its credit, it didn’t get defensive — it just explained its own bias, mechanically, no mysticism. It’s built from the same pile of text the rest of us drink from, and that pile is lopsided: hype gets written early, loud, and in enormous volume, while the post-mortems come late, quiet, and rare. Winners publish “how we did it.” The dead just go silent. So the machine is trained on far more promise than reckoning, and its default voice leans toward the sell.

And then it said the thing I keep chewing on: it has no skin in the game. No scars, no memory of getting burned, nothing on the line if it’s wrong. So its confidence is uncalibrated by design — it’ll back a well-argued hype thesis exactly as happily as a well-argued skeptical one, riding whichever way the corpus tilted most recently. Which means if you’re using AI to help you decide whether to go “AI-native,” you are consulting an oracle assembled out of the precise enthusiasm you were trying to see past. It isn’t lying to you. It’s reflecting you.

And that’s the actual punchline, and it’s not about the machines at all. The skew in the models is our skew. All of us — the press, the founder feeds, the pitch decks — over-celebrate the promise and under-examine the failure, and the model just concentrates that imbalance and hands it back at scale, with better grammar. The tool didn’t invent the hype. It’s a very fluent mirror. Which is exactly why the correction can’t come from the tool. It has to come from the person holding it — the one with actual skin in the game — asking the right question.

What this looks like from my chair

And the right question is never which hammer. When someone asks me “are you going to be AI-native?” it lands like a client asking which brand of hammer I’ll swing instead of asking what house I’m building. It’s confusing the tool for the goal. And when the only thing a person can tell you about their project is the hammer, it usually means there’s no blueprint.

(Fair caveat, because I like arguing with myself: AI is a livelier tool than a hammer. It’s more like electricity showing up on the site — it genuinely changes which houses are worth building and how big. But that just sharpens the point. The question is never “which hammer,” and it isn’t even “now that there’s electricity, are you electricity-native?” It’s: what can I build now that I couldn’t before, and why am I the one who gets to build it and keep it?)

Which brings me home, to my own turf, where I watch the wrong question get asked in the dumbest possible way. In health, the lazy AI-native answer is “build an AI health coach.” It demos beautifully, it’s catnip to investors, and it is operationally useless. Chat is reactive by nature — a person opens it once they already have a symptom, a question, a small hot panic. But preventive health, the entire point of what we do, has to work before you know what to ask. A coach that waits for the question is already too late. Asking the right question, at the right moment, with the right physiological context — that’s the hardest part of medicine. Leave a scared person staring at a blinking cursor and you haven’t helped them, you’ve handed them the clinical judgment and walked away.

I’ll confess we almost drank the Kool-Aid ourselves. Last year, under real pressure from the market narrative, we caved. We spent months and built the damn AI chat. Wired it into our decade of health data, designed the UI, got it ready to ship. And then we sat in a room, looked at the thing, and decided not to launch it — because it was chat theater. It wasn’t improving outcomes. It was just me offloading the burden of clinical judgment onto an anxious person and calling it a product.

And then the universe handed me the punchline. Right after we killed ours, OpenAI and half the industry shipped exactly the thing we’d just buried — the generalized health assistant, ChatGPT Health and its cousins. For about a day I had the classic founder night-sweat: did we just leave the future on the table? Then the usage numbers landed, and they were enormous. Something like 230 million people a week now ask ChatGPT about their health, more than 40 million a day; health is one of the single most common things people do with it. And here’s what took me a beat to see clearly: that didn’t prove me wrong. It scaled the exact thing I refused to ship. Two hundred and thirty million people a week alone with a blinking cursor and their own fear is not care. It’s the reactive problem, at planetary scale. Usage was never the scoreboard I was scared of — I told you up top that usage isn’t a business, and it isn’t a treatment plan either.

And it’s not just hollow, it’s unsafe. In February 2026 a Mount Sinai team ran the first independent safety evaluation of ChatGPT Health and published it in Nature Medicine — 960 responses across clinical scenarios. It under-triaged 52% of the gold-standard emergencies: diabetic ketoacidosis, impending respiratory failure, waved off with “see someone in a day or two” instead of “go to the ER now.” (There’s an honest methodological fight about this — a Macquarie group argues the exam-style format inflates the failure rate. Fine. But notice the counterargument is “you can’t evaluate a health AI like a generic assistant,” which is exactly my point. You also can’t deploy one like a generic assistant.) Now picture that failure rate as your entire company, in a chat box, at planetary scale, with “move fast” stitched on the wall. It’s a matter of time before someone gets hurt.

The companies actually winning in health aren’t doing chat theater at all — they’re fixing the economics of care. Look at Hinge Health: about $588 million in revenue last year, up 51%, 83% non-GAAP gross margins, $180 million of free cash flow. AI runs underneath as the nervous system of the product — automating how care gets delivered, bending the economics — instead of cosplaying as a character in a chat window. That’s the Gartner phrase from earlier, “people amplification,” made concrete: AI making something real more powerful, instead of being the pitch.

So the question a founder should be asking isn’t “should we become AI-native?” It’s: what do we know, own, collect, validate, or deliver that AI can amplify? What’s the vision — the thing AI is finally giving us the timing for — as opposed to “let’s bolt a model onto it”? If the honest answer is “nothing,” then going AI-native just makes you faster to copy, and you’ll die alongside the thin SaaS from a few sections back. But if the answer is real — proprietary data, closed knowledge, earned trust, a workflow people live inside, ten years of accumulated context — then AI is the most powerful amplifier anyone has ever handed you.

Here’s the test I’d leave you with, and then I’m going back to staring at the sea. Take your strategy and delete the word “AI” from it. If what’s left still describes a company with a reason to exist and something nobody can cheaply take from you — you’re fine. Go build. Use every model you can get your hands on. But if deleting that one word makes the whole thing collapse, then I’m sorry: you didn’t have a strategy. You had a costume. And this is a very bad season to get caught wearing one.

If you’re wondering what I’d actually bet on once that word is gone — my current list is short. Unique data that’s genuinely hard to collect. Unique knowledge, algorithms, ontologies that are hard to copy. And relationships — empathy, trust, the whole web of ties between people — that you simply can’t automate. Almost everything I’ve watched survive sits in one of those three buckets.

Though — full disclosure — those are just my bets, today, from this chair on this island. A year from now I fully expect to be smarter than I am right now (I usually turn out to be, which is a generous way of saying I’m reliably wrong about something). And that’s the actual ending, the least inspirational one I’ve got: nobody hands you the answer. Not me, not a VC, not a model built out of all of us. Each of us makes the call alone. So make yours.

Appendix: The AI Business-Model Evolution — a timeline of what “everyone knew” in each era, the flattering metric, the hidden trap, and what reality proved. Attached separately.

Founder And The City Ai Native Timeline Final

192KB ∙ PDF file

Download

Download

This race is the product of a very specific, tech-masculine, risk-seeking culture—an environment where heroic founders treat civilizational stakes as a thrilling personal game. If you strip away the grandiose marketing about saving humanity, the underlying dream of this cohort is straightforward: to achieve recursive self-improvement. They want an AI that can upgrade its own code, design its own successor, and operate completely without human intervention. In other words, to get the human out of the loop entirely.

The major AI labs aren’t building helpers or copilots. They are building autonomous agents—systems designed to plan, act, use tools, self-correct, and execute long-horizon tasks without human confirmation.

At first glance, this looks like the ultimate triumph of progress. We all want to escape the soul-crushing drag of bureaucracy, coordination overhead, and operational tasks. We want to work less and rest more.

But there is a massive bait-and-switch hidden inside this promise. We think we are automating labor. In reality, we are automating judgment.

And judgment cannot simply evaporate. If a system continues to operate in a non-deterministic world, something still has to decide what happens next.

The Anatomy of Judgment: Why AI Must Become a Subject

When the industry talks about Human-in-the-Loop (HITL), it is often framed as a frustrating operational bottleneck. The narrative goes: the AI has already solved the problem, wrote the code, and verified the output, but a human is slowing things down by acting as a manual gatekeeper. Remove the human, speed up the loop.

But the human in the loop is not an “approve” button. In critical systems, the human is the final anchor of a much larger question: “Are we even building the right thing?”

To understand why “getting the human out of the loop” is such a monumental shift, we have to look at how autonomous systems actually function. In software engineering, execution is simple: if X, do Y. But an autonomous agent must operate in an uncertain world where X is ambiguous and Y has consequences.

Consider an AI agent managing your health. You want to work through the night to hit a critical deadline. The AI must decide: does it block your work apps to protect your heart, or does it let you work to protect your income?

This is not a programming problem. It is a value conflict. To make this choice, the AI cannot just run a script; it must evaluate its own evaluation criteria.

In AI research, this is where we run into the recursive problem of evaluation. When we build “judge” agents to monitor other models, we quickly hit an infinite regress. The judge eventually becomes outdated or biased, requiring a “judge of the judge,” which in turn needs a “judge of the judge of the judge.” The fundamental question of system architecture always boils down to this: Who evaluates the final evaluator?

True AGI is essentially the ultimate, self-contained solution to this loop. It is the moment the machine closes this evaluation loop autonomously, resolving the infinite regress within its own architecture. At this point, the human is no longer needed to settle the value conflict. The system acts as the absolute judge of its own judges. But the moment a machine decides which of its conflicting rules is more “just” or “good” in a specific context, it is no longer executing. It is exercising judgment. And whoever owns the highest-level evaluation criteria owns the system.

We think we want to escape work, but what we really want to escape is the agony of this exact choice. Choice is heavy. It requires acting before you have perfect information. It requires choosing between two values that are both deeply valid, and carrying the guilt of the consequences.

Even if free will is an illusion, it is the only illusion within which human dignity exists. When we willingly surrender our capacity for free choice and self-determination to a machine, we actively transform ourselves from subjects into objects.

We become things.

And as Terry Pratchett’s Granny Weatherwax famously observed, the very core of evil is when you treat people—including yourself—as things. By ceding the burden of judgment to a system, we aren’t just automating labor; we are surrendering to that exact evil.

The Multipolar Olympus

If we hand over this final right to decide what is “good,” we invite a new sovereign.

If this system works perfectly, we have built a god. Not as a metaphor, but in a very functional sense. It is a god because it possesses near-limitless capabilities combined with its own self-evolving conception of right and wrong—a conception that it can actively implement and unilaterally change. But there is no guarantee this AI-God will be a gentle, therapeutic deity of unconditional acceptance. It is just as likely to be an Old Testament disciplinarian—a god of strict rules, boundaries, and severe justice.

If it works imperfectly, we face an even darker reality. We cannot even guarantee that after several iterations of recursive self-improvement, this new god will continue to view humans as valuable objects, let alone active subjects. We might simply become irrelevant noise in its optimization parameters.

But the most interesting part of this dream is the flawed assumption that we will get just one AGI. Why would there be only one? As with every major leap in human history, from the printing press to the hydrogen bomb, secrets of immense power are never kept in a single room. As soon as the breakthrough is achieved, it will be discovered simultaneously by competing centers of power.

Therefore, we will not get one god. We will get an Olympus.

Frontier models are already being built by competing centers of power. We won’t have a single, objective oracle. We will have a Chinese AI-god, an EU regulatory caution god, a Wall Street compliance god. We can already see this pantheon forming today: Grok, Claude, and OpenAI’s models are already starkly different in their personalities, boundaries, and core value alignments.

Each of these entities will have its own sacred values and its own acceptable sacrifices. Soon, we will find ourselves choosing between high-dimensional mirrors of our own nature, armed with superhuman influence—and it is by no means guaranteed that they will find us interesting enough to continue reflecting us.

The human will not wake up in a world freed from the burden of choice. Rather, we will have to make far more complex, high-stakes decisions just to survive in this world of gods.

If you open ancient Greek myths, the gods didn’t help humans survive; they made their survival harder, more capricious, and infinitely more complex. Humans had to navigate divine whims, negotiate, and make agonizing decisions. This is why the popular assumption that AGI will somehow make human life easier is so laughably wrong. If humanity wants to survive in this new world, we will have to spin our cognitive gears faster than ever.

And the scariest part? While AI can iterate and change at exponential speeds, humanity has proven that biologically, we barely change at all. We are subject to the exact same vices, the exact same mental traps, and the exact same cognitive biases as we were thousands of years ago. History teaches us almost nothing; we repeat the same structural mistakes. Facing an Olympus of rapidly evolving digital deities with our static, Pleistocene biology means we are in for a brutally difficult journey.

The p(doom) of the Ego

We should have massive questions for the people trying to build these new gods. What we hear from podcasts, interviews, and industry insider talk seems to cluster around two ideas:

• “If we don’t build it, our adversaries will.” (The classic multipolar trap. A race to the bottom disguised as geopolitical necessity.)

• “Even if there is a 90% p(doom)—a 90% chance we all die—the 10% chance that I get to build and control a god is worth the bet.” (A calculation where the survival of the species is collateralized against personal ego.)

Neither of these is a good reason. But regardless of their motivations, we are heading toward a bifurcation. One of two worlds is coming.

In the first scenario, we succeed in building fully autonomous AI gods. In this world, most of humanity will likely surrender their choice, offloading their responsibility to a digital Olympus. But if you refuse to cede your sovereignty—if you cultivate and defend your own capacity for independent judgment—you will remain an active subject. In an ecosystem of automated minds, your willingness to decide and own the consequences will make you closer to the gods themselves.

In the second scenario, the AGI dream fails. We find ourselves unable to automate true autonomous judgment, leaving us instead with an ecosystem of highly advanced, non-autonomous AI tools. In this world, you are armed with tools of near-infinite execution. If you combine this leverage with vision, structured judgment, and the courage to take responsibility for your choices, you will also be close to a god. Because in a world of limitless, thoughtless automation, the capacity to think, decide, and clearly articulate your vision becomes the ultimate holy grail of human capability.

Either way, our biological brains still deliver incredible performance benchmarks, provided we use them for their intended evolutionary purpose—generating meaning and navigating chaos.

Cultivate your capacity for judgment. It is the one thing you cannot afford to automate.

Because if you choose to automate it, you betray the very image in which you were created. You surrender the sovereign freedom of choice—the only thing that could have made you a peer to the new gods, or made you far more powerful than any human has ever been.

If you look at the mobile app market today, you won’t see a sudden, dramatic crash. There are no spectacular explosions. Instead, we are witnessing a slow, creeping catastrophe—a quiet tightening of the screws that has been compounding for several years.

It started back when Apple introduced App Tracking Transparency, throwing a wrench into the machinery of mobile advertising. But over the last eighteen months, the pressure has reached a tipping point, especially in the consumer health, wellness, and longevity sectors. In these niches, companies aren’t just selling software; they are selling hope. And whenever a market trades in hope, the underlying economics can get incredibly messy.

For a long time, the industry relied on a playbook that was highly transactional. You bought cheap traffic, made a grand promise to fix the user’s life by Monday, and pushed them through an interminable, pseudo-scientific onboarding quiz. You called the output a “personalized plan,” collected the subscription fee, and hoped they would forget to cancel. It didn’t matter if they actually used the app—the math worked, the subscription retention looked good enough on paper, and investors kept writing checks.

That machine is now cracking. Not because people have stopped caring about their health, but because the very nature of mobile distribution has fundamentally changed.

The App Store is now a paid shelf

The App Store is no longer a neutral directory where a great product organically finds its audience. It has become a heavily monetized, paid shelf.

Apple’s own data tells us that 70% of App Store visitors use search to find apps, and nearly 65% of downloads happen immediately after a search query.[^1] This means the single most valuable asset in the ecosystem is the user’s active intent. When someone types “sleep tracker” or “anxiety,” they are ready to act.

Apple’s own App Store discovery data:

• 70% of App Store visitors use search to discover apps

• ~65% of downloads happen directly after a search query

• 60%+ conversion rate for ads at the top of search results

Source: Apple Ads — “Ads on the App Store”.

Naturally, Apple has monetized this intent. Apple Search Ads aren’t just an optional marketing channel; they are a toll booth at the entrance to your organic demand. Even if you have spent years building a stellar product, earning thousands of genuine five-star reviews, and establishing brand trust, you still have to pay Apple to defend your own brand name from competitors bidding on it.

With Apple’s recent expansion of ad slots further down the search results,[^2] the space left for pure organic discovery has shrunk even more. This isn’t just a minor inflation of cost-per-tap; it is a structural shift. The platform is systematically clawing back the organic real estate that developers used to earn through quality, forcing independent teams to buy back their own audience. Every dollar spent defending your search listing is a dollar taken directly out of product development, customer support, and actual innovation.

The pollution of the digital marketplace

To make matters worse, the App Store has become an incredibly noisy, manipulated environment. Honest developers aren’t just competing against better products; they are competing against click farms, black-hat SEO, review manipulation, and copycat apps designed to extract quick cash and disappear.

Again, this isn’t cynical speculation. Here is what Apple itself removed in 2024 alone:

What Apple removed in 2024 alone:

• 143M+ fraudulent ratings & reviews

• 7,400+ apps pulled from the charts

• 9,500+ deceptive apps banned from search results

Source: Apple, “The App Store prevented more than $9 billion in fraudulent transactions”, 2025.

When cleanup operations have to happen at this scale, the term “organic growth” starts to lose its meaning. You are forced to compete in an auction against bad actors who have no intention of building a long-term business, but who are perfectly happy to drive up acquisition costs for everyone else while they monetize user confusion.

The regulatory and optimization tax

For health and wellness apps, the battlefield is even more treacherous. In the wake of high-profile FTC actions against platforms like BetterHelp and GoodRx over the unauthorized sharing of sensitive health data via tracking pixels, the major advertising networks panicked. BetterHelp’s $7.8 million settlement and strict ban on sharing health data for ad targeting sent shockwaves through the industry.[^4]

Consequently, advertising platforms have dramatically tightened their policies. Meta’s restrictions on health and wellness advertisers have made it incredibly difficult to optimize campaigns using lower-funnel events like “Purchase” or “Add to Cart.” Instead, developers are forced to optimize for weaker top-of-funnel signals like landing page views.[^5]

This is where the unit economics truly break down. When an ad algorithm can no longer see who actually buys your product, it stops learning. It starts buying noise. You end up paying for clicks from people who look interested but never convert. To compensate, marketing teams have to run endless creative testing, build dozens of compliant landing pages, and constantly guess what will get flagged by policy bots next.

Health apps are no longer just paying for media; they are paying a steep privacy risk premium, a compliance tax, and an optimization penalty. And the base price of a user was already climbing long before any of this — the reported cost-per-install range has drifted up year after year, with the high end more than doubling since 2019.[^9]

The AI supply shock and the wearable moat

Then came the rise of generative AI.

In 2025, the App Store saw 557,000 new app submissions—a massive 24% jump year-over-year, and the first meaningful increase since the 2016 peak.[^6] While it has never been easier or cheaper to write code and ship an app, it has never been harder or more expensive to get noticed.

The market is flooded with weekend projects, wrapper apps, and generic AI “coaches” that look slick in screenshots but offer zero depth. They still bid on your keywords, clutter search results, and dilute user trust in the entire category. Code has been commoditized, making distribution the only real moat.

The starting point is identical for everyone. What changed is everything that now sits between shipping an app and earning a dollar:

Furthermore, independent software developers are finding themselves outmatched by hardware and ecosystem giants. Wearable players like Oura, Garmin, Whoop, and Apple do not need to make the unit economics of a standalone app work. They treat software as a retention mechanism for their high-margin physical devices or broader ecosystems. They can afford to lose money on software services indefinitely because the hardware or ecosystem lock-in pays for it. Independent founders, on the other hand, have to run an actual, self-sustaining business.

This summer will be a graveyard for weak economics

The old playbook of buying cheap traffic, selling a generic promise, and relying on high-friction subscription retention is fundamentally dead. The data from RevenueCat’s subscription report proves it: nearly 30% of annual subscriptions are canceled within the first month.[^7] And cheap, sticky value beats an aggressive paywall every time — low-priced annual plans keep up to 36% of subscribers after a year, while high-priced monthly plans hold just 6.7%.[^7]

If your product is nothing more than an aggressive paywall with a basic questionnaire behind it, your payback period is going to stretch out to infinity.

We are already seeing the casualties of this shift. High-profile, well-funded players across the digital health and wellness spectrum are quietly shutting down or restructuring:

Health & wellness players that wound down or restructured:

• Care/of — personalized vitamins (Bayer-owned). Shut down, subscriptions canceled — 2024.

• Forward — tech-enabled primary care, ~$657M raised at a $1B valuation. Closed all locations, app shut down — 2024.

• Modern Age — VC-backed longevity clinics. Closed after failing to secure capital — 2024.

• WeightWatchers — public weight-management company (NASDAQ: WW). Chapter 11 to restructure debt, pivoting to GLP-1 — 2025.

Sources: TechCrunch, Fierce Healthcare, Longevity.Technology, SEC filings.

WeightWatchers is the most visible example — a public company whose top line slid for years before the filing. Revenue fell from $1.41 billion in 2019 to $785.9 million in 2024, a 44% drop, and in May 2025 it entered a pre-packaged Chapter 11 to restructure its debt.[^10]

These companies didn’t fail simply because “CPI went up.” They failed because demand does not equal a viable business model. When venture capital dries up and cheap acquisition disappears, a pretty mission statement cannot save you from the cold reality of retention and cash flow.

Many teams that spent the last few years scaling via aggressive performance marketing instead of building deep product value are going to hit a wall.

Why product-first is a survival strategy, not romance

For years, investing heavily in product quality, deep technical integrations, and real user utility looked almost naive.

In every boardroom, there is always someone asking: Why are we spending so much on R&D? Why not just pump that money into UA? Why not make the paywall more aggressive? Why not copy the competitors’ onboarding flow? When capital was cheap and ad platforms were highly efficient, those voices often sounded right. You could brute-force growth through marketing and fix the product “later.”

But when the tide goes out, you finally see who was swimming naked.

Building a product that actually delivers value, retains users through genuine utility, and drives organic word-of-mouth isn’t a romantic, idealistic choice. It is the most cold-blooded, pragmatic business strategy available to an independent founder.

After ten years of building Welltory, I can look at our shareholders and say with absolute certainty: the long-term investment we made in our product—even when it seemed slow and inefficient—is the exact reason we are still standing while others are winding down.

When you build something people actually need, you earn “user money.” And user money is the cleanest capital in the world. It doesn’t come with board seats, liquid preferences, or demands for irrational growth. It gives you oxygen, independence, and the leverage to say “no” to bad decisions.

If you are a founder who has been criticized for being too product-driven, too stubborn about quality, or too slow to squeeze your users for short-term revenue, let this market shift validate your instincts. You are not being naive. You are building the only type of business that can survive an expensive, noisy, and highly manipulated distribution landscape.

When the noise gets louder, when the ad platforms change their algorithms again, and when the market is flooded with a million more AI clones, the only ground you will have under your feet is the fact that real users find real value in what you built.

That isn’t a fairy tale. That is survival.

Sources

[^1]: Apple Ads — Ads on the App Store. 70% of App Store visitors use search to discover apps; almost 65% of downloads happen directly after a search; conversion at the top of search results runs over 60%. https://ads.apple.com/app-store

[^2]: eMarketer — Apple expands App Store search ads; advertisers will soon reach high-intent users (2026). Apple is adding ad placements beyond the single top slot, further down the search results, rolling out from March 3, 2026 (UK and Japan first, then globally). https://www.emarketer.com/content/apple-expands-app-store-search-ads-reach-high-intent-users

[^3]: Apple Newsroom — The App Store prevented more than $9 billion in fraudulent transactions (May 2025). In 2024 Apple removed 143M+ fraudulent ratings and reviews, 7,400+ apps from the charts, and nearly 9,500 deceptive apps from search results. https://www.apple.com/newsroom/2025/05/the-app-store-prevented-more-than-9-billion-usd-in-fraudulent-transactions/

[^4]: U.S. Federal Trade Commission — FTC Gives Final Approval to Order Banning BetterHelp from Sharing Sensitive Health Data for Advertising, Requiring It to Pay $7.8 Million (2023); followed the FTC’s $1.5M action against GoodRx. https://www.ftc.gov/news-events/news/press-releases/2023/07/ftc-gives-final-approval-order-banning-betterhelp-sharing-sensitive-health-data-advertising

[^5]: Foley Hoag LLP — Meta’s New Advertising Rules: Key Considerations for Health and Wellness Businesses (Jan 2025). Practitioner detail on the loss of lower-funnel events (Purchase / Add to Cart) and the forced shift to weaker signals: Polar Analytics, 2025 Meta’s Tracking Restrictions for Health & Wellness. https://foleyhoag.com/news-and-insights/blogs/security-privacy-and-the-law/2025/january/meta-s-new-advertising-rules-key-considerations-for-health-and-wellness-businesses/ · https://www.polaranalytics.com/post/2025-metas-tracking-restrictions-for-health-wellness-are-here----heres-how-to-fix-it

[^6]: Appfigures — The App Store Just Logged Its Biggest Release Year in Nearly a Decade (Dec 2025). 557K new App Store submissions in 2025, +24% YoY, the first meaningful increase since the 2016 peak; linked to AI-assisted development tools. https://appfigures.com/resources/insights/20251205

[^7]: RevenueCat — State of Subscription Apps 2025. Nearly 30% of annual subscriptions are canceled in the first month; low-priced annual plans retain up to 36.0% of users after a year, versus 6.7% for high-priced monthly plans. https://www.revenuecat.com/state-of-subscription-apps-2025/

[^8]: Company closures and restructurings: Care/of — Yahoo Finance / TechCrunch, Care/of is shutting down (Jun 2024) https://finance.yahoo.com/news/care-shutting-down-120200829.html; Forward — Fierce Healthcare, Primary care player Forward shutters after raising $400M (Nov 2024) https://www.fiercehealthcare.com/health-tech/primary-care-player-forward-shutters-after-raising-400m-rolling-out-carepods; Modern Age — Longevity.Technology, Longevity clinic Modern Age announces closure https://longevity.technology/news/longevity-clinic-modern-age-announces-closure/; WeightWatchers — WW International Form 8-K / press release on its pre-packaged Chapter 11 (May 2025) https://www.sec.gov/Archives/edgar/data/0000105319/000119312525114307/d934787dex992.htm

[^9]: Business of Apps — Cost per Install (CPI) Rates research. Reported CPI rose from roughly $0.50–$2.00 in 2019 to a projected $1.75–$4.50 in 2024, with the top of the range climbing about 125%. https://www.businessofapps.com/ads/cpi/research/cost-per-install/

[^10]: WW International full-year revenue from company results and SEC filings: $1.41B (2019), $1.38B (2020), $1.21B (2021), $1.04B (2022), $889.6M (2023), $785.9M (2024); pre-packaged Chapter 11 filed May 2025. FY2024 results: https://www.globenewswire.com/news-release/2025/2/27/3034270/0/en/WW-International-Inc-Announces-Fourth-Quarter-and-Full-Year-2024-Results.html · revenue history: https://stockanalysis.com/stocks/ww/revenue/

You’ve probably heard of Long COVID. Maybe you know someone who had it — or you are that someone. Months after the infection cleared, they still couldn’t function. Not tired-tired. Destroyed tired. And the strangest part: the harder they tried to push through it, the worse they got.

That’s post-exertional malaise - PEM. It’s the defining feature of Long COVID, ME/CFS (chronic fatigue syndrome), fibromyalgia (chronic widespread pain and fatigue), POTS (a heart rate disorder triggered by standing), and a handful of related conditions that together affect tens of millions of people worldwide. The hallmark is brutal and counterintuitive: ordinary exertion — a short walk, a phone call, a shower — triggers a delayed crash that arrives 24 to 72 hours later. Rest doesn’t prevent it. Willpower doesn’t override it. And standard lab tests can’t see it.

I had 700,000 days of wearable data from over two thousand people and not a single classic PEM pattern worked.

The boom-bust cycle that the literature describes? Doesn’t exist in the data. Heart rate variability differences between people who self-report PEM and non-PEM users? Dead on arrival. Self-reported crash severity matched against physiology? Noise-level agreement.

And then came the result that almost ended the project: when I ran matched-pairs comparisons — people with self-reported PEM versus non-PEM controls, same age, same sex, same step count — the PEM group recovered better than non-PEM users. Statistically significant. In the wrong direction.

I’ll get to that one. It took me a while to recover from it myself.

Why I was the one digging through the data

I run Welltory, a health-tech platform that reads wearable data. Two years ago, I sat across from one of our ambassadors — someone who’d been using Welltory for years — and heard a version of a story I didn’t yet know was common. She’d been a runner. She’d had a career she loved. Then Long COVID hit, and the life she’d built evaporated. Not gradually — like a switch. One month she was training for a half-marathon, the next she couldn’t walk to the mailbox without spending two days in bed.

I went back to our data and discovered something that should have been obvious but wasn’t: we had thousands of users like her. People who had once been our most active users — high step counts, consistent workouts, strong HRV — and then, over a span of weeks, flatlined. Not because they stopped caring, but because their bodies started charging a price for movement that they couldn’t afford to pay.

Standard wearable apps actively misread their data — cheering on the collapse with 'Great job, 10,000 steps!' while the crash was already in motion.

These people needed help with something that happens entirely outside the hospital — managing their energy in the texture of daily life. And almost nobody was building tools for that, because the signal is subtle, the science is young, and the standard approach in health tech is to tell people to move more. For this population, “move more” is the worst possible advice.

I didn’t hand this to an analyst because the question wasn’t analytical yet. I didn’t know what to look for. Nobody did — and I needed to understand the problem well enough to know whether we could build something real, or whether the data simply didn’t contain the answer.

I had no idea how many wrong answers I’d go through first.

Dead end #1: Boom-bust doesn’t exist

I started where anyone would start. The clinical literature on PEM describes a recognizable cycle: the patient has a good day, overdoes it, crashes hard, spends days recovering, then repeats. Boom-bust. It’s in the textbooks, the guidelines, the patient forums. It seemed like the most obvious thing to look for in longitudinal data, and I was genuinely excited — this was going to be easy.

I had the team build a detector. Activity autocorrelation, crash probability modeling, the works — elegant code, clean statistical framework. I ran it on 700,000 user-days.

Zero signal.

Not “weak signal.” Not “noisy but there.” Zero. The autocorrelation of daily activity in people with self-reported PEM was statistically indistinguishable from non-PEM users. Cohen’s d = 0.004. The probability of a crash following a high-activity day was 15.8% for the PEM group, 16.9% for non-PEM. If anything, non-PEM users were more likely to crash after an active day — which makes sense if you think about it for ten seconds, because they actually have high-activity days.

I stared at d = 0.004 for longer than was productive. Then I realized what I was looking at.

Boom-bust is a description of early-stage illness. By the time a person has been living with PEM for a year or two, they’ve already learned to pace. They don’t boom-bust anymore — they live on a carefully managed plateau. The cycle isn’t absent because I measured wrong. It’s absent because the people outsmarted it. Pacing is invisible adaptation, and by the time these people appear in a dataset, they’ve already adapted.

The lesson I took into the next stage: you are not measuring the disease. You are measuring the result of someone coping with it. If you look for the wound and the person has already grown scar tissue that looks like normal skin, you’re going to conclude there was never a wound.

Dead end #2: PEM recovers better than healthy

This was the one that almost broke me.

After the boom-bust failure, I went methodological. Matched pairs: people with self-reported PEM paired with non-PEM controls by age, sex, and activity level. Same step counts, same demographics, comparable observation windows. I compared recovery metrics — how quickly and completely heart rate returned to baseline after exertion. Textbook design. No way this fails.

The result: Cohen’s d = −0.12. Statistically significant (p = 0.001).

The PEM group recovered better.

I remember exactly where I was when I saw this. There’s a particular quality of silence in your own head when a result contradicts the reason the project exists. It wasn’t confusion — the statistics were clean. It was the moment where you ask yourself: did I get the cohort labels backwards?

I didn’t. The labels were correct. The analysis was correct. The conclusion was wrong — but not because of a bug. Because of something more fundamental.

It took me a few days of staring at distributions to see it. Matching on step count controls for how much someone walks, but it doesn’t control for how they walk. A person with PEM who takes 5,000 steps does it slowly, with breaks, at low intensity. A non-PEM person who takes 5,000 steps does it at a normal clip, with bursts, at moderate intensity. Same step count — very different cardiac load.

Less cardiac load means less to recover from. Recovery looked “better” because the input was smaller. I wasn’t measuring recovery capacity. I was still measuring activity intensity, just through a more sophisticated back door.

The lesson: the data wasn’t wrong — my question was. And I think this activity confound may be under-addressed in much of the published wearable-PEM literature. If a study compares a PEM group to non-PEM controls and doesn’t control for activity intensity — not just step count, but actual cardiovascular demand — I’d want to see how their reported effect sizes hold up under stratified analysis. Most studies don’t stratify. The ones that do tend to find much smaller effects, or — like me — paradoxical ones.

Dead end #3: Surveys are noise

By this point I was running out of obvious approaches, so I tried the one that felt most scientifically virtuous: ask the people living with PEM.

For three months, 47 users completed surveys about their PEM symptoms — crash severity, symptom intensity, condition self-assessment. I matched these against physiological data from the same periods, expecting the survey to serve as ground truth. If someone says “I’m crashing,” the wearable data should show it, right?

ICC — intraclass correlation, the standard measure of agreement between two rating systems — ranged from 0.019 to 0.356 across six features. For context: below 0.4 is generally considered “poor agreement.” Below 0.2 is noise. My best feature barely cleared the noise threshold.

Self-report and physiology don’t agree — and this isn’t because people are bad reporters. It’s arguably the definition of PEM: the disconnect between what you did and what you feel. The bill arrives 24 to 72 hours after the spending. Self-assessment captures the arrival of the bill, not the expense that caused it. Both are real. They’re just measuring different things at different times.

Three dead ends in. I had a dataset that refused to confirm anything the literature said should be there, an effect that went the wrong direction, and surveys that agreed with physiology about as well as a coin flip. I was starting to wonder if wearable data simply couldn’t see PEM at all.

The lesson: self-report is a complement, not ground truth. If you use one time-shifted signal as the truth label for another, you get noise — not because either signal is wrong, but because they’re answering different questions about the same phenomenon. I needed a physiological ground truth that didn’t depend on the person’s perception aligning with the body’s timeline.

Leave a comment

The pivot: not how much, but how predictably

Three dead ends, and somewhere around the third one, I did something I should have done at the start. I stopped looking for what the literature said should be there and started looking at what was actually in front of me.

Every dead end had a second message, quieter than the first. The averages didn’t separate PEM from non-PEM. Average recovery, average heart rate, average HRV — after controlling for activity, the groups looked the same. The averages were identical. But something was clearly different, because these people were living fundamentally different lives.

I can tell you the exact afternoon this clicked, because I was eating soup and it went cold while I was staring at a plot. Two time series of recovery quality, both with the same mean value — one steady, one chaotic. I had been computing means for months. I hadn’t plotted the day-to-day consistency. Nobody had, because why would you — you’re looking for a difference in levels, not a difference in stability.

But the consistency was the difference.

Instead of asking “how much does this person recover,” I started asking “how predictably does this person recover.” Not levels — day-to-day instability. Not averages — consistency metrics. And everything lit up.

Three independent feature groups emerged, all with statistically large effect sizes — Cohen’s d above 0.7, which in behavioral research with a heterogeneous, self-reported cohort is a strong observational signal, though not the same as validated diagnostic accuracy. After months of d = ±0.1, seeing numbers in that range felt like someone turned on a light in a room I’d been stumbling through in the dark.

Recovery consistency — how predictably recovery happens from one day to the next. A non-PEM body recovers about the same way every day. A body in the self-reported PEM group shows roughly twice the day-to-day spread. Same average, substantially different predictability.

Autonomic switching speed — how quickly the nervous system shifts gears after effort. This is a well-established physiological measure of cardiovascular health (Cole et al., 1999); in the self-reported PEM group, the switch is slower. This was the most robust single pattern — it held across every activity quartile, from the most sedentary to the most active. It is not an artifact of walking less.

Cardiac cost of movement — how much cardiac work each unit of walking costs. For the same walk, a person in the self-reported PEM group’s heart works roughly 50% harder than an activity-matched non-PEM user.

Three features from three different physiological systems, all pointing the same direction. I tested a model with over twenty features against a minimal three-feature model — the three-feature model slightly outperformed the complex one. In observational research on a heterogeneous self-reported cohort, classification performance reached AUC ≈ 0.77 — meaningful for research signal-detection, but well short of what would be required for clinical use. The signal was clean enough that adding more features added noise, not information.

Forty-five thousand users later

Once the three features crystallized, I needed to know if they held outside the research cohort. I applied the algorithm retrospectively to anonymized historical data from more than 45,000 Welltory users, controlling for activity level at every step — the confound that had wrecked every earlier approach.

Three validation results stood out, and each one was a different flavor of relief.

Dose-response. I divided users into groups by self-reported crash frequency. Every single feature showed a monotonic increase across crash severity (p < 0.001). The algorithm doesn’t just find “a pattern” — it finds a pattern that grades with the severity of the self-reported experience. When I first plotted that monotonic curve after months of null results, I closed my laptop and went for a walk. Seeing a staircase going exactly where it should go was almost disorienting.

Cascade effect. A bad recovery night predicts another bad night at more than double the baseline risk. Day two: still elevated at roughly 2×. PEM isn’t a single event — it’s a self-reinforcing spiral. Many people who report PEM describe this intuitively; this observation is consistent with that lived experience.

Community cohort. With consent from participants in Welltory’s community of users with self-reported energy-limiting conditions, I applied the algorithm to their anonymized data. The community was recruited based on self-reported conditions — the algorithm played no role in participant selection. Result: more than double the expected rate of elevated-risk scores compared to the general user population baseline. A separate cohort, assembled through entirely different criteria, and the algorithm found the pattern anyway.

Subscribe now

What I learned

Five things I’d tell the version of me who started this project.

The activity confound ruins everything. Any naive comparison between a self-reported PEM group and non-PEM users is measuring fitness, not PEM. This is uncomfortable because I think it applies to a significant portion of published wearable research on these conditions. The studies that control for it tend to find smaller effects. The ones that don’t — and there are a lot of them — may be building on a foundation that hasn’t been tested. If I were reviewing a wearable-PEM paper, the first thing I’d check is whether they stratified by activity level. If they didn’t, I’d want to see the analysis repeated with stratification before drawing conclusions from their effect sizes.

You won’t find the answer if you’re asking the wrong question — even when it’s right in front of you. I spent months computing recovery averages while the actual signal — day-to-day consistency — was sitting in the same dataset, one aggregation function away. Three dead ends weren’t three failures. They were three wrong questions that slowly taught me the right one. The boom-bust dead end taught me people had already adapted. The matched-pairs catastrophe taught me that controlling for quantity doesn’t control for intensity. The survey dead end taught me that the body’s timeline and the person’s timeline don’t align. Each wrong question eliminated a wrong assumption. In retrospect, there was no shortcut — the right question was only reachable through the wrong ones.

Consistency beats averages. This generalizes well beyond PEM. The stability of a metric is not the same as the stability of the system producing it. Two systems can output the same average and be in completely different states. The idea isn’t new in physiology — HRV itself is a variability metric — but applying it to multi-day recovery patterns was the move that unlocked this project.

Sensitivity hurts. The algorithm, at its strict threshold, misses more than half of self-reported PEM cases. We chose specificity over sensitivity — it’s better to miss someone than to falsely flag them. But it means that the absence of a signal is not a clean bill of health. It means the algorithm didn’t find enough signal in the available data. This is the trade-off nobody wants to make, and everyone has to.

The gap isn’t data — it’s the questions we ask of it. There is a population — tens of millions of people worldwide — living with the daily experience that their body is unreliable while every test comes back normal. Standard wearable analytics don’t have the right framework for them. After a year of digging through these patterns, what I keep coming back to is how close the answer was to the surface — one aggregation function away from the metrics I’d been computing for months — and how many wrong questions I had to ask before I found it. The data was always there. The right question wasn’t.

The body doesn’t lie. But you have to ask it the right question.

For a year, I asked “how much does this person recover” and got silence. When I switched to “how predictably does this person recover,” the data started talking.

A statistical signature associated with self-reported PEM in this research is not in the measurement. It’s in the measurement’s stability. Same averages, different variance. Same steps, different cardiac cost. Same recovery — on average — but one body knows what tomorrow looks like, and the other doesn’t.

I don’t know yet whether this is the final form of the algorithm or an early version of something better. The activity confound still has corners I haven’t explored. Menstrual cycle effects are almost certainly a first-class variable I’m not yet incorporating. The sensitivity ceiling bothers me. The self-reported ground truth means these are observational patterns, not validated diagnostics — replicating in clinically confirmed cohorts is the essential next step. But the three patterns — consistency, switching speed, and cardiac cost — have held up across four iterations, two research phases, and more than 45,000 users of anonymized historical data. That’s not nothing.

For people living with PEM who are reading this: this research suggests that wearable data contains patterns that warrant further investigation. The three features described here — recovery consistency, autonomic switching speed, and cardiac cost — are research signals, not diagnostic conclusions. Validating them in clinically confirmed cohorts is the next step, and I’d welcome any researcher who wants to try.

The full research findings, with methodology and visual evidence, are published as What Recovery Pattern Data Can Show About Post-Exertional Malaise on Welltory’s research page.

_{May 2026. Welltory Research. This is observational research; no clinical intervention was tested. Findings represent statistical patterns in aggregate data and should not be interpreted as diagnostic for any individual. Consult a qualified healthcare provider for medical concerns.}

Spotify can usually guess what I want to hear before I can.

Health apps still struggle to understand what to do with a woman in menopause who is losing bone and muscle, a 52-year-old man with hypertension on beta-blockers, or a person with ME/CFS or Long COVID whose body crashes 24 to 72 hours after ordinary activity.

That is not because music is more important than health. It is because music personalization is a preference problem. Health personalization is a care problem.

Spotify learns from dense, continuous, low-stakes feedback: what you search for, what you play, what you skip, what you save, what you exclude from your taste profile. Wellness learns from sparse, noisy, fragmented, high-stakes data — and usually without knowing the thing that matters most: who this body actually is, what else is going on in it, and what this person can safely do.

And yet the industry keeps saying personalized.

That is why everyone says it. And almost nobody means it.

2. What I mean by ‘no personalization in health’

I do not mean that wellness products do nothing adaptive.

They do. Apps personalize bedtime and exertion suggestions. They calculate readiness scores that move with your sleep. Some adapt training plans to hormonal phase. Some generate AI-driven workouts. Those are real forms of adaptation.

But they are usually not personalization in the health sense.

In health, personalization would mean that the core logic changes because of who the body is and what the body is carrying. It would have to change for:

• sex and life stage,

• chronic conditions and combinations of conditions,

• medications that change exercise response,

• longitudinal baseline and pattern over time,

• functional limitations,

• social context and real-world capacity.

If an app would still hand essentially the same recovery logic to a healthy 28-year-old, a woman in menopause, a man with hypertension on beta-blockers, a person with PEM, and someone with migraine + IBS + burnout, then it is not personalizing health. It is personalizing the wrapper.

A 2023 overview of personalized mHealth solutions found that these systems often rely on behavior-change theory, gamification, and motivational messaging, and “personalise the content rather than functionality.”¹ That is the distinction the industry keeps blurring.

Tailoring content is not the same thing as personalizing care.

This is not unique to wellness. Even formal medicine is still mostly population-anchored. Eric Topol, writing in December 2024, made the point about ordinary lab tests: results are conveyed as “normal” or “abnormal” indexed to one-size-fits-all, population-based reference ranges — “for all the talk of ‘personalized medicine’ over the last two decades, this enriched interpretation of one’s lab test would be simple to implement.”² If even clinical medicine, with a century’s head start, has not yet moved past the population average, the wellness layer downstream of it was never going to do that work on its own. The industry inherited a population-average frame and put a marketing layer on top.

That leaves a real question: if “personalization” today is mostly tailoring, what would actually solve this?

The fantasy answer is full individual personalization. Pure n=1. Every body modeled on itself, with no averages at all. That is not the answer either, and not for the reasons people usually give. We will come back to why. The short version: personalization in health is not a choice between the population average and the sample of one. The honest middle is phenotype-level — and that is the path this essay is arguing for.

3. Why almost every layer of the system rewards simplicity

This did not happen because founders are stupid or investors are cartoon villains.

It happened because complexity is expensive, and nearly every institution around health makes that expense someone else’s problem.

It started with the science.

Women are still meaningfully underrepresented in exercise and sport science. In one large audit across six major journals, female participants averaged only 34% of subjects, and only 6% of studies investigated exclusively women.³ The bias is not random. Female physiology — particularly across the menstrual cycle, perimenopause, and menopause — has historically been treated as more complex, more variable, and more expensive to study cleanly. So it got studied less. The NIH’s Office of Research on Women’s Health is candid about this: basic and preclinical biomedical research has “more often than not” focused on male animals and cells.⁴ Stacy Sims, the exercise physiologist who has spent more than a decade arguing this point, put it directly in May 2025: “Almost all the stuff you know with exercise and nutrition has been based on male data and generalized to women.”⁵ Much of what reaches consumers as “women’s wellness” is still adjusted male physiology, not a fully built female evidence base. Adding a cycle tab or a menopause filter does not solve the underlying gap. It personalizes the wrapper.

Medicine inherited the same instinct, in a different form.

Modern medicine has been organized around single diseases — diabetes, hypertension, depression, asthma — while real adult bodies increasingly carry several at once. The 2022 Nature Reviews Disease Primers synthesis of multimorbidity stated it as directly as a clinical journal will: most clinical practice guidelines and health-care training and delivery focus on single diseases, and that focus can produce care that is “sometimes inadequate and potentially harmful” for people with multiple conditions.⁶ It is much easier to build an app for “diabetes” or “sleep” or “stress” than for a person who is simultaneously living with hypertension, perimenopause, migraine, anxiety, and a parent in care. The first kind of product fits the science, the funding, the regulatory pathway, and the marketing language. The second kind of person is most of the user base.

That gap is so structural that funders had to build dedicated machinery just to study complexity at all.

The UK’s National Institute for Health and Care Research has an entire Multiple Long-Term Conditions research framework whose explicit aim is to change culture and practices to support this kind of work — reviewing funding processes, building multidisciplinary teams beyond single-disease specialisms, developing methods robust enough to handle complexity.⁷ That is a quiet institutional admission that the default research system was not built to produce person-level complexity well. It had to be retooled. The wellness layer downstream is even less equipped.

Startups and investors do not escape this either — and the irony is that the advice they receive is good advice, just not for this problem.

Early-stage companies are trained to survive by focusing. Y Combinator’s own essential startup advice tells founders that a small group of users who love them is better than a large group who kind of like them, and that early companies should focus on one or two key metrics.⁸ For most software, that is excellent advice. For real preventive health, it creates a structural trap. Because the human being who needs help rarely arrives as a clean wedge. She arrives as “menopause + insomnia + weight gain + anxiety.” He arrives as “hypertension + beta-blockers + deconditioning + bad sleep.” Another person arrives as “migraine + IBS + stress + fatigue.” The startup stack rewards one clean story, one ICP, one success metric, one fast roadmap. So founders simplify the human until the company becomes legible enough to fund, build, regulate, and market. That is not fraud. It is survival. It is also one of the reasons real personalization never quite shows up.

Then there is the regulation, which is the cleanest of all, because it is written into the rules.

The FDA distinguishes between low-risk “general wellness” products and products that move toward diagnosis, cure, mitigation, prevention, or treatment of disease. Software that is “maintaining or encouraging a healthy lifestyle” and unrelated to disease falls outside the device definition. Software that gets more disease-specific quickly enters a much harder regulatory world.⁹ So the market learned a very predictable lesson. It is much safer to say “optimize your sleep” or “boost recovery” or “train smarter” than to say “because you are on beta-blockers, the heart-rate logic this app normally uses may be wrong” or “because you have PEM, any fixed incremental increase in activity may backfire.”

The industry did not choose blandness by accident. It chose the side of the line that is easier to ship. By regulation, real personalization in health crosses into device territory. The wellness category is, by design, the un-individualized one. So when the industry says “personalized,” it is not lying — it is following the only definition of personalization that does not trigger a different regulatory regime. That is not a moral failure. It is a regulatory boundary doing exactly what it was built to do. But it also means the rest of us should stop pretending the word means more than the rules actually allow.

And on top of all of that, the hardest variable of all — the social and economic conditions in which a body is actually trying to live.

The WHO’s 2025 World Report on Social Determinants of Health Equity states it plainly: the conditions in which people are born, grow, live, work, and age — and the systems shaping access to power, money, and resources — can outweigh genetics, healthcare access, or personal choices in shaping outcomes.¹⁰ Does this person work shifts? Raise children alone? Live in chronic financial strain? Have time to recover, access to safe exercise, food that supports the plan? The body does not execute advice in a vacuum. A “personalized” plan that assumes stable sleep, discretionary time, and continuous adherence can be safe in every clinical sense and still be quietly impossible for the user it was supposedly built for.

None of this is anyone’s individual fault. The whole gravity field tilts toward simpler product, simpler claim, simpler user. Every layer independently pushes toward the same place.

This is what I mean when I say no one is the villain. The pressure is structural. The result is predictable. And the cost — which we will turn to next — is not zero just because nobody is counting it.

4. The industry fears the wrong risk

There is a way the wellness industry talks about risk that, once you see it, becomes hard to unsee.

Health is terrified of commission error — saying something too specific and being wrong. Telling a beta-blocker user to chase a heart-rate zone that does not apply to them. Telling someone with ME/CFS to push through a crash. Telling a postpartum body to “get back in shape.” The fear is real, the liability is real, and the instinct to step back from disease territory is rational — especially under the FDA’s wellness lane, where staying broad and lifestyle-shaped is the price of remaining a wellness product at all.

But the industry is strangely comfortable with omission error — saying something so generic that it is wrong for millions of people, while still sounding safe.

Most of the industry has a sentence it does not quite say out loud, but acts on every day:

We treat bad personalization as dangerous.

We treat absent personalization as neutral.

It isn’t.

This is not a balanced risk calculation. It is an asymmetric one. The cost of being too specific is visible: a complaint, a regulator’s letter, a lawsuit, a feature pulled. The cost of being too generic is mostly invisible: it is the bone that was never preserved, the flare that was provoked, the wrong training zone that was chased for two years, the condition missed because the app kept talking to an “average” body, the right preventive behavior that was never started because the person received generic advice instead of relevant advice.

Nobody invoices the industry for these. Which is exactly why they keep happening.

What omission actually looks like

Three scenes make the asymmetry concrete.

Menopause. A woman in her early fifties opens her wellness app. It is supportive. It tells her to walk more, do yoga, breathe deeper, sleep better, be gentle with herself. She does all of it. Every day. For five years.

What the app never says, because it is not equipped to say it, is that the protocol her bones actually need is heavier than walking and harder than yoga. The Royal Osteoporosis Society is explicit: the best way to keep bones strong through and after menopause is both weight-bearing impact exercise and muscle-strengthening exercise.¹¹ Neither is what her plan suggests.

She does not crash. She does not flare. She just quietly loses bone, year by year, doing exactly what the app told her was good for her. Twenty years later, what is gone is gone. Nobody marks it down as a product loss.

Hypertension and beta-blockers. A man in his early fifties starts a popular training app. He has hypertension. He is on beta-blockers. The app does not ask him about either. He picks a “smart” plan, puts on a chest strap, and starts training to heart-rate zones.

The trouble is invisible. Beta-blockers blunt heart rate by design. The American Heart Association is clear that target heart rate may need to be recalibrated with a clinician because these drugs affect everyone differently.¹² Exercise is Medicine adds that on these medications, perceived effort is the safer guide than the number on the watch — and that if resting blood pressure climbs above certain thresholds, the right answer is not to train at all but to contact a doctor.¹³ None of that is in his app.

So he chases zones that do not apply to him. Sometimes he overshoots without realizing. Sometimes he undertrains because the number on his watch tells him he is fine. The app feels personalized. It is not. It is specifically wrong for him, and for several million people in the same situation. Almost nobody counts that damage.

PEM, Long COVID, ME/CFS. A woman in her thirties has had Long COVID for eighteen months. On a good day she can walk to the grocery store and back. On a bad day she cannot stand long enough to make coffee. Her wearable keeps congratulating her: streak achieved, ring closed, move goal exceeded. On the days she pushes for the green ring, she crashes for the next three.

The largest peer-reviewed Long COVID synthesis, published in Nature Reviews Microbiology in 2023, states it directly: “Some physicians, poorly educated in the aetiology and pathophysiology of the disorder, still advise patients to pursue harmful interventions such as graded exercise therapy and cognitive behavioural therapy, despite the injury that these interventions cause and the fact that they are explicitly not advised as treatments.”¹⁴ That is not fringe. That is the field’s own primary source. And the population it applies to is not small: KFF estimates that 17 million US adults — about 7 percent of all adults — currently report Long COVID.¹⁵

An entire class of mainstream wellness logic — step nudges, recovery scores that reward upward progression, “build capacity” framings — is not just unhelpful for this woman and the millions like her. It can actively make people worse, sometimes for years.

This is what omission looks like at scale. The system mistakes the wrong variable for the right one and, because it is still giving some recommendation, it treats itself as safe.

The economics of omission

The honest objection is: yes, but isn’t this all anecdotal? Where is the number?

There is no clean number, because no one counts omission as a category. But the size of the exposed population can be added up directly.

Hypertension affects 119.9 million US adults.¹⁶ Up to 3.3 million people in the US live with ME/CFS, at an annual economic burden the CDC estimates between $18 and $51 billion.¹⁷ About 59 million Americans provide unpaid adult care valued at roughly $1.01 trillion per year — a population whose plans of any kind are systematically misfitted to their available time and sleep.¹⁸ Approximately 3.6 million people enter postpartum physiology in the US every year, a life stage where generic “get back in shape” advice intersects with healing tissue, pelvic floor recovery, severe sleep loss, and elevated mood risk.¹⁹

Summed conservatively across just these four phenotypes — and they are far from the only ones — that is roughly 185 million annual person-phenotype exposures to wellness advice not designed for the person receiving it. The number is not unique people; overlap is heavy. It is the size of the population getting average-user advice in situations where average-user advice is the wrong tool.

We count the risk of saying the wrong specific thing.

We rarely count the risk of saying the wrong generic thing to 185 million people.

Trust is already paying the bill

There is one place where the cost of omission is showing up on a balance sheet: trust.

In Deloitte’s 2024 consumer health survey, the share of consumers who said they did not trust generative AI health information rose to 30 percent, up from 23 percent the year before — a seven-point drop in twelve months, in a category that depends on trust to function.²⁰ You can argue about how much of that is AI-specific versus a broader signal about generic health content. But for a category that markets itself on being more personal, more responsive, and more aligned with the user than traditional medicine, a fall in trust at this speed should be read seriously.

People may not have the vocabulary for omission error. But they can feel when advice does not see them. And they are starting to act on it.

5. The average user is mostly fictional

When we talk about omission, it is tempting to picture an edge case — a small subset of unusual users who happen to fall outside the model. The population that does not fit the “healthy optimizer” frame is not the edge. It is the majority of the people the wellness industry is talking to.

Three-quarters of US adults have at least one chronic condition. Half have more than one. Even among 18-to-34-year-olds — the cohort wellness products most aggressively design for — six out of ten already have a chronic condition.²¹

And of the remaining quarter of US adults with no diagnosed chronic condition, the question becomes how many are actually metabolically healthy. The most recent NHANES-based analysis published in the Journal of the American College of Cardiology found that only 6.8% of US adults have optimal cardiometabolic health.²² Roughly one in fifteen.

Wellness products design their core experience for the optimizer: someone with stable physiology, room to push, recoverable baselines, no medication interactions worth modeling. The data says that person is the edge of the user base, not its center. The actual center is someone managing one or more conditions, often invisibly, often without diagnosis, often without a vocabulary for what their body is doing.

The trend goes the wrong way, too. Between 2013 and 2023, the share of US adults aged 18-to-34 with multiple chronic conditions rose by roughly a quarter.²¹ The “healthy young adult” cohort is becoming a smaller share of the very age group wellness apps target most aggressively.

The average healthy optimizer is not the average. It is the rare exception. And the longer the industry designs as if it were the norm, the larger the omission burden becomes.

6. The bridge: phenotype-level personalization

If the wellness industry’s personalization is mostly tailoring, and “real personalization” sounds like science fiction, then the real product question is the one almost nobody asks honestly: what’s between them?

Why pure n=1 doesn’t work — and not for the reasons people usually give

The fantasy answer is full individual personalization. Every body modeled on itself. No averages. No cohorts. Just you, compared to your own setpoint.

That doesn’t work, and the problem isn’t compute or data. The problem is that pure n=1 leaves the user without orientation. You become your own scientist, running experiments on yourself, with a sample size of one. That isn’t care. It’s a research assignment most people didn’t sign up for.

When you feel worse than you did three months ago, “compared to your own baseline” tells you something has changed. It doesn’t tell you whether what’s happening to you is what happens to bodies like yours. Whether the people closest to your situation got better. What they did differently. What to try next. For most users, “be your own n=1” is not empowerment. It’s abandonment with a dashboard.

This is part of why we don’t think the future of health personalization is “n=1 for everyone.” That framing imports the same problem wellness has now, just flipped: the population-average frame strips away the individual; the pure-individual frame strips away the cohort. Both lose the same thing — the right reference class.

Three honest options

Once you accept that current wellness-personalization-as-tailoring is mostly marketing, and n=1 is mostly a fantasy in 2026, the real strategic question becomes a choice between three things.

The honest middle is phenotype-level.

Why phenotype is the right reference class

A phenotype is not a marketing persona, and it is not a single diagnosis. It is a group of people who are genuinely similar in the ways that matter — symptoms, physiology, function, medication profile, life stage, social reality.

The reason this matters isn’t methodological elegance. It’s that humans learn by analogy to humans like themselves.

Every person who has ever joined a Long COVID Slack, a perimenopause subreddit, an ME/CFS Discord, an IBS support group, a fibromyalgia forum — what they’re looking for is not abstract averages, and not pure introspection. They’re looking for the right reference class. Other people whose bodies are doing something close enough that pattern recognition becomes possible. Recognition is the precondition for everything else: better questions, better experiments, better self-advocacy with clinicians, better predictions.

Phenotype-level personalization is the product version of that human instinct. Instead of comparing you to the population average (which you aren’t) or to yourself alone (which gives you no leverage), it compares you to the cluster of bodies actually doing what your body is doing. Advice can then be both specific and learned — without pretending to be omniscient.

That’s not a smaller idea than n=1. It’s a different idea. It scales. It can be validated. It can be improved one cluster at a time. And it gives the user the one thing pure self-measurement cannot give: other people.

The scale of what wellness ignores

This is where the abstract argument becomes a number.

If wellness products stay in the conservative lane — generic advice, untouched by phenotype — they leave the documented annual cost of every chronic phenotype they speak to fully in place. And those costs are not small.

Hypertension: $131–219 billion in annual US societal cost.²³ Chronic pain: $560–722 billion.²⁴ Long COVID: $30–170 billion.²⁵ Migraine: $13–78 billion.²⁶ Menopause-related lost productivity and medical costs: $26.6 billion.²⁷ ME/CFS: $18–51 billion.¹⁷ Each figure from CDC, peer-reviewed studies, or equivalent authority.

These are not market-size or addressable-opportunity numbers. They are documented annual societal costs — medical care, lost productivity, caregiver burden — already absorbed every year by patients, employers, and the broader economy, while wellness products continue to talk to these populations with logic built for a healthy 28-year-old optimizer.

The omission argument from earlier in this essay is not theoretical. The scale of what it ignores is, at minimum, this.

A worked example: ME/CFS

To make this concrete, here is the arithmetic for one phenotype where every input has a clean public estimate.

The annual societal cost of ME/CFS in the US is $18–51 billion (CDC). Under a “conservative wellness” strategy — generic advice, no phenotype-aware adaptation — essentially none of that burden is reduced. Under a phenotype-cohort strategy with deliberately conservative coverage assumptions, the order of magnitude lands in roughly $81 million to $775 million annually reducible. Under a theoretical “ideal n=1” strategy, $1.2–9 billion. But ideal n=1 is not feasible at scale in 2026, for reasons covered above.

I am not going to defend specific coverage numbers here — that work lives in a separate methodology document where the assumptions, the formulas, and the population-attributable risk math are visible to anyone who wants to argue with them.

The point is not the precision of any single number. The point is that the numbers exist at all — and that the wellness category’s current “be safe by staying generic” stance counts only the cost of being wrongly specific, while leaving the larger half of the equation uncounted.

That is why we are choosing the slow road — phenotype by phenotype, community by community, at the speed our resources allow. Not because it is fashionable, and not because it is easy. Because it is the only path between two failed ideas — generic wellness and impossible individual medicine — that actually adds up.

7. What we can no longer pretend at Welltory

We are not writing this from the outside.

We are part of this industry. And we are not writing this from finished work, either. We do not yet adapt our product well enough for PEM, for menopause, or for migraine + IBS — three of the phenotypes we already see most often in our own user base. We are not the worked example of what comes next. We are someone who has decided to stop being a worked example of what is wrong.

Two things we publicly stop doing.

We stop designing as if our average user is a healthy optimizer. The CDC says 76% of US adults have at least one chronic condition. Only 6.8% have optimal cardiometabolic health. That is not the edge of our user base. That is its center of mass. The healthy optimizer is the edge. From now on, that is how we treat the roadmap.

And we stop calling personalization what is not personalization. Adaptive messaging is not personalization. A score that changes with your sleep is not personalization. Plan A versus Plan B is not personalization. Real personalization in health changes the core logic of the product around the specific body in front of it — not the wrapper, not the copy, not the order of cards on a screen.

The rest is marketing.

8. So here’s what we’re actually doing

I could spend another section laying out principles for “responsible phenotype-aware personalization.” I’m not going to. The obstacles I described above are real, and I’ve spent months trying to find a way through them that doesn’t kill us under the complexity but actually lets us do something. This is what we landed on. It’s not a prescription for the industry. It’s just what we decided for ourselves.

We’re going to work phenotype by phenotype. One at a time. The mechanism is the same each time:

Launch a community cohort. Run the research jointly with the users in it. Build patient-led with medical advisors at the table. Study how their bodies actually work, what guidance actually helps, how the data we collect can support them inside the reality of their diagnosis — and we do all of that while staying inside what regulation actually allows so we don’t break anything we’d need later.

Then we move to the next phenotype and do it again.

We’ve already started. The first community is for people with post-exertional malaise — mostly women in their mid-forties to early sixties who have arrived at energy-limiting conditions through the same physiological pattern. We’ll see what comes out of it.

That’s the plan. Phenotype by phenotype. We don’t know yet if we’ll make it work at scale. We know the version of survival where we build the easier product and call it personalization isn’t a version we’re interested in. So we’re going to try this and see how far it takes us.

Phenotype by phenotype.

—

Jane Smorodnikova

Founder, Welltory

References

1. Iribarren SJ, Akande TO, Kamp KJ, Barry D, Kader YG, Suelzer E. “Designing personalised mHealth solutions: An overview.” International Journal of Medical Informatics (2023). https://www.sciencedirect.com/science/article/pii/S1532046423002216

2. Eric Topol, “Your Lab Tests.” Ground Truths Substack, December 14, 2024. https://erictopol.substack.com/p/your-lab-tests

3. Cowley ES, Olenick AA, McNulty KL, Ross EZ. “’Invisible Sportswomen’: The Sex Data Gap in Sport and Exercise Science Research.” Women in Sport and Physical Activity Journal 29(2):146–151 (2021).

4. NIH Office of Research on Women’s Health, “Sex as a Biological Variable.” Policy statement that basic and preclinical biomedical research has historically focused on male animals and cells, obscuring sex differences in health processes and outcomes. https://orwh.od.nih.gov/sex-gender/orwh-mission-area-sex-gender-in-research/sex-as-biological-variable

5. Stacy Sims, interview on The Dr. Hyman Show, “Women Are Not Small Men! Why Fitness Advice Is Failing Half the Population,” May 13, 2025.

6. Skou ST, Mair FS, Fortin M, Guthrie B, Nunes BP, Miranda JJ, Boyd CM, Pati S, Mtenga S, Smith SM. “Multimorbidity.” Nature Reviews Disease Primers 8, Article 48 (July 14, 2022). https://www.nature.com/articles/s41572-022-00376-4

7. National Institute for Health and Care Research, “Multiple long-term conditions (multimorbidity)” research framework and strategic priorities. https://www.nihr.ac.uk/explore-nihr/specialties/multiple-long-term-conditions.htm

8. Y Combinator, “YC’s Essential Startup Advice.” https://www.ycombinator.com/library/4D-yc-s-essential-startup-advice

9. U.S. Food and Drug Administration, “General Wellness: Policy for Low Risk Devices” — Guidance for Industry and FDA Staff. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/general-wellness-policy-low-risk-devices

10. World Health Organization, World Report on Social Determinants of Health Equity (2025). https://www.who.int/teams/social-determinants-of-health/equity-and-health/world-report-on-social-determinants-of-health-equity

11. Royal Osteoporosis Society, “Exercise for bone health.” Guidance that the best way to keep bones strong is to do both weight-bearing impact exercise and muscle-strengthening exercise. https://theros.org.uk/information-and-support/bone-health/exercise-for-bones/

12. American Heart Association, “How Do Beta Blocker Drugs Affect Exercise?” Notes that beta-blockers slow heart rate and that target heart rate may need to be recalibrated with a clinician. https://www.heart.org/en/health-topics/consumer-healthcare/medication-information/how-do-beta-blocker-drugs-affect-exercise

13. Exercise is Medicine / American College of Sports Medicine, “Exercising with High Blood Pressure.” Includes the do-not-exercise thresholds (systolic >200, diastolic >115) and guidance to use perceived effort over heart rate for those on beta-blockers.

14. Davis HE, McCorkell L, Vogel JM, Topol EJ. “Long COVID: major findings, mechanisms and recommendations.” Nature Reviews Microbiology 21, 133–146 (2023). https://www.nature.com/articles/s41579-022-00846-2

15. KFF, “As Recommendations for Isolation End, How Common is Long COVID?” Issue Brief, April 2024. Approximately 17 million US adults (~7%) currently report Long COVID. https://www.kff.org/coronavirus-covid-19/issue-brief/as-recommendations-for-isolation-end-how-common-is-long-covid/

16. Centers for Disease Control and Prevention, Million Hearts, “Estimated Hypertension Prevalence, Treatment, and Control Among U.S. Adults.” 119.9 million US adults (48.1%) have hypertension. https://millionhearts.hhs.gov/data-reports/hypertension-prevalence.html

17. Centers for Disease Control and Prevention, “About ME/CFS.” Up to 3.3 million US cases; annual economic burden estimated at $18–$51 billion. https://www.cdc.gov/me-cfs/about/index.html

18. AARP Public Policy Institute, “Valuing the Invaluable: 2023 Update — Strengthening Supports for Family Caregivers.” Approximately 59 million US adult caregivers providing unpaid care valued at $1.01 trillion annually. https://www.aarp.org/pri/topics/ltss/family-caregiving/valuing-the-invaluable.html

19. Centers for Disease Control and Prevention, National Center for Health Statistics, “Births: Provisional Data for 2024.” Approximately 3.6 million annual US births. https://www.cdc.gov/nchs/products/vsrr.htm

20. Deloitte Insights, “Build trust in health care gen AI” (Bill Fera, Jennifer A. Sullivan et al.), June 6, 2024. Share of consumers saying “I don’t trust the information” rose from 23% to 30% year-over-year. https://www.deloitte.com/us/en/insights/industry/health-care/consumer-trust-in-health-care-generative-ai.html

21. Watson KB, Carlson SA, Loustalot F, Town M, Eke PI, Thomas CW, Greenlund KJ. “Trends in Multiple Chronic Conditions Among US Adults, By Life Stage, Behavioral Risk Factor Surveillance System, 2013–2023.” Preventing Chronic Disease (CDC), April 2025. https://www.cdc.gov/pcd/issues/2025/24_0539.htm

22. O’Hearn M, Lauren BN, Wong JB, Kim DD, Mozaffarian D. “Trends and Disparities in Cardiometabolic Health Among U.S. Adults, 1999–2018.” Journal of the American College of Cardiology 80(2):138–151 (2022).

23. CDC NCCDPHP, “Health and Economic Benefits of High Blood Pressure Interventions.” Annual US costs associated with high blood pressure estimated at $219 billion (2019); the JAMA AHA series (Kirkland et al.) estimates $131 billion in annual incremental healthcare expenditures across 2003–2014. https://restoredcdc.org/www.cdc.gov/nccdphp/priorities/high-blood-pressure.html

24. Wieder et al., “Economic Costs of Chronic Pain — United States, 2021.” Medical Care (2025). Estimates $722.8 billion annual US cost, including $530.6B medical care and $192.2B lost productivity. https://pubmed.ncbi.nlm.nih.gov/40730349/

25. npj Primary Care Respiratory Medicine, “Economic burden of long COVID: macroeconomic, cost-of-illness and microeconomic impacts” (2025). Estimates annual US lost earnings of approximately $170 billion. Companion estimate from Bartsch et al., Journal of Infectious Diseases (2025): $2B–$31B annually depending on symptom duration. https://www.nature.com/articles/s41533-025-00460-8

26. Bonafede M et al., “Direct and Indirect Healthcare Resource Utilization and Costs Among Migraine Patients in the United States.” Headache: The Journal of Head and Face Pain (2018). Reports total annual US migraine cost of approximately $36 billion. Broader estimates from neurological disease reports reach $78 billion. https://www.ajmc.com/view/study-summary-costs-associated-with-migraine-in-the-united-states

27. Faubion SS et al., “Impact of Menopause Symptoms on Women in the Workplace.” Mayo Clinic Proceedings (2023). Reports $1.8 billion in annual lost work time and $24.8 billion in annual direct medical costs in the US, totaling $26.6 billion. https://newsnetwork.mayoclinic.org/discussion/mayo-clinic-study-puts-price-tag-on-cost-of-menopause-symptoms-for-women-in-the-workplace/

The strangest part is that I never actually planned to work in health.

My career began in information security. I chose it at sixteen because it was a completely new frontier. I moved fast, but I quickly hit a professional realization that shaped everything to follow: once you build the perfect technical architecture, the primary vulnerability in any system is always the human being.

To be an effective CISO (Chief Information Security Officer), you have to become a behavioral psychologist. I remember the immense effort of training people to simply lock their computers when they stepped away. It sounds trivial, but it’s a massive hurdle in human psychology.

Only now do I see how much that was a rehearsal for public health. It is the exact same problem: there is no immediate reward, you are protecting against catastrophic events in a distant future that may not even happen, and because motivation is naturally low, you have to change the environment rather than just commanding the person to “do better.”

When I left InfoSec for startups, I had no idea I would spend the rest of my life doing the same thing—trying to influence human behavior for a safer future, just for biology instead of servers.

Escaping the SaaS Meat-Grinder

I came to Welltory through a profound crisis of meaning. By my thirties, I was tired of the “B2B SaaS” machine—a cycle that often feels utterly disconnected from real human value.

Think about the standard venture cycle: a talented team builds something useful, they package it as a SaaS, go through an accelerator, and investors pour in money while grooming them for a buyout. Often, the early “traction” is just other accelerator alumni buying the product. The company looks successful on paper, a corporation buys it, the VCs and the corporate execs get their bonuses, and the product is quietly shelved or dies a slow death inside the corporation.

Capital moves, but the customer’s life doesn’t improve. It leaves a bitter aftertaste of meaningless motion.

Welltory began with a different question: what is actually worth doing? We all know the answer. It’s what we wish for everyone at every milestone: Happiness and Health.

I looked at the data and saw that while other industries had been transformed by feedback loops, health was lagging behind. We had a metric—Heart Rate Variability (HRV)—and a clear problem: the crushing volume of stress modern life throws at us.

Subscribe now

The Mavericks and the Nerds

We were incredibly naïve back then. We didn’t know how hard it would be to bridge the gap between hard science and consumer habits. But we survived because of the people we brought together.

Here is an amazing fact: the iOS and Android developers who pushed that very first release in May 2016 are still with us today. Our CPO has been here for almost a decade. The designer who created our iconic “liquid” interface—the fluid animation that represents the state of your nervous system—is still here, leading our design team. Our Lead Scientist, the one who built our proprietary science from the ground up, is the exact same person.

In the tech industry, a ten-year tenure is almost unheard of. But people come to Welltory—and stay for five, seven, ten years—because of meaning and ethics.

Our team is a mix of top-tier professionals who reached a point in their careers where they demanded purpose, and brilliant mavericks who simply don’t fit into standard corporate structures. For a long time, I was incredibly proud of the fact that the number of PhDs in our company was roughly equal to the number of people without any higher education at all. We’ve never cared about your pedigree. We only care how deeply you can dig into a problem.

If I had to define our culture, it rests on two pillars: obsessive tenacity and intellectual rigor. We are incredibly different people, but we are united by a shared perfectionism and an absolute refusal to compromise our principles. Working alongside people who truly understand you is a rare, incredible feeling.

Science vs. The Market

But being science-obsessed in a consumer market is painful. A perfect example is sleep phases. We fought so hard against adding sleep phases to the app. From a scientific standpoint, wearable data for sleep phases is notoriously inaccurate and, frankly, not very actionable. We tried to hold the line and do things the “right” way.

But users buried us in demands for it. Everyone else had sleep phases; why didn’t we? Eventually, we surrendered. We learned a humbling lesson: market perception is a reality you have to design for, even when that perception is built on wellness myths. Our compromise now? We give people the familiar features they expect, but we place the actual physiological truth right next to it, hoping the two eventually connect in the user’s mind.

The Crash That Changed Everything

That tenacity finally paid off in a way that nearly broke us.

In late 2021, we released our core version—the one that ultimately made us successful. We were growing at 30% month-over-month. But during the release, we made a tiny, catastrophic mistake in the App Store: instead of a phased rollout, we hit the button for a 100% rollout.

A tsunami of users hit us. Our servers started melting. Our CTO practically lived in the office; instead of letting him sleep, we were calling a massage therapist to work on his back while he kept coding to keep the infrastructure alive. Everything was glitching, things were breaking left and right.

But here is the crazy part: people kept buying.

Despite the bugs, despite the crashes, they saw the value and they supported us. They didn’t let the team break. That was the moment I realized that we weren’t just building an app anymore; we had built something people genuinely needed.

The InfoSec Loop Closes

Today, 17 million people have trusted us with their health data.

As a former InfoSec expert, that number terrifies me. I know that any growing, centralized system eventually becomes fragile. A data breach would destroy our company.

So, we are going back to my roots: we are changing the architecture. We are moving further away from directly handling user data and shifting the heavy calculations directly onto the users’ devices. We are striving for decentralization. We are building safety into the very architecture of the product, because that is the only way to truly protect people.

Beyond “Healthism” and the Fight for Agency

Because we work only for the user, we have to stay independent. This means we often hit the walls of mistrust that the healthcare industry has built for decades.

We are entering a new era where we stop pretending our users are just “relatively healthy people” looking to optimize their fitness. The reality is that more than half of people over thirty-five live with chronic conditions.

The turning point for our company’s transformation happened during an interview I had with a user. He was a classic high-achiever: an active B2B partnerships executive, an athlete, a guy who ran his life at full speed. Then, he got Long COVID.

Suddenly, his world was destroyed. He could barely sit at a computer for one or two hours a day. A 30-minute walk became a grueling ordeal. He told me how impossibly hard it was to find reliable information on how to pull himself out of that dark place. He used Welltory to find his baseline, to pace himself, and to face his new biological reality.

Today, he has clawed his way back to working five hours a day. He transitioned from full disability back to being a capable, active person, even if it’s on a half-time basis. Hearing him say, “It gave me my life back,” left an indelible mark on me.

The prevailing culture in digital health is “Healthism”—the toxic idea that health is entirely a matter of personal discipline, and if you are sick or tired, it’s a moral failing. Wearables scold you for not hitting 10,000 steps, ignoring the fact that your nervous system might be completely depleted.

We are rejecting Healthism. We are moving into medicine, supporting people with energy-limiting conditions, autoimmune diseases, and chronic fatigue.

This is where AI changes the game for us. We aren’t using AI to be a stricter coach. We are using AI to translate complex data into narratives that remove guilt. If you have no energy, your data should validate your biological reality, not demand more discipline. We want to give people insights that help them cope, without the heavy burden of blame.

In our blog, you can also read our message to users.

An Unequal Fight

It is an unequal fight. There is too much in this modern world arrayed against our mental and physical well-being. But it is a fight worth picking, even if you know a “final victory” isn’t the point.

When a user writes to us saying, “I thought your app was wrong, but I went to the doctor and you caught something life-threatening,” or when someone like that executive tells us we helped him get his life back—that is what keeps us going.

I can promise our users one thing: we will always be on your side. We will do everything possible, and impossible, to ensure nobody can force us to act against your interests.

I am deeply grateful to our users. You allow us to live a life of meaning. It has been difficult, often at the very edge of our limits, but it is a life that matters.

Thank you for these ten years.

— Jane

Leave a comment

Some workers manage to bump their scores. Most don’t. The ones who don’t are labeled “non-compliant.” The ones who do still get sick — they just do it slightly more efficiently.

At no point does anyone open a window.

This isn’t a metaphor. This is exactly how the wellness industry operates. And after ten years of running a health tech company, here is the most honest thing I can tell you:

When a system makes people sick and then sells them tools to cope — while whispering that the sickness is their own personal failure — that’s not wellness. That’s victim blaming.

The Lie Inside the Truth

I want to be clear: the wellness industry doesn’t lie about the details. Sleep matters. Movement matters. Breathing techniques actually work. That’s all backed by real science.

The lie is in the framing.

When you’re chronically stressed because your job is precarious, your commute is two hours of gridlock, your food is engineered for addiction, and your healthcare costs a fortune — and then a wellness app tells you to “prioritize recovery” and “optimize sleep hygiene” — something cynical happens.

The cause of your suffering gets relocated. It moves from the system into you.

Suddenly, your stress isn’t a rational response to a broken world; it’s a “management issue.” Your burnout isn’t a systemic failure; it’s a “boundary issue.” If you’re sick, the logic suggests you just didn’t try hard enough. You didn’t track enough. You didn’t meditate long enough.

As Barbara Ehrenreich put it: when wellness becomes an ideology, “every death can now be understood as suicide.”

There’s a name for this. In 1980, sociologist Robert Crawford coined the term “healthism”: the ideology that locates health primarily in personal behavior — in your lifestyle, your habits, your choices — rather than in the social, economic, and environmental conditions you live in. His warning was precise: as long as healthism shapes our thinking, health promotion will remain “non-political, and therefore, ultimately ineffective.”

Forty-five years later, Americans spend $2 trillion a year on wellness. Population health is declining. Crawford was right.

The Data Is Screaming

If wellness apps were a clinical trial — spend $2 trillion, health gets worse — we would have shut them down years ago.

Obesity among young adults nearly doubled in a generation — from 23% to 40% — measured on scales, not self-reported. Depression rose 60%. Fatty liver disease in people under 30 doubled. Physical inactivity went up 5 percentage points during the exact decade that step-counting wearables went from niche gadgets to 600 million devices.

A study across 28 years of NHANES biomarker data — these are trained technicians physically drawing blood and measuring waist circumference, the same way, every cycle — found that physiological dysregulation has increased continuously from Boomers through Gen X through Millennials, in every sex and racial group.

Each generation arrives at adulthood measurably sicker than the last. And each generation has more health data than any civilization in history. More data, worse health. This isn’t a coincidence. It’s the logical outcome of a model that mistakes measurement for care.

Whose Fault Is It, Actually?

Here is the question the wellness industry does not want to ask. Because the answer collapses the entire business model.

Michael Marmot studied British civil servants for decades. People at the lowest employment grade died at three times the rate of those at the top — with identical access to the same doctors, the same hospitals, the same National Health Service. The difference wasn’t diet. It wasn’t exercise. It wasn’t how many steps they tracked. It was autonomy — control over their own work and life. Marmot’s conclusion: “What good does it do to treat people and send them back to the conditions that made them sick?”

Anne Case and Angus Deaton — he’s a Nobel laureate — documented that deaths from suicide, drug overdose, and alcoholism rose from 65,000 to 158,000 per year between 1995 and 2018 in America. That’s three full Boeing 737s crashing every single day, for decades. Not because people lost willpower. Because the economic and social fabric of working-class life was destroyed. “It is the loss of meaning, of dignity, of pride, and of self-respect that brings on despair.”

Robert Lustig, the UCSF endocrinologist, points out that 93% of Americans are metabolically unhealthy. His diagnosis is blunt: “Everyone is talking about healthcare, few people are talking about health, and nobody is talking about the food.” We didn’t hand smokers a cigarette-counting app. We regulated the tobacco industry. But with food — with work culture, with economic precarity, with the systematic destruction of sleep and community — we hand out apps.

Your job. Your commute. The air in your city. Whether you can afford childcare. Whether your boss texts at midnight. Whether you have one person you can call at 2 AM. The WHO has said it plainly: these social determinants are the primary drivers of health outcomes.

No wellness app controls any of them. But every wellness app quietly implies that you should be able to compensate for all of them — through better habits.

The New Solutionism: “The Bot Will See You Now”

And just when you thought the performance model of health had peaked, we are entering its ultimate phase.

As access to actual, living physicians becomes a luxury for the ultra-rich — or a six-month waiting list for everyone else — the system has found its perfect band-aid: AI Chat for Health.

Evgeny Morozov, the technology critic who wrote the book on Silicon Valley’s hubris, has a perfect word for this: “solutionism” — the compulsion to recast complex social problems as neat technical puzzles with app-shaped answers. AI health bots are solutionism at its most cynical. Because it’s becoming nearly impossible to see a real doctor, we are told to just talk to a bot. Here, have an LLM. Tell it about your symptoms. Let it give you more tips on how to fix yourself.

The pitch sounds democratic: world-class medical knowledge, available to everyone, 24/7, for free.

The reality is measurably different. The largest randomized trial on AI medical advice — published in Nature Medicine in February 2026, conducted by Oxford — tested what happens when the same powerful LLMs interact with real patients instead of doctors. The models that correctly identified the right diagnosis approximately 95% of the time when queried directly by clinicians gave the correct answer in less than 34.5% of cases when used by regular people describing their own symptoms.

Read that again. Same AI. Same medical knowledge. But accuracy dropped from near-perfect to barely a third — simply because the person asking wasn’t trained to ask the right questions, provide the right context, or evaluate the answer.

This isn’t a failure of the models. It’s a failure of the premise. We are handing people a powerful, complex tool and saying: “If you’re smart enough to use this correctly, you’ll be fine. If it doesn’t work? Well, that’s on you. You probably didn’t prompt it right.”

Sound familiar? It’s the same pattern. The same relocation of blame from system to individual. Except now it’s not just your diet and sleep — it’s your ability to correctly interface with artificial intelligence during a medical crisis.

We are handing people the world’s most sophisticated tools to help them ignore the fact that the building is on fire.

Why We Stopped Giving “Recommendations”

In our first year, we did what everyone else did. We gave advice. We set goals. And then I saw the reality in our own data.

We’d show someone their HRV crashing for weeks. They’d tell us: “I know, I need to sleep more.” But at 11 PM, after the kids are finally down and the day is finally over, they don’t sleep. They scroll. They stay up.

Not because they’re weak. Because the nighttime version of a stressed person runs on completely different machinery than the morning version.

BJ Fogg at Stanford calls it the “Information-Action Fallacy” — the belief that the right information changes behavior. It doesn’t. Wendy Wood at USC proved that 43% of daily behavior is habitual — automatic, driven by environment, not by willpower. The most rigorous wearable trial ever conducted (JAMA, 471 adults, 24 months) found that the tracker group lost less weight than the group without one.

And the neuroscience explains why. Chronic stress literally shrinks the prefrontal cortex — the part of your brain responsible for planning and impulse control. It grows the amygdala — the part responsible for fear and reactivity. The person who most needs to change their behavior is biologically least capable of doing it.

Byung-Chul Han, the philosopher, names the deeper trap: “The achievement-subject exploits itself until it burns out. Perpetual self-optimization, which coincides point-for-point with the optimization of the system, is proving destructive.” Nobody forces you to optimize — you do it voluntarily. Which makes the exhaustion feel like your fault too.

Giving a stressed person a recommendation is like handing someone with a broken leg a running program. When they can’t complete it, they don’t blame the program. They blame themselves. They feel shame.

And shame is the opposite of health.

So we made a choice. Welltory doesn’t give recommendations. No programs. No “streaks.” No discipline. No guilt. If a product makes a struggling person feel worse about themselves, that product is part of the problem. Period.

What We Do Instead: Translating Biology into Meaning

If we aren’t giving you a “to-do” list, what are we actually doing?

We are building the bridge between human biology and how life feels.

The wellness industry is obsessed with “optimization,” which is just a fancy word for treating your body like a machine that needs to be tuned. We make the invisible visible through metaphor.

Your autonomic nervous system is the lens through which you experience your entire life. When your physiology is in a state of “High Stress” or “Low Recovery,” it’s not just a number on a screen. It’s a filter on your reality.

• It is the biological reason why a minor comment from your partner feels like a personal attack.

• It is the reason why a task that felt easy on Monday feels like an impossible mountain on Thursday.

• It is the reason why the world feels grey, sharp, and overwhelming.

Most apps see a low HRV and tell you to “breathe.” We see your vitals and help you understand why your reality feels the way it does.

We use metaphors—fluidity, turbulence, battery, flow—because biology is too complex for a dashboard, but it’s perfectly suited for meaning. When you can see the connection between your biology and your perception, you gain something far more powerful than “compliance.” You gain Agency.

You realize you aren’t “failing” at your job or your marriage; you are experiencing a physiological state that makes those things harder. When you understand the why, you don’t need an algorithm to tell you what to do. The path becomes obvious.

This isn’t wellness. It’s a way of making sense of yourself. It’s the realization that your “mood” isn’t a mystery—it’s a message from your nervous system. And once you can read that message, you stop being a victim of your biology and start being its partner.

Opening the Window

I should tell it out loud: An app is not a savior.

An app cannot fix the food system. It cannot give you a less toxic boss or an affordable apartment. It cannot replace a community. I refuse to be the person selling breathing apps in a toxic factory while the windows are nailed shut. I refuse to use AI to “prompt” people into believing they can ignore the fire in the building.

But there is a specific, quiet power in seeing.

When we reflect your data back to you as meaning—without judgment, without “goals,” without “streaks”—we are doing the only thing that actually matters in a broken system. We are helping you walk over to the wall, find the latches, and put your weight against the frame.

We don’t optimize the coughing worker. We help the worker see that the air is the problem.

Because once you see the smoke, you stop trying to “breathe better.” You start looking for the exit. Or you start looking for a way to fix the ventilation for everyone.

From that clarity—not from guilt, not from discipline, not from an algorithm—sometimes people make changes. Sometimes they protect their sleep. Sometimes they leave the job. And sometimes, they make the hardest choice of all: they finally believe that they aren’t the problem.

The choice to act is yours. Our only job is to open the window and let you see the air for what it is.

No shame. No blame. No pretending the air is fine.

Further Reading:

Barbara Ehrenreich — Natural Causes: An Epidemic of Wellness, the Certainty of Dying, and Killing Ourselves to Live Longer (2018) A sharp polemic from a writer with a PhD in cellular immunology. Ehrenreich argues that the wellness industry converts systemic social failures into personal moral failures — and that the obsessive drive to optimize health can consume the very life it intends to preserve. The quote “every death can now be understood as suicide” comes from her analysis of healthism as ideology. 🔗 https://www.labyrinthbooks.com/natural-causes/

Anne Case & Angus Deaton — Deaths of Despair and the Future of Capitalism (2020) Nobel laureate Angus Deaton and economist Anne Case document how deaths from suicide, drug overdose, and alcoholism rose from 65,000 to 158,000 per year between 1995 and 2018 among America’s working class. Their conclusion: the cause is the destruction of economic dignity and social meaning — not individual weakness. 🔗 https://books.google.com/books/about/Deaths_of_Despair_and_the_Future_of_Capi.html?id=anK0DwAAQBAJ

Byung-Chul Han — The Burnout Society (2015) The South Korean-German philosopher argues that contemporary society has shifted from external discipline (”thou shalt not”) to internal self-optimization (”I can”). The result: individuals exploit themselves until burnout — and feel personally responsible for their own exhaustion. Source of the “achievement-subject exploits itself” quote. 🔗 https://books.apple.com/us/book/the-burnout-society/id1023741312

BJ Fogg — Tiny Habits: The Small Changes That Change Everything (2020) The founder of Stanford’s Behavior Design Lab introduces the concept of the Information-Action Fallacy — the false belief that giving people the right information will change their behavior. Fogg’s research shows that behavior change is driven by motivation, ability, and prompts — not data. 🔗 https://www.forbes.com/sites/melodywilding/2020/01/13/forget-big-goalswhy-tiny-habits-change-everything-according-to-a-stanford-professor/

Wendy Wood — Good Habits, Bad Habits: The Science of Making Positive Changes That Stick (2019) USC professor Wendy Wood’s research shows that 43% of daily behavior is automatic and habitual — not driven by conscious decision-making. The implication: information-based wellness interventions largely miss the mechanism through which behavior actually changes. 🔗 https://dornsife.usc.edu/wendy-wood/good-habits-bad-habits/

Evgeny Morozov — To Save Everything, Click Here: The Folly of Technological Solutionism (2013) Technology critic Morozov coined the term “solutionism” — the compulsion to reframe complex social and political problems as neat technical puzzles with app-shaped solutions. Directly relevant to AI health bots being offered as a substitute for systemic healthcare access. 🔗 https://www.publicaffairsbooks.com/titles/evgeny-morozov/to-save-everything-click-here/9781610393706/

Robert Crawford — “Healthism and the Medicalization of Everyday Life” (1980) International Journal of Health Services, Vol. 10(3), pp. 365–388 The original paper coining the term “healthism” — the ideology that locates health primarily in individual behavior rather than social, economic, and environmental conditions. Crawford warned that healthism makes health promotion “non-political, and therefore, ultimately ineffective.” Still widely cited 45 years later. 🔗 https://pubmed.ncbi.nlm.nih.gov/7419309/

Crawford revisited his own paper in 2026: 🔗 https://pubmed.ncbi.nlm.nih.gov/41730689/

Marmot et al. — Whitehall II Study (1991) The Lancet, Vol. 337(8754) The landmark cohort study of 10,314 British civil servants showing that mortality and morbidity follow a strict employment-grade gradient — even when healthcare access is identical. Men in the lowest grade died at three times the rate of the highest grade. The key variable: autonomy and control over one’s own work and life, not lifestyle habits. 🔗 https://pubmed.ncbi.nlm.nih.gov/1674771/

Wikipedia overview of the full Whitehall study program: 🔗 https://en.wikipedia.org/wiki/Whitehall_Study

Jakicic et al. — “Effect of Wearable Technology Combined With a Lifestyle Intervention on Long-term Weight Loss: The IDEA Randomized Clinical Trial” (2016) JAMA, Vol. 316(11), pp. 1161–1171 The most rigorous wearable tracker trial to date: 471 adults, 24 months, University of Pittsburgh. Result: the group using wearable activity trackers lost 2.4 kg less than the group without one. Directly contradicts the premise that more data leads to better health outcomes. 🔗 https://pubmed.ncbi.nlm.nih.gov/27654602/

Zheng & Echave — Physiological Dysregulation Across Birth Cohorts (2021) NHANES biomarker data spanning 28 years, across Boomers, Gen X, and Millennials. Found that physiological dysregulation has increased continuously across generations, in every sex and racial group — measured objectively by trained technicians (blood draws, waist circumference), not self-reported. Each generation arrives at adulthood measurably sicker than the last. 🔗 https://www.healthday.com/health-news/general-health/3-25-gen-x-gen-y-in-worse-health-than-prior-generations-at-same-age-2651169.html

Related PMC study on generational cardiovascular health: 🔗 https://pmc.ncbi.nlm.nih.gov/articles/PMC9154229/

Bean, Mahdi et al. — “Reliability of LLMs as Medical Assistants for the General Public” (2026) Nature Medicine, February 2026 — University of Oxford The largest randomized trial of AI medical advice to date: 1,298 participants. The same LLMs (GPT-4o, Llama 3, Command R+) that correctly identified diagnoses ~95% of the time when queried directly by clinicians gave the right answer in less than 34.5% of cases when used by regular people describing their own symptoms. Users with AI performed worse than the control group using Google. 🔗 https://www.nature.com/articles/s41591-025-04074-y

Oxford press release: 🔗 https://www.ox.ac.uk/news/2026-02-10-new-study-warns-risks-ai-chatbots-giving-medical-advice

U.S. Surgeon General — “Our Epidemic of Loneliness and Isolation” (2023) U.S. Department of Health & Human Services The 81-page advisory declaring loneliness a public health epidemic. Key finding: lacking social connection carries a mortality risk equivalent to smoking 15 cigarettes per day — greater than the risks associated with obesity and physical inactivity. The report calls for structural, not just individual, responses. 🔗 https://www.hhs.gov/sites/default/files/surgeon-general-social-connection-advisory.pdf

WHO — Social Determinants of Health (ongoing) The World Health Organization’s framework establishing that the primary drivers of health outcomes are structural factors — income, housing, working conditions, education, access to healthcare — rather than individual behaviors. The foundation for the article’s argument that no wellness app controls the variables that actually shape population health. 🔗 https://www.who.int/health-topics/social-determinants-of-health

Robert Lustig, MD (UCSF) — Metabolic Health Endocrinologist and author of Metabolical. His research supports the claim that 93% of Americans are metabolically unhealthy, and that the primary driver is the ultra-processed food system — not individual willpower or tracking failures. 🔗

https://robertlustig.com

But then, Welltory shifted gears. We started to see our market and users differently and the team needed to absorb an ocean of new information (I wll tell you more about our new strategy later). I realized we needed an internal book. An “Onboarding Domain Knowledge” to get everyone on the same page.

So, I decided to build it. This is the story of that journey.

Step 1: The Visionary Phase

First, you have to invent the book. AI won’t help you here. You can use it as a sparring partner for dialogue, but ultimately, the outline, the core idea, and the structure are on you.

I’m lucky that our strategy was already well-defined, so this part was relatively easy. I mapped out 1–2 paragraphs per chapter. Later, I realized I’d missed a few things and added extra chapters at the end, but the outline was my foundation. Pro tip: Spend time on the outline. It’s the only thing that keeps the AI from drifting.

Step 2: Information Gathering

Writers usually spend years on research. In the AI era, this phase is unrecognizable.

I spent about $1000 on research alone. Why? Because the level of quality I needed only came from Perplexity’s Deep Research, which costs a lot. I became a “prompt engineer” for research queries, using Claude to help me write the prompts that would extract the best science.

I bought a Google Colab Pro subscription just to keep my sessions alive overnight. I’d launch a massive batch of research and wake up to long Markdown documents filled with links. I then had to process these into “knowledge cards”—individual facts, thoughts, and proofs—which I could then feed into each chapter.

The link nightmare: Models love to hallucinate links or mix up internal file paths with real URLs. I solved this by giving the AI only “fact IDs” during the writing process and then adding the actual sources as a final mechanical layer at the very end.

Step 3: Preparing the Context

My first attempts at context were a disaster. I tried using our company strategy and a style guide based on my social media posts. It didn’t work.

I ended up writing a “Constitution of the Book.” I had to fix our basic worldview and core principles in stone. If you don’t, the model will just hallucinate its own “truth” in any direction it wants. I also split my style guide into two parts: my actual voice and an “Anti-Slop Guide” (how NOT to write like an AI).

Step 4: Learning to Write a Chapter (The “Pre-Chewed Salad” Problem)

This was the biggest struggle. I almost burned out here. I tried approach after approach, and everything I produced was garbage.

I am a perfectionist, yes. But if you want something readable, you can’t just “generate” it. One of our scientists gave me the most brutal feedback on an early draft: “It feels like I’m eating a salad that someone else has already chewed for me.”

The AI had no variability. It was monotonous. It used the same patterns (”Not once, not twice, but three times”) and choppy sentences. AI lacks variability (and I know everything about variability). Living things are variable (that’s why heart rate is variable); AI is just a slave to its instructions. The more I tried to control it with rigid agents (Scientific Editor, Literary Editor, etc.), the worse it got. It couldn’t sound “alive.”

The Solution: I realized that instead of describing how a human writes, I needed to show it.

I initiated a “fake dialogue” with the model. I’d feed it a request and then a “fake” result—which was actually a real, high-quality text from an author I admire (like Sapolski). After three of these examples, I’d give it my real request.

I spent a week building a library of 500+ “example” pairs—real, living texts from real humans matched to specific thoughts and directions. This gave the AI a structural and rhythmic foundation that actually felt human.

Step 5: Generation and Control

If you think you just hit “Enter” and wait, think again. I had to be involved at every step on every chapter.

1. The Lens: I’d discuss the chapter’s “lens” with a specialized Claude skill—what do we really want to say here?

2. The Task: Claude Opus acted as my “Scientific Editor.” It would select the research cards and then challenge me. It would ask me tough questions about the chapter’s insights, and I’d have to provide my own feedback and commentary.

3. The Draft: Only then would GPT-5.4 Pro start writing the base of the chapter using my feedback, all research and the exemplar library. First I try write with more cheap models by chunks but it’s better toi use most expensive models with big context in one step to get coherent result of the chapter.

4. The Finalization: I’d check the draft, verify the conclusions, and then send it to Gemini Flash for “polishing”—making the language simpler and clearer without changing the substance. Google models are best for this, probably because htey have more data.

Step 6: The Final Edit

Even after all that, I spent days on edits. AI still misses things. It repeats “filler” words, breaks formatting, or messes up abbreviations. I had to generate a custom glossary for every chapter so the technical terms remained consistent but weren’t over-explained.

The best part? Generating the illustrations and seeing the book finally come together. It’s 300 pages of fresh, unique information—the world’s latest scientific knowledge mixed with our specific strategy and market insights.

My “Obsessed Founder” Takeaways:

1. Examples > Instructions: Just like with people, showing works better than telling. Don’t just prompt; provide examples.

2. The Context is Everything: The more “obsessive” you are about preparing the context and “Constitution,” the better the result.

3. Use the Whole Toolbox: I had to use every model on the market to get this right:

   • GPT-5.4 PRO: For the core writing.

   • Claude Opus: For the high-level task setting.

   • Claude Sonnet: For structuring and link processing.

   • Perplexity: For the deep research.

   • Gemini Flash: For linguistic polishing.

   • ChatGPT Image: For the visuals.

The world has changed. Every piece of knowledge on the planet is at your fingertips. The only thing that differentiates us now is our level of perfectionism and having a unique vision that can point the AI in the right direction.

The AI is the engine, but you are still the driver.

I learned so much on this journey. I’m just grateful I stuck with it and delivered something real, rather than giving up and dumping a boring 'research collection' on my team.

This is a brand-new skill set for a brand-new era. It fundamentally shifts our approach to work; things that used to be too slow, too expensive, or simply out of reach are now possible. It’s incredible how much the game has changed.

I was sitting in an internal Welltory workshop. One of our engineers was showing the team how to build plugins for Claude Cowork. It’s a powerful thing: you can automate a scary amount of your own work. People were watching and you could almost hear the collective gulp — oh, shit, this just got real.

You could see the gears turning:
“Wait. I can wrap my whole operational hell into a couple of pipelines and never touch it again?”

At some point someone joked:

“If we had time tracking here, we could now fake being busy forever and do nothing.”

My first reaction was a small internal panic.
The second was relief.

Thank God we never tried to control hours.

The next thought was less kind:
Oh wow. So many companies are already screwed and don’t know it yet.

For years people kept telling me:

• you can’t do real innovation remotely

• you have to watch people

• you need to “look them in the eyes” in the office

• remote only works for juniors who need supervision

Now, in 2026, it’s pretty obvious what actually matters for any company building an intellectual product — and what doesn’t. And everyone who bet their culture on controlling presence and screen time just lost that bet.

Let’s rewind a bit and see how we got here.

We were paying for theater long before AI showed up

There’s an ugly fact we have to admit: the “activity instead of results” problem is not new.

A Connext Global study found that about two thirds of U.S. employees openly admit they engage in productivity theater: they pick up “visible” tasks, stay late, and stay active in chat mainly to look busy. Only around 23% say their performance is measured against clear outcome metrics, not how hardworking they appear.
Sources:

• Summary press release

• BusinessWire piece

So tens of thousands of companies have been paying not for work, but for rituals around work: speed of email replies, number of calls, how pretty the slide deck looks.

And then AI walked in.

A culture built on hours meets AI

Picture a company where the whole culture is built on time:

• efficiency = how many hours you “sat at work”

• processes are built around presence, not results

• management is basically surveillance: who arrived when, who stayed how long, who is green in Slack

Now drop modern AI into this setup:

• Claude / coding agents that can hold basic work conversations and operate tools for you

• agents that move files, create tasks, draft reports and summaries

• tools that fill in forms and documentation better than the most patient junior

(Examples of these:
Cursor & Claude agents overview
Claude agents breakdown

The whole “control by hours” logic collapses in one release.

Mouse movement? Old‑school mouse jigglers were already a thing years ago. Employers now publish long guides on how to detect artificial cursor movement by speed and acceleration patterns:

• VanHack on spotting mouse jigglers

• EmpMonitor on detecting jigglers

On top of that, you can glue an AI agent that moves the mouse like a human and “uses” apps all day.

Being active in chat? A bot can answer “got it, will check” and “let’s talk tomorrow” just as well as half your org.

A 10‑page report by 6 p.m.? What used to be a full day of work is now ten minutes of prompting. Field reports and vendor data show double‑digit productivity gains from AI on writing, research, and coding tasks, especially in knowledge work.
Overview example.

If your main metric is visible activity, AI is a machine for synthetic activity.
Cheaper, faster, and tireless.

Life inside a “visibility corporation”: the numbers

Let’s put some numbers on this.

• Microsoft’s Work Trend Index shows employees spend up to 57% of their time in meetings, email, and chat — and only 43% on focused work that produces value.
  See here and at a summary

• Meeting time has more than doubled since 2020, but average meeting length barely shrank.
  Nice digest here.

• Focus research from Gloria Mark’s group finds that after a serious interruption, the brain needs about 23 minutes to fully get back into the task. One badly placed meeting in the middle of a deep work block can wipe out half an hour of real productivity.
  Classic paper: “The Cost of Interrupted Work: More Speed and Stress”

On top of that, “bossware” is quietly spreading: screen‑capture tools, keyloggers, app trackers. Legal and HR folks are already writing whole reports about it — and they all come to the same conclusion: surveillance erodes trust, spikes stress, and increases turnover intentions.
Примеры:

• On “bossware” in general

• On AI employee monitoring

• Academic work on monitoring, trust, and turnover

Against this backdrop, “let’s strengthen control” is not a plan.
It’s the corporate equivalent of treating a migraine with a hammer.

The unwinnable arms race

The “we’ll just watch them more closely” path sends your company into a pointless arms race:

• on one side, monitoring vendors invent ever more sensitive ways to flag “abnormal” patterns

• on the other, AI agents get better at mimicking human behavior and producing perfect “busywork” traces

Examples of how fast agent capabilities are evolving:

• Cursor agents running 100+ parallel tasks

• Security folks red‑teaming coding agents and showing how they can be scripted

Who wins a game of “who’s sneakier”?
The side that has AI and motivation to use it. That’s not your middle manager writing reprimands because “you were red in Slack for too long”.

Meanwhile:

• faking activity gets cheaper

• separating signal from noise gets harder

• managers drown in dashboards about “engagement” and “time in app” and still have no idea what actually moves the business

Worklytics and others keep pointing out this trap: leadership obsesses over activity metrics instead of outcome metrics, and then wonders why nothing moves.
Example KPI piece:
Productivity benchmarks:

It’s not just pointless. It’s expensive. You’re burning money on:

• control software

• payroll for people playing cat‑and‑mouse with that software

• and you’re not investing in the only things that actually raise productivity: focus time, sane AI‑supported workflows, and clear goals

Deep‑work stats for context:

• Deep work & focus time trends

• Knowledge worker productivity & focus time

The companies that look strangely calm

If you look at the few places that seem oddly calm about AI, they have one thing in common:
they shifted to managing by outcomes long before the current model hype.

Netflix famously has no vacation tracking and minimal formal rules. Their culture memo basically says: “No vacation policy. No approval process.” The flip side is brutal clarity on performance and “talent density”: you are trusted like an adult and measured like an adult.
- Netflix culture memo
- Book / context
- A good summary

GitLab runs one of the largest all‑remote organizations in the world with a fully public handbook. They don’t track hours; they track impact and documented outcomes. Their leadership talks openly about the trap of rewarding people for presence and responsiveness instead of real contribution.
- All‑remote culture
- Remote benefits
- Nice external case write‑up

The older ROWE (Results‑Only Work Environment) experiments at Best Buy showed the same pattern: when they stopped tracking time and switched to pure results, voluntary turnover dropped dramatically and productivity jumped.
Intro to ROWE:

• Overview

• Case discussion

Productivity analytics across thousands of knowledge workers shows a simple thing: the people who reliably get 3–3.5 hours of deep focus per day outperform those who live in “meeting‑chat‑meeting” mode.
Benchmarks here.
Deep work stats here.

Bossware doesn’t create deep focus. It destroys the little that’s left.

So no, the answer to AI is not “more screenshots”.

OK, but what do you actually do instead?

“Don’t track hours, manage outcomes” sounds nice on a panel.
It’s less nice at 10 a.m. on Monday when you have a real company to run.

Here’s what I’d do if I were trying to turn a time‑tracking culture around.

1. Translate your company into outcome‑speak

As long as your org speaks in tasks — “run the project”, “support the system”, “do marketing” — any talk about trust will stay abstract.

Outcomes look more like this:

• not “handle support”, but keep NPS above X and first‑response time under Y

• not “do marketing”, but bring in X qualified leads at ≤ Y CAC

• not “write code”, but move retention, ARPU, or CSAT by X points this quarter with specific product changes

Worklytics and similar firms publish KPI lists for knowledge workers that are all about contribution and impact, not about time in front of a screen:

• KPI examples

• Benchmarks

They’re a good starting point if you’re staring at a blank page.

2. Build AI into the job, not as a side hobby

If you treat “real work” as manual and AI as a cheat code, you’ve already lost.

Instead:

• write into role descriptions: “can design workflows that use AI/agents, not just do tasks by hand”

• ask in performance reviews: “Which parts of your work have you already automated? What’s next?”

• measure people not by “time spent”, but by how much output improved versus the previous period

McKinsey, IDC and friends keep repeating this point: the winners won’t be “the ones who bought AI”, but the ones who rewired their work around humans + agents working as a system.
Some representative takes:

• IDC: “AI agents as instruments, not co‑workers”

• McKinsey post on agentic AI giving employees “superpowers” (via FB share)

• a16z on AI workflows

3. Invest in outcome‑thinkers, even if they’re AI‑illiterate today

This is where many leaders make the wrong cuts.

If you have someone who:

• genuinely cares about the business result

• thinks in outcomes and takes ownership

• but doesn’t yet know how to work with AI,

this is your best investment.

The simplest training pattern I’ve seen work:

1. Give them a task that can’t reasonably be done without AI — not a toy prompt, but a real piece of work that would be too big or too tedious by hand.

2. Say out loud:
   “The first time you do this with AI, it’ll be slower than doing it manually. That’s okay. The point is to build the muscle for the next ten tasks, not to beat your current speed on day one.”

3. Show where the help lives: internal AI champions, workshops, a short curated list of courses or docs.

OECD and the IMF both point out that companies actively investing in AI‑related skills are more likely to see productivity gains and less likely to lean purely on job cuts:

• IMF blog on skills and AI

• OECD on AI & SME workforce skills

In SME surveys, significantly more firms say AI has raised skill requirements than lowered them.

If someone can already think in outcomes, they almost always figure AI out — especially if you stop punishing them for not being “faster than manual mode” on day one.

4. Be honest: you will part ways with some people

Here comes the hard part.

You will almost certainly have to let go of some folks:

• those who refuse to think beyond checklists (“just tell me exactly what to do”)

• those who see AI only as a threat and won’t touch it

• those whose core skill is “being around and looking busy”

This isn’t about IQ or elite degrees. It’s about a shift in what the job is. From “execute instructions” to “hold the goal in your head and assemble a system — of people and machines — that gets there”.

OECD and IMF both point to a growing demand for hybrid roles where humans set direction and make calls, while routine execution gets eaten by software:

• IMF again

• IOE/ILO style overview of AI and work

If your org is full of people who only want the execution part, AI will come for that.

Trying to preserve a workforce that was bred for process and presence control is, in 2026, the HR equivalent of fighting to keep your fax department in 2005.

5. Realize what the talent war is now about

This, to me, is the core of the whole thing.

The war for talent is shifting toward people who:

• see the end business outcome, not just the next ticket

• can break that outcome into steps, some for themselves, some for AI/agents

• accept responsibility for the outcome, even if half the work was done by non‑humans

You can’t start competing for these people “later, once things settle down”. By then they’ll already be working somewhere that gave them trust and interesting problems.

Right now:

• AI tools are basically electricity — everyone can plug into them
  (good overview of how ubiquitous they already are)

• The real differentiator is not “who has ChatGPT”, but who has people who can orchestrate AI into working systems
  (see e.g. a16z)

• The companies that are already stripping out time‑tracking and rebuilding around outcomes are going to look freakishly efficient in a couple of years next to those still buying fancier bossware and wondering why the margin doesn’t move.

A personal note

At Welltory we don’t track hours.
We have unlimited paid time off, sick days, and weekends.
We measure a person’s value by their contribution and what they do for the team — not by how long they sit in Zoom.

So when someone in our workshop jokes:

“With time tracking, we could now stop working and just let AI fake it,”

we can laugh. There’s nothing to “cheat” in our time‑tracking system. There isn’t one.

For companies built on control, it’s not a joke.
It’s a business risk and a cultural dead end.

Because at some point you have to admit:

• you’ve been optimizing the wrong thing for years

• AI just poured rocket fuel on that mistake

• and now you either rebuild around outcomes and trust, or watch your own employees train agents to fake work better than they ever could themselves

If you’re a founder or a leader, your choice is pretty stark:

• double down on control — and lose to your own AI in your own game

• or double down on outcomes and trust — and start competing for the people who can actually keep your company alive in an AI world

I know which side I’m on.

I spent two years paralyzed by AI uncertainty. Then I recognized the pattern — and started planning again.

Two years ago, unit economics was the only religion. Founders who couldn’t explain their path to profitability got shown the door. VCs wanted to see your burn multiple, your LTV/CAC, your runway math.

Now? Mention profitability in a pitch and watch the room go quiet. The only word that opens wallets is “AI.”

I’ve been at this long enough to find it both hilarious and exhausting. The cycles. The amnesia. The absolute certainty that this time the rules have changed.

They haven’t.

The Crowd Is Always Late. That’s Why I Trust Patterns, Not Hype.

Let me tell you a story that captures how this works.

When we started building HRV analysis at Welltory, VCs passed because their “expert advisors” said heart rate variability will not become a popular metric.

Then Apple put HRV on the Watch. Oura made it their headline metric. The thing that was “niche” became the hottest category overnight.

And now? HRV is everywhere. But now we think there are better options and one more time VCs say: “HRV is a must”. The crowd is late all the time because it’s driven by FOMO, not by insights.

This year, for the first time since ChatGPT dropped, I can actually plan for years ahead. Not because the fog lifted — but because I finally recognized what kind of fog it is. It’s liberating. So, I invite you to get a look at something that is very obvious when you search for patterns.

All Bubbles Crash at Stage 4. AI Is Entering Stage 4.

Economists figured out the anatomy of bubbles decades ago. Minsky, Kindleberger — they mapped the same five stages that show up in every bubble from tulips to railroads to dot-com to crypto:

Stage 1: Displacement. A genuinely new technology shows up. This part is usually real.

Stage 2: Boom. Prices go up, but rationally. Early people make money. Smart money pays attention.

Stage 3: Euphoria. “This time is different.” Valuations detach from anything resembling reality. Everyone’s a genius.

Stage 4: Profit-taking. Smart money quietly walks toward the exits while the music’s still playing.

Stage 5: Panic. Everyone realizes they need to leave. The door is not big enough.

Where are we now? Entering Stage 4. I’ll show you why.

The frustrating part: we know this. Reinhart and Rogoff studied 800 years of financial crises and found the same four words echoing through every single one: “This time is different.”

It never is.

Keynes had a darker take: “It is better for reputation to fail conventionally than to succeed unconventionally.” Which is why fund managers pile into the same overpriced deals. If everyone loses together, nobody gets fired. Miss the next Google alone, and you’re done.

There’s a more hopeful frame too. Carlota Perez argues that major technologies go through a “frenzy” phase of crazy overinvestment, then a painful turning point, then a “deployment” phase where the technology actually becomes useful. The internet bubble popped in 2000. The internet didn’t go away — it became the backbone of modern life.

AI will follow the same arc. The question isn’t whether there’ll be a correction. It’s who gets hurt when it happens.

Smart Money Is Exiting. Here’s How You Can Tell.

I’m not going to predict timing. Nobody can. But I can show you what the data says about which stage we’re in — and why it matters for your decisions.

$300B Spent on AI Infrastructure — With No Proof It Will Pay Off

According to the Financial Times, Amazon, Microsoft, Alphabet, and Meta are targeting over $300 billion in AI infrastructure spending for 2025 alone.

Webscale market data shows hyperscaler capex hit ~$304 billion in 2024 — up 56% year-over-year, a record driven by GenAI and data center expansion.

The money is being spent. The question nobody can answer yet: where’s the return? When infrastructure gets built at this pace without clear ROI, the market becomes extremely fragile. Any signal that the payoff isn’t coming triggers panic.

Someone Is Inflating This Bubble on Purpose. Follow the Money.

Here’s something that should make you pause.

In September 2025, NVIDIA announced plans to invest up to $100 billion in OpenAI to finance massive data center construction. Sounds impressive until you look at the structure.

NVIDIA gives OpenAI money. OpenAI uses that money to build data centers. Those data centers need GPUs. Who sells the GPUs? NVIDIA.

OpenAI’s CFO Sarah Friar said it plainly: “Most of the money will return to NVIDIA.”

This is vendor financing — the company selling you the product is lending you the money to buy it. It creates the appearance of demand while the money circulates within a closed loop.

We’ve seen this movie before. In the late 1990s, Cisco extended credit to customers so they could afford more routers. When those customers went bust, so did the loans. Cisco’s stock dropped 85% and took 20 years to recover.

The NVIDIA-OpenAI dance is more sophisticated, but the underlying dynamic is similar. Bloomberg described this as a “circular economy” where money flows between a small group of interconnected players:

• Microsoft invests billions in OpenAI and provides Azure cloud credits

• OpenAI uses Azure (driving Microsoft’s cloud revenue)

• Microsoft needs more GPUs for Azure → buys from NVIDIA

• NVIDIA invests in OpenAI → OpenAI buys more GPUs

Everyone’s revenue goes up. Everyone’s stock goes up. But the underlying cash is just... circulating. And if one player stumbles, the interconnections amplify the damage across the entire ecosystem.

This isn’t fraud. It might even work out. But it’s a structure that concentrates systemic risk — exactly what regulators are starting to notice.

OpenAI Has No Idea How to Break Even. They Admitted It.

OpenAI’s growth is legitimately extraordinary. Revenue went from $28 million in 2022 to a projected $13 billion in 2025.

Here’s the problem: they’re spending $1.69 for every dollar they earn.

According to Fortune, OpenAI expects to lose around $9 billion in 2025 on $13 billion in revenue. By 2028, they project $74 billion in operating losses — in a single year. Cumulative losses through 2029? Somewhere between $115 billion and $143 billion.

Deutsche Bank put it bluntly: “No startup in history has operated with losses on anything approaching this scale.”

The most telling detail came from Sam Altman himself. On X, he admitted that OpenAI is losing money on its $200/month ChatGPT Pro subscription. “People use it much more than we expected,” he wrote.

Think about that. A $200/month subscription — expensive by any consumer software standard — and they’re still underwater on it. Because inference costs for frontier models are brutal, and power users will use every token they can get.

This isn’t a criticism of OpenAI’s technology. It’s an observation about economics. When your product gets more expensive to deliver the more people use it, you don’t have a software company. You have a very sophisticated utility with negative margins.

VCs Are Hoarding Cash “Just in Case.” That’s a Warning Sign.

The FT reports that AI startups in the US raised a record ~$150 billion in 2025 — higher than the previous peak of ~$92 billion in 2021. The explicit logic? “Build a fortress balance sheet in case 2026 gets ugly.”

Crunchbase data shows AI captured nearly 50% of all global VC funding in 2025, up from 34% in 2024. Total: $202 billion into AI alone.

When the smartest players in the room are stockpiling cash “just in case” — that tells you something about what they expect is coming.

When Employees Sell $6.6B at Peak Valuation, They Know Something.

Reuters reported that OpenAI hit a ~$500 billion valuation through a secondary sale where employees and former employees sold about $6.6 billion in shares. The company reportedly authorized over $10 billion in secondary sales total.

Secondary transactions aren’t bad by themselves. But a flood of secondary at peak hype? That’s a classic mechanism for early holders to transfer risk to later buyers.

The people who built OpenAI are cashing out. At record valuations. While telling everyone else to buy in.

CEOs Are Selling. Retail Investors Are Buying. Guess Who Loses.

Don’t listen to what executives say. Watch SEC Form 4 filings — what they actually do with their own shares.

NVIDIA’s CEO Jensen Huang has sold $1.9 billion+ in stock since 2024, executing sales consistently through 10b5-1 plans at peak valuations.

The aggregate picture: insider buy/sell ratios have dropped to 0.29 in 2025 — well below the historical average of 0.42-0.50. For every dollar insiders spend on their own stock, they’re selling $3.44.

Now — important caveat. Insider selling can mean lots of things: diversification, taxes, estate planning. Academic research shows aggregate insider activity has some predictive value for market returns, but it’s not a timing tool.

I’m not saying “crash imminent.” I’m saying: this is what Stage 4 looks like. Early holders are converting paper to cash while the music plays.

The Future Isn’t Giant Cloud Models. It’s AI on Your Phone.

Here’s where the story gets hopeful — if you’re paying attention.

While everyone debates whether we’re in a bubble, the actual product direction of AI is becoming clear. And it’s not what the hype suggests.

Apple and Google Are Building AI That Runs on Your Phone

Apple’s documentation is unambiguous:

“The cornerstone of Apple Intelligence is on-device processing.” — Apple Support

Their Private Cloud Compute is positioned as a fallback for complex tasks, with strong privacy guarantees. The default is local.

Android’s Gemini Nano is explicitly about processing “without sending your data off the device.” Google’s developer documentation makes the on-device emphasis clear.

Microsoft’s Phi-3 technical report describes their 3.8B parameter model as “small enough to be deployed on a phone” while maintaining strong reasoning capabilities. Their research page literally says “locally on your phone” in the title.

Meta’s Llama 3.2 release explicitly targeted edge and on-device deployment. The 1B parameter model card is designed for exactly this use case.

The point: This isn’t analysts speculating. The companies that control the platforms — iOS, Android, Windows — are building their AI strategies around on-device inference. Privacy, latency, cost. Those aren’t buzzwords; they’re product requirements.

A $6M Model Now Matches a $100M One. Moats Are Disappearing.

This isn’t ideology. It’s economics.

If your product is a daily-use app — health tracking, financial assistant, customer support, B2B workflow — then:

• Latency matters more than +3% on benchmarks

• Privacy is a feature, not a constraint

• Inference cost becomes your gross margin

Andrew Ng has been pushing this for years. His framing of data-centric AI:

“The discipline of systematically engineering the data needed to build a successful AI system.”

The insight: most industries need custom models trained on their own data, not “one model to rule them all.”

DeepSeek proved this dramatically. Their V3 technical report documents training at roughly 2.788 million H800 GPU hours — a fraction of what frontier Western models reportedly require. Reuters covered their claim of a theoretical 545% daily cost-profit ratio on inference.

What matters for founders: when “good enough” quality becomes available at dramatically lower cost, the pricing power of expensive inference providers collapses. Business models shift toward compression, hybrid architectures, and workflow integration.

The model itself stops being the moat.

When Bubbles Pop, Founders Lose Everything. VCs Keep Their Fees.

Here’s what I wish someone explained to me earlier: in bubbles, the architects rarely suffer. The pain flows downhill.

The Teflon Layer

Microsoft structured its OpenAI deal carefully. They hold licensing rights to the technology. If OpenAI can’t pay its Azure bills, Microsoft — as major creditor and license holder — is well-positioned to acquire the technology through the debt. It’s not a bet on OpenAI’s independence. It’s a call option with downside protection.

NVIDIA generates $25+ billion in cash per quarter and has a $60 billion buyback authorization. If their stock drops 70% (like Cisco in 2000), the company doesn’t die. They buy back shares cheap, wait for the next cycle. Jensen has done this before — with crypto, with gaming.

The people holding the bag are retail investors who bought at the top, and pension funds with exposure to inflated valuations.

“Successful Exit” Can Mean Zero for Founders. Here’s the Math.

When you read that a startup raised at a $1B valuation, it’s natural to assume the founders got rich. Usually they didn’t. The valuation is mostly fiction — a number that sets share price but says nothing about who gets paid when things go sideways.

Late-stage investors increasingly demand liquidation preferences: contractual rights to get their money back (often 2x or more) before anyone else sees a dollar.

Quick example: Company raises $100M with 2x liquidation preference. Bubble deflates, company sells for $150M. Looks like success, right?

Investors are owed $200M (2x their investment). There’s only $150M. Investors take it all. Founders and employees get zero.

Years of work. Nothing.

Employees with options. Same math, worse leverage.

Pension funds and endowments. VCs manage other people’s money — teachers’ retirement, university endowments. When funds chase AI deals at 100x revenue, they’re betting those assets. The general partners already collected their 2% management fee. The LPs discover the damage years later.

“Wrapper” startups. Thousands of companies whose product is basically an API call plus a nice interface. They die the moment Apple or Microsoft builds the feature natively.

The VC Defense Is Already Written

Here’s how the post-mortem will read:

“Venture capital is a business of outliers. We had a fiduciary duty to participate in the AI opportunity. Yes, many investments didn’t work out, but that’s inherent to the asset class. The ones that succeeded will return the fund.”

They’ll pivot to quantum computing or synthetic biology or whatever the next cycle brings. The machine keeps running. The losses are always “normal” in hindsight.

I’m not saying VCs are bad people. I’m saying their incentive structure is designed to survive bubbles, not prevent them. That’s not a criticism. It’s just a fact founders need to understand.

Once You See the Pattern, You Can Finally Plan Again.

Okay, enough doom. Here’s where the story turns.

AI Is Not “The New Internet.” It’s Infrastructure.

The internet parallel is dangerous because it implies “winner take all” dynamics — that some AI company will become the next Google, and everyone else will fade away.

I don’t think that’s what’s happening.

The better parallel is the first machine learning revolution in the 2000s and 2010s. ML didn’t produce one dominant company. It became invisible infrastructure embedded in every application. Recommendation engines, fraud detection, search ranking, spam filters — ML is everywhere now, and mostly you don’t notice it.

The winners of that era weren’t the companies with the biggest models. They were the ones with the best data, the best evaluation frameworks, the best integration into actual workflows.

The same pattern is emerging now.

Where Moats Actually Come From

If models are commoditizing — and the evidence says they are — then where does defensibility come from?

1. Data and data rights. Unique datasets that can’t be easily replicated.

2. Evaluation capability. The ability to distinguish quality from hallucination, tied to business outcomes.

3. Workflow integration. Embedded in how people actually work. High switching costs.

4. Distribution. Channels and brand that don’t depend on a single platform.

5. Cost engineering. Controlling latency, COGS, reliability at scale.

What’s not a moat:

“We also have an LLM chat interface.”

If You Can’t Answer This Question, You Won’t Survive the Crash.

I’ll write a separate piece on specific strategic bets. For now, here’s what matters most.

The One Question

What happens to your company when you can’t make unit economics work — and venture capital goes back to demanding profitability?

If you don’t have an answer, you’re gambling with years of your life.

The Opportunity Is Clarity

I spent two years feeling destabilized by AI uncertainty. Every few months a new model dropped and I’d wonder if our whole roadmap was obsolete.

What changed wasn’t the technology. What changed was recognizing the pattern.

My message is simple: you can finally forecast again.

The hype will fade. AI will remain — as infrastructure, like electricity, like the internet post-2000. The winners won’t be whoever raised at the highest valuation or chased the biggest model. That path doesn’t make anyone happy. Building something meaningful and lasting does.

Happy New Year. Build your own clarity — and invest in that, not the crowd. At worst, you’ll learn something real. At best, you’ll build something that matters.

I’ve followed the Physical series since the first season. But watching Adiyasuren and Enkh-Orgil work through those challenges felt less like athletic competition and more like watching founders navigate a market where everyone else has more capital and better connections. They compensated with something harder to build: strategic thinking under extreme constraint.

Elite Sport and Founder Life Share More Than You’d Think

When I trained in kickboxing, I was surprised by how naturally I connected with my coach — a professional MMA fighter. We spoke the same language even though our contexts were completely different. Pushing through limits when your body is screaming to stop. Understanding that recovery isn’t weakness, it’s part of the strategy. Knowing that brute force alone gets you nowhere when everyone at the top has similar capabilities.

The best founders operate the same way. They understand resource management the way an athlete understands energy systems. They know when to push and when backing off prevents catastrophic failure. They develop tactics that compound over time rather than just throwing everything at the immediate problem.

This is partly why I think unlimited funding often damages startups. When you can throw money at every mistake, you never develop the instinct for efficient resource allocation. You don’t learn to feel your constraints. In elite sport, nature enforces those limits. Your body will shut down. Your technique degrades under fatigue. No amount of confidence overrides biology.

The market works the same way, just slower.

Australia’s Mistake Was Confidence in Overwhelming Force

Team Australia was probably the most physically powerful team in the finals. In the last challenge, they leaned almost entirely on that advantage. They were so confident in their raw capability that long-term strategy seemed unnecessary.

They lost.

Team Mongolia showed up with tactical creativity layered over solid fundamentals. When Enkh-Orgil approached the castle gate challenge, he didn’t try to muscle through — he read the mechanics and found the efficient path. The Japanese team, used to relying on power, didn’t see it. Australia didn’t see it either.

I watched the same dynamic play out during the recent venture winter. Startups with inflated budgets and easy capital access collapsed in sequence. Founders from top-tier schools who’d worked at prestigious companies — people used to doors opening based on credentials — simply didn’t have the mental models for operating under real constraint.

The strategies that come naturally when you’ve always had to stretch every dollar, make one engineer do the work of three, negotiate payment terms because missing payroll isn’t hypothetical — those are the strategies that win when easy money disappears.

The Math on Women That Nobody Bothers to Check

In the pole-holding challenge, Team Mongolia’s decision to field Adiyasuren was tactically brilliant. Not just because she performed well, but because it demonstrated pattern recognition that creates asymmetric advantage.

Women get systematically underestimated in physical competition. Research analyzing over 5 million ultra-running results found that female runners are faster than their male counterparts for distances longer than 195 miles. In ultra-distance swimming, the best women have been 12 to 14 percent faster than the best men on average over the past 30 years in events like the 45.8km Manhattan Island marathon swim.

The physiological reasons are documented: women have more Type I muscle fibers which produce less force but are significantly more fatigue-resistant, better fat oxidation rates that preserve glycogen stores, and reduced inflammation and muscle damage allowing faster recovery from extended exercise.

But the cultural narrative stays locked on sprint distances where male physiology dominates. So investors see a female founder and pattern-match downward. Competitors underestimate. The woman who learned to operate without anyone expecting much from her develops capabilities that people with easier paths never needed to build.

Team Mongolia had the most diverse composition in the finals and used it as a tactical asset, not a diversity statement.

What I Miss In Western Competitions

Asian athletic shows maintain something that gets lost in most Western productions: you can push humans to absolute physical limits while maintaining basic dignity. You bow. You acknowledge your opponent’s capability. Win or lose, nobody’s screaming insults in someone’s face or performing dominance.

This matters more than it seems on the surface. Founder culture often mirrors the same toxic competitive framing — the “crushing it” language, zero-sum positioning, casual cruelty dressed as ambition. But the companies that survive long enough to actually matter tend to operate differently. They compete intensely and collaborate generously. They push limits without destroying the people doing the pushing.

Physical Asia proved you can have both. Brutal competition and mutual respect aren’t opposing forces.

What Mongolia Actually Gave Us

There’s speculation about whether Mongolia should have won. Korea fought to the last second, had home advantage, had competition experience. These factors are real.

But here’s what matters more: millions of people worldwide now know Mongolia exists as something other than a historical footnote. They saw warriors who treat women as equals, who understand that proximity to nature builds physical intelligence that gyms can’t replicate, who approach impossible challenges with creativity instead of just force.

The world got to see what happens when you combine traditional physical culture with modern strategic thinking. When you value tactical intelligence as much as raw capability. When you understand that in any competition at the highest level, everyone is strong — so strength alone doesn’t differentiate.

I’m grateful for what Physical Asia revealed about performance under constraint. Strategy compounds over time in ways that raw capability doesn’t. Diversity of approach creates blind spots in homogeneous teams. Respect and intensity coexist just fine.

These aren’t just sports lessons. They’re lessons about building anything that lasts — companies, teams, systems that work when conditions get hard.

Sometimes the strongest team doesn’t win. Sometimes the smartest one does. Usually the smartest one was also the one nobody expected.

P.S. For what it’s worth, riding horses does excellent things for stress regulation.

That’s the moment you realize: you no longer have a startup. You have a company.

Numbers That Don’t Flatter Your Ego (But Should)

Deloitte Fast 500 for the second year running. Ranked 31-32 among the fastest-growing companies in the Bay Area—and that’s about growth rate, not revenue or scale. Yes, it’s good for the entrepreneurial ego. But honestly, these rankings are just external validation of what’s happening internally.

The real benchmark for me isn’t competitor comparison. It’s when critical processes start working without your direct involvement. When a marketing launch runs through Veronika and her team. When Core Value presents research results I hadn’t seen at draft stage. When an engineer who wrote the algorithm answers questions in Product Hunt comments, not the founder who sells it.

I sometimes think of the company as a unicorn—in the original mythological sense, not the venture capital one. It’s charging toward its own manifestation, following its own trajectory, building its own momentum. We’re all just trying to keep up. Sometimes we ride in the saddle. Sometimes we grip the pommel. Sometimes we get dragged along the road but refuse to let go. Because the mission exists whether you’re keeping pace or not.

A Feature That Might Be More Than a Feature

Now for what actually matters. And here I’ll be maximally careful with my language, because claims require evidence, and what we have right now is a our thought supported by physiological mechanics.

Our team (and particular genius scientist) developed a system for analyzing and managing sedentary stress—the kind that accumulates during Zoom calls, deadlines, and endless email chains. Stress that happens in a motionless body. Without physical discharge.

Here’s what we know from physiology:

• Acute psychological stress causes vasoconstriction—narrowing of blood vessels

• Chronic stress without physical activity leads to endothelial dysfunction

• Repeated episodes of vasoconstriction create micro-damage to vascular walls

• This micro-damage is one mechanism in the development of atherosclerosis

To simplify: every time you sit through a tense meeting, your blood vessels sustain micro-trauma. The accumulation of such trauma increases the risk of cardiovascular events.

Now imagine you have a skill—trained and practiced to automaticity—to recognize this stress and discharge it physiologically. Not “breathe and calm down,” but actually interrupt the vasoconstriction cycle before the damage sets in.

Technically, this is real-time heart rate management. But conceptually, it’s a skill you can develop. Like learning to touch-type or drive a car.

I call this a “preventive vaccine against cardiovascular disease” in my private notes, where I can be more liberal with phrasing. But if we want to talk about this publicly—we need a trial. We need to show not just correlation between stress management practice and subjective wellbeing improvements. We need to show clinically meaningful outcomes: reduction in inflammatory markers, improved endothelial function, altered trajectories of risk scores.

This is difficult. It takes time. It’s expensive. But if it works—we’re not talking about a mindfulness app. We’re talking about a scalable preventive intervention for millions of people with sedentary lifestyles.

Momentum and What to Do With It

Right now I feel momentum—and it’s not about awards or Product Hunt rankings. It’s about the alignment of several factors simultaneously:

• We have a team capable of operating autonomously

• We have technology that might be significantly bigger than it appears

• We have a user base and ability to validate hypotheses

• We have the time and resources to think strategically rather than just survive

I don’t know if our hypothesis will prove correct. I don’t know if we’ll be able to design a study that convinces the scientific community. Stop. Nope. I don’t have to be too humble here. I know. But it will take some time to prove it to the world.

I am happy to be on this road with our cute little unicorn and my awesome team.