Tag: writing

Categories
AI

The Shape of the Question

Marc Andreessen made two claims recently that don’t quite fit together, and I haven’t been able to stop pulling at the seam.

The first: for almost any topic, the top AI systems now give him better answers than the world-class experts he could call on the phone. And he can call basically anyone. This isn’t a casual observation from someone without access — it’s a meaningful data point about what AI is actually doing to the value of expertise.

The second: the only real skill left in using AI is knowing what to ask. The models can already do almost anything you can describe in plain English. The bottleneck lives in your own head.

Hold those two claims next to each other. If the AI beats the experts, then the quality of your question only has to clear a low bar — good enough to unlock what the system already knows. You don’t need to ask like a cardiologist to get a cardiologist-quality answer. You just need to ask.

Except that’s not how it works in practice. And the gap between the two claims is where something important lives.

The better the question, the better the answer — even from a system that already knows more than any human alive. Expert-level interrogation of a superhuman system produces something qualitatively different from naive interrogation of the same system. The gap between a good question and a bad one doesn’t shrink because the underlying capability grows. It may widen. A sharper instrument in an unskilled hand doesn’t close the distance — it just makes the skilled hand more lethal.

What the AI has done is commoditize answers. What it has not done — cannot do — is commoditize the ability to know which question to ask.

There is a concept from epistemology that keeps surfacing here: the unknown unknown. Donald Rumsfeld made the phrase famous and then spent years living down the mockery, which was unfair, because the underlying idea is genuinely important. There are things you know you don’t know — the gaps you can name, the questions you can form. And there are things you don’t know you don’t know — the territory you can’t even see the edge of. The naive user of AI operates almost entirely in the second category. They ask what they already suspect. They get answers that confirm the shape of what they already believe. The system is brilliant and they are using it as a mirror.

The sophisticated user has learned to ask the AI to challenge their assumptions. To find the holes. To steelman the opposing view. To identify what’s missing from the framing. That second posture requires a kind of intellectual self-awareness — an ability to stand outside your own thinking and interrogate it — that is neither common nor easily taught.

Here is the uncomfortable implication: that self-awareness is not randomly distributed. It correlates with education, with reading, with having thought carefully about hard things for a long time. The people best positioned to ask good questions are, largely, the people who already had access to good answers through the old system. The gate moved. It didn’t disappear.

There’s a democratic story told about AI and I believe parts of it. The kid in rural South Dakota with a good question now gets an answer that rivals what the partner at McKinsey gets.

But access to information was never really the binding constraint. The binding constraint was always the ability to know what information you need — to feel the shape of your own ignorance precisely enough to ask for what fills it. That skill wasn’t distributed by the old system and it won’t be distributed by the new one. It has to be built, slowly, through years of reading and thinking and being wrong and trying again.

What AI may actually be doing is widening the gap between people who ask well and people who don’t — making the former dramatically more capable while leaving the latter approximately where they were, just with a faster way to get answers to questions they already knew to ask.

Somewhere right now, someone is sitting with the most capable thinking tool in human history, asking it to write a cover letter. The tool will do it beautifully. And the gap will quietly widen.

Categories
AI Living

The Threshold

There is a specific feeling. You are trying to understand something — a medical term in a lab report, a clause in a contract, how a particular piece of software actually works under the hood — and you hit the edge of what you know. The territory beyond is unfamiliar and the path is unclear, and something in you decides, quietly and almost without announcement: I don’t know how to figure this out.

And then you move on.

Marc Andreessen, talking to Joe Rogan recently, buried something important inside a longer riff about AI prompting tricks. Most of his list was the kind of thing you’d read in a productivity newsletter — ask it to steelman both sides, pretend it’s a panel of experts. Useful, not revelatory. But one observation was different: pay attention to the exact moment you think “I don’t know how to figure this out.” That’s the moment you should open the AI.

He said it almost offhandedly. I haven’t been able to stop thinking about it.

What he’s really describing isn’t a technique. It’s a behavioral pattern that most of us developed so gradually we don’t recognize it as a choice. The feeling of epistemic overreach — of arriving at the edge of one’s competence — became, over decades, a stopping condition. We learned to treat not-knowing as a wall rather than a door because, most of the time, it functionally was one. The library was closed. The expert was unavailable. The research was paywalled. You moved on.

The habit calcified. Now it persists even when the conditions that produced it no longer apply.

I notice it in myself, and I’m someone who is genuinely curious — who likes knowing how things work, who will follow a thread further than most people bother to. That’s not modesty; it’s relevant context. Because even with that disposition, I still hit the wall. I’ll be reading something and encounter a concept I only vaguely follow — some nuance in immunology, some historical episode I’ve only half absorbed — and I feel the familiar slight contraction, the small withdrawal. I read past it. The curiosity was there. The friction was higher.

Curiosity alone was never enough. What determined whether I pushed through wasn’t how much I wanted to understand — it was whether understanding felt retrievable at all. Most of the time, it didn’t. So I moved on, and the curiosity found something else to chase.

There’s a darker version of this worth sitting with. The people who never developed the quit reflex — who hit not-knowing and felt compelled rather than defeated — are, disproportionately, the ones who built things. The intellectual persistence wasn’t incidental to their contributions; it was probably constitutive of them. Curiosity as stubbornness. The refusal to accept the wall as final.

Elon Musk is the limit case. When he decided he wanted to go to Mars and found the rockets prohibitively expensive, he didn’t defer to the aerospace industry’s consensus about what was possible. He started reading propulsion manuals and cold-calling engineers. The quit signal either never fired or got overridden so fast it made no practical difference. The result was reusable orbital rockets, which the industry had largely decided weren’t worth pursuing. The dig reflex, taken to its extreme, rewrote what was considered feasible.

But the trait is undifferentiated. It doesn’t come with a calibration mechanism. The same refusal to accept expert consensus that produced SpaceX also produces a certain amount of confident wrongness — the Twitter decisions, the Covid takes, the occasional foray into geopolitics with the certainty of someone who has read a lot of Wikipedia. The dig reflex, unregulated, has no obvious stopping condition.

AI doesn’t change that underlying trait. What it changes is the access cost for everyone else.

For most of human history, the friction wasn’t random. It selected for people whose drive was strong enough to overcome it regardless of cost — the right connections, the right institution, the time to burn. Now that friction is lower for everyone, nearly to zero, for an enormous range of questions.

What I’m trying to build is the opposite of the quit reflex. Not the Musk version — boundless, uncalibrated, occasionally catastrophic. Something more modest: the habit of checking before giving up. Noticing the moment of not-knowing and treating it as a question rather than a verdict.

It requires noticing the moment. Which is harder than it sounds, because the reflex is fast and the moment is brief.

The contraction happens. You’ve already moved on. Somewhere behind you, the question is still there.

Categories
Haiku Living Reading

The Presence We Keep Deferring

I have so many unread articles saved to Instapaper that I’ve stopped checking the count. Each one felt, in the moment of saving it, like something I needed. A long piece on urban planning, a profile of someone interesting, a reported essay I fully intended to sit with.

The app is beautifully designed for exactly this — the frictionless capture, the clean reading interface waiting patiently on the other side.

What it can’t do is manufacture the attention I didn’t have when I saved it and still don’t have now. The articles aren’t the problem. The premise is: that presence is something you can bank.

There’s a haiku I keep returning to, from Natalie Goldberg’s Three Simple Lines. It’s by a poet named Fumiko Harada:

Morning chill
I savor this moment —
one meeting one lifetime

Eleven words. No verb in the third line, which makes it feel less like a thought and more like a verdict.

The Japanese concept underneath it is ichi-go ichi-e — loosely, “one time, one meeting.” It’s a Zen idea with origins in the tea ceremony, the understanding that each gathering is singular and therefore irreversible. You cannot archive it. You cannot search for it later. When it ends, it doesn’t go anywhere you can retrieve.

This is what the Instapaper queue is, at scale: an archive of moments I decided to experience later. The article about urban planning was written by someone who spent months reporting it, on a day when some editor thought it was ready, and landed in my feed on a morning when something about the headline caught me. That constellation doesn’t reassemble. Later is a different article.

The tools I use every day are getting astonishing. There are systems that can summarize, translate, recall, explain, anticipate. I use them. I find them genuinely useful.

But there’s a habit of mind they reward — a kind of perpetual deferral of full attention — that I haven’t fully reckoned with. The promise, always, is that you can engage more completely later, once the summary is ready, once the transcript exists, once the notes have been taken. Presence becomes a productivity tax you pay while waiting for a deliverable.

Harada’s haiku doesn’t moralize. The speaker isn’t lecturing herself into awareness. She’s just cold, and awake, and choosing to notice. I savor this moment. The word “savor” does a lot of work. It implies effort. You savor things that could be missed.

The pivot in the third line is what stays with me. One meeting one lifetime. Not “this meeting will last a lifetime” — that would be sentiment. It’s more like a mathematical statement: the cardinality of this encounter is one. There is exactly one of them. This morning, this particular chill, whatever conversation or solitude is happening inside it — that set has one element. By tomorrow it has zero. No amount of documentation changes that arithmetic.

I’m working on believing that.

Categories
AI History

The Arrival

Yoshua Bengio spent forty years building the foundation of modern artificial intelligence. He won the Turing Award for it. And he didn’t think he’d live to see it work.

That’s the quiet fact buried inside Stephen Witt’s New Yorker profile of him. Bengio — one of the three researchers whose decades-long bet on neural networks eventually became the architecture underlying every large language model running today — had made peace with the idea that the thing he was building was a multi-generational project. Something for his successors to finish. Then Witt writes: “one day in late 2022, the technology had simply arrived. He compared it to meeting an extraterrestrial.”

Hemingway once described bankruptcy happening two ways: gradually, then suddenly. He meant ruin. Bengio experienced something harder to name — not ruin but arrival, which carries its own vertigo. The gradually was four decades of work that most of his peers considered quixotic. The suddenly was a Tuesday in November when a chat interface went live and the world quietly changed.

What unsettles me about the extraterrestrial comparison isn’t the strangeness it implies. It’s the distance. You meet an alien; you don’t meet something you made. The metaphor suggests that even its creator couldn’t fully recognize it — that the thing, once arrived, belonged to a category that exceeded its own origins.

We don’t have good language for this. Breakthrough, inflection point, paradigm shift — these are words people reach for after the fact, when they’re building timelines. What Bengio seems to be describing is the experience of standing in front of a threshold you spent your life approaching, and finding it already behind you.

The technology didn’t ask permission. It didn’t announce itself.

It arrived.

Categories
AI Technology

The Bathwater Problem

Gary Kamiya was writing about the Tenderloin when he said it, but the line has been following me around: “The problem is that by saving the baby, you also save the bathwater.”

The pattern is remarkably consistent across every major information technology. Each one arrives promising to liberate the deserving — the faithful, the learned, the civic-minded — and each one immediately, inevitably, arms everyone else too. Gutenberg’s press was understood by its champions as a device for spreading the true Word; within decades it was the primary infrastructure for Protestant schism, Catholic counter-propaganda, astrological almanacs, and pornography. The reformers got their Bible. They also got their pamphlet wars.

The telegraph was greeted as a force for peace — shared information would make war irrational, commerce would bind nations. It also became the nervous system of commodity speculation, financial manipulation, and the first truly industrial-scale news hoaxes. The telephone: connection and the crank call, the crisis line and the threatening voice in the dark. Radio: FDR’s fireside chats and Father Coughlin. Television: Murrow taking down McCarthy, and also fifty years of manufactured consent. The internet: the largest library ever assembled and the largest sewer.

The pattern isn’t coincidental. It’s structural. Each technology expands what’s possible for human expression and coordination — and human expression and coordination contain both the noblest and the worst of us in roughly fixed proportion. The tool doesn’t change the ratio. It scales both sides of it.

What’s interesting historically is how each generation believes their technology will be different — that this time the architecture can be designed to select for the good. The internet era produced the most elaborate version of this belief: algorithmic curation would surface truth, network effects would reward quality, the wisdom of crowds would outcompete misinformation. Instead it turned out that engagement was the attractor, and outrage was the highest-engagement content. The bath got hotter.

The AI moment is the same belief system, restated with more technical sophistication. But the Kamiya line stands. You are saving a baby, and you are saving bathwater, and no one has yet designed a tub that can tell the difference.

The question isn’t whether the bathwater comes with the baby. It always does. The question is whether you turn on the tap.

Categories
AI Books Writing

The Tax We No Longer Have to Pay

When Carol Coye Benson and I sat down to write Payments Systems in the U.S., one of the first problems we had to solve wasn’t about payments. It was about history.

To understand why the ACH network works the way it does, or why checks persisted decades longer than anyone expected, you need the institutional sediment underneath — the regulatory decisions, the failed experiments, the path dependencies baked in by choices made in the 1970s that nobody thought would still matter in the 2000s. The history is the explanation. Strip it out and you have a description of current practice with no account of why it exists or what it cost to get there.

But history takes pages. And pages test a reader’s patience. So you compress. You make judgment calls about what survives the cut and what gets left behind, and you make those calls knowing that every omission is a bet — a bet that the reader can follow without it, that the thread holds without that particular knot.

Writing it taught me something. The act of compressing, of finding the minimum sufficient version of a complex thing, forces a clarity that living inside the complexity never quite delivers. You don’t fully know what you understand until you have to say it precisely enough for someone else to follow.

But compression is always a loss. You feel it as you write. The version in the book is thinner than the thing you know.

Garry Tan uses a term — “tokenmaxxing” — that initially sounds like jargon from a performance optimization thread. The idea is simple: don’t be stingy with context. Give the model everything. Every source document, every relevant article, every piece of background that a human reader would never sit still for. Let it synthesize rather than guess.

The instinct it runs against is deep. We have spent decades building information systems around compression — search engines that retrieve rather than ingest, executive summaries that stand in for reports, one-pagers that distill months of work into something a decision-maker can absorb in four minutes. All of it was a rational response to a real constraint: human attention is finite and expensive. You couldn’t afford to read everything, so you built filters. The whole architecture of how organizations manage information was designed around that limit.

Tokenmaxxing is a bet that the limit has moved.

The model can read everything. The cost of giving it full context — the uncompressed history, the original sources, the institutional sediment — is low enough now that filtering before the model sees it may introduce more error than it prevents. You’re potentially discarding signal when you summarize for the model the way you’d summarize for a human. The model doesn’t need the one-pager. It can handle the report.

This doesn’t dissolve the need for curation entirely. More context isn’t always better — models can lose the thread in noise the same way humans do, just differently. The skill shifts from summarizing to selecting: not what’s the minimum version of this but what’s actually worth including. Different judgment, still essential.

But the deeper change is upstream of any particular project. The compression we built into every research process, every briefing, every book — that was never the goal. It was the tax we paid for human cognitive limits. Part of the process doesn’t pay that tax anymore.

When I think about writing that payments book today, I don’t think the book itself would change much — it still has human readers with finite patience. But the map we drew before writing it, the synthesis work, the “what connects to what across fifty years of regulatory history” work — that could happen at a different depth now. The understanding you bring to the writing can be informed by everything, not just the subset you had time to read.

The payments book was written entirely for humans, with all the compression that implies. But Tyler Cowen just published what he calls a “generative book” — 40,000 words released free online, paired on the same screen with a Claude interface so readers can discuss, interrogate, and extend it in real time. He’s writing for both audiences simultaneously now. The human reader and the model that will help that reader go deeper. The text is optimized not just to be understood but to be used — as context, as a jumping-off point, as raw material for a conversation that the author won’t be in.

That’s a different kind of writing. Not better or worse. Different. The compression decisions change when one of your readers has no patience to protect.

Writing still clarifies thinking. That part hasn’t changed. But what you’re clarifying, and who you’re clarifying it for, is quietly expanding.

Categories
Business Creativity Space SpaceX

Test like you fly!

There’s a phrase in the SpaceX documentary that keeps coming back to me: “Test like you fly.” It sounds like a slogan. The kind of thing that gets painted on a factory wall and eventually stops meaning anything. But the more I sit with it, the more I think it’s actually a philosophy that reaches well beyond rocket engineering.

The video — a 25-minute documentary SpaceX released last week — is ostensibly about Starship Version 3. New ship, new booster, new engines, new pad, new test site. Everything rebuilt. And they’re not shy about framing it as a reset, not an upgrade. One description I read called it “a quiet violence in progress.” That phrase stopped me cold, because it’s exactly right. Progress that looks violent from the outside — all that fire and metal — but is somehow quiet in its inevitability.

What moved me watching it wasn’t the engines. It was the engineers. SpaceX put the people on camera: the ones running cryogenic pressure tests at 80 Kelvin, stress-testing tank structures at 70% proof, explaining their failures and their data with the flat affect of people who have made peace with how long hard things take. There’s something almost monastic about it. You choose a problem that will not yield easily. You accept that the work will outlast any individual sprint of enthusiasm. You go back to it anyway.

I keep thinking about that in the context of what we’re doing with AI — the other enormous, fast-moving project that I spend so much of my mental energy on. The development arc is different: iterative releases, weeks not years between jumps, demos that blur into deployment. But the same principle is buried in there somewhere. The best AI teams I read about aren’t the ones shipping the most polished demos. They’re the ones building infrastructure for failure — evals, red-teaming, structured feedback loops. Test like you fly.

The Raptor 3 engines now produce 280 metric tons of thrust each. Thirty-three of them on a Super Heavy booster means over 17 million pounds of liftoff force. I have no intuitive frame for that number. What I do have a frame for is what those numbers represent: three years of iteration on top of five years before that, on top of a theoretical foundation laid by people who didn’t live to see any of this. There’s a compounding in that which I find genuinely moving. Nobody built the Raptor 3 in isolation. It came from everything that broke before it.

The hardest part of the documentary isn’t the engineering. It’s the implicit acknowledgment of how much remains undone. No Starship has yet achieved full orbital velocity with both stages intact. In-space refueling is still untested. The thermal protection systems need more work. And yet — SpaceX talks about unmanned cargo missions to Mars before the end of this year like it’s on the roadmap, not the wish list. That sentence used to sound like marketing. Watching the footage, it doesn’t anymore.

I’m not sure what to do with that feeling exactly. It’s something between awe and vertigo. We’re living in a moment when the audacious has started to have quarterly milestones. When the impossible keeps showing up on timelines and then — bewilderingly, uncomfortably — meeting them.

Test like you fly. Fail with rigor. Build the thing you actually need, not the thing you could more easily explain.

I keep turning that over. There’s a post in there somewhere about writing, too — about the drafts nobody sees, the structural tests that fail, the versions that taught you the one that worked. But that’s for another day.

For now I’m just sitting with the footage of those 33 engines lighting up, and the quiet weight of how much went wrong before they could do that.

Categories
Writing

The Unfinished Note

I’ve been sitting with a Susan Orlean line for a few days now, the way you sit with a splinter you can’t quite locate.

“Stories don’t need a ‘conclusion,’ a flourish of finality. It’s better to leave readers falling forward, tumbling through the piece and beyond it, finishing the tune in their heads.”

What strikes me isn’t the advice — plenty of writing teachers have said something like it — but the verb she chose. Tumbling. Not drifting. Not lingering. Tumbling. There’s a loss of control in that word, a small helpless momentum, the way you take one more step than you expected on a dark staircase and your body has to catch up to itself.

I’ve always been suspicious of endings that arrive wearing their own bow. You can feel them coming, those last paragraphs — the rhythm slows, the sentences get more declarative, the writer seems to straighten up and clear their throat. And then comes the lesson, the restatement, the turn toward uplift or hard-won wisdom. The piece explains what it was about. You close the browser tab and that’s the end of it.

But some pieces don’t end so much as they stop, at the right moment and the right angle, and something in you keeps moving. You find yourself thinking about them in the shower two days later. You’re not remembering the conclusion because there wasn’t one. You’re still inside the piece, finishing the tune, as Orlean says. The writer handed you the melody and walked off mid-phrase.

I think about this with music. Jazz, especially. The best solos don’t resolve — they suggest a resolution and then leave the air charged with it. Miles Davis understood that the note you don’t play is still a note. The silence after the phrase is part of the phrase.

I’m not sure I’ve ever actually written an ending this way. Most of my pieces come in for a landing; I can feel myself starting to circle and descend. Maybe that’s the real lesson in Orlean’s line — not a craft note about structure, but a challenge to trust the reader enough to leave the door ajar. To believe the piece was good enough that they’ll want to keep walking around inside it.

I’m still not sure I do.

Categories
Storytelling Writing

The Nerve of the Opening Line

For years I wrote first paragraphs that explained what I was about to say, which is a little like a joke that begins by describing how funny it is.

Susan Orlean has a better idea. In her book Joyride, she writes that a lede doesn’t need to preview the story or summarize what the rest of the piece will be about. What’s important is that it captivates readers and holds them fast to the page so they keep reading.

The conventional wisdom about ledes is that they exist to tell readers what they’re about to read. The billboard theory of the opening. Here is what this story is. Here is why it matters. Here is what you’ll find if you continue. The lede as table of contents, compressed.

Orlean is saying something stranger and more honest: the lede’s job is not to inform. It’s to hold the reader.

There’s a distinction there worth sitting with.

Informing a reader is a transaction — you transfer content, they receive it. Holding a reader is something else entirely. It’s closer to what a magician does in the first thirty seconds of a performance, or what a stranger does when they say something at a party that makes you turn and face them fully. You haven’t learned anything yet. You’ve just been made to stay.

The ledes that have held me longest tend to have almost nothing to do with the stories they open.

Joan Didion begins “The White Album” with a single sentence — “We tell ourselves stories in order to live” — that takes the entire essay to even partially fulfill.

Gay Talese opens his Frank Sinatra profile not with Sinatra’s voice or his legend but with a man going silent: “Frank Sinatra, holding a glass of bourbon in one hand and a cigarette in the other, stood in a dark corner of the bar between two attractive but fading blondes who sat waiting for him to say something. But he said nothing.”

Tracy Kidder opens The Soul of a New Machine not with computers but with a boat in a storm, Tom West awake for four straight nights while everyone else is seasick, the rest of the crew left wondering what on earth this man does for a living.

None of these ledes summarize. All of them hold.

What they share, I think, is a quality of disturbance. They’ve moved the ground slightly underfoot. Something is tilted.

Didion’s first sentence argues that we tell ourselves stories in order to live, and you feel the vertigo in it immediately — wait, is that true? Is that a good thing or a desperate thing?

Talese gives you a man diminished by illness and silence, and everything that follows is measured against that diminishment.

Kidder’s boat goes somewhere that prose about minicomputers wouldn’t, and by the time you’ve crossed that dark water with West, you’re already a different kind of reader than you were on page one.

I think about this when I try to write.

I grew up reading ledes the billboard way — I thought the first paragraph was a promise about what the reader would receive. And sometimes I still write them that way, which is to say I write them first and delete them later, because they’re stage fright disguised as generosity. Here is what I’m about to tell you really means please don’t leave before I find my footing.

The Orlean formulation — captivate, hold, keep reading — shifts the pressure off the writer’s anxiety and onto the reader’s experience. The question is no longer what do I need to tell them? The question is what will make them unable to leave?

That’s a harder question. It requires knowing something about what people can’t resist. Strangeness. Motion. A body in trouble. A door left open. The suggestion that someone knows something you don’t.

The best ledes I’ve ever written didn’t come first. They came after I’d already written the whole piece and finally understood what it was actually about — which turned out not to be the thing I thought it was about at the start. You can’t write the sentence that makes someone stay until you know what you’re asking them to stay for.

The lede isn’t a promise. It’s a wager. You’re betting that the reader will follow disturbance into the dark — and the only way to make that bet is to trust the disturbance yourself first. Most of us don’t. Most of us write the billboard because we’re afraid that if we don’t explain what’s coming, the reader will leave.

But the reader doesn’t leave because they’re confused. They leave because nothing reached out and held them.

The explanation never does that. The strangeness might.

Categories
AI Thinking Tools

Outsourcing Thinking but not Understanding

There’s a line mentioned in a recent discussion by Andrej Karpathy that I keep turning over: You can outsource your thinking but you can’t outsource your understanding.

It sounds like a warning. Maybe it is. But the more I sit with it, the more it feels like something older — a distinction philosophers have been trying to draw for centuries, suddenly made urgent by the fact that we now have a tool that makes outsourcing thinking almost frictionless.

Here’s what I notice when I use AI well: I get the answer, and I feel satisfied. There’s a small dopamine tick. Task closed. But if someone asks me an hour later to explain the reasoning, I often can’t. The thinking happened — somewhere — but not in me. I was a conduit. A confident one, too, which is the dangerous part.

This is different from looking something up. When I Google a fact and paste it into a document, I know I’m borrowing. The seam is visible. But when I ask an AI to reason through a problem with me, the output arrives in first person, in fluent prose that matches my own register, and something in my brain says I worked this out. The seam disappears. That’s new. That’s the thing we don’t yet have good instincts for.

Karpathy’s deeper point is about construction. He’s a builder by temperament — his mantra, which he traces to Feynman, is that if you can’t build it, you don’t understand it. What you can’t yet construct, you merely think you understand. There are always micro-gaps in your knowledge, invisible until you try to arrange the pieces yourself and find they don’t quite fit. The AI doesn’t change that equation. It just makes it easier to mistake the map for the territory — and to feel strangely proud of a map you didn’t draw.

Hesse understood this, in a different century and a different idiom. In Siddhartha, the young seeker travels to meet the Buddha himself — the most perfectly articulated wisdom in the world, delivered by the man who actually found it. Siddhartha listens, acknowledges that the teaching is flawless, internally consistent, the most complete account of liberation ever assembled. And then walks away. Not from arrogance, but from recognition: even the Illustrious One cannot hand you his liberation. The path was his. He walked it. That walking is not transferable, no matter how perfect the description of the destination. Received knowledge, however exquisite, is not the same as earned knowledge. The gap between them is exactly the size of your own unlived experience.

That’s the same argument, made across two and a half millennia. Feynman says you have to build it. Hesse says you have to live it. Karpathy says the AI can do neither for you.

He’s also made a related observation about educational video — that a lot of content on YouTube gives the appearance of learning but is really just entertainment, convenient for everyone involved. Nobody has to do the hard part. AI-assisted thinking has the same shape, just more intimate. You’re not passively watching — you’re actively typing, prompting, engaging. It feels like cognition. But engagement isn’t understanding. Typing a question is not the same as wrestling with it.

I don’t think the answer is to use AI less. That’s not Karpathy’s argument either — he’s spent the last year building a school premised on AI tutors expanding what people can learn. The lesson is about custody. When I hand a problem to an AI, I need to stay in the loop as a learner, not just as a reviewer. There’s a real difference between asking give me an answer and asking help me build the reasoning. The first outsources thinking. The second — if you insist on it, if you refuse to be a passenger — can still leave the understanding in you, where it belongs.

But insisting is the work. And the work is now easier to skip than it has ever been.

Understanding isn’t a product you receive. It’s a residue — what settles in you after genuine struggle, after the confusion and the dead ends and the small hard-won moments of clarity. Siddhartha couldn’t get it from the Buddha. You can’t get it from the AI. Karpathy’s line is a custody argument: the thinking can travel, but the understanding has to stay home.

What unsettles me is that we’re building tools that make the borrowing invisible — that dress outsourced reasoning in the first person, that let us feel like we’ve understood something we’ve only processed. Siddhartha at least knew he was walking away from the teaching. He felt the gap. We might not even notice ours.