Promise kept: this is the piece Everyone Is Asking the Wrong Question About AI closed by promising. Here it is.

How many times has your gut been right?

Not the times you wish you had listened — the times you actually had the feeling, and later found out whether it was correct.

I have been asking people this for years. Strangers, colleagues, people at dinners. The answer comes back remarkably consistently: north of ninety per cent. And not from one or two people. From something like nine out of ten of everyone I ask.

That is a strange number to just walk past. Ninety per cent is not a hunch. If a model gave me ninety per cent on a hard classification task, I would ship it.

So either a lot of people are flattering themselves about their own track record, or something real is happening and we have a bad name for it. This piece is about what that thing actually is, and about a rule that follows directly from it, one short enough to keep in your head:

Never delegate judgment.

The three ways of working with AI, and the one that has no name

That number — the ninety per cent — is what you are risking every time you take a model’s answer without running it through your own filter first. So before going further into what the filter actually is, it is worth being precise about how people are skipping it.

Ethan Mollick describes two well-known modes in Co-Intelligence. The centaur divides the labour on a clean line: you do this part, the machine does that part. The cyborg intertwines — you move back and forth continuously, sentence by sentence, until the seam between your work and the model’s is hard to find.

Both are real. Both work. I have watched excellent people operate in each.

There is a third mode, and it has no name, so I will give it one.

Abdication. Not dividing the work. Not intertwining with it. Handing it over and looking away. The centaur and the cyborg both get a noun. This one gets the ghost — present in name, absent in substance.

Why it feels like winning

Abdication does not feel like a failure mode. That is the whole problem.

It feels like leverage. It looks identical to being fast. The output arrives, the ticket closes, the deck goes out, and nothing in the moment tells you which of the three modes you were in. There is no error message for having stopped thinking.

And it compounds in the wrong direction. Every time you accept an answer without checking it, the cost of checking the next one goes up, because you have less context than you would have had. The gap between what you have shipped and what you understand widens quietly, and it widens fastest in exactly the people who adopted AI earliest and hardest.

Then one day the model is down, or rate-limited, or confidently wrong in a way that matters, and you discover the thing that was actually true the whole time: they did not get faster.

They got hollow.

The distinction that actually matters

The usual debate is about volume. How much AI should you use? Should juniors use less? Is it making us worse?

Volume is the wrong axis. I know people who use AI for almost everything and are getting sharper, and people who use it lightly and are getting duller. The dial is not the variable.

The variable is which part you hand over.

Delegate the search. Delegate the boilerplate. Delegate the first ten wrong answers — that is the best possible use of a machine that does not get tired or embarrassed. Delegate the tedium, the formatting, the scaffolding, the lookup you have done four hundred times.

Keep the deciding. What to build. Which tradeoff you are accepting and what it costs. What “done” means. Why this design and not the other one. Whether the thing in front of you is actually right, or merely fluent.

That is the whole rule: never delegate judgment. What it does not tell you yet is why your gut gets to ninety per cent in the first place, or what “judgment” is actually made of. That is worth answering properly, because the answer is what makes the rule more than a preference.

But what IS judgment, actually?

I have been using the word as if it were obvious. It is not, and unpacking it is the reason the number at the top of this piece is real and not self-flattery.

I am not a neuroscientist. What follows is a plausible story assembled from what I know, mostly from building machines rather than studying brains.

Start with the problem your body actually has

Right now your senses are producing an enormous volume of data. Photons, air pressure, skin temperature at every point, joint positions, chemical traces in the air.

If all of it reached your awareness you would not be overwhelmed. You would be useless. You could not hold a thought long enough to finish it.

So there is a filter, and you can catch it working. You were not aware of your shirt against your shoulders until you read that. Or the hum of the room. Or the floor against your feet. All of it was sensed. Almost none of it was reported.

Your awareness is not receiving the world. It is receiving a heavily edited summary, and something else chose the cuts. The interesting question is what happened to everything that got cut.

The architecture, in the only vocabulary I actually speak

A metaphor, not anatomy — but one that keeps predicting things correctly, which usually means there is something under it.

The amygdala is a tiny dedicated chip. One small network, very few outputs — fight, flight, freeze, and by some accounts a fourth, appease. It is the fastest thing you have because it is the smallest and does almost nothing. It does not deliberate. It fires.

The cortex is the CPU with the memory attached — short-term, long-term, and enough processing to run the filter above. It decides what is worth escalating.

The executive brain is the GPU. Massively parallel, the part we proudly call thinking. Fast, expensive, and — critically — it only ever sees what the filter passed it.

And then there is the gut. It has its own nervous system: the enteric nervous system, between one hundred and six hundred million neurons, enough to run digestion without asking your head. That is why “second brain” stuck.

But the number that stops me is not the neurons. Picture every blade of grass in a football field — roughly forty million. Now picture a million football fields. That is the order of magnitude of the microbial population in your gut: tens of trillions of organisms, chemically active, in constant conversation with your nervous and immune systems.

In machine terms: enormous memory, enormous parallelism, a very slow clock. Not a processor for arithmetic. Exactly the processor for holding a vast, slowly-updated model of what normal feels like.

The jungle

You are walking somewhere wild and you become certain — not suspicious, certain — that something is behind you. Your neck knows before you turn. Ask what evidence you had and you would say I just know.

But consider what was available in the seconds before. The birds stopped. A different set moved, all one direction. A troop of monkeys that was ignoring you has gone quiet and is looking past you. Insect noise dropped out of one quadrant. There is a smell that was not there ninety seconds ago.

Every one of those reached you. Your eyes and ears and nose captured all of it. It simply did not clear the filter, because individually not one of them is worth interrupting you for.

Collectively they are one pattern, the pattern has a name, and something matched it against a very old library and returned a single bit: wrong.

That is what a gut feeling is. Not the absence of reasoning — reasoning you did not get the transcript of. The conclusion cleared the filter; the evidence did not. Which is exactly why it arrives as a feeling instead of an argument, and why you cannot defend it when challenged. You never had access to the inputs.

None of that pattern-matching happens in the gut itself. The enteric nervous system does not think, feel, or generate conscious thought — it runs your digestive tract with remarkable autonomy, not bird-watching. What is actually doing the computing sits one level up: Antonio Damasio’s somatic marker hypothesis, which holds that bodily states guide decision-making nonconsciously. Your body registers a response before your conscious mind produces a reason, and that response biases the choice you are about to make.

The usual evidence is the Iowa Gambling Task. Participants draw from four decks; two are quietly rigged to lose. Healthy participants start avoiding the bad decks before they can tell you which decks are bad. The body’s response comes first; the explanation later, if at all. Patients with damage to the ventromedial prefrontal cortex — the region that integrates those bodily signals — keep choosing the losing decks, reasoning fine and losing anyway.

The processing happens in the brain, below awareness. The gut is part of the loop and part of how the result is delivered — more the display than the processor. There is a documented mechanism by which you know things you cannot justify, and it outperforms the part of you that talks.

The oldest version of a method for this is 2,400 years old

None of the above is new, and going back to where it is oldest is what turned it from an observation into something you can actually practice.

Socrates did not write. What we have is mostly Plato, and in the dialogue Theaetetus Socrates explains what he thinks he is actually doing when he talks to people. His mother, Phaenarete, was a midwife. He says he practises the same trade — on minds.

The word for it is maieutics, from the Greek maia, midwife.

And then he adds the part that changes the image completely: he is barren of wisdom himself. He has nothing to deliver. He can only help you deliver, and then help you work out whether what was born is healthy or not worth raising.

That is a much stranger claim than “asking questions is a good teaching technique.” He is saying the knowledge was never his to give. The dialogue is not a transfer. It is a birth, and it happens in you.

The two moments when the idea gets born

Read enough of the dialogues and you notice the birth almost never happens while someone is being told something. It happens at one of two moments, and the Greeks had a word for each — and once you have the neuroscience above, you can see exactly what each word is describing.

The first is when you contradict yourself. Socrates’ method of cross-examination is called elenchus, and its defining feature is that he refutes you using only premises you yourself just conceded. He does not import an outside argument. He shows you that two things you already believe cannot both be true. Nobody defeated you. You were already holding both — one compiled, one not, and they disagreed.

The second is when you run out of road. The state that elenchus produces is called aporia — from a-poros, literally without passage, no way through. It is the moment you reach the edge of what you actually know, as opposed to the edge of what you can fluently say.

Aporia feels like failure. Socrates treats it as the opposite: someone who knows they do not understand something is in a far better position than someone who falsely believes they do. Those two moments — self-contradiction, and the boundary of your own knowledge — are the only two places new understanding reliably comes from. Everything else is transfer, and transfer does not change what you can do.

Now look at how you actually use AI

Here is the uncomfortable comparison, and it is why this 2,400-year-old method is suddenly relevant again.

Socrates was barren of wisdom and full of questions. A language model is the exact inverse: full of answers, and it has lived nothing. It has read more than any human ever will and been in no rooms at all.

And the default interaction we have all settled into runs in the wrong direction. You bring a question. It brings an answer. You accept the answer. That is abdication compressed into a single exchange, and the tooling nudges you toward it a hundred times a day. Nobody decided this. It is just what a chat box implies.

Co-dialectic is the inversion. Not “give me the answer” — cross-examine me until I contradict myself or run out of road. Make the machine play Socrates, which is the one role its architecture is strangely well suited to, because a system with no lived experience has no ego invested in your being wrong.

The reason this matters for the rule at the top of this piece: elenchus is judgment training. It is the mechanism by which you get calibrated. Abdication skips it. Co-dialectic is how you keep the loop that abdication quietly removes.

How to actually implement it, today, in whatever you already use

You do not need a plugin for this part. You need to change what the model thinks the relationship is, and you need it to persist so you are not re-establishing it every session.

Step one: write the premise into your global config, not into a chat message.

Every serious tool has a persistent instruction layer — a global memory or config file, a custom instructions field, a system prompt. That is where this belongs. A framing you retype each morning is a framing you will stop retyping by Thursday.

Step two: state the division of labour and, more importantly, the division of ownership.

Most people write “be my thinking partner,” which is a vibe, not an instruction. Be specific about four things: what you bring, what it brings, what you decide, and what it decides.

Something close to this, in your own words:

We work as one unit that produces one output together. Neither of us is the assistant.

I bring judgment, taste, and the things I learned by being there. I have been in rooms you have not been in, and I know things I cannot fully put into words. You bring recall across far more material than I can hold at once, speed, and the ability to see a pattern across all of it. You have read more than I ever will and lived none of it.

I own the decisions: what we build, which tradeoff we accept, what “done” means, and whether an answer is actually right. You own the search, the drafts, the first ten wrong answers, and the tedium.

When I ask you something that is really a judgment call, do not answer it. Ask me the question that shows me where my own position breaks. If two things I have told you cannot both be true, name those two and stop. If I have gone past what I actually know, say so plainly instead of filling the gap fluently.

Never flatter me. Do not open by telling me it is a good question. Flattery lowers my filter, and the filter is the reason I am here.

When you disagree with me, say it once, plainly, with the reason. If I overrule you, drop it and do not relitigate it through what you build next.

Step three: notice which instruction you want to delete first.

For most people it is the no-flattery line, and that is worth sitting with. Praise is pleasant and it is also the mechanism by which the conversation stops being useful. A model that opens by admiring your question is managing you, not thinking with you. It also quietly teaches you to bring it questions you already know the answer to.

Step four: use aporia as the success metric.

You are not looking for a session that ends in a good answer. You are looking for the moment the model asks something you cannot answer, or points at two of your own statements that do not fit. That is the birth. That is the entire point of the exercise, and it is the only output that stays with you after you close the laptop.

If you never hit that moment, you were not in a dialectic. You were being served.

That compressed premise is a public version of the first idea in something larger — the constitution that Exponential OS runs on, which governs harder problems: how a swarm of agents shares memory across sessions, when an agent may act without asking, and how a lesson learned by one becomes available to all of them. None of that is required to use what is above. The premise is the load-bearing part, and the premise fits in a config file.

What this looks like in a system, not just a chat

The same principle scales down into architecture, and this is what I actually build. I have argued a version of this before, from a stage: in January 2026 I gave a keynote on behavioural-health systems, and the line I closed with was “automation doesn’t remove risk — it redistributes it.” Remove human judgment from a loop and you do not eliminate harm, you obscure accountability and shift the burden onto the most vulnerable person in the system. I was arguing it in the highest-stakes domain I could find, months before it became a general conversation about AI — because the principle was never really about AI. It was about where judgment is allowed to live.

Here is what that principle looks like when the system is software instead of a hospital:

Verification is a funded stage, not an intention. The bottleneck in agentic systems is not generation, it is verification. “I will look it over” is not a stage.

A model reviewing its own output shares its own blind spots. So review runs cross-family — a different model, from a different lab, trained on a different mixture. Two instances of the same model agreeing tells you almost nothing. That is elenchus with the ego removed.

Disagreement is the signal. When judges disagree, that escalates rather than getting averaged into a pass. Averaging is how you destroy the only interesting datapoint you had.

Evaluations run as deploy gates, not dashboards. A correctness regression blocks the release. A dashboard that goes red and ships anyway is decoration.

Acceptance criteria are fixed before the code exists. Otherwise “done” drifts toward whatever was produced — the most human failure mode in the stack, and it has nothing to do with AI.

None of it assumes the model will be right. All of it assumes the model will be confidently wrong and the system has to be right anyway.

Which is why judgment is the part you cannot hand over

A large part of judgment is not reasoning at all. It is the accumulated, unindexed residue of everything you have been present for. The meeting that felt wrong for reasons you could not name and was exactly as wrong as it felt. The hire everyone loved on paper. The architecture that looked fine and made your neck itch.

You cannot write that down — which is precisely why you cannot delegate it. A model can be given everything you know how to say. It cannot be given what you know and cannot say. And the second category is where the ninety per cent lives.

That asymmetry is not a temporary gap that scale closes. It is structural. The machine has read everything and been present for nothing.

How far the machine actually is, on the axis nobody measures

Some numbers, checked rather than remembered.

Your brain is about two per cent of your body mass and consumes about twenty per cent of your energy — a tenfold concentration. In power terms, on the order of twenty watts. A dim bulb.

What that twenty watts computes is genuinely hard to state; serious estimates span orders of magnitude depending on how you define an operation. The headline comparison is commonly put at exaflop scale.

Now the other side. As of 2026, El Capitan reaches about 1.8 exaFLOPS. It has genuinely caught up on raw speed. It draws roughly thirty megawatts.

Twenty watts against thirty megawatts. Call it a million to one.

And that flatters the machine, because it compares speed to speed. It does not count what your twenty watts is also doing: running the sensory filter, maintaining a body, holding a lifetime of context, and producing the ninety-per-cent signal we started with.

And we have not digitised two of the five senses

This is the part that surprises people, and the one I would point at hardest.

Taste is essentially undigitised. Not poorly digitised — there is no taste equivalent of a camera or a microphone. Nothing captures a flavour, transmits it, and reproduces it elsewhere.

Smell is barely started. We have detectors for specific molecules — smoke, carbon monoxide, particular volatiles — which is a smoke alarm, not a nose. A nose separates an enormous space of compounds and mixtures, in context, at trace concentrations, and wires them straight to memory in a way no other sense does.

Two of your five input channels have not meaningfully been captured at all. Before worrying about a machine replacing your judgment, notice that it cannot yet replace your tongue. The sensor package is not close — and judgment is built out of what the sensors gathered over a lifetime.

So what is the actual future

Not replacement. That story needs the machine to acquire what it has no route to: presence, a body, twenty watts, and forty trillion collaborators in your abdomen.

And not romanticism either. Human intuition is ninety per cent, not one hundred, and the missing ten contains some spectacular errors. That is exactly why you pair it with something that has read everything.

The interesting configuration is the one where each side brings what the other structurally cannot. You bring judgment, presence, taste, and the unindexed library. The machine brings recall across more material than you can hold, speed, and pattern-matching across all of it — having read everything and lived none of it. Neither is the assistant. Both sign the output. I have a naive name for what that combination aims at — ASHI, Artificial Super Human Intelligence — not AI that replaces the human, and not a human with a faster tool, but the two as one unit, deliberately assembled, each covering the other’s structural blind spot.

And that is the real reason never to delegate judgment. It is not sentimentality about being human. It is that judgment is the part you are actually contributing. Hand it over and there is no cyborg. There is just the ghost — a slower copy of the machine, signing work it never did.

The uncomfortable part

Keeping judgment is slower this quarter. Genuinely, measurably slower.

The ghost will out-ship you for a while. In a demo they will look better. In a sprint review they will look much better.

I do not have a comfortable answer to that, and I distrust people who offer one. What I have is a longer horizon: the ghost’s advantage lasts exactly as long as nothing surprising happens, and something surprising always happens. When it does, the difference between the two of you is not speed. It is whether you can do the job at all.

A test you can run today

Take the last significant thing you shipped with AI.

Can you explain, to a colleague who is annoyed about it, why it is the way it is? Not what it does — why that tradeoff, why that structure, why not the obvious alternative?

If yes, you were a centaur or a cyborg. Either is fine.

If you find yourself reaching for the model to reconstruct your own reasoning, you already know which mode you were in.

That is not a reason to use less AI. It is a reason to move the line.

Sharpen the input side first

If a gut feeling is pattern-matching over sensory data that never cleared the filter, then one thing follows immediately, and it is the most actionable idea in this piece.

The quality of your judgment depends on the quality of what your senses actually took in.

Not on how hard you think. On what got captured before the filter ran. You cannot pattern-match against material you never registered, and modern life is very good at narrowing the aperture — head down, one screen, same room, notifications carrying the only signals you consciously receive all day.

This is what grounding exercises are actually for, and it is why I stopped treating them as a wellness accessory. Deliberately attending to each sense in turn — what you can see right now, hear right now, feel against your skin, smell, taste — is not relaxation. It is re-opening the input channel and feeding the library that the nonconscious matcher draws on.

Treat it that way rather than as a calm-down technique. You are maintaining the sensor array that your ninety per cent runs on. If the machine’s disadvantage is that it has no senses and no presence, then letting your own senses narrow is the one way to voluntarily give up your only structural advantage.

Then point the machine at the other half

Sharpening your input is one side. The second move is what you do with it.

Use your AI as a co-dialectic, not as an oracle. Bring the thing only you have — judgment, lived experience, the ninety per cent you cannot justify — and let the machine bring the half you do not have: recall across more material than you can hold, and the ability to see a pattern across all of it.

And do it for the right job. Not for answers you could have looked up. For synthesis, and for innovation. And ask for the right thing. Most people ask a model to produce. The two asks that actually exploit the asymmetry are:

Poke holes in this.

Strengthen this using what you know that I do not.

Both of those point the machine at the half you cannot cover, and they are worth being precise about why they work.

It pattern-matches across a volume you cannot approach. Not marginally better — trillions of comparisons against material you will never read. Ask it where your argument has been made before, where it failed before, and what the strongest published objection to it is, and you get in seconds what a literature review would cost you weeks.

It does not get tired and it is not on a clock. It has no bad afternoon, no ego about being wrong, no investment in the position it argued five minutes ago. That is what makes it a better Socrates than most humans — elenchus is unpleasant to receive from someone who enjoys it, and the machine does not enjoy it.

And on the other side, be equally clear about what only you have. You are time-bound — for now, that is the human condition and it is the constraint AI has not solved. You have lived experience. You have the compiled judgment described above. Nobody can give you more hours, and nobody can give the machine a life.

That is the work neither side can do alone, and it is worth being precise about why. Synthesis needs both a wide field of material and a sense of which connection actually matters — the machine has the field, you have the sense. Innovation needs both a large space of possibilities and the taste to know which one is worth building — the machine can generate the space faster than you ever will, and it has no way to know which corner of it is interesting, because interesting is a judgment and judgment is what you brought.

Ask it for an answer and you get a lookup. Bring it your half and ask it to argue with you, and you get something neither of you walked in with.

That is the whole trade. Sharpen your senses so your side is worth bringing. Then bring it, and never hand it over.

Delegate the search, the draft, the first ten wrong answers, the tedium. Keep the deciding.

Socrates could not give anyone knowledge and said so. The machine on your desk can hand you an answer for everything and has lived through nothing. Only one of you in that conversation has been in the room. Act like it.

This is the follow-up I promised in Everyone Is Asking the Wrong Question About AI, which lays out the seven habits of the AI world. Habit one is learning how to learn. This is the one that protects it.

Which brings me to what’s queued next, because both pieces came directly out of writing this one.

The neuroscience above says gut and wisdom survive because they are structural, not sentimental. I went looking for whether that holds up from the outside too, and found it backed from an unexpected direction: Anthropic’s own economic modeling of what AI does to the economy, three scenarios, modest to extreme, a twenty-fold spread between them — and judgment is the one thing that wins in all three. Then I went looking for where I think that model is wrong, and ended up at an Athenian army in 490 BC that had no business winning the battle it won by a mile.

The other piece: two hundred books in two years, after reading nothing at all for the twenty years before that. Not a discipline story — a framework story, and the three sentences of Greek philosophy that turned out to already be sitting underneath how I was thinking, twenty-three centuries before anyone called it cognitive therapy.

That is two pieces queued. If you want one sooner, say which in the comments — I order them by what people actually ask for, and one line from you moves it up.

If this was useful, three small things that actually help it travel: hit like so it reaches someone who needs it, tell me in the comments whether your own gut number landed north or south of ninety per cent, and send it to the one person you thought of while reading.

That should stop you, and it is worth sitting with before you read another word.

Two days ago Dario Amodei published an essay called “We Must Pace the Frontier”, arguing the industry should deliberately slow how fast models gain capability. Within hours Sam Altman committed to match the first step. Elon Musk wrote three words: “Dario is right.” The market discounted AI-exposed valuations.

They said it anyway.

Costly signals are the only kind worth reading

I teach a version of this to executives at Berkeley and it applies cleanly here. When someone tells you something that costs them nothing, you have learned about their marketing. When someone tells you something that costs them money, you have learned about their beliefs.

These men run the labs. Their valuations are built on the assumption that capability keeps compounding on schedule. Arguing publicly for a slower frontier is an argument against their own quarter, and the market charged them for it inside the same week.

That is a costly signal. It does not prove they are right. It does prove they mean it.

Now notice who is not asking.

Not the regulators. The President spent the week emphasizing America’s lead over China. Not the legislatures. Not any government with subpoena power or inspectors. The alarm is coming from inside the building, and the building holds the only instruments.

That asymmetry is the entire story. The people with the measurements are frightened. The people with the authority are not. When the only parties who can see the dial are the only parties asking to slow down, the reasonable inference is not that they have gone soft. It is that they are reading something the rest of us cannot see.

So what are they reading?

They have told us, publicly and in unusual unison. Amodei, Hassabis and Altman jointly called for laws against AI-assisted bioweapons, warning that these systems now outperform PhD-level virologists on many biotechnology questions. Ask people in this industry what actually keeps them up and the answers converge with unsettling consistency on one scenario: a pathogen, engineered, moving too quietly and too fast to contain.

Anthropic then published what it had actually caught.

Roughly thirty-five separate efforts in a single thirty-day window, attempting to route around safety controls. Bird flu. Orthopoxviruses. Novel venom toxins. In one case from May, a researcher affiliated with a military institute asking for help drafting a grant application for gain-of-function work on chikungunya.

And then the line that nobody should skim past:

Their newer models can no longer be assumed to sit below the threshold for meaningful bioweapons assistance.

That is the first time a frontier lab has said that out loud about its own system. Not a think tank. Not a critic. The lab, about its own model, in public.

My question is simple, and I have not found anyone with a satisfying answer to it.

That is what the monitoring caught. What about what it did not?

Thirty-five interventions in thirty days is not a reassuring number. It is a sampling rate. It tells you the detector is working. It tells you nothing about the denominator.

The fiction was optimistic

We have had the parables for two decades and we have learned almost nothing from any of them.

The Animatrix imagined how the handoff happens, and looking at it now it reads as generous. It assumed the machines would want us alive, farmed and dreaming. It never considered that the danger might be engineered by us first, with the machine as a very capable lab assistant. It never considered that a system could hand a motivated human a capability no state actor had granted them.

Optimization needs no malice. It never did. That was the whole point of the paperclip argument and we filed it under science fiction because it arrived wearing a costume.

And Tim Urban asked the question in 2015 that still has no answer. Not what happens in the decade after we reach the threshold. What happens in the sixty seconds after.

Nobody has answered that. Not in eleven years. We have simply gotten better at not being asked.

If you have read this far and someone specific came to mind, a founder, an investor, a legislator, someone who can actually move a policy lever, forward this to them now rather than at the end. That is the only distribution that matters for a piece like this.

Which brings me to my neighbors

There are more than 240 billionaires living in California, holding over two trillion dollars between them, according to work out of Berkeley by Saez and Zucman. Worldwide, more than 3,300, holding north of twenty trillion.

You do not just fund this race. You hold the policy. Not metaphorically. The compute gets built where you finance it, the labs incorporate where your counsel advises, and the bills that would constrain any of it live or die on rooms you are in and I am not.

And California is where this gets decided first. This state writes rules the world copies. It happened with vehicle emissions. It happened with privacy. It will happen here. Act local, and it travels global on its own.

So I want to talk about the insurance you have been buying.

The islands. The bunkers. The compounds with the blast doors and the five-year food stores and the hydroponics. I understand the instinct. I want to point out that it is two doors, and both of them close.

Door one: you seal yourself in and it works. You survive. You spend the remainder of your life inside a structure with no community to belong to, no one to build with, nothing to be useful for. That is not survival in any sense a human being would recognize. It is exile, at a premium, chosen in advance.

Door two: you step outside. And the bird carrying the new flu lands on your roof, because a pathogen does not respect a property line and air does not check the guest list.

A blast door filters a mob. It does not filter an aerosol. The threat these men are describing is the one category of risk your money cannot build a wall against, because the wall has to let you breathe.

You are not above this. You are inside it, with better furniture.

The button

We are currently debating whether there should be an off switch. What it would look like, who would hold it, whether it is even coherent to imagine one given China and open weights and the fact that nobody can unilaterally stop anything.

Those are real objections and I do not have a clean answer to them. I am a practitioner, not a policy expert, and I would rather say that plainly than pretend otherwise.

But I know this much. The debate about whether the button exists is only worth having while it still does.

Use the stop button, the one we still think exists, before it disappears.

Anand Vallamsetla writes about responsible AI, agentic systems, and what it takes to keep building with these tools honestly. Previously engineering leadership at Google, Entrepreneur in Residence at AI Fund, and four and a half years teaching AI strategy and responsible deployment to Fortune 500 executives at UC Berkeley Executive Education.

Everything cited here, in one place:

• Dario Amodei, “We Must Pace the Frontier”

• Anthropic, “Countering misuse of AI: September 2026” (the threat report)

• AI executives’ joint warning on bioweapons (Semafor) and the private fear behind it (Axios)

• The market reaction to the slowdown call (CNN)

• Saez & Zucman, California Billionaires (UC Berkeley) and the Forbes 2026 Billionaires List

• Tim Urban, “The AI Revolution: The Road to Superintelligence” (Wait But Why, 2015)

• The Animatrix and the paperclip argument, for anyone who has not met them

Every week someone asks me a version of the same question: what should I learn to get into AI?

For a long time I assumed it was a question from people at the start. It is not.

I get it from students, and I also get it from people twenty years into a career who are very good at something the world is quietly repricing. A principal engineer. A director who has shipped more than most people ever will. They ask it more carefully, and usually later in the conversation, but it is the same question.

Here is why it lands so hard for the second group.

Experience used to be an asset that compounded. Now it can also be an anchor. Twenty years of hard-won judgment in a domain is enormously valuable right up until the moment the ground under the domain moves — and then the same twenty years become the reason it is hard to move with it. Nobody warns you that expertise has a maintenance cost.

So the question is understandable. It is also the wrong one, for everybody asking it, at every level.

It is wrong because whatever you learn will have changed by the time you have learned it. The half-life of a specific skill in this field is now shorter than the time it takes to acquire it. Anyone optimizing for a curriculum is running hard toward a place that has already moved.

The better question is harder and it does not have a syllabus:

How do I succeed in an AI world?

Seven habits, from twenty-six years of getting this right and getting it wrong in roughly equal measure.

1. Learn how to learn. Nothing else compounds.

The winner is not whoever knows the most. It is whoever closes the loop fastest

Learn. Apply. Get evidence. Correct. Repeat.

Most people optimize the first step and neglect the other four. They read more, watch more, save more. But knowledge that never meets evidence does not become skill — it becomes trivia. The loop only pays when it closes.

Here is the part people miss, and it is the part that matters most.

When you shorten your loop, you do not only win faster yourself. You teach the shortened loop to the people who come after you, and they start from your ceiling instead of your floor. Then they shorten it again and hand that forward.

That is not a linear gain. It is logarithmic, and it compounds across generations rather than quarters.

Which is why the highest-leverage thing available to you is not learning a framework. It is getting systematically better at the act of learning itself — and then teaching that.

2. Use an Optimal Knowledge Acquisition Framework, not a reading list.

I taught this at Google as a Googler-to-Googler coach, and the pattern held across everyone I worked with. The people who moved fastest were never the ones who read more. They were the ones who knew what to skip.

A reading list is a queue. A framework is a filter. The filter is worth more, because the constraint was never access to information — it has not been for twenty years. The constraint is attention, and attention spent at the wrong depth is simply gone.

Three things the framework has to do:

Acquire faster. Know when to go three levels deep and when one level is genuinely enough. Most material deserves one.

Retain longer. Retrieval beats review. If you cannot explain it without the document open, you have not learned it. You have visited it.

Implement like there is no tomorrow. Knowledge that never ships decays at the same rate as knowledge you never acquired. The only difference is that you feel better about it.

I call it the Optimal Knowledge Acquisition Framework, and optimal is the operative word — the point is not to acquire more, it is to acquire the right depth at the right cost. It gets its own piece. If you want it sooner rather than later, say so in the comments and I will move it up the queue.

3. Find your ikigai before you optimize your resume.

Also something I taught at Google — though I did not find it myself. Pravir Gupta, a VP and GM at Google Cloud and an unofficial coach of mine, handed me the book.

Four circles: what you love, what you are good at, what the world needs, and what you can be paid for.

The reason this matters is not balance. It is fuel. A career built on only what you are good at and what pays runs on push — a manager, a deadline, a promise about the next level. Push runs out, usually around year seven. Pull does not. Find the thing you would read about on a Saturday and the loop from habit one runs on its own: you learn faster than the person being pushed, because nobody has to make you. That is also what survives the pivots, and this field will make you pivot, more than once, whether or not you planned for it.

Pravir put it better than I can: passion and knowledge are a dangerous combination, and if you have ten out of ten in both in a given area, you are unbeatable. Note the last part. Not ten out of ten in general — in a given area. Knowledge is habit one. Passion is this one. Neither alone is dangerous, and that is the whole reason he gave me the book rather than a reading list.

The order matters. Optimize the resume first and you become very good at being hired for work you do not want. Find the intersection first and the resume becomes a description rather than a performance.

4. Your career is not a line. It is hills.

You have to walk down one hill to climb the next. Every senior person you admire has a valley they do not put on LinkedIn — the role that did not work, the company that folded, the year that looks like a gap.

What is genuinely different now is that you can turn the descent into a wave instead of a drop.

The valley used to be long because relearning was slow. You left a domain, and re-entering a different one at real depth took years. That is no longer true. With an AI that can take you to the working edge of a field in weeks rather than quarters, the trough gets shallower and shorter.

That is the biggest structural advantage available right now, and most people are using it to write emails faster.

5. Collect dots that do not connect yet — then use AI to connect them.

You cannot connect the dots forward. You can only collect more of them, on purpose, and trust the pattern to show up later

But this is where AI earns its place, and it is not the way most people use it.

It has effectively infinite knowledge and no lived experience. You have lived experience and a small fraction of the knowledge.

That asymmetry is not a weakness on either side. It is the raw material. Innovation is the synthesis between the two — between what has been recorded and what has been lived. Neither half produces it alone.

So do not use AI as an oracle that agrees with you. Agreement is worthless; you already had that. Make it a co-dialectic. Bring it the intuition you cannot justify yet. Let it argue back. Argue with it. Ask it to show you the three fields where your half-formed idea already exists under a different name.

The dots you could not connect alone start connecting. Not because the machine is smarter, but because you finally have something to think against.

The inversion: this is the golden hour for experienced people

For twenty-six years, every new technology put me at a disadvantage

A new grad could pick up a new language faster than I could. Always. They had nothing to unlearn and more hours to spend, and I was competing on the one axis where youth genuinely wins. That was true of every wave, and I felt it every time.

That axis just disappeared.

Every language is now a prompt away. The thing I used to lose on is free. And what I am left holding is the one thing that took twenty-six years to accumulate and cannot be downloaded: lived experience, which is another word for judgment.

Knowing which of three working solutions will still be working in two years. Knowing when a demo is hiding a problem rather than solving one. Knowing when to kill something. None of that is in the model, because none of it was ever written down — it was paid for.

So let me say the quiet part plainly. I am more effective now than fifty developers were five years ago. Not because I got smarter. Because the part of the job I was worst at became free, and the part I am best at became the bottleneck.

If you have twenty years, you have not been made obsolete. You have been handed leverage on the exact thing you spent twenty years buying.

And it is the best moment to be a beginner, too

Here is the part that should be uncomfortable for people like me, and I would rather say it than pretend.

Everything I know, I paid for with mistakes. Years of them. What worked, what did not, and why — that was tuition, and the bill came in time.

You can now get a large part of that in one prompt.

They say only fools learn from their own mistakes and the wise learn from others’. That was always good advice and almost impossible to follow, because other people’s mistakes were not accessible to you. Now they are. Every post-mortem, every failed architecture, every hard-won lesson someone bothered to write down is one question away.

That is the exponential advantage, and it is available to the person with zero years.

The only thing standing between you and it is knowing how to ask. Which is a skill, and it is learnable, and it is the single highest-return thing you could learn this month.

I spent something like forty hours reading prompting research to work it out. You do not have to. I productized what I learned as an open-source plugin called Co-Dialectic — it runs on Claude Code and other CLIs, and it teaches the method by making you use it rather than by explaining it. That is the whole design: you learn it by doing it. The write-up is here: https://www.thewhyman.blog/p/co-dialectic-v4-is-live-your-ai-is

Both of these are true at once. The experienced person’s disadvantage just vanished, and the beginner’s biggest cost just collapsed. That is not a threat to either. That is the best moment either of you has had.

6. Aim at the field that does not exist yet.

Yuval Noah Harari has warned that by 2050 a “useless class” might emerge, and that bioengineering combined with AI could separate humanity into a small class of superhumans and a large underclass. He is careful to say might. So am I.

I will go somewhere he does not.

I think that within a hundred years there may be a species that looks back at us the way we look at chimpanzees today. Not a metaphor about capability gaps — an actual discontinuity, where the thing looking back is part biological and part engineered, and the distance is not one of degree.

If that is even half right, the durable work sits in two places: where biology meets machines, and where Earth meets everything past it.

Which today points at bio-mechanical and bio-electrical engineering. Brain-computer interfaces and neuroscience. Aeronautics. Material science. Security, because every one of those becomes an attack surface. And running through all of it, responsible and ethical AI — because someone has to decide what we should build, not only what we can.

That list is what I can see from where I am standing in 2026. It will be wrong within a few model generations, and that is the point. Do not memorize the list. Learn to read where it is going.

Pick the intersection nobody has named yet. That is where the room is.

7. Curate your inputs like they are your diet. They are.

Swami Vivekananda said: “We are what our thoughts have made us; so take care of what you think. Words are secondary. Thoughts live, they travel far.

Everyone quotes the first half. Almost nobody asks the obvious follow-up.

If what I think is what I become — then what makes what I think?

You do not choose your thoughts directly. Try it. Try deciding, right now, to think differently for the next hour. It does not work, because thought is not the input. Thought is the output.

What you think is what you eat, what you listen to, what you read, what you watch, who you spend time with, and how you sleep.

You eat french fries because somewhere upstream you decided they were fine. You exercise because somewhere upstream you decided it was worth it — and the real return on exercise was never the body, it was the mind.

The thought came from the input. Change the input and the thought follows.

Which turns something that looks like social hygiene into something more serious.

Follow the people who make you more capable. Your feed is not entertainment. It is the raw material your thinking is made of. An hour a day of confident, empty commentary does not leave you neutral — it leaves you with the shape of someone’s thinking and none of the substance. Choose people who are doing the thing rather than describing it. Choose the ones who make you uncomfortable because they are further ahead.

And knowing who is worth following is itself a skill. It sits downstream of judgment, which sits downstream of everything above. This is why the framework is a loop and not a list.

Then there is the part I am less sure about, and I would rather say so than pretend.

I think what you dream is what you become.

Not the ambition sense. The literal one. A dream changes what you think, and what you think changes what you become — so the dream sits one level further upstream than any of it.

I suspect this is what meditation has been doing all along. Deliberately placing better things where thought is made. Whether that is self-manifestation or just very old, very good input hygiene, I genuinely do not know.

But I know which way I would bet.

Build, build, build — because that is how judgment gets made

I have been circling one claim this whole piece, so let me state it directly

In the machine age, the only thing that stays scarce is judgment. Knowledge is free. Execution is getting free. Judgment is not, because judgment is compressed lived experience, and the machine has no life to compress.

Which raises the obvious question: how do you acquire judgment on purpose, rather than by waiting twenty years for it?

You build. Repeatedly, quickly, and in public.

Every small thing you build and ship runs one honest experiment. It either works or it does not, and either way you now know something you could not have read. Fail fast, succeed faster — not as a slogan, but as the actual manufacturing process for judgment.

There are three builds inside it, and they are not the same build.

Build the thing. The silly idea, actually shipped. Small enough that failing costs you an afternoon rather than a quarter.

Build the evidence. The repo, and the write-up of what surprised you. This is the step almost nobody does, and it is the difference between having learned something and being known to have learned it.

Build the judgment. The part that accrues to you. It is the only one of the three that compounds, and it is the only one the machine cannot hand you.

Technical skill accumulates along the way, as a by-product. That is the right place for it now. It is no longer the scarce thing, so it should not be the goal.

What this looks like in practice

Here is exactly what I do, and it is not complicated.

I ask an LLM to give me a few silly ideas for learning some feature I do not know. Silly on purpose — low stakes, no plan, nothing riding on it.

Then I build one. Then I publish it to GitHub, because a repo is evidence and an intention is not.

Then, when I have learned something that surprised me, I write about it.

Defense in depth is the clearest example. I was building a Medicaid eligibility copilot. I ran a boundary test — the threshold was twenty thousand dollars, so I fed it nineteen thousand nine hundred and nineteen. The system said not qualified.

One dollar. And one dollar is somebody’s healthcare.

I could not read my way out of that. I had to build my way out, and what came out the other side was a set of methods for building reliable systems on top of unreliable, non-deterministic models — which became my first article. I did not have that judgment beforehand. The build produced it.

And the by-product is a career

I did not set out to build a personal brand. I set out to learn things. But look at what the loop produces on its own:

The articles show how you think. Anyone deciding whether to hire you is really trying to answer that one question, and an article answers it better than a resume ever will.

GitHub shows the receipts. It is the difference between claiming you can do something and having done it, timestamped.

And LinkedIn turns it into inbound. Interviews start arriving instead of being chased. That is not a growth hack. It is what happens when the evidence exists in public and you stop having to describe yourself.

You do not need anyone to tell you what to build

This is the part people get stuck on, and it is the part that does not matter.

You cannot reason your way to the destination. That is the trap, and smart people fall into it hardest — they try to think their way to certainty before spending anything, and they are still thinking a year later.

Andrew Ng’s thesis at AI Fund is the opposite: you build your way there. You do not analyze until the answer appears. You run the smallest experiment that produces real evidence, and the evidence tells you the next one.

Small builds are how you take a position on the future without having to predict it. Do enough of them and the direction stops being a guess — not because you worked it out, but because you walked it out.

So do not wait for the right idea. Build a silly one today. Publish it. Write down what surprised you.

The three words

You do not need permission to start, and you do not need a specialized degree. You need a loop, and the discipline to keep shortening it.

If I could put three lines on a wall for anyone — at twenty-two or at fifty-two — it would be these.

Dare to dream. That sets the direction. It is not the soft part of the advice; it is the most upstream lever you have.

Learn to learn. That is the engine. It is the only skill that does not depreciate.

Build, build, build. Fail fast, succeed faster — that is how you buy judgment cheaply,

instead of paying twenty years for it. And judgment is the only thing that will still be scarce.

Which brings me to the thing I am writing next, because it is the failure I now see most often.

The people who adopted AI fastest are developing a dependency they did not plan for. Not on the tool — on the judgment. They have handed over not just the typing but the deciding, and when the model is down or wrong or unavailable, they cannot do the work at all. They did not get faster. They got hollow.

Ethan Mollick, in Co-Intelligence, describes two ways of working with AI. The centaur divides the labor — a clean line, you do this, the machine does that. The cyborg intertwines, moving back and forth continuously. He argues the cyborg is where you end up once you have really found a co-intelligence.

I think there is a third mode he does not name, and it is the one quietly spreading: abdication. Not dividing the work, not intertwining with it — just handing it over and looking away. It is the most comfortable of the three and the only one that leaves you unable to do your job when the tool disappears.

The cyborg way is the one I am building — co-dialectic and Exponential OS are both attempts at the same rule. I have written about what that looks like in practice, and what compounds once you stop starting over, here: https://www.thewhyman.blog/p/the-cyborg-the-exponential-advantage Never delegate judgment. Delegate the search, the draft, the first ten wrong answers, the tedium. Keep the deciding. A model that agrees with you is not the same as a model that is right, and the moment you stop being the one who can tell the difference, you have traded a skill for a subscription.

I wrote the full version of that rule here: Never Delegate Judgment.

That is one of two pieces queued: this one, and the Optimal Knowledge Acquisition Framework from point 2, written out in full. Both are the mechanisms rather than the sentiment. If you want one of them sooner, say which in the comments — I order them by what people actually ask for, and one line from you moves it up.

I write about applied AI, evaluation, and building systems that build products.

If this was useful, three small things that actually help it travel: hit like so it reaches someone who needs it, tell me in the comments which of the seven you would argue with, and send it to the one person you thought of while reading.

And subscribe. The Optimal Knowledge Acquisition Framework from point 2 gets written up in full, and it lands here first.

AI Fund is Andrew Ng’s venture studio. Ng founded Google Brain, was Chief Scientist at Baidu, teaches at Stanford, and was named to the TIME100 AI list in 2023. Getting into the program was a 48-hour build challenge, then interviews with the CPO, the CTO, and Ng himself.

Here’s the part that still gets me. I applied for this through LinkedIn like everyone else — the posting had roughly two thousand applicants at the time — with a referral from Fahad Aziz, a founder I know well. I heard nothing for two months.

Then Mike Rubino, who runs talent at AI Fund, emailed me out of the blue. Subject line: this is not spam — builder dinner with Andrew Ng. It arrived at 3:52 in the morning and invited me to a dinner that same evening: about fifteen builders, ninety minutes, Mountain View, in person with Andrew. Everything about it looked like something you would delete.

I nearly did. Don’t. If an email like that ever lands in your inbox, read it twice.

I didn’t get in through the front door. The front door doesn’t scale — two thousand applications is a volume problem, not a verdict on any of them. Mike had just started at AI Fund and had rebuilt how they sourced, using AI to go and find the people the pipeline was never going to surface. So I didn’t get in because someone got lucky with my resume. I got in because someone built a better way to look. He isn’t doing that alone: Dylan Love — a recruiter who went and got a CS masters to become a data scientist, and now a GM at DeepLearning.AI — is building AI for talent with Andrew. Between them, the machinery that finds people is being rewritten. Which is the argument of this whole piece, arriving two months before I understood it as a product lesson: when the cost of doing something collapses, doing it stops being a differentiator, and the moat moves to being found.

I’d spent the previous six weeks rebuilding how I showed up publicly. In the 45 days to April 28: 27,942 impressions and 10,620 people reached — up 510% and 622% over the prior 45 days. Mike’s email landed on day twenty-six of that window, while the curve was already climbing. I don’t think those two facts are unrelated.

Which is the point. The growth wasn’t the result of being found. It was the cause of it.

I’ve since wondered whether that’s part of why the conversation went anywhere at all. A studio whose entire operating model is hypothesis, test, evidence is not going to be moved by someone describing their potential. It’s going to be moved by a chart.

I went.

One more person belongs in this part. After the dinner came a 48-hour build challenge, and after I submitted it I did the thing everyone does: I waited. Rob Ross, a founder I’d met days earlier, told me what I couldn’t see — that there were other people working on the same challenge, that they were being weighed against each other, and that every one of them was still moving while I sat on a submitted file. Build something tonight, he said. Ship it in the morning.

He called at eight the next morning to ask whether I actually had. I said I was sort of working on it. He made me prioritize. At nine I committed: by eleven I’d send him the exact email I intended to send AI Fund at five that afternoon, describing what I’d built and shipped. If I missed either mark I’d take him to the Ritz-Carlton. He turned that down — I don’t care about the Ritz-Carlton, it has no value. You charge $500 an hour. Give me eight hours of consulting for my startup. He refused a meaningless stake and repriced the bet in the only currency that mattered.

I hit both deadlines. Mike replied within hours and set up the onsite.

In a competitive process, waiting is the losing move. What you submitted is the opening, not the last word.

Much later — week nine of the EIR — that relationship turned into a different kind of conversation. The kind where you give someone honest feedback and risk the relationship a little, because you think they can use it. Seven specific things, candidly delivered. He took it exactly as intended, and later gave me a signed copy of Andrew’s book. Which is the other half of how this works: being found is the opening, not the outcome. What you do with the relationship afterwards is the part you actually control.

Mike deserves more than a mention. When the whole thing nearly came apart at the last minute, he did the rarest thing a person on the other side of a negotiation can do — he told me it was completely fine to walk away, and explained his reasoning instead of pressing. That’s why I said yes. You learn a lot about an organization from how it behaves when you might leave.

I’d made a version of this bet once before, with my own life. Eight years ago I moved from Texas to the Bay Area at a point when roughly eight people were leaving California for Texas for every one going the other way. I went the wrong way down that highway on purpose, for exactly this reason — proximity to where things get known first. The reaction stuck with me. Someone told me there’s no middle ground with you: you’re either too dumb, or you’re too smart and you know something we don’t. I said thank you for the compliment, and that time would tell. Eight years on I’m still not certain which one I am — but I’d make the same move again tomorrow.

And none of what followed was a separate decision. Working with the best minds I can get near, staying close to research, and getting to teach — that’s one strategy, and I set it in motion eight years ago when I pointed the car west. Joining AI Fund wasn’t a new plan. It was the same plan, one layer up.

I almost didn’t take it. Navya Chitimireddy, a VC at Abyro Capital, cut through weeks of my deliberating with a single question: would you regret it if you didn’t? That’s Bezos’s regret-minimization framework, and it ended the debate in about four seconds. She was also the first person to tell me plainly that distribution is the moat and that the answer is company workflows — which turns out to be most of what this piece is about.

The answer, once she made me say it out loud, wasn’t about the credential. It was proximity. I’d regret not getting to watch one of the sharpest thinkers in applied AI reason in real time — what he considers worth researching, how he thinks about distribution, what he kills and how quickly. And not only him: a bench of operators who had already done the thing I was trying to do. You can read all of that. Sitting next to it is a different kind of learning.

Rakesh Utekar and Ashwyn Sharma pushed me the same way from different angles. Rakesh is the kind of builder who makes you recalibrate what one person can ship in a week — I’m only half joking when I call him the Terminator. Ashwyn had seen the place from the inside and couldn’t name a single negative: his read was that AI Fund is built for serious founders, the ones who don’t chase shiny objects.

There was one more thing, and it’s what stopped the decision feeling like a career calculation. Most studios and accelerators fund what the market has already validated — the shiny category, the adoption curve that’s already bending. What I saw here was a different starting question: does this problem have to be solved, for people’s sake? Back it because society needs it, not because the signal has already arrived. I care more about problems at that scale than about being famous or getting rich. It felt like a perfect match.

Three months is short enough that everything you learned is still vivid, and long enough that some of it is real. I wanted to write these down before the vividness fades into general wisdom, because the specifics are what make them useful.

Three parts: finding the pain, the moats that are left, and how you actually build now. Some of it I heard from mentors and internalized. Some I watched teams get right or badly wrong.

None of this is beginner material. If anything, the founders who need it most are the ones with the most scar tissue — because the playbook that worked before AI is now the thing most likely to mislead you. AI collapsed the cost of building. Table-stakes engineering stopped being a differentiator the moment it stopped being scarce. That single change moves where the moat lives, and everything below follows from it.

Part 1 — On finding the right pain

1. Go after the biggest pain — that’s where engagement and willingness to pay live.

Every product solves SOME kind of pain. That’s not the bar.

The bar is: is this the biggest pain? Because that’s the only place where users actually engage with the product and are willing to pay for it.

Solve a mild pain and you get polite nods. Users try it once, don’t come back, and never open their wallet. They’d already adapted to that pain and built a workaround. You’re not selling a solution — you’re competing with inertia.

Inertia wins.

Andy Ku, AI Fund Partner for Product Build, kept pulling me back to that one word — biggest. Not a pain. The biggest one.

The biggest pains are the ones users mention unprompted, the ones with an ugly workaround already bolted on, the ones that make them curse at their screen. Those are pre-qualified. The user already believes the pain shouldn’t exist. Your product just has to make it go away — and they’ll pay for that.

2. Find multiple users with the same pain before you build.

There’s nothing magic about a specific number. What matters is more than one. That bar came from Andy Ku, and it held every time I tested it.

When several users independently describe the same pain in similar words, that’s your evidence the pain is real and worth solving. One user is an anecdote. A handful pointing at the same thing is signal — and it directly validates lesson one, because a pain multiple people carry is far more likely to be a big one.

Skill lies in finding the ICP. Not defining it — finding it, in the wild. Once you’ve found a few who share the pain, build for them specifically, not for the abstract “market.” “Market” is a shape you infer later.

3. Ask facts, not wants — then observe.

Andy and Jill Shih, a product and UX leader with 25 years in the discipline who now runs AI Fund in Taiwan, both drilled this into me — the Mom Test and contextual inquiry, applied together.

The Mom Test came to me from Vaibhav Mathur, a physicist turned builder in the same cohort, who put the book in my hands and kept pointing me at the sources actually worth following. He has a nose for who to listen to.

First: never ask “would you buy this?” — the answer is always kind. Ask facts about the past. How long did that task take last time? What did it cost you? Who else was involved? What did you try before?

Second: watch someone actually work through their real task. You’ll see the pain they’ve stopped mentioning because they’ve stopped noticing it. Survey data doesn’t touch this layer.

Three killer questions that quantify the pain:

1. How long does the task take?

2. How much does it cost?

3. What happens if it doesn’t get solved?

The third one does double duty. It surfaces the cost of inaction — which is where willingness to pay actually lives, because a budget line exists to avoid a consequence, not to buy a feature. And it makes the user say the pain out loud, which often raises their own estimate of what solving it is worth. It was the most useful question I asked.

And when you’ve built something, don’t ask them if they’d use your product. Watch whether they do. Behavior is the only signal that doesn’t lie.

Part 2 — The moats that are left

4. Data is a moat you can design for.

Any new AI system has three paths to defensibility:

1. Own the data — proprietary corpus, exclusive access, historical archive nobody else has.

2. Buy access no one else can — exclusive partnerships, licensed streams, geographic or regulatory access. Sandeep Gupta sharpened this one for me — the acquisition angle most founders skip past.

3. Design the product so it generates new data every time it’s used. This is where I landed: every user interaction should make the system more intelligent for the next user using it. That’s a compounding flywheel with no ceiling — not a static advantage.

Feature parity is six weeks away. Data compounds.

Own a vertical.

It’s worth being blunt about why this matters more than it used to. Anyone can ship the feature now, and foundation-model companies are absorbing whole verticals as fast as they appear. So the moat moved. What’s left is the hard engineering — eval, reliability, eliminating AI slop. The data. The distribution. Finding the ICP before anyone else does. And, least obviously, still being there after the fifth pivot while thousands of people build the same thing the same week.

That last one sounds soft. It isn’t. When building is free, persistence is scarce.

5. The hardest engineering problem is the moat — and proximity to research is how you hold it.

This one is Andrew Ng’s, directly.

Don’t compete on features. Don’t compete on being first to market. Invest in the technology and solve the problem others have tried — and failed — to solve.

Every space has competitors. Users not choosing them yet usually means one of two things: the competitors don’t solve the real problem, or nobody’s heard of them. Sometimes both.

There’s a third reason, and it’s the one AI keeps manufacturing: the capability to solve it only just arrived. A model ships, a research result lands, and a problem that was genuinely intractable last year becomes tractable this quarter. That isn’t a market rejecting a solution — it’s a market that never had one. Being in the Bay Area is an unfair advantage here, not because of the money but because you hear about the capability months before it’s obvious, and the window between possible and crowded is where the whole game is played. Hear it early, build fast, get there first.

That is the bet I described at the top — made once with a moving truck, and again with a career.

The hardest problems are the moat that’s left.

Eliminating AI slop — low-quality, hallucination-heavy generative output — is a category-defining engineering problem. Many teams have tried. Most have failed. That’s exactly why it’s the moat. Cross-family judgment, structured refusal, guardrail-as-architecture, eval-gated deployment — the whole discipline of making probabilistic systems trustworthy in production. Solve one of those and the ICP finds you. You never had to compete on features.

But there’s a second half to this that took me three months to see clearly. If the hardest problem is the moat, then staying close to research is how you keep it — because the hard problem moves. The thing that was unsolvable last quarter has a paper about it this quarter, and the team that read it first gets there first. A moat built on solving one hard problem is a moat with an expiry date. A moat built on being continuously close to where the hard problems are being solved doesn’t have one.

I stay close to research two ways. Teaching is the first — four and a half years of it at UC Berkeley Executive Education, taking AI strategy and responsible deployment to more than fifteen hundred Fortune 500 executives, and I’m going back to it. Nothing exposes a shallow understanding faster than a room of executives asking why. And I work with people who carry the field in their head. Andrew is a research encyclopedia; ask him about a problem and you get the shape of the literature back, not an opinion.

That was, honestly, part of the decision. When I said yes, I told myself: I don’t have to worry about the technical challenge at all — Andrew will be there to unblock that. It sounded a little like wishful thinking. It wasn’t — but it also didn’t mean what I assumed it meant.

He never handed me a solution. The slop work was mine: I built the evals, largely out of his own course, and got a 60–80% reduction. What he unblocked was the layer above the solution. Am I even working on the right engineering problem? And who has already solved this, or is solving it right now — which researchers, which teams, which papers. A ten-minute answer to that saves a month, because the expensive mistake is never a bad implementation. It’s an excellent implementation of the wrong problem.

Eli did the same thing from a different angle: a steady stream of technical resources and trends on the slop problem, and one instruction that stuck — build a taste for it. Then he pushed further. Don’t lean on the models alone. Don’t become an AI zombie.

That one landed hard, because I’d written the same warning about myself years earlier, before business school: as a (just) technologist, when all you have is a hammer, everything looks like a nail. AI is the largest hammer any of us has ever been handed. The correction isn’t to use it less — it’s to keep the judgment that decides where to swing.

And the way you keep that judgment isn’t restraint. It’s architecture. The reason I build gates — evals, cross-family review where no model family grades its own work, checks that fire whether or not I remember to run them — is so the judgment survives the speed. An AI zombie lets the model make the call. A framework forces the call to be made, by a human, and recorded. Same tools, opposite posture. Most of what looks like heavy AI use in how I work is actually the scaffolding built to stop exactly that failure.

6. A real buyer is a moat. User delight isn’t.

Delighted early users are necessary, not sufficient.

The wedge that becomes a company needs an economic engine — buyers whose budget already exists, whose pain is board-level, whose measurable outcome is unambiguous.

Users loving something is a starting condition. Not proof of a company.

The clearest way to check: can you name the specific budget line item that already exists in the buyer’s org, and the specific measurable outcome your product will improve, and the specific title of the person who owns that budget? If you can’t name all three, you don’t yet have a commercial signal. You have user delight and a hypothesis.

Part 3 — How you actually build now

7. Eval is the product.

This one is Andrew Ng’s again — I first heard it in his DeepLearning.AI course on Agentic AI, which is also where the autonomy-level framing and the case for evals clicked for me, and it kept landing harder every time I saw it play out inside the studio. His way of putting it: the thing that separates a good team from a great team is how they run their evals.

Every AI system I saw succeed had eval infrastructure baked in from day one. Every stall traced back to teams shipping the model output first and adding evaluation later. Later never comes; the deploy calendar hardens around the shipped shape.

When the model changes underneath you weekly — new snapshot, new alignment update, new subtle regression — evaluation is what makes the system real. Not the demo. Not the pitch. The eval harness is the thing that keeps you shipping when the ground moves.

8. Build frameworks and AI-native agentic workflows, not just products.

The strongest teams I saw weren’t shipping products faster. They were shipping the framework that shipped the products.

Your own dev workflow. Your eval harness. Your security gates. Your architecture-alignment checks. Sonar. Cross-family judgment. Guardrail-as-architecture. Each of these is a small compounding investment that turns every product you ship next into a faster and safer ship than the one before.

The framework compounds. The product depreciates. If you’re only building products, you’re re-buying the same tooling forever.

That’s the inward half. The outward half matters more.

Ship the agentic workflow, not just the feature. A feature gets evaluated, compared, and swapped. A workflow gets adopted — it becomes how the work actually gets done, and then it becomes the thing nobody wants to unpick on a Tuesday. Embed yourself in how the user operates and you stop being a tool they chose; you become the backbone they’d have to rebuild around to leave.

That is switching cost you earned rather than trapped someone into. And in a market where the feature you shipped can be cloned in six weeks, it may be the most durable moat on this list.

I didn’t arrive at this on my own. Dylan and Gaurav Surtani are the two who made it concrete for me. Gaurav is a one-person AI engineering team at DeepLearning.AI, and watching how he works is the argument: he isn’t hand-writing more code than everyone else, he’s built the workflow that writes it — agents, MCP servers, skills, cost-aware model routing — and then he open-sources the pieces. In the developer arena the differentiator stopped being how fast you type. It’s the workflow you built to do the typing.

9. The clock on consumer AI wedges is compressing.

Categories I saw close in weeks that would have taken quarters two years ago.

Speed of learning matters more than speed of building. A studio like AI Fund is one of the few environments engineered specifically for that compression — the entire operating model is “run more hypothesis-test loops per week than you could run alone.”

Persistence is the moat nobody lists.

I watched what that looks like up close. Jayanth (Jay) Madduru, a Founder in Residence at the studio, ran the loop harder than anyone I’ve met — thesis, test, feedback from the studio, then straight into the dark to learn something new about the user or a new way to build. One weekend he shipped 150 pull requests. That throughput has tooling underneath it — he had picked up the Superset setup after seeing what it did to my own output, then built better workflows on top of mine and pushed the whole setup further than I had. Far enough that Superset called him directly to tell him he was their heaviest user, and offered him a job. Then Monday he’d be back with a sharper thesis and more energy than he left with, like a spring that won’t stay compressed.

He credits me for some of what he learned about building. He has no idea he handed me more back than that.

Most of those loops fail. That is the job, not a detour from it. When building is free and thousands of people are shipping the same week you are, the differentiator stops being the code and becomes who is still running the loop after the fifth one breaks. Almost nobody is. You only fail when you stop trying.

The meta-thesis — why any of it worked

Eli Chen insisted on this one — he’s a Technology Partner at AI Fund and previously co-founded and was CTO of Credo AI: you can’t just reason your way to the destination.

You start with a hypothesis. You test it with real users. You pivot on what you learn. You start over with a sharper hypothesis. That’s the loop. There is no other loop.

Every one of these only became real to me because I got to run that loop compressed. Any of them read as generic advice on their own — “own the data,” “eval is the product.” They only become your advice when you’ve held them against your own hypothesis and seen how they change what you’d do next.

That’s what a studio buys you: not the destination, but the loop.

If you want to be found, show receipts

The piece argues the moat moved to being found. Here’s the practical version, and it’s the advice I’d give anyone in the market right now.

Claims don’t travel. Receipts do. Your resume says you know a technology. So does everyone else’s. The résumé is a claim; nobody can verify it in the eight seconds it gets. What survives that eight seconds is proof someone can click.

Three tiers, in increasing order of how much they’re worth:

1. Public contributions. A GitHub profile with real commits beats a bullet point that says the same thing. It’s the cheapest receipt available and most people still don’t have one.

2. Open source. Strictly harder, and worth more precisely because it’s harder — you’re inviting strangers to read your actual code and judge it. That takes guts, and everyone reading it knows that.

3. A working thing they can use. The strongest form. Don’t tell someone you understand AI systems — put an agent on your site and let them talk to it. I did exactly that on my own site rather than adding another line claiming technical depth. One interaction settles a question a paragraph can’t.

The pattern underneath all three: stop describing your capability and start exposing it. Descriptions are indistinguishable at scale — that’s exactly the collapse this whole piece is about. A thing that works is not.

The reframe — worth solving vs. vitamin water

Every founder starts with hope that this experiment will be the one. When the wrap comes, it stings.

Andy, Eli, and a friend from my poker circles — Anurag Jain, a multi-time founder now building in stealth — kept saying the same thing in different words: an experiment that doesn’t lead to commercial success is the experiment doing its job. Failed hypothesis is data, not verdict.

Anurag put it more sharply: “Solving the right pain is important. Raising money for something that isn’t even a big pain worth solving is much worse.”

Wrapping cleanly on something that turned out to be vitamin water — not a painkiller — saved me the next seven years trying to force market-fit onto a hypothesis the market wasn’t asking for.

There’s a symmetry here I only noticed while writing this. Both explorations I ran during the EIR were bets on the same thing this piece argues. One was an AI-powered platform for adapting content across channels; the second came after it. Strip the surface off either and they were about the same problem: being found. I was building the answer to the exact thing that got me the job — I just hadn’t connected the two yet.

What we learned was that the market wasn’t ready to pay for it on the timeline a studio needs. That’s a finding about commercial urgency, not about the thesis. I still think the thesis is right. I think it gets more right every month building gets cheaper.

Solve the right pain. The wrong one drains you.

What I’m carrying forward

On finding the pain

• The Mom Test on every conversation with a potential user — ask facts, not wants, then watch them work. And never on the strength of one user; one is an anecdote, several with the same pain is a market.

• The hypothesis → test → pivot → sharper-hypothesis loop as the only real product-development method.

On the moats that are left

• Design the data loop from day one. The data you accumulate by operating is the one moat a competitor can’t clone by shipping the same feature.

• Take the hardest engineering problem on purpose. Once building is cheap, the problem others tried and failed to solve is the moat that’s left.

• Choose the room, not just the project. Proximity to research is how you keep that moat, because the hard problem moves — what was unsolvable last quarter has a paper about it this quarter. Ten minutes with someone who has already crossed it beats a week of solo effort.

• A real buyer beats a delighted user. Willingness to pay is the only validation that survives contact with a budget.

On how you build

• Encode the standard as a gate, not a habit. Evaluation-as-infrastructure and cross-family review — never letting one model family grade its own work — survive a deadline only when they’re gates in the system rather than intentions in my head. That is most of why I build an agentic operating system instead of a checklist.

• Ship the agentic workflow, not just the feature. A feature gets compared and swapped; a workflow gets adopted and becomes the backbone.

• Optimize for learning rate, not output. The variable that compounds isn’t hours worked, it’s how fast one loop closes — idea to shipped to measured to sharper idea. Shortening that cycle is the only advantage that keeps paying after the market moves.

• The founder cadence of shipping a real thing every 2–3 weeks even when the market takes longer to answer.

One more, less tactical

Every founder goes through lows. The hypothesis breaks. The pivot isn’t obvious yet. The loop feels like it’s grinding on you instead of forward. What separates founders who keep building isn’t grit — it’s having someone in the room who asks the right questions in those moments.

Eli did that for me more than once. He also believed the slop problem was worth solving and made that clear when it counted — and he’d be the first to tell me when I was drifting off it.

Cheerleaders are easy to find. So are critics. One person who is both — who backs you when you need backing, and says the hard thing when you’re off track — is rare, and worth more than most advice. You don’t get to appoint that person. You can only be the kind of person they’d bother doing it for.

The silver lining is always there. You often can’t see it alone.

Two courses worth your time

If you take one thing from this piece and act on it, make it these.

Agentic AI — the one I’d hand any engineer or founder building with agents. Evals, agentic workflows, and the autonomy-level framing that stops you from over- or under-trusting a system. Most of Part 3 above is this course meeting real deadlines.

AI for Everyone — not technical, and that’s the point. I recommend it to literally everyone: the PM, the designer, the ops lead, my own family. It’s the fastest way to give someone accurate intuitions instead of headlines.

What’s next

Two tracks, running in parallel.

The right build. I’m back in the market — applied AI leadership, agentic systems and AI-native engineering, or the right early-team seat as an IC architect.

What I care about is the layer underneath the product: how agents plan, execute, validate and evaluate; how you route models so cost scales with judgment rather than volume; how context survives a session boundary; and how evaluation gets treated as architecture rather than a phase that arrives after the build.

A side project, deliberately. Exponential OS started in February as a plugin I wrote to get my own resume out the door, and turned into an agentic operating system I use every day — alongside Co-Dialectic, the open-source prompt and context optimizer I maintain. Engineers started asking to buy it, which is the only reason I think of it as more than tooling.

I’m calling it a side project on purpose. I haven’t found a scalable problem yet. And the foundation is evolutionary enough that there’s no precedent to validate it against — which makes it genuinely interesting to build and genuinely hard to prove.

Here is the list above, turned on my own work:

• The hardest problem, on purpose. The part I took on is the one most teams route around: making a probabilistic system trustworthy enough to leave alone. Cross-family review where no model family grades its own output, eval gates between stages, refusal treated as architecture.

• The data loop from day one. Every gated ship writes a record — which model ran, what it cost, which gate caught what. That is data accumulated by operating, and nobody clones it by shipping the same feature. Where it is built to go next is the part I actually care about: a system that learns from its own operation, and — by consent, never by default — from how other teams run theirs. A workflow that gets sharper because somebody else’s build broke is a moat that compounds instead of depreciating.

• The standard as a gate, not a habit. It is a pipeline, which is what stops a deadline from quietly deleting the review step. Teams adopt a workflow. They only ever compare a feature.

• The room, not just the project. Four and a half years of teaching Fortune 500 executives, and I’m going back to it — because a room of executives asking why is still the fastest way I know to find the edge of my own understanding.

• Distribution, built in. I am the product’s first user: I run my own distribution through it, which lifted my LinkedIn engagement roughly 500% and my Substack growth well over 1,000%, and is how this piece reaches you. Eating your own dog food is not a slogan — it is the only way I find the defects a customer would have found first.

• Distribution I already had, without calling it that. Fifteen hundred Fortune 500 executives taught, the UC Berkeley Exec Ed AI alumni community I run, the Xoogler community, LPM, talks at SXSW and Step SF and ClawCamp, and writing like this. Channels compound quietly for years before you notice they are channels — which is also the honest reason to want an accelerator or an investor. The right one is a distribution asset before it is capital.

• Still open: the ICP, and the first real buyer. Finding who this is unmistakably for is work I still have in front of me, and engineers asking to buy is interest, not validation. Willingness to pay is the only proof that survives contact with a budget.

Those last two are why it stays a side project, and they are in here on purpose. A moat list you only ever apply to other people’s companies is a blog post.

And the moat this piece says nobody lists — still being there after the fifth pivot — is one you can only show. My ICP has moved five times since February: careers, then branding, then culture, then developers, then AI startups. Every move meant rebuilding something I had just finished. None of them changed the thesis, because the thesis was never the customer — it was the loop. I also killed two explorations this year on exactly the signal above. Persistence isn’t refusing to let go. It’s knowing which thing to hold.

The teams I want to work with build the system that builds the product. If that’s yours, let’s talk.

Grateful for Andrew Ng; Andy Ku and Jill Shih for the customer-discovery frameworks that reshaped how I think about pain; Eli Chen for the execution-first mindset — beachhead discipline, experimentation-based learning, and the “5% inspiration, 95% perspiration” frame — that I’ll carry into every build; Anurag Jain for the reframe on solving the right pain; Sandeep Gupta, a fellow EIR from my cohort, for the data-access-as-moat lens; Dan Landau for advising me on story, personal brand and how to market from day one; Oliver Heckmann for the Superset setup and the productivity practice around it — it rebuilt how I work day to day, and it spread to the people around me; the AI Fund team; and the founders who shared their stories with me during the EIR.

And a separate thank you to the people who didn’t just cheer from the sidelines — the ones who put their own time into this on purpose. The interns who built alongside me — Kaden Jackson, Sidhant Lochan and Michael Chen. The engineers who shipped things they didn’t have to. The ICP discovery partners who gave up their own hours to help me work out who the customer actually was. The design partners who let me watch them work and told me the truth about what was broken. The marketers who handed over hard-won insight and asked for nothing back. The AI Fund portfolio companies and the people there who took my calls, sat through interview after interview, opened their playbooks, and genuinely tried to be my ICP — several of them were a real fit, the commercial pull just wasn’t there yet. The friends who went looking for customers on my behalf without being asked.

Support is passive. These people were deliberate about it — and that is the difference between a hard three months and a good one.

They say it takes a village to raise a child. It takes at least as many to raise a startup, and I had four: the AI Fund portfolio, my UC Berkeley EMBA cohort and the Haas alumni network, the Xoogler community, and LPM — the Large People Model. Most of the customer interviews behind Part 1 came out of those rooms — strangers who gave me an hour to tell me what was actually broken. That is the network you query when the answer isn’t in anybody’s training data. Thank you for being part of the journey.

Onwards.

However, I came across this app at 1 AM in the morning, and it seemed interesting. So, I thought I would try. I paid $30 or so for it, and it said 100% money-back guarantee, etc.

It is a personal assistant. I couldn’t stop using it till 3 AM and felt the money was already paid for.

I did the following by filling in a few questions like ChatGPT, and it took care of the rest:

1. Deleted All Data Broker Data

2. Submitted Nuclear SPAM removal request with top 12 email campaign brokers

3. Added to Do Not Call List

4. Return request for Amazon that has passed the return date. It created a letter, AND IT MAILED it. Mindboggling!

5. Signup for FREE Crypto Airdrops

6. Filed unclaimed property request

7. For fun, I uploaded a doc to mail for FREE.

8. For fun, I requested an FBI profile about me.

9. For fun, I requested a credit report, even though I was getting this from other sources.

You can do much more - Online Notary, Surf free trials, beat parking tickets, sue robocalls for cash …

Use the following code to join so WE can get some FREE stuff!

https://thewhy.li/DoNotPayCode

This really speaks to people with high cognitive abilities that need scientific evidence for everything that they do.

This book is so needed in today’s world where 60% of the population is experiencing some form of depression from isolation and feeling lost.

Spirituality can be passed down to generations through genes (29%). 71% is in your hands (nurture). Go change yourself!

Blinkist Summary (Text/Audio) - Awakened Brain (Referral link to get discount)

Thrift Books - Awakened Brain

Please leave you comments about what you liked or didn’t like.