Tag: artificial intelligence

Categories
Living

Why the Light Leaves Mornings First

I woke up at my usual time this morning and noticed the room was a little dimmer than it had been just a week or two earlier. The light felt more reluctant to arrive. It wasn’t dramatic — just enough to make me check the clock and wonder whether the days were already turning.

They were. And the change had begun earlier than I expected.

Here on the San Francisco Peninsula, the earliest sunrise of the year came around June 12. The summer solstice — the official longest day — arrived on June 21. Yet the latest sunset didn’t occur until June 28. That gap means the mornings started shortening while the evenings were still lengthening. Nature didn’t wait for the solstice to begin reclaiming the light on the side of the day I care about most.

As a lifelong morning person, I’ll admit this feels slightly unfair. I treasure those early hours when the light arrives gently and the day still feels open and possible. Learning that those hours began to shrink first, while the evening light held on a little longer, feels like a quiet trick played on people who love the dawn.

The total amount of daylight still reached its peak on the solstice itself, when the Sun stands highest in our sky. But sunrise and sunset don’t move in perfect step with each other. Two subtle astronomical effects are responsible: the gradual shift in the Sun’s declination and the Equation of Time — the small irregularity in the Sun’s apparent motion caused by Earth’s tilted, elliptical orbit. Together they create this gentle asymmetry. The mornings give ground first.

By today, July 10, both ends of the day are shortening. Sunrise here is around 5:54 a.m. and sunset around 8:32 p.m., bringing us down to roughly fourteen hours and thirty-eight minutes of daylight and losing about a minute each day. The process, though, began weeks ago — quietly, on the morning side.

I’ve been reflecting on how often change arrives this way: unevenly, and rarely all at once. One part of a season or a life begins to shift while another part still feels steady. It’s easy to grumble when the change touches something we love. But there’s also an invitation in noticing it — to pay closer attention to these transitions, to savor what remains abundant, and to stay curious about the small asymmetries that shape our days.

This morning the light is arriving more quietly than it did just a few weeks ago. Later I’ll step outside earlier than usual, just to meet it while it’s still generous. The day is already growing shorter on my preferred side, yet it remains long enough for a walk, for gratitude, and for whatever the remaining hours might bring.

Nature doesn’t negotiate. It simply keeps showing up — sometimes a little earlier, sometimes a little later than we’d prefer. And perhaps that, too, is part of what we’re meant to notice.

Categories
AI China Youth

The Arithmetic of Youth

The first meeting was at one of the banks on a high floor somewhere in Shanghai, the kind of view that turns a city into an abstraction. It was 2005, and I was there the way American investors were there that year — curious, a little jet-lagged, trying to read a country that was rewriting itself faster than anyone could print the new edition. Across the table sat a management team, and what struck me wasn’t anything they said. It was how young they were. Not junior-young. Running-the-company young.

Afterward — in the hallway or the car, early in the trip, when I still had the confidence of someone who thought he could just ask — I put the question to one of our local colleagues. Casually, expecting a casual answer. Something about a young country, a young economy, energy meeting opportunity.

The answer I got instead was the Cultural Revolution.

There was a generation, she explained, that simply wasn’t there. Sent to the countryside, pulled out of universities, handed shovels instead of textbooks. By the time China opened back up, that cohort had a hole in it — a rung missing from the ladder. So the young people I’d just watched run that meeting weren’t there because anyone had bet on youth. They were there because there was no one older left to put in the chair. Youth, in that boardroom, wasn’t a strategy. It was a vacancy dressed up as one.

I have thought about that answer, off and on, for twenty years, without knowing what to do with it. Then a few weeks ago I read a summary of a conversation with Nathan Lambert — an AI researcher who’d just spent time visiting the frontier labs in Beijing and Hangzhou — and I found myself back in that room, except everything about the youth in it had flipped.

He describes teams at places like Moonshot AI as almost absurdly young, tight-knit, close to giddy about the work — “the best vibes,” he calls it. Zhipu AI, he says, has built something close to an AGI showroom, a physical space engineered to perform confidence for whoever walks through the door. These aren’t companies with a hole where the experienced people should be. These are companies that went looking for twenty-five-year-olds because twenty-five-year-olds move at the speed frontier AI research demands, and installed them at the center of the room. The showroom isn’t hiding a vacancy. It’s staging a choice. That’s panel two.

Same demographic. Same first city — Beijing both times — with a high-speed rail line now running to Hangzhou instead of whatever second city I’d have named twenty years ago. Opposite cause. In 2005, youth in the room meant a generation had been taken from the labor force involuntarily. In 2026, youth in the room means a generation has been selected for it, deliberately, competitively, because being young is now the qualification rather than the disqualifier. The Cultural Revolution left a gap that youth filled by default. The AI boom left a door that youth is filling by design.

I would have stopped there, satisfied with the irony, except for a number I couldn’t get out of my head once I went looking for it: 15.6 percent. That’s China’s urban youth unemployment rate — ages sixteen to twenty-four, university students excluded — as of May 2026, and it counts as good news, down from 16.3 percent in April. A year earlier it had spiked to nearly nineteen percent in a single August, the month twelve million university graduates walked out of commencement and into a labor market that had no idea what to do with them. Some will sit for civil service exams, chasing what people there still call the iron rice bowl — the illusion of permanence a state job used to guarantee, back when your grandparents didn’t choose their careers so much as get assigned them. Others will enroll in another degree, not because they want one, but because a classroom is a more dignified place to wait than an unemployment line.

So there is a third panel now, and it doesn’t fit neatly next to the other two. It isn’t a vacancy, and it isn’t a showroom. It’s just a very large number of young Chinese people who did everything they were told to do — studied hard, got the degree — and are standing outside a door that isn’t opening. And somewhere behind that door, in a much smaller room with much better lighting, another group of young Chinese people, maybe the same graduating class, are building the technology that a Silicon Valley researcher travels overseas to admire for its vibes.

I don’t think those two rooms are as separate as they look. I think the showroom is real, and I think the twelve million are real, and I think the mistake — my mistake, sitting here in Menlo Park two decades removed from that conference table — is letting either one stand in for “Chinese youth” as if it were a single sentence instead of a population. The Moonshot AI team is not a representative sample. It’s the visible sliver of a generation, selected with a precision that turns the unemployment numbers into part of the same mechanism — one sorting process, not two unrelated stories. The best vibes in that lab and the worst numbers in that economy might just be describing the two ends of the same funnel.

I keep coming back to that hallway in 2005, and to how confident I was in the question I asked — as if a generation’s youth could only ever be telling one story. It couldn’t then, and it can’t now. I got a true answer that day and thought I understood something. I understood one panel of a triptych I hadn’t seen the rest of yet — and I’m still not sure I’ve seen all of it.

Categories
AI

Context Rot

Here is a small, possibly embarrassing confession: I have never, not once, gone looking for the best AI model.

I have a model. It lives in a browser tab — Safari, usually, on whichever device is nearest, occasionally Chrome if I happen to be at the desktop. It does what I need — drafts an email, untangles a sentence, tells me what a Norwegian emigration record from 1856 probably says — and then I close the tab and go on a walk.

Somewhere out there, presumably, a much smarter, much more expensive machine is doing something extraordinary with protein folding or hedge fund arbitrage or the outer edges of mathematics I will never visit. I have made my peace with never meeting it.

This did not used to feel like a confession. For a while there — a year, eighteen months — it felt like the central drama of the whole industry: which model was “best,” who had it, who had lost it, whether some lab’s quarterly earnings call would reveal that the frontier had quietly moved sixty miles down the road while everyone was looking the other way. Benchmarks were released like box scores. People argued about them the way people argue about batting averages, with the same weird intensity, the same conviction that a two-point difference in some abstract reasoning test settled something important about the future.

And then, at some point I can’t quite date — it crept up, the way these things do — I noticed I had stopped caring.

Not because the frontier stopped moving. It didn’t. It’s still moving, arguably faster than ever, in ways that occasionally show up in the news with all the drama of a soap opera (a delayed launch, a researcher poached, a stock down five percent in an afternoon, always something).

I stopped caring because none of it touched me. My model — whatever it was, this week — had long since crossed some invisible threshold past which more didn’t register as more. It was already better than I needed. It has been better than I needed for a while now. I suspect I am not unusual in this. I suspect most people, doing most things, most days, are operating comfortably inside a capability surplus so large they’ve stopped noticing it’s there, the way you stop noticing a room is warm.

If the top of the model isn’t for people like me — and it increasingly isn’t — then who, or what, is it actually for? I went looking for one piece of the answer and found, instead, a metaphor.

It’s called “context rot.” I have to admit, before I go further, that I’m not sure I’ve ever felt it myself — which, on reflection, is its own small piece of evidence. My sessions close in minutes, not hours. I ask, it answers, I leave. Whatever happens to a model over the fourth or fifth hour of sustained, dependent work is a country I simply don’t visit.

But other people do, increasingly — entire teams do, for entire projects — and what they’re finding out there is worth understanding, even secondhand. It describes something that happens to AI models when they’re asked to work for a long time on something complicated — not five minutes, but five hours; not one question, but a hundred small decisions stacked on top of each other, each one depending on the last.

You’d think the limiting factor would be room. Models have a “context window” — a stated capacity, like a gas tank, measured in tokens, and for a while the marketing numbers on these were the whole story: two million tokens! A library! And you’d think, as with a gas tank, that the thing runs fine until it’s empty and then it stops.

That is not, it turns out, what happens. What happens is closer to what happens to your desk.

You know the desk. Everyone has the desk. It starts the morning clean — an aspirational, almost insulting cleanliness — and by four in the afternoon it is a geological record of the day: three coffee cups, a stack of things you meant to file, a Post-it with a phone number you no longer need, the good pen buried under a printout of something you already dealt with an hour ago. The desk is not full. There is, technically, room. You could clear a space if you tried. But you don’t try, because functionally, cognitively, the desk has stopped being usable long before it ran out of surface area. You start looking for the stapler and forget what you were stapling. This — and I did not make this term up, I want to be clear, though I wish I had — is context rot. The window hasn’t run out. The signal has just drowned in its own debris.

Researchers watching this happen to long-running AI agents have found something almost cruelly elegant about how it fails: it doesn’t fail gradually, the way you’d expect a desk to get gradually messier. Errors compound. A task that takes twice as long doesn’t get twice as likely to go wrong — the failure rate roughly quadruples. Two mistakes early in a long chain of dependent steps don’t add up to a slightly worse outcome. They multiply into something close to total collapse, four hours in, for reasons that trace back to a single bad assumption made in the first twenty minutes and never revisited.

Here is where the frontier comes back in — not as the whole answer, but as a piece of one.

It is not that frontier models are smarter in the way a benchmark measures smart — better at a single hard math problem, a cleverer turn of reasoning. Plenty of models can do that now; the “good enough” tier has crept remarkably high.

It’s that frontier models are apparently, marginally, meaningfully better at not rotting. At keeping the desk usable at hour six. At knowing which of the forty things on the desk actually still matters and which is a coffee cup that should have been thrown out an hour ago. This is a genuinely different kind of intelligence than the one benchmarks were built to measure, and it is almost invisible from the outside — you don’t see it in a single exchange, you see it only in the difference between a project that holds together over three days and one that quietly, subtly, stops making sense somewhere around Tuesday afternoon and nobody notices until Thursday.

If that’s true — if the frontier’s real edge is durability rather than raw cleverness — you’d expect to see it show up in how the labs actually deploy their own models: saving the sharpest tools for the tasks that need to survive the longest.

I went looking for a real-world example and found one closer to home than I expected: Anthropic’s own Slack tool, the one where you tag the AI into a channel the way you’d tag a coworker, and it works alongside a whole team over days, learning the channel as it goes. It runs on a serious, capable, thoroughly frontier model — but not, it turns out, on the company’s very best one. That one is held back, reserved for a smaller and stranger set of problems nobody has solved before at all. I sat with that for a while. The tool built to survive a whole team’s whole week, in public, under the most sustained pressure any of their products face, wasn’t handed the sharpest blade in the drawer. It was handed the second-sharpest — which was apparently, entirely, enough. Which tells you something about where the two kinds of intelligence actually diverge: the merely-very-good model handles the desk staying clean for a week, in public, in front of a whole team, where one bad assumption made Monday and never revisited would be visible to everyone by Thursday. The truly new capability is being held in reserve for something else altogether.

I don’t have a tidy place to land this, and I’m suspicious of anyone who does. But here’s the closest I can get.

Imagine a three-Michelin-star chef — the kind of person who has spent thirty years learning to coax something transcendent out of a single scallop, who can tell you, by smell, that a stock has forty more minutes in it — standing at your stove on a Tuesday night making you a grilled cheese sandwich. It will, I promise you, be a very good grilled cheese sandwich. The bread will be evenly golden. The cheese will have reached some ideal, fully-considered state of melt. But almost none of what makes that chef extraordinary is actually being used to make it — none of the thirty years spent learning to hold forty things in mind at once without losing track of any of them, the exact skill, it occurs to me, that keeps a long, complicated project from quietly falling apart on day three. The technique is idling. The thirty years are in the room, present, available, and almost entirely beside the point, because a grilled cheese sandwich was never the place where thirty years shows up. It shows up somewhere else — in a dish you will never order, on a night you weren’t there.

What you got instead, on your ordinary Tuesday, was simply more than enough.

Categories
AI AI: Inference Semiconductors Uncategorized

5 Critical Management Lessons from the Founders at Etched

How two young founders are building what could become one of the most important companies in the AI era — and what their story teaches about leadership, execution, and building at the edge of the possible.

I recently listened to the latest Invest Like the Best podcast from Patrick O’Shaughnessey which was a remarkable conversation with Gavin and Rob, the founders of Etched, the company building specialized AI inference hardware that’s aiming to be radically better than existing solutions. Their story — starting as very young founders against massive skepticism, raising serious capital, and now shipping full rack-scale systems — is packed with hard-earned wisdom.

One of the comments Patrick makes at the beginning was how during his due diligence on the company he kept being told that semiconductor technology wasn’t a place for young people. You need seasoned, middle age experts to master this domain. Exactly not these founders.

Note: the following is based upon an AI’s analysis of the conversation transcript with me asking “What are the five most important management lessons from this conversation?” These lessons are relevant whether you’re leading a team, building a product, or simply trying to do meaningful work in our fast-moving world.

1. Velocity Compounds — Prioritize Speed Ruthlessly

In hardware, and increasingly in any deep-tech endeavor, speed isn’t just an advantage; it’s often the deciding factor.

Etched didn’t just design a chip — they built the full inference solution (chip, board, power delivery, interconnects, cold plates, and production processes) in parallel. They sent engineers to live in Bangalore for months to unblock vendors. They ran 24/7 shifts and did massive pre-work (including putting full chip designs on FPGA clusters) so that when the silicon finally arrived, they had working inference in racks in just 40 days.

Key takeaway: Look for every opportunity to parallelize. Accept higher short-term costs if they buy meaningful time. As they put it, “You win by shipping.” The best part is often no part — and the best vendor is no vendor, when vertical integration lets you move faster. Velocity, velocity, velocity.

2. Build Teams with Legends + High-Drive Talent

One of the most distinctive parts of their approach is how they recruit. They seek out “Legends” — people who have done the hardest versions of the problem before (like the engineer who built Nvidia’s HGX and DGX systems) — and pair them with exceptionally driven, somewhat naive high-performers who refuse to accept conventional limits.

They use “project-based recruiting,” mapping the hardest technical problems ever solved and persistently pursuing the actual people who did the real work. Their culture self-selects for people willing to move their families to San Jose to bet on two young founders taking on the world.

Key takeaway: For breakthrough work, average talent doesn’t suffice. The combination of deep experience and raw, first-principles energy creates magic. Invest heavily in finding and retaining these people — even if it takes 20 conversations. You can also learn a lot if the best in the world talent turns down the opportunity to work with you!

3. Assume It’s Possible, Then Solve the “Unsolvable” Problems

Repeatedly in their story, experts told them certain things were impossible. Their response? Assume it is possible and figure out how.

The most striking example was a clock domain crossing issue that required aligning signals to within 50 picoseconds — something many engineers said couldn’t be done. People quit. They solved it in about two weeks during a very dark period.

Key takeaway: When you hear “impossible,” treat it as the beginning of the investigation, not the end. Cultivate a “find a way” mindset across the team. The moments when things feel hopeless are often when the most important progress happens. I’m constantly struck by how often persistence results from simply realizing (or assuming) that something is actually possible.

4. Production Is the Real Product

Etched’s mantra is “Production is the product.” They obsess over not just technical performance but manufacturability, supply chain resilience, serviceability, and the ability to scale to gigawatts.

They made deliberate choices around process nodes and memory to avoid zero-sum competition. They built their own factory processes and test infrastructure early. Future designs are being simplified specifically for faster production cycles and higher reliability at massive scale.

Key takeaway: In any business that hopes to reach real scale, think end-to-end from the beginning. Technical excellence without production excellence is just a prototype. Optimize for output (tokens, units, whatever your metric is) at volume. There’s a lot of “zero to one” thinking here.

5. Bet Big and Stay Existentially Focused

Building in semiconductors requires enormous capital. Etched raised roughly $100 million early on when they were still very young and pre-tapeout — after most traditional investors had passed. They knew half-measures wouldn’t work.

This existential focus (this one product determines whether the company lives or dies) creates a different level of intensity that attracts talent, suppliers, and customers who believe.

Key takeaway: Match your ambition with appropriate resources and commitment. Clear existential stakes help filter for the right people and partners. In a world of distractions, singular focus on what truly matters is a superpower.

Final Thoughts

Gavin and Rob’s story is the combination of technical sophistication and deep human resilience. They faced a tough personal battle with cancer (in Rob’s case), widespread doubt, brutal technical challenges, and fundraising pressure — and kept moving forward with curiosity, determination, and humility.

In an age of AI and accelerating technology, the ability to build teams that can solve seemingly impossible problems at speed may be one of the most valuable capabilities a leader can develop. Their example reminds us that the future belongs not just to the smartest, but to those who can execute with urgency while maintaining clear principles. Velocity, velocity, velocity.

Categories
AI Learning Photography

Autopilot

“Superb photographs are not just taken with cameras. They come from within you, your eyes, your mind, your heart, not ice cold equipment.” Fan Ho

There’s a half-second on the street, somewhere between seeing a frame and shooting it, that used to take me whole minutes. Early on, with a camera in my hands on the streets of San Francisco or on the subway platforms in New York, I’d see something — light falling a certain way, a gesture about to resolve into a gesture — and I’d think my way through it. Assess the composition or the angle. Worry about the background. By the time I’d worked it out, the moment might be gone, replaced by some lesser version of itself.

That doesn’t happen to me anymore, and I couldn’t tell you when it stopped. Somewhere along the way the thinking disappeared and the shooting stayed. I see the frame and the shutter goes, and only afterward, looking at the file, do I understand what I saw. I didn’t explicitly decide to skip the thinking. It just stopped showing up, the way a habit eventually stops asking your permission. Or how driving a car becomes second nature.

I think about this because of a problem the AI labs have been calling continual learning. The AI models we use are like brilliant interns. They can solve a hard problem at nine in the morning and a harder one by five, and they’ll astonish you doing it. But every session starts over from zero. Whatever they got right on Tuesday evaporates by Wednesday, the way a dream is gone by the time you’ve found your slippers.

The industry’s first answer was to give them a longer memory — let the window hold the whole case file in front of them, all the time. This works for a while, the same way it would work for me on the street if I stopped and re-derived the exposure math for every frame. But that isn’t how I shoot anymore. I don’t have the math open. I have what’s left after thousands of frames did the math for me and then got out of the way.

Based on some exploration I did this morning using AI I found three different AI research efforts that are now chasing that gap, from different angles, none of them all the way there.

A team out of Stanford and NVIDIA built something called TTT-E2E, which lets a model keep adjusting its own internal weights while it reads — not just holding the page in front of it, but being changed by the page, a little, as it goes. It runs thirty-five times faster than the brute-force method of remembering everything, because it isn’t remembering everything.

Google’s research arm published something called Nested Learning around the same time, built on the idea that a mind isn’t one system learning at one speed, but several systems nested inside each other — some updating by the minute, some by the year.

And a scrappier strand of work called self-distillation has models teaching cheaper versions of themselves, not by handing over a transcript, but by training the cheaper model to arrive on its own at whatever the well-informed version would have concluded.

None of this is what happens when I make a photo. Not yet. But it’s aimed at the same gap I live in every time I shoot before I understand what I’m shooting. The gap between having the math and having the eye.

I once asked Doug, a good friend who’s spent as many days on the street as I have, how he knew when to press the shutter. He didn’t have an answer, not really — just a shrug, and something about the moment feeling complete before he could explain why. That shrug took him years to earn. He didn’t keep the years. He kept the shrug.

And then a few years ago Doug did something I still don’t fully understand. He abandoned digital and went back to film. Not for any project, not for the look of it — he could get that in post if he wanted it. He went back to the actual mechanics: loading a roll, metering by hand, often using a tripod, etc. I needled him about it some, the way you’d needle a cigarette smoker who’d taken up a pipe instead, as if the inconvenience were the point. He told me he wanted to slow down, and that film was the only thing that reliably made him do it. Twelve frames and then you stop and reload and you can’t fix it later. The very friction he’d spent decades shooting his way out of, he went looking for again, on purpose.

I don’t know what to do with that, except to notice that he’s the same man who can give me the shrug and also the man who walked back toward the thing the shrug had replaced. Maybe that’s the part the labs haven’t gotten to yet, underneath all the vocabulary of weight updates and meta-learned initializations. Compression is the whole point, until the day it isn’t.

Note: This line of thinking started with a recent essay by Dwarkesh Patel on what he calls continual learning. It’s become a real focus of his thinking about how we get to a better future with AI.

See: https://www.dwarkesh.com/p/the-next-paradigm

Categories
AI

What the Lessor Keeps

Two airlines can fly the same airplane. Not airplanes of the same type — the same airplane, serial number and all, handed back at the end of a lease and reassigned, sometimes within weeks, to a competitor on another continent. AerCap owns more commercial aircraft than any airline on earth, and it leases them to airlines that spend their advertising budgets convincing passengers that flying them is a distinctive experience. The 737 MAX that wears Ryanair’s livery this year might wear Lion Air’s the next, repainted, recertified, its avionics untouched, its airframe indifferent to the change of ownership. The lessor does not care who is flying its asset. It cares that the asset comes back in airworthy condition and that the lease payments clear.

What the airline owns, in the sense that matters, is never the aircraft. It is the route network built up over decades of slot negotiations at constrained airports. It is the maintenance log — every inspection, every part swapped, every anomaly a mechanic in Singapore flagged in 2019 that turned out to predict a fatigue crack nobody else had seen yet. None of that travels with the airplane when the lease ends. It stays behind, compounding, in systems the airline built and the lessor never touches.

Karl Mehta, who has spent a career inside enterprise software watching this kind of asymmetry repeat itself, put a version of it plainly: a model is a brain you rent, and you and your competitor rent the same one. The formulation has the compression of something that has been tested in a few dozen meetings before it found that sentence. It is also, structurally, the airplane story. Anthropic and OpenAI and Google are AerCap. They retain residual value on enormous capital assets — clusters of GPUs depreciating on a schedule, weights trained at a cost that only a handful of balance sheets in the world can absorb — and they lease access to those assets by the token, to anyone who can pay, including, in the same afternoon, two companies trying to put each other out of business. The model does not know whose prompt it is answering. It has no loyalty file. It has, in fact, no memory at all, in the ordinary sense of the word — each call begins exactly where the last one ended for everybody, which is nowhere.

The asymmetry that airlines exploit is the one available here too, and it sits one layer up from the engine. Call it the embedding store, the vector database, the fine-tuning corpus, the retrieval index — the terminology varies by vendor, but the function is constant. It is the accumulated, indexed residue of every customer interaction a company has had, structured so that the rented brain can be handed the relevant fragment of it at the moment of each new call. A bank’s fraud model and a competing bank’s fraud model can call the identical foundation model, route through the identical API, and arrive at entirely different verdicts on the identical transaction, because one of them is retrieving against eleven years of labeled chargebacks specific to its own card portfolio and the other is retrieving against four. The intelligence rented by the hour is, for practical purposes, a commodity, priced down toward marginal cost the way jet fuel is priced — everyone pays close to the same number per unit. The memory is not a commodity. It cannot be, because it is not for sale; it is the institutional record of what has already happened to you, and no amount of capital lets a competitor buy a copy of your chargeback history any more than it lets them buy your maintenance logs.

This produces a particular kind of corporate vertigo, which Mehta’s sentence is really addressing. For three or four years the industry conversation about artificial intelligence has been a conversation about models — which lab’s was larger, which benchmark moved, which release cycle a company should anchor its roadmap to. That conversation rewards being an early and aggressive lessee. But a lessee relationship, however aggressive, does not compound into anything a competitor cannot eventually also lease. The compounding, when it happens, happens in the layer below the API call: in how cleanly a company has structured the record of its own customers, its own failures, its own edge cases, so that the rented brain, plugged in fresh every morning with no memory of yesterday, can be handed exactly the right fragment of yesterday and made to look, for a few hundred milliseconds, like it has been there all along.

A hospital chart has two kinds of entries. There is the vital-signs strip clipped to the bed rail — temperature, pulse, blood pressure, checked every four hours and replaced every four hours, because a reading from yesterday tells the night nurse nothing about the patient in front of her right now. And there is the permanent record in the file downstairs: the allergy that nearly killed him in 2019, the surgery, the medication history going back a decade, written once and never overwritten, because that record is exactly as valuable ten years from now as it is today. Nobody confuses the two charts. Nobody staples last Tuesday’s blood pressure into the permanent file. The hospital figured out, long before anyone digitized it, that memory is not one problem. It is two, and they fail in opposite directions if you run them through the same system.

Most teams building the layer Mehta is describing make exactly that mistake — they staple everything to the same chart. The shorthand for it is dumping everything into a vector database and praying, and it is worth asking why that particular error is so popular. The answer is that it feels like progress: embeddings go in, something resembling memory comes out, and the team moves on to the next sprint without confronting the harder question, which is what kind of memory it just built.

Short-term memory is the vital-signs strip — everything the model needs to finish the task in front of it and nothing it needs after. A customer-service exchange in progress, the order number already mentioned, the fact that this is the second call today, belongs here. So does the scratchpad of a multi-step agent: the search results just pulled, the file just opened, the partial answer being assembled before it commits. The test is not how important the information is but how long it stays true. A customer’s mood this minute is real and gone in twenty minutes; storing it permanently is like stapling yesterday’s temperature reading into the permanent file, undated, until the chart tells you nothing about fever and everything about clutter. Short-term memory should live in the context window itself, or a session-scoped cache, and it should be allowed to die when the session ends. The sin is not forgetting it. The sin is remembering it forever.

Long-term memory is the file downstairs, and it does not come in one shape any more than that file does. The first shape is semantic memory — facts. A customer’s account tier. The chargeback history that decides, in fractions of a second, whether this morning’s transaction clears. Facts belong in a database with a schema, not a vector store, because a fact has a right answer and a vector store gives you an approximate neighbor. Ask a vector index what tier a customer is on and it hands you the five most semantically similar sentences in the corpus — one correct, four merely correct-sounding. Ask a schema the same question and it tells you, because that is what the schema is for.

The more sophisticated shops are already building the seam between the two, rather than picking one and living with its blind spot. A knowledge graph keeps the relationships a schema is good at — this customer, that account, this chargeback, in fixed and queryable connection to one another — while still letting a retrieval layer search across it by meaning rather than by exact key. The approach has a name now, GraphRAG, and the name matters less than what it concedes: that facts and resemblance are different operations, and the honest fix is to run both and let each one answer the kind of question it’s actually suited for, not to force a single index to pretend it can do both jobs at once.

The second shape is episodic memory — what actually happened. The specific conversation last March in which the customer explained, at length, why the previous fix didn’t work. The exact sequence of an agent’s failed attempt at a task, preserved so the next attempt doesn’t repeat it. This is where the vector store finally earns its keep, because an episode isn’t an exact-match lookup, it’s a resemblance — has anything like this come up before — and a vector index, built to find the nearest thing to a fuzzy question, is the right tool for that question and almost no other. The error was never using a vector store. The error is using only a vector store, for facts as well as episodes, on the theory that one hammer with sufficient cosine similarity can stand in for the whole toolbox.

The third shape is the rarest, and the one teams forget to build at all: procedural memory, which is not a fact and not an episode but a skill — the model’s learned sense of how this company writes a refund email, escalates a complaint, formats an invoice. Style is the visible half of it. The other half is harder to see and matters more: the rails the model is forced to run on before it ever gets to choose a word. A refund above some threshold routes to a human, no exceptions, because the workflow says so, not because the model was persuaded to think so on this particular call. An agent that touches a production database does it through a reviewed function with a fixed set of permitted calls, not through whatever query it improvises in the moment. None of that lives in a prompt, and none of it lives in the model’s weights either. It lives in code — the orchestration layer, the permissioning, the state machine the agent is required to pass through — and it is procedural in the oldest sense of the word: not a memory of what to say but a memory of what is and isn’t allowed to happen, enforced whether or not the model that day feels like remembering it. It doesn’t live in a database at all. It lives in fine-tuning, in carefully maintained house-style examples, and in the surrounding scaffolding of guardrails and permitted actions, and it changes slower than the other two, the way a surgeon’s hands carry both technique and caution years after the specific patients are forgotten. A company that has built rich semantic and episodic memory but skipped this layer has a model that knows everything about its customers, writes in exactly the right voice, and is one well-crafted prompt away from doing something the company never agreed to.

The real argument here is not which database serves which layer — that part is plumbing, and plumbing changes every eighteen months. The argument is that memory has to be triaged the way the hospital triages it, with something deciding on purpose what survives the session and what doesn’t, rather than writing every token of every interaction into the same undifferentiated store and trusting retrieval to sort it out later. A vector database with no triage in front of it is not a memory system. It is a landfill with a search function, and it will retrieve the wrong eleven-month-old conversation with the same confidence it retrieves the right one, because nobody wrote the part of the system whose only job is deciding what belongs on which chart.

The lessor’s airplane, repainted, will fly for someone else next year. The route network will not. Neither will the schema that knows a customer’s tier on contact, nor the index that remembers the conversation from last March, nor the fine-tuned hand that knows, without being told twice, how this company writes a refund email. These are the things that do not come back at the end of the lease, because they were never on it.

Categories
AI AI: Large Language Models AI: Transformers Authors Podcasts Writing

The Billboard

The fog was still sitting on the hills when I put in my earbuds and headed out.

Sebastian Mallaby was talking about billboards.

Tim Ferriss had asked him the question he asks everyone: if you could put anything up there, for millions of people to see, what would it be? Mallaby has spent years inside the minds of the people who shaped modern finance — the hedge fund managers, the venture capitalists, the builders of things that changed how the world moves money. He has more material than most people accumulate in a lifetime. He could have said anything.

He said: Prepare your mind.

I kept walking. The houses were quiet in the particular way they get when school lets out for summer — no buses, no car doors, no kids at the corner. Somebody’s sprinklers were running.

The phrase comes originally from Louis Pasteur, who understood something that most people don’t: that chance is not democratic. It does not distribute itself evenly among those who wait. It finds the people who are ready. Chance favors the prepared mind. Pasteur said it, and then he proved it, and then the rest of us spent a century and a half learning it was true.

What struck me about Mallaby’s answer wasn’t the phrase itself. It was the way he said it had kept appearing in his research, surfacing in different decades and different worlds, like a message the material kept trying to send him.

He told the story of Arthur Patterson at Accel Capital. Before a new technology arrived, Accel would work through the implications — what company needs to be built, what founder fits the moment, what the right pitch looks like. So when an entrepreneur finally walked in, when the situation was live and competitive, they already knew ninety percent of what they were hearing. They could move fast because they had already moved slow.

That’s preparation as institutional practice. But Mallaby found the phrase again in a different register entirely, embedded in a single human moment that has always seemed to me like one of the hinge points of our era.

He was interviewing Ilya Sutskever, asking him why he had seen it so quickly.

In 2017, a paper called Attention Is All You Need appeared online. It described a new architecture for neural networks — the transformer — that would eventually rewrite the terms of what artificial intelligence could do. On the day the paper went up, Sutskever read it. And then he ran. He went down the corridor to find his collaborator Alex Radford and told him to stop what he was doing. Everything. Stop. We are going to build a language model on this architecture.

Not someday. Now.

Mallaby asked him how he had seen it so clearly, so fast. And Sutskever’s answer, in its essence, was the same two words: prepared mind.

He had been thinking about the problem of modeling sequential data since his PhD in Canada. For years he had been carrying a question the field hadn’t answered yet. And when the answer appeared — when the transformer showed up on a website one ordinary day — he didn’t have to reason his way toward it. He recognized it. The solution arrived and found a mind that had been waiting for it, that had already cleared space for it, that was already arranged around the shape of exactly this kind of answer.

This is what preparation actually is. Not the accumulation of facts. Not readiness in the generic sense, the vague self-improvement sense. It is the long, patient cultivation of a specific question, held close and kept alive until the answer has somewhere to land.

Mallaby chose that phrase for his billboard because it kept finding him — in the venture capital world, in the AI world, across decades and disciplines and very different kinds of genius. The prepared mind is not a personality trait. It is a practice. It is the work you do before the work arrives.

The sprinklers had clicked off by the time I turned back toward home. The fog was starting to lift off the hills. I was thinking about what I had been preparing for, whether I even knew.

Categories
AI: Diffusion Models Art and Artists

An Algorithm by Hand

He didn’t know what he was looking for the first time he walked into the Alhambra. You come in from the heat and the light hits the walls and for a moment you just stand there, your mind doing something it doesn’t have words for yet.

That was 1922. Escher was twenty-four years old, recently graduated. The Moorish artists who made these walls had been dead for six centuries. They had left no notes. No theory. Just the walls.

He came back in 1936. Some things you have to see twice.

What the artists in the Alhambra had discovered — without algebra, without proof, working in a tradition that forbade them from drawing a single living creature — was that there were exactly seventeen ways to tile an infinite plane with a repeating pattern. The Russian mathematician Fedorov would articulate this formally in 1891, centuries after the fact, the way mathematics sometimes chases art home and explains what it already knew.

Escher took the problem and made it harder. He asked: what if the edge between two tiles is also the outline of a fish? What if the sky is made of birds and the birds are made of sky? He would move a single line and the whole system would tremble. He did this for years. Revision after revision, in small notebooks, by hand.

There is a word for what he was doing. We just didn’t have it yet.

The word is algorithm.

An algorithm is a set of rules, followed in sequence, to solve a problem. We think of them as things that live in machines, in data centers drawing enough power to light a city. We think of them as fast. Escher’s algorithm was not fast.

He would begin with a grid. Hexagons, maybe, or the interlocking diamonds of a pattern he had traced from the Alhambra walls. Then he would ask the question that made everything hard: what lives here? Not what shape — what creature? What thing with a spine and a purpose and an outline that a human eye would recognize before the brain caught up?

The constraint was absolute. Every point on every edge had to satisfy two animals simultaneously. Change one line and you changed everything downstream, the way a single altered fact in a long investigation suddenly makes you reread everything you thought you knew.

He failed constantly. The notebooks are full of it. Half a lizard becoming nothing. A bird whose wing destroyed the fish below it. He would back up and try again, the way you back up on a road that has stopped being a road.

He was doing, neuron by neuron, what a diffusion model now does in milliseconds.

But here is the thing about milliseconds. They don’t leave notebooks.

Categories
AI Bicycles History

The Bicycle Shop

Part 2 of 3…

It is eleven-thirty on a Tuesday night and she is arguing with a language model about a spreadsheet.

Not arguing, exactly. That’s not the right word. She is coaxing. She is debugging. She is reading error messages that tell her almost nothing and rewriting prompts that almost work, and she has been doing this for two hours, and the spreadsheet still isn’t right, and she is going to try one more thing before she gives up and does it by hand. She is a data analyst at a mid-sized logistics company in Columbus, Ohio. She is not a researcher. She is not a founder. Nobody is writing about her. She is just a person trying to get a machine to do something useful, and the machine keeps almost doing it, and she keeps learning, in the gap between almost and done, something she couldn’t have learned any other way.

She doesn’t know what she’s learning. That’s the important part.

In 1892, two brothers opened a bicycle repair shop on West Third Street in Dayton, Ohio. The bicycle craze was at its peak — the safety bicycle, with its two equal wheels and chain drive, had just replaced the penny-farthing, that absurd high-wheeler everybody called loose change and the riders, with complete seriousness, called the ordinary. The brothers fixed flats and adjusted brakes and built custom frames and ordered parts from Coventry and kept the books and swept the floor. It was ordinary work. Nobody was writing about them either. What they were doing was accumulating, without knowing they were accumulating, a physical understanding of how machines move through space — the gyroscopic principles, the weight distribution, the thousand small calibrations that kept a rider from falling. They were learning in their hands what no university taught and no book fully contained.

Eleven years later they flew.

We tell the Wright Brothers story as a story about flight. It makes sense — flight is the thing, the miracle, the moment the world changed. But the actual story, the one that explains how Kitty Hawk was possible, is a story about a bicycle shop. It is a story about unglamorous preparatory work, about the education that hides inside the constraint, about what you learn in the gap between the machine that exists and the machine that should exist. Orville and Wilbur didn’t go to Kitty Hawk despite the bicycle shop. They went because of it. The shop was the point. They just didn’t know it yet.

We are in the bicycle shop right now.

The people building with AI today — the prompt engineers, the fine-tuners, the agent builders, the data analysts in Columbus arguing with spreadsheets at midnight — are doing work that looks, from the outside, like mere tinkering. Unglamorous. Iterative. Full of failure. The tools are awkward. The models hallucinate. The context windows run out at the wrong moment. Every solution opens three new problems. It feels like the penny-farthing: powerful enough to be useful, constrained enough to be maddening, requiring a kind of practiced vault just to get started.

But that awkwardness is the education.

Every time a prompt fails, the person writing it learns something about how the model thinks — about what it responds to, what it resists, where it gets confused, where it surprises you. Every agent that breaks in production teaches its builder something about the gap between what a model can do in a demo and what it can do under load, with real data, with users who don’t behave the way you expected. Every context window that runs out forces a decision about what actually matters, what is essential, what can be cut. These are not just technical lessons. They are epistemic ones. They are lessons about the nature of intelligence, about how meaning gets encoded and retrieved, about what it means for a machine to understand something versus to pattern-match on the surface of understanding.

The people learning these lessons right now don’t have a name for what they know. They just know it in their hands.

This is how it always works. James Starley’s craftsmen in Coventry bent and brazed bicycle frames by feel and experience, knowing things in their hands they couldn’t fully explain on paper. That embodied knowledge — the tight tolerances, the interchangeable parts, the discipline of making things that had to work — migrated into every bicycle shop that followed, crossed the Atlantic, and ended up in a shed in Ohio. The Wright Brothers didn’t invent precision manufacturing. They inherited it, absorbed it, and applied it to a problem nobody else had solved because nobody else had brought those particular hands to that particular problem.

The chain drive was the hinge. Before it, the bicycle’s design was locked — bigger wheel for more speed, higher and higher off the ground, until the machine teetered at the edge of what a human could survive. The chain drive broke the constraint. It decoupled the pedals from the wheel, let the gearing do what only size had done before, brought the rider back to earth. What had been a machine for athletes became a machine for everyone. What had been the ordinary became, almost overnight, something new.

We are waiting for the chain drive.

Not waiting passively — it is being built right now, in a hundred places at once, by people who mostly don’t know they’re building it. It might be the interface that finally makes AI genuinely accessible to people who can’t do the running vault. It might be the memory architecture that lets a model carry context the way a human carries context, not in a window but in something more like experience. It might be something nobody has named yet, something that will seem obvious afterward, the way all elegant solutions seem obvious after the fact.

What it will not be is the product of people who stayed away from the bicycle shop.

The analyst in Columbus closes her laptop at midnight. The spreadsheet is still not right. She has learned three things about how the model handles date formatting, two things about how it interprets ambiguous column headers, and one thing about her own assumptions that she didn’t know she was making. Tomorrow she will try again. She will get closer. At some point — not tomorrow, maybe not this year — she will get it right, and the thing she learned in the gap will be available to her for the next problem, and the one after that, and she will carry it forward without knowing she’s carrying it, the way craft always travels, in hands that have done the work.

She doesn’t know what she’s riding toward.

That’s the ordinary part. That’s always been the ordinary part.

Categories
AI

The Transit Authority

Today SpaceX went public. The valuation target was $1.77 trillion — already the largest IPO in history, surpassing Saudi Aramco — and the market wanted more.

I was curious about the S-1, so I read the TAM section. SpaceX claims a total addressable market of $28.5 trillion. Rockets and Starlink together account for about $2 trillion of that. The rest — $26.5 trillion — is artificial intelligence. Enterprise AI applications alone: $22.7 trillion.

IDC analyst Arnal Dayaratna said the quiet part out loud: “To be crystal clear, its positioning there right now is basically nonexistent.”

That is an honest sentence. It describes most TAM claims in most S-1 filings. The market did not care. The stock was up 25% anyway.

But the $22.7 trillion number is interesting regardless of whether SpaceX captures it. It asks a real question: how large is the enterprise AI opportunity, really? And what does capturing it actually require?

The answer has something to do with transportation.

We do not travel the same way for every trip.

Walk to the coffee shop. Take a scooter to the office. Ride share to the airport. Commute by train. Drive your own car on weekends. Fly when you need to get somewhere fast and far.

Each mode has a different cost structure, a different latency, a different quality profile. Nobody takes a plane to buy milk. Nobody walks to a meeting in another city. We allocate the mode to the trip, instinctively, without much thought. The routing decision is invisible.

AI inference is arriving at exactly this moment. Until recently, there was one mode: you called the big frontier model. GPT-5.5. Claude Fable. Gemini 3 Pro. You paid the tolls, you waited, and you got what you needed. It was like renting a plane for every trip. Expensive, but simple. There was nothing else on the road.

That is no longer true.

The walk tier is a model running on your phone or laptop — no network, no cost, no data leaving the device. Google’s Gemma 4 and Microsoft’s Phi-4 now handle classification, autocomplete, document summarization. You do not even notice you are using AI.

The bike tier is a small model running on your own hardware — a workstation, a private server. Fast, cheap, data stays on-prem. These models can handle tasks that required GPT-4-class APIs eighteen months ago.

The rideshare tier is cheap cloud inference. You are not driving, not owning, but you get there quickly and cheaply. What cost $22,500 a month in 2025 runs for $405 today. That is not a gradual erosion. That is a structural break.

The car tier is dedicated hosted compute — reserved capacity, predictable performance, always available.

Frontier models are the airplane. Dense reasoning, long-context synthesis, genuinely hard problems. You use them when you need to get somewhere fast and far. You do not use them to classify customer support tickets.

Here is the problem nobody had two years ago.

Picture the IT director at a mid-size insurance company. She deployed a frontier model API last year. Smart decision at the time — one vendor, one contract, everything works. Now she’s gotten the quarterly invoice and done the math. Roughly 80% of the queries hitting that API are things like: extract the date from this document, categorize this claim, summarize this email thread. Tasks a much cheaper model handles just as well. She has been flying everyone to a meeting across town.

She is not alone. Most organizations that built on frontier APIs in 2023 and 2024 are now discovering they over-provisioned for the average query and under-thought the distribution. The expensive mode works. That’s the trap. You don’t look for alternatives when the thing you’re doing works.

The routing layer is where this resolves. A routing layer is need that sits between the application and the model tier and asks, for each incoming query: what does this actually require? Simple queries go to the cheap tier. Hard queries escalate to frontier.

Route 90% of requests to the cheap tier, 10% to frontier. You cut costs by 86%. The quality loss on the 90% is negligible, because most production queries are not frontier-hard. Most trips, you walk.

Back to the $22.7 trillion.

The number is real in the sense that enterprise software currently costs a lot. The global market — CRM, ERP, HR systems, supply chain, all of it — runs roughly $700 billion annually. If AI agents eventually do much of the work those systems mediate, and if the value gets priced into the AI layer, you can arithmetic your way toward very large numbers.

But the routing story embeds an uncomfortable question: if inference costs are collapsing, and if smart organizations route most of their traffic to free or near-free edge compute, who actually captures the value?

The model providers need volume. But enterprise routing gives sophisticated buyers a systematic exit from frontier pricing for the bulk of their workload. You call the expensive plane only when you need to cross an ocean.

This is why the routing layer matters more than it looks. The company that becomes the transit authority — the entity that sits between all the modes and makes the dispatch decision — is structurally positioned to matter as much as any individual model provider. The transit authority does not own the planes or the trains. It knows where you are going and picks the right mode. That intelligence, at scale, is a moat.

SpaceX is not that company. IDC is right about that. But the $22.7 trillion figure, even as a promotional artifact of an S-1, is pointing at something real: the opportunity is large enough that the infrastructure for consuming AI efficiently may be as valuable as the AI itself.

The frontier model providers are the airlines. Necessary, impressive, expensive to operate, essential for the long haul. Emerging routing solutions are building the booking platforms — the systems that decide when you actually need a plane, and make sure you are not buying a first-class ticket to go ten blocks.

In transportation, the booking platforms eventually captured enormous value. Expedia, Booking.com, Google Flights. The airlines, which had all the brand and all the infrastructure, found themselves competing for placement in someone else’s interface.

That story may be ahead of us in AI. The models are the planes. Someone else may be Expedia.