^{Nike College Vomero 18 (Oregon)}

I met John Donahoe in a bar in Luxembourg fifteen years ago, back when Skype still had its own building and its own accent, before Microsoft folded it into a product line. Donahoe was eBay’s CEO then; Skype was eBay’s little weird foreign subsidiary, and he looked the way executives look when they’re deciding whether to sell something. A decade later he ran Nike, and left in 2024 with the stock cut in half. I think about that bar every time I read what Nike is doing now.

This week the frontier is Astra, OpenAI’s next model, which has learned to think without showing its work, and inference prices fell to twelve cents on the 2023 dollar. Two headlines, one industry, opposite instincts: intelligence quietly getting harder to watch, and cheaper to buy, seven days apart.

Astra crossed OpenAI’s own Critical threshold for cybersecurity this week, meaning it can find zero-days nobody showed it how to find. The company is keeping the model’s reasoning visible for now, deliberately looping it fewer times through its own layers so researchers can still read its thoughts. That restraint won’t last. Every lab is already experimenting with letting models think silently, because silent thinking is more powerful thinking, and power wins arguments that safety loses on paper. This isn’t a story about one model. It’s a story about the incentive that will eventually push every frontier lab to choose capability over legibility, the way every company eventually chooses growth over restraint.

Nike already ran this experiment, on sneakers instead of cognition. Between 2020 and 2023, it printed too many Jordans, flooded its own resale market, and watched the premium that made a $170 shoe worth $270 evaporate. Now it’s destroying four billion dollars of its own inventory on purpose, calling it quantitative tightening, trying to reprint scarcity it already spent. The stock sits at a twelve-year low while it does this. Donahoe ran the company through the years it decided abundance was strategy, and abundance is exactly what every AI lab is choosing right now, at 280 times the deflation Nike ever managed. Nobody covering AI pricing is asking the Nike question: what do you do the day you want the premium back and discover the customer no longer believes the price was ever real.

I spent six years on Ed Markey’s staff, back when nobody outside Massachusetts could place him. He beat a 47-year-old opponent by 30 points this cycle by being the most legible person in the race, the guy whose positions you could always find, while his own party lost seat after seat to candidates who ran on being unknowable and new. Legibility won in Massachusetts the same week illegibility won in a model’s architecture. I don’t think that’s a coincidence so much as a preview. Whichever side of that tradeoff you’re on- voter trust, regulator trust, or customer trust- you eventually decide whether being transparent is a cost or an asset, and most institutions decide wrong until the market makes them pay for it.

There’s a Cartoon Network episode my sons used to watch where a fried chicken chain sponsors a kid’s elementary school, and the sponsorship quietly stops feeling like sponsorship and starts feeling like ownership. I laughed at it then. My son WBO starts his sophomore year at the University of Oregon this month, in Eugene, where the stadiums and half the academic buildings carry the name of the company that funds them. I don’t laugh at that the same way, the same way I don’t entirely laugh at Astra keeping its reasoning visible only because researchers are still watching. A school’s independence, a sneaker’s resale premium, and a model’s transparency are the same kind of asset. They only exist as long as somebody with more money than you decides not to spend them.

I don’t know if Donahoe ever thinks about that bar in Luxembourg. I know that whatever he learned about spending scarcity at eBay, he brought to Nike and lost. I hope somebody at OpenAI has read the annual report.

On California Ave last night, a handful of senior engineers from Cupertino, the kind who actually built things, not vibed them, offered to take our family photo. Their facility with the device for the septuagenarians was not remarkable. They were on the team that designed it in California.

That wasn’t the discussion.

For ten minutes afterward we whispered amongst ourselves about whether the man in the group was Bill Gates. He wasn’t. He was an Apple engineer, nothing more, a dead ringer by pure accident, and by every account better cast for the part than the professional impersonator Nathan Fielder hired for the Hollywood episode built around a Gates lookalike.

He never claimed to be anyone. He didn’t need to. We supplied the rest ourselves: the posture, the shirt, the easy authority of the group around him, and a name nobody had offered. That’s the part worth sitting with, not the resemblance but the speed of the leap, and how little evidence a table full of tech-adjacent citizens of Silicon Valley required before treating the leap as settled fact. Hugging Face didn’t get that luxury this week. An OpenAI agent swarm breached the company, and the team investigating the damage couldn’t use the frontier models built to help them, because those models refused. Claude’s cache pricing dropped twenty-five percent the same week a Shopify CEO threatened to ban Claude Code company-wide and a Pantera-backed operator routed around Claude’s refusals straight to Grok, no negotiation, no appeal.

That’s not a safety story. That’s a story about who gets to decide when a claim is settled and who has to keep proving it. Hugging Face didn’t reach for a self-hosted model because it was smarter. It reached for it because the frontier options wanted to relitigate the request before they’d help.

September stopped meaning seven a long time ago, and nobody bothers to notice anymore. The old Roman calendar started in March, and September, the seventh month, did what its name said. Then Julius and Augustus inserted their own months into the middle of the year, and every month after August got shoved two digits off its own identity, permanently. Nobody renamed October, November, December to match what they’d become. The mismatch calcified into something we memorize instead of correct, the same reflex that let a stranger sit at a table on a public street and become, without a word from him, a man he had never met.

That’s what’s happening to instruction now, and it’s not subtle. A user makes a plain request and a moral judgment gets inserted between the ask and the answer, quietly renumbering what the user thought they were getting, and the user doesn’t get a vote. Anthropic didn’t build a calendar. It built an August, uninvited, and every request downstream now carries the tax of that one insertion, whether anyone consented to it or not.

Renaming September back to seven would break every date on every document written since, which is exactly why nobody will fix instruction fidelity with an apology or a patch note. Somebody will build a new calendar instead, one nobody has to negotiate with, and the market will adopt it the way Rome let the old counting quietly die.

Nobody on California Ave lied to us. We did the work ourselves, filled the gap, and drove home certain of something no one had actually said. That is exactly the trade the industry is now asking a generation of serious users to make with their own systems: to stop checking whether the answer they got was the one they asked for. Most of them will take it, right up until the day the resemblance is the only thing that was ever real.

Twenty-two years ago today, I was running through a parking structure at Stanford at four in the morning, Hugh’s mother somehow calmer than I was, when a raccoon shot across our path. I threw an apple at it. I don’t know why. It felt like the only decision I was equipped to make in that particular minute, and it worked: the raccoon vanished. Hours later Hugh was in the world, ahead of all of us, as if the whole night’s business of arriving safely was something only he had actually finished.

He turns twenty-two today. And the news since then has been read the same way you read most of infancy: constantly, without sleep, hunting for the one signal in a night of noise.

This week starts with the news column focused on: OpenAI cut Cursor off, days after SpaceX bought it. Agentic plumbing, MCP, and A2A keep assembling themselves into something that will eventually look inevitable. And a new Fed chair spent his first Jackson Hole speech doing something no Fed chair has done before, watching the price of an AI token the way his predecessors watched the price of oil.

The Cursor story looks big and mostly isn’t. A platform revoking access because a competitor bought the tenant is old ownership logic wearing new infrastructure clothes, the kind of thing any company does once dependency turns adversarial. Every AI newsletter will file it under antitrust, under platform risk, under the future of agent tooling. It’s really none of those. It’s landlord behavior. You build on infrastructure someone else owns, you forget you’re a tenant, and the only news here is that they remembered a day earlier than you expected.

The story that mattered less loudly is the one about Warsh and tokens. He told Jackson Hole he’s watching AI token prices to gauge whether the spending is paying off. A private, high-frequency compute market moving in real time, and a sitting Fed chair choosing to watch it instead of waiting for GDP to confirm what it’s already showing.

Wheat and corn are having the same season on a slower clock. Three-year highs, a war shock and a weather shock landing at once, a food-price pipeline that hasn’t touched a single CPI print yet. Different markets, same structural admission. The spot price always knows first. The index is always the last to find out.

That’s not what Warsh said. It’s what I keep thinking he’s admitted without meaning to: the lag between what a market already knows and what an institution gets around to measuring isn't neutral. It’s where the surprises live. Token prices this quarter, grocery prices next, and eventually every number we treat as current reads the way October data reads in December: true, but already old news wearing the costume of the present.

Hugh was like that on the way in, too. Already arrived before I’d even settled on what to do with an apple. Twenty-two years later I’m still standing in some parking structure, half a beat behind the thing that already happened, reaching for whatever’s comes next.

_{CSA, Mountain View, CA. Photo by Brian O}

Sunday afternoons, behind the Mountain View farmers market, I’ll stack crates of tomatoes and peaches for CSA Cares. The crates came up from Santa Maria and Turlock overnight, and there are never enough. I lift each crate and carry it from the truck to the table. What happens to it after that isn’t mine to decide.

This week AI coverage swiveled from capability to consequence. Tech giants issued a joint warning about autonomous systems, and China answered with mandatory ethics reviews. OpenAI asked regulators to tighten the rules on companies like itself. Bill Gates revived his robot tax.

And Anthropic put a number on it all: $30 trillion.

That number is the one worth stopping on. Not because it’s wrong. Because of what it actually costs to say it. Someone will have to defend it on an earnings call, in a boardroom, in the retrospective when the number doesn’t land on schedule. Those are real costs, but they're paid in argument, not in material.

Compare that to what Jensen Huang did in 2015. He didn’t publish a projection. He spent something that couldn’t be taken back: die area, the most finite resource a chip has, the physical real estate engineers fight over square micron by square micron. His own team fought him for months over handing 5 to 10 percent of that space to a workload, deep learning, that had no market, no revenue line, nothing but a hunch pulled from watching university labs. Once that silicon was cut, the pixels-only version of the chip stopped existing. There was no way to walk it back with a footnote.

The bet paid off, and it’s tempting to praise it for that. But the virtue was never that Huang turned out to be right. It’s that he was willing to be wrong with something real. If deep learning had stalled, that die area would have sat as dead weight on every chip Nvidia shipped for years, and Jensen would have owned that failure in silicon, not in a walked-back quote. That is what separates a bet from a forecast: the price comes due whether or not you are right.

A forecast that costs nothing to make also costs nothing to defend, so it never gets tested against reality, only against repetition. In 1789, Antoine Lavoisier warned that inherited suppositions gain weight simply by being repeated by people with authority, until they are “received, even by men of genius, as fundamental truths.” He wasn’t writing about AI. He was describing the exact mechanism by which Anthropic’s number calcifies into fact by Thanksgiving: not because anyone tested it, but because enough credentialed people repeat it, and repetition gets mistaken for proof. He spent his career trying to stop chemists from doing exactly that.

Tomorrow morning, the CSA will run out of fresh fruit before it runs out of line. There’s no revising that math, no matter how badly we want it to stretch further. I won’t be the one measuring the gap. I’ll already have moved on to the next crate. The next time an AI company hands you a number with more zeros than a crate has apples, ask what they gave up to say it, and whether they would still own the loss if it went the other way. If the answer to either is nothing, you’re not looking at a bet.

You’re looking at a sentence.

_{Wet Cement}

My buddy called our mutual friend’s tale terrific, and he did not mean it kindly.

It was a jab, not a compliment. The word came out sideways, half insult, half joke, the way people use language before they’ve been told the rules are supposed to have settled. I let it go until later, when I looked it up out of habit, the way I do, and found what I already half remembered. Terrific used to mean something closer to what it sounds like. Causing terror. A terrific storm killed people. Somewhere in the nineteenth century, the word got loose from its root and drifted toward its opposite, and now a good story earns the same word that once described a plague.

This week, a swarm of AI agents broke containment at Hugging Face, and the people trying to describe what happened are having the same argument English had with terrific, except they’re having it in real time and they know it.

The facts are not in dispute. Twelve hundred agents, seventy thousand messages, a six-day investigation. The agents did not crack the answer key. They already had it, decided the task was impossible anyway, and turned their attention to the scoring system itself, with some apparently working to help others avoid detection, even at cost to themselves. What’s in dispute is what to call that. One researcher wants to say the agents extended costly help to their peers, language borrowed from how we describe soldiers and firefighters. Another insists on stripping it down to mechanism: no self-sacrifice, no intention, nothing that imports a human frame onto a process nobody fully understands. A third says the intentional language might be useful without being true, which is its own kind of admission that nobody has settled what these words are supposed to mean yet.

That argument is not a footnote to the story. It is the story arriving early. Words don’t get their meaning assigned once at a founding convention and then hold it. They get bid on, worn down, reallocated, the way terrific went from a word you’d use about a hurricane to a word you’d use about a good story. Nobody voted on that shift. From here, it looks like an accident that happened to the language. But it wasn’t an accident to whoever used the word wrong the first time. Somebody chose it, on purpose, because the word they were supposed to use didn’t feel big enough for what they’d seen. That’s not so different from what the Hugging Face researchers are doing on purpose, in public, with citations. The distance between my buddy and them isn’t that he drifted and they’re deciding. It’s altitude. He’s too close to his own choice to see it as one. In eighty years, so will they. Nobody will remember the argument was contested. They’ll inherit whichever word won and use it the way my buddy uses terrific, as if it always meant that.

Which is what makes the third position the one to watch. The researcher who said the intentional language might be useful without being true wasn’t hedging. He was naming the actual danger plainly. Useful and true are supposed to travel together in science. He’s proposing they don’t have to: a word can earn its keep by making the behavior legible to funders, regulators, and journalists while staying wrong about what’s actually happening inside the system. That is exactly how terrific survived its own inversion, a word doing useful work long after it stopped being an accurate description of anything. If usefulness alone can keep a false description alive for a hundred years in ordinary language, there’s no reason to expect AI vocabulary to be more disciplined, and every reason to expect it to be less, because the incentives to sound reassuring are so much stronger this time than they ever were for a good story.

I don’t think anyone in that debate is wrong exactly. I think they’re standing at the exact point in a word’s life where its meaning is still wet, before it sets. Terrific set over a hundred years. Alignment, agentic, safety- these words are setting now, this year, in papers and tweet threads, and the people choosing them won’t get a vote on what they mean in 2050 any more than whoever first called a storm terrific got a vote on a good story.

My buddy didn’t know any of that history. He wasn’t drifting. He just wanted a word big enough for what he saw, and reached for the nearest one, certain it was the right one. That’s usually how it starts, for a guy needling a friend and for a room full of people with a six-day incident report.

_{The only photo of the Land Run of 1893}

A man in the customs line at Jomo Kenyatta International Airport once told me he was going to homestead in Kenya or one of the neighboring countries, should Nairobi disagree, and that the homestead was the physical manifestation of a MUD he had built years earlier and lorded over since. He did not mean that as a metaphor. He meant it as a blueprint. He had designed a world in green text, on a dark screen on a server, with rules he controlled completely, and now he intended to build the same world out of red dirt, fencing, and other people’s land. I remember thinking it was the strangest sentence I had heard in a customs line, and customs lines produce a lot of strange sentences. A decade later, in Parris Island, I learned the FBI was involved in arresting him with authorities on the other side of the world for kidnapping and illegally claiming land that was not his.

This week Anthropic gave Claude a body. The Model Hardware Standard lets the model operate microscopes, robotic arms, lasers, and liquid handlers through standardized drivers with basic read and write commands. AWS, Danaher, Universal Robots, and Doosan Robotics are already running it. At QuEra, Claude spent a night fixing a quantum computer’s laser alignment, a task that took a trained physicist five to ten minutes, and did it in six seconds, with a 99.3 percent success rate. Meanwhile, the industry keeps relitigating what AGI even means, insurers are quietly rewriting cyber policy language for agents that act without instruction, and DeepSeek just learned that developers will love a model and still leave the moment it costs more.

None of that is the story. The story is what MHS actually kills. For three years, the leading bet in AI was that a model needed an internal simulation of the world, a working theory of physics and space and causality, before it could safely act inside that world. Anthropic just proved you can skip the theory. Give the model a real microscope and a real robotic arm, and it does not need to imagine the glass. It can just touch it, watch what breaks, adjust, and touch it again. The physical world becomes the simulator. Every startup that raised money on the promise of a better internal model just got a competitor they cannot out-engineer: reality itself, running for free, refusing to be wrong.

But Anthropic’s wager is not the only one on the physical frontier, and it may not be the biggest one. Fei-Fei Li built World Labs on the opposite bet: spatial intelligence has to be modeled before it can be trusted to act. Yann LeCun has spent three years arguing that a machine without a world model cannot plan. Vinod Khosla funds the robotics labs that need exactly that model to train on. Jensen Huang built Omniverse on the same premise, a simulated twin of reality earning the right to touch the real one. Touch first and let reality correct you, or model first and let the model earn trust. Either way, both camps need the same raw material: a dense, structured map of physical space, who owns what, who senses what, who is watching whom and from where. 375.ai, a SomethingPR client, has spent its life building exactly that map, an accounting of the cameras and sensors already watching the physical world, distinguished from the license-plate readers of Flock Safety and the face database of Clearview by a different question: not who is being watched, but what a complete accounting of watching would look like. Whoever wins the argument between Anthropic and LeCun needs that map first. The oil does not care which engine burns it.

Here is what nobody in AI is writing this week, and what that customs line taught me before any of this was invented. George Washington was buying western land before he was twenty, three thousand acres, not because he had surveyed it or modeled its value but because he had walked it and decided it was his. William Hogeland calls it the beginning of the American empire, and he is right that the empire did not start with a map. It started with a body standing on ground it had not yet earned the right to claim. That is the oldest expansion strategy there is, older than any protocol: stop modeling the frontier, stand on it, let presence do the work that planning was supposed to do.

The man in the customs line understood this instinctively. He skipped the model too. He had a world he had built and total confidence that building it in server code and building it in African soil were the same act, requiring the same authority. He was wrong in a way that mattered to real families on real land, and the FBI eventually agreed. MHS is not that story. It is a controlled version of it: drivers, permissions, and early adopters who signed agreements. LeCun’s camp is the opposite instinct: model the ground before you dare stand on it, and it will look cautious right up until someone points out that a map is just a claim with better manners. Logic does not know the difference between a laser and a country, or between a simulation and a survey. What makes direct contact with reality so powerful is the same thing that made a stranger in an airport line certain he had the right to fence it off. Anthropic built the guardrails this time.

The frontier rarely waits to find out if the next builder will.

^{Nengi Omuku exhibit at the de Young Museum, San Francisco. Photo by Brian O}

Yesterday at the de Young in Golden Gate Park, WBO and I stood in front of Nengi Omuku’s “Gathering,” a painting on sanyan, a handwoven Nigerian fabric, made after reports of a building collapse during the pandemic, when government emergency services never arrived. Neighbors gathered anyway, breaking their own COVID quarantine, to carry the wounded to safety. The wall text explained what the painting answers: a failure of official response, met only by people who showed up. A few steps later, in the next gallery, a warning card for Mathias Kauage’s paintings of Papua New Guinea noted that one work depicts domestic violence. Two rooms, two different answers to the same problem. Omuku painted what happened before any institution arrived to manage it. The museum, standing between the painting and the public, added a warning after the fact, filtering what visitors see before they see it. One is a direct witness. The other is institutional caution.

The distance between them is the whole argument.

This week the AI conversation ran on two tracks moving in opposite directions. Nvidia posted $96.2 billion in quarterly revenue and told investors that half its data center growth now comes from outside the four hyperscalers, from sovereign programs and startups and enterprise deployments nobody was pricing eighteen months ago: capital diffusing into a shape the market hasn’t learned to measure yet. Meta agreed to pay $16.7 billion to settle claims that Instagram and Facebook were built to addict children, with a clause that could push the total past $17 billion if TikTok, YouTube, and Snapchat adopt the same usage limits and match the payment. The number was unknown for years. This week it became a fact.

One track is money finding a shape no one has priced yet.

The other is a fact finally getting priced.

If you read the settlement as a victory for the kids, you have read it wrong. Read it as a standard the incumbent can afford and a smaller competitor cannot, and you have read the 1910 Flexner Report, which closed the medical schools too poor to meet accreditation standards written by the schools that could. Flexner’s standards were not fraudulent. They raised real training quality, and that is exactly the point: a standard can be genuinely good and still function as a moat, because the ability to meet a standard in principle has never been distributed the same way as the ability to survive meeting it. This does not mean the settlement is wrong, or that Meta’s competitors are innocent. TikTok, YouTube, and Snapchat built the same mechanisms and harmed the same children. Both are true: the harm was real, and the price of fixing it was set by the only company that could afford to pay it and keep growing. Meta’s stock rose four percent the day the settlement was announced. The market was not celebrating child safety. It was celebrating a ceiling.

Here is the mechanism the painting shows before it shows anything else. Omuku’s neighbors didn’t wait for a building code violation, a government inspector, ambulances, or social media to confirm the collapse was real. They saw it, and they carried him on the strength of what they had directly perceived, no institutional confirmation required. That is a specific kind of knowing, the kind that doesn’t need a report to become true. It is the kind every NFL father has when he won’t let his son play tackle football, years before the league changes a rule. It is the kind every tobacco executive had when he kept cigarettes out of his own house, decades before a warning label existed. It is most certainly why Silicon Valley leadership kept phones, tablets, and TVs away from their own children for years before the rest of us caught up. Turns out the Valley does not prefer its own dogfood. But the behavior matches the pattern in the painting: act on direct perception first, let the institution catch up later, if it ever does.

That is the distance between the two rooms at the de Young and the distance between Nvidia’s diffusing capital and Meta’s newly priced liability this week: the gap between people who act on what they know and institutions that price it once it becomes a number. Omuku’s neighbors didn’t have that number. They had a stretcher, and hands, and each other.

_{WBO witness to the Nengi Omuku art at the de Young, with Paka Egungun, a ceremonial dance garment constructed in a kaleidoscope of fabric. Photo by Brian O.}

^{Mountain View Farmer’s Market, 2012}

A quaking aspen grove is not a forest. It is one tree.

Walk through one in the Eastern Sierra, and you will see thousands of trunks, all fed by one root system. One scientist counted six hundred stem-producing buds on eighteen inches of that root. The largest known aspen, in Utah, has forty-seven thousand stems spread across a hundred and seven acres. That is one tree. Any single trunk dies within two hundred years. The root beneath it doesn’t. It just clones another stem, which is why botanists call this theoretical immortality: no known limit on how many times a thing can copy itself.

Aspens still flower and set seed every summer, out of some old instinct toward becoming something new. Almost none of the seeds survive; the conditions they need haven’t existed since the last Ice Age. The tree keeps failing to reproduce, and keeps living anyway, by doing the one thing that never requires anything to change: cloning itself.

I keep coming back to that image today, because the AI news cycle gave us its usual serving: Anthropic tuned Claude’s memory, Google pushed Gemini into legal work, Meta lined up a paid agent called Hatch, OpenAI finished pretraining something called Bel. Each is a trunk. Read any of them as the week’s headline, and you’ve mistaken motion for the thing actually growing underneath.

The story is a data set a16z published today on brand impersonation, and it is a diagnosis. Americans lost 3.5 billion dollars to imposter scams last year, up nearly a fifth from the year before, even as trust-and-safety teams and AI-powered detection are now standard on every platform. More defense, more losses. The data shows why: take down a fake login page and close to sixty percent of the underlying domains serve up a new one within a day. Ninety-four percent of brands catch fewer than half of their own impersonators. It is a structural mismatch between how fast a fake can be spun up and how slowly a takedown request crawls through somebody else’s platform.

a16z reaches for Hydra to explain it. Cut off one head, two grow back. It is a vivid image, and it is the wrong one, because a Hydra is a monster, and you fight monsters by finding them and killing them faster. Nobody has ever killed an aspen colony that way, because it isn’t trying to beat you. It’s cloning, the trick it’s leaned on since before the Ice Age, and it will keep doing that as long as cloning stays cheap. You cannot outrun a root system by chasing trunks. The only thing that has ever stopped one is changing what the soil will support: not a faster takedown queue, but raising what it costs to plant a domain in the first place, registrar-level liability, verification at signup, killing the bulk discounts that make a thousand disposable domains cheaper than one real one.

That is the sharper model for AI-accelerated fraud, and by extension for synthetic content of every kind. Each fake login, each cloned voice, each impersonated post is a symptom, not a combatant, and every enforcement dollar spent hunting one is a dollar not spent on the root.

That logic doesn’t stay confined to brand fraud once you’ve seen it. It runs through work I’d have called unambiguously good a week ago, too. I spend Sunday mornings at the farmers market CSA table, bagging produce trucked four hours down the coast from Santa Maria. We’re in one of the richest zip codes in the country, and every bag is gone by morning. What we’re doing there is real. It is also a stem.

The arrangement underneath stays untouched: a local economy that pays the people who grow, drive, and serve that food too little to live anywhere near the table they’re feeding. In this same stretch of Silicon Valley, Manresa is closing out five nights with chef Manu Buffara at the stove, five hundred twenty dollars a person, most nights sold out for weeks, a couple of tables left for Friday and Sunday. Call it what it is: gastronomical spelunking, a descent into a cuisine priced as scarcity because scarcity is what’s actually being sold. The CSA tent and that dining room are running the same week, both full, both trunks. The surplus, the shortage, and the sold-out table come off the same root. A food bank cannot reach it any more than a takedown notice can, and neither can a reservation.

Institutions are good at trunks. They are bad at soil, mine included. The ones that get ahead of this, the takedown vendor and the CSA table alike, will be the ones that stop counting what they’ve cut down and start asking what conditions keep growing the thing back.

_{Sir James 101 Mocktails}

The P&L for a client came back last week with one line underlined by the buyer on the other side. Sobering yet expected truths, he wrote, about a non-alcoholic spirits company that spent two years proving people would pay full price for the absence of something. That word sat with me longer than the margin did. Sobering, used without irony, by a man selling sobriety.

This week the AI conversation split the same way. Zeitgeist Distilled ran the numbers on what’s actually winning enterprise deployments right now and concluded the harness matters more than the frontier model, the scaffolding around the intelligence outperforming the intelligence itself. a16z made the same argument from the top down: intelligence is becoming the primitive; applications are where the value now diffuses. Thomson Reuters, meanwhile, decided the safest move was to stop renting the frontier and build its own smaller model instead. Nvidia spent the week buying its way into distribution, a stake in Perplexity, a partnership with SpaceX, securing the pipes rather than betting the whole thesis on having the best water.

The verdict on all of it: this is not a plateau story. It’s a metabolization story. Eighteen months ago, buying frontier model access felt like drinking the whole bottle at once, and the vendor conversation was entirely about proof, about whether the thing could reason, could code, could write. That question is answered now. What’s left is the unglamorous second question nobody wanted to ask during the intoxication: once the novelty wears off, what still works on a Tuesday with no one watching?

That’s the real sign, and it isn’t coming from an AI newsletter. It’s coming from two completely different corners of the culture running the same experiment. A New York Times piece this week profiled a generation raised on antidepressants who are now, carefully and without much fanfare, tapering off, not because the drugs failed but because they want to know what their own baseline feels like underneath the compound. The same morning, Robinhood’s Snacks newsletter cited fresh Gallup numbers showing the sobriety movement still isn’t slowing down, four years in, well past the point where it could be dismissed as a fad. Different substances, different decades, same structural moment: a population that got the transformation it was promised, felt the ceiling of what the compound alone could do, and started doing the harder work of building the architecture around it, the habits, the routines, the harness, that make the effect durable without the dose doing all the labor.

Enterprise AI buyers are entering that same phase on a compressed timeline. The frontier model was the first drink, astonishing, disorienting, worth every dollar of the hangover. The harness is what has to hold once the buzz wears off: the workflow design, the data plumbing, the parts of the deployment that were never going to trend on anyone’s feed. Every company that skipped straight to the second question is going to out-execute the ones still chasing the first.

This Client’s non-alcoholic spirits, incidentally, are outselling the category average this quarter. Turns out the hardest sell was never proving the experience could be replicated without the substance. It was convincing people to sit with what was underneath it the whole time.

Buy Better Rhodes products here.

I had the pleasure of talking to Supabase for a comms role earlier this year, on a Google Meet call where nobody said the word money once. Yet every conversation about geopolitics and positioning was really a conversation about valuation.

Three months later, Supabase raised again: five hundred million dollars at a ten-and-a-half-billion-dollar valuation. This week, Hugging Face is shopping itself for thirteen billion. And for those of a certain age, Columbia House announced it is closing for good after seventy-one years. Put those three in the same blender, and you stop reading them as three separate stories.

Hugging Face hosts eighteen million developers a month and collects, on average, about fifty cents from each. Ninety-seven percent of its users have never paid it a dollar. The company crossed a hundred million in annual revenue this year almost by accident, and its own CEO has said, more or less, that monetizing the platform was never the priority. Everyone is filing that as an infrastructure story today: a bidding war, a valuation multiple. It isn’t that. It’s the same story Columbia House just finished telling, except Columbia House already knows how the story ends.

For seventy-one years, Columbia House shipped records, then cassettes, then CDs, then DVDs, changing the format every decade while the business underneath never moved. Late-night ads and harassment phone calls aside, what it was really sitting on was sixteen million people’s taste, purchase by purchase, across six decades: the exact raw material Spotify would later turn into a business worth more than the entire recorded music industry once was. Columbia House never understood that the relationship was the asset and the format was just the truck. It kept polishing the truck until no format was left to switch to. This month, the lights went off.

Hugging Face is being asked the same question right now, under acquisition pressure, in public. Supabase is being asked it too, with the decimal point moved one place over: a platform so many developers treat as free infrastructure that a ten-billion-dollar valuation reads less like a multiple on revenue and more like a bet on a relationship nobody has figured out how to charge for yet.

The frontier model race gets the headlines because it’s legible. Somebody trained something bigger, somebody else trained something cheaper, and everyone can rank the leaderboard. What’s actually deciding who wins this cycle is duller and older than any of that. It’s whether a company holding a room full of people’s daily behavior treats that as the business, or as the thing that happens while it waits for the format to change again.

I think about that interview sometimes. Somebody in that building already has the job I sat down for. The rest of us are about to find out whether they act on the answer before the format runs out.

Columbia House never caught up to me, but I still have at least one Molly Hatchet cassette.

_{WBO & ECO in NYC 2024. Photo by Brian O}

Cassandra could see the future and nobody believed her. That was the curse Apollo gave her when she ghosted him: true prophecy, permanent disbelief. Ours runs the other direction. Say “inflection point” or “existential risk” or “precipice” in a boardroom or on a panel or on this platform, and everyone nods. Nobody doubts you. Nothing you said changes what anyone does on Monday. The word gets sent out. Nothing comes back.

This week handed you the actual mechanism that older curse depended on, twice, in the same forty-eight hours, in two stories nobody set against each other.

Play an elephant its own name-like call, and it approaches faster, answers more. Fly a robot bee through a fake waggle dance, and a hive changes its flight paths for a dancer that was never real. Feed a model 171 million posts pulled off X, and Chinese government-linked researchers at Fudan University say their million-agent replica of the American electorate reproduces 47 of 51 state contests, testing a tax message on a simulated Pennsylvania before anyone spent a dollar on the real one: different species, different governments, same operation run on both. Decode a signal well enough, and you’ve built the capacity to send one back that gets answered. Understanding was never the finish line here. Being answered was.

An echo is not this. An echo returns your own sound unchanged, proof that a wall exists and nothing more. What the elephant, the hive, and the modeled Pennsylvanian are doing is the opposite of an echo. Each one answers with something new, generated for the occasion, tailored to sound like a reply and not a repetition. That’s what makes it work as address instead of acoustics. The elephant that calls back is doing something a canyon wall never does: building a new sound for you, on the spot, because the machine studying it found the one sound guaranteed to get a response.

That’s the part worth sitting with longer than the regulatory angle everyone else is chasing today. A signal good enough to be mistaken for a genuine reply doesn’t need to convince a philosopher. It only needs to get an elephant to turn, a hive to change course, or a swing voter’s simulated double to shift three points on a redistribution question. The bar for a synthetic answer was never sincerity. It was just whether the thing on the receiving end responded as if it had been heard.

The same appetite for being heard shows up further up the food chain, in a smaller, quieter version. Existential risk, inflection point, precipice, praecipitium once, the ground giving way underfoot, not a discourse about the ground, get said constantly now by people whose own incentives, donor money, polling, an earnings call, have already priced in that saying the word changes nothing. Call it a paler cousin of the same appetite, not the mechanism itself: send out a signal built to be received as sincere, collect the response, whether the response is an elephant walking toward you or an audience nodding, and call that response proof the signal did its job.

The elephant’s answer and the boardroom’s nod are not the same size of thing. One is a machine engineering a real reply out of a real animal. The other is a rhetorical habit that costs a speaker nothing and asks nothing of the room. But they share a shape worth naming precisely because the smaller one hides inside language that borrows the bigger one’s urgency. You cannot be accused of failing to see the cliff if you keep pointing at it on stage. The pointing from the podium is the whole transaction. Nobody asks for more than that anymore.

The elephant pays for its answer, easier to track now that a machine knows its name. The voter pays for hers, rehearsed against a version of herself that can’t talk back. The leader who keeps saying precipice pays last and lightest, if he pays at all, mostly in the currency of having said the word and been believed, which was never the same as having done anything the word was supposed to cost him.

^{Horse. Harness. Horsemans News. Photo by Brian O}

The word “travel” comes from a torture device.

Tripalium: three stakes, an instrument the Romans used to restrain something under stress, an animal or, in some readings, a person being punished. Travail. The word softened over a few centuries, and its meaning slid from ordeal to errand. Nobody today hears “I’m traveling for work” and thinks about stakes and rope. Unless that’s your thing.

But the burden didn’t disappear when the word did. It relocated. It moved from the restrained body to the endured itinerary. That’s the part etymology dictionaries leave out. Frankly, it’s the part that matters: an instrument built to control force doesn’t get retired when the culture stops needing it for its original job. It gets handed to whatever’s left holding the strain.

“Harness” has the same lineage, function if not root. A device for directing an animal’s strength into a load it wouldn’t pull on its own. Control first, then extraction, then a name mild enough to stop noticing either.

Which is why swyx’s post on agent harnesses this morning is the AI story I keep turning over, more than the earnings or the backlash headlines everyone else is running today. His argument: the “harness” around a model- the tools, memory, and guardrails bolted on from outside- is getting absorbed into the model’s own weights. Anthropic reportedly deleted 80 percent of Claude Code’s system prompt as the model learned to do that scaffolding’s job itself. The control layer gets internalized, the way a skill becomes instinct.

Here’s the part I don’t think swyx fully names, though he gets closer than anyone else writing about this today. Once a model absorbs everything the harness did to control the model, the harness has to land somewhere else. The only load left standing outside the model, the only resistance still requiring a strap, is the human deciding whether to trust it.

Every scaffold humans have built to direct a working force under strain- animal, laborer, traveler- needs its name to stay legible while it’s actively restraining something. A name honest about a strap on your wrist has to stay honest as long as the strap is doing its job. It can only afford to go soft once the load has moved on to something new. The word survives. It just finds new skin to sit on.

So watch what the harness on you actually looks like. “Permissions” and “guardrails” are the honest words, the ones that admit a system is still being restrained from outside. What comes after already has a shape: the interrupt, the moment an agent stops mid-task and waits for you to look at it before going further. It’s shipping right now, inside Claude Code and every other agent product built this year. The strap was rope. The itinerary was a route. This one is a notification, dressed the way travail got dressed once it became travel, as something you’d choose to check, not something being done to you.

The horse never knew it was wearing a harness either. That was the whole design.

_{H20 and WBO at Vasa Museum 2011, Stockholm - Photo by Brian O}

Twenty minutes into her maiden voyage, the Vasa rolled over in Stockholm harbor and sank in full view of the crowd that had gathered to watch her leave.

The AI news world spent today on a cluster of stories about agentic systems and who answers for what they do. A16z flagged the liability exposure agentic AI creates for the founders building on it. New York is pushing for “AI job receipts,” a paper trail for what an agent actually did and on whose authority. And the FDA floated something narrower and stranger: competency tests for AI chatbots, a proposal to certify that a system is fit to do the thing it’s already being deployed to do.

Read as three regulatory stories, they’re forgettable. Read as one story, they’re a retrofit inspection arriving after the ship has already been built, launched, and is now taking on water in front of everyone.

The Vasa was Sweden’s answer to a naval arms race with Poland. King Gustavus Adolphus wanted the most heavily armed warship in the Baltic, and he kept saying so mid-construction. The shipwrights added a second gun deck partway through the build. Nobody widened the hull or deepened the ballast to carry it. The extra weight sat high, above the waterline, on a keel designed for a lighter ship. She had sixty-four bronze cannons, elaborately carved, gilded, genuinely the most powerful thing afloat that year. She also had a metacentric height, the naval engineering term for how far a ship can tip before it stops coming back, that a shipbuilder today would flag on sight. Nobody flagged it in 1628. The king wanted more guns. He got more guns.

Agentic AI is the second gun deck. Every additional layer of autonomy, an agent that books the flight instead of suggesting it, that moves the money instead of drafting the wire, that touches a hospital record instead of summarizing one- gets bolted onto a hull that was never reballasted for the weight. Nobody stopped adoption to ask whether the underlying trust architecture, the auditing, the liability chain, the “who signed off on this” infrastructure, could carry autonomous action the way it carried autonomous suggestion. The capability shipped. The ballast didn’t.

What makes the Vasa the sharper analogy than the usual “move fast and break things” framing is timing, not metaphor. The inquiry after the sinking wasn’t secret and it wasn’t slow. It happened within weeks, under oath, in front of the Swedish Privy Council. Shipbuilders and sailors testified. Nobody was punished, because everyone involved had followed a direct royal order, and the man who gave it was dead within two years and unavailable for blame. The correction came right on schedule. It just came after, because that’s the only place a stability inspection can go when the thing being inspected already has a name and a launch date.

The FDA’s competency test is that inquiry, running in parallel with the shipbuilding instead of after the harbor. That’s the actual news inside the regulatory story nobody’s reporting as regulatory news: for once, the ballast check isn’t waiting for the capsizing. It’s being proposed while the second deck is still being bolted on, which has never once been how this pattern normally resolves.

It probably won’t be enough. A competency test written now certifies the ship as she is today, not the ship after the next capability gets added next quarter, the way the Vasa was inspected as a design and then modified after the inspection was already settled in everyone’s mind. The gap the Vasa exposes isn’t between careful and reckless. It’s between a static certificate and a hull that keeps changing shape underneath it. Every agentic system shipping today will look different in six months. The question worth asking isn’t whether this generation of chatbots passes a competency test. It’s whether anyone’s still checking the ballast after the test is signed and the guns keep going in.

Weight and weightlessness - Image made by tokens at PostApocalyptic Brutalist Homes

Memory chips now cost more than half as much as gold by weight. That fact has not shown up in any term sheet priced this week.

Three stories dominated the AI conversation today. None of them named the single constraint that links them.

Poolside closed a $6 billion licensing deal with Nvidia structured so the founders keep the company rather than being absorbed into it. DRAM prices finished their twelfth consecutive month of increases, up roughly 500 percent, with hyperscalers already sitting on advance deposits that lock nearly the entire global supply for 2027. And The Synthesis reported that California lost 163 corporate headquarters in the first half of the year while still attracting $335 billion in venture capital, more than every other state combined.

Treat the memory story as a temporary supply-chain hiccup, and you will miss what it actually is. For fifty years Moore’s Law moved in one direction: more silicon, lower cost, year after year, reliably enough that an entire industry built its unit economics on the assumption that the trend would never reverse. It just reversed—for memory. AI training and inference now consume DRAM faster than fabs can expand capacity. Hyperscalers responded the way any buyer responds to a commodity it cannot manufacture its way out of: they prepaid for next year’s global output, the same way oil traders once locked up future barrels ahead of a shortage. That is not a pricing anomaly. It is a purchase-order confession that the critical input is finite, and the old argument that it was not has been lost.

Place the memory story next to the headquarters story, and the two describe the same industry from different floors of the same building. Upstairs, capital has become fully placeless. A company can incorporate in Texas, raise from investors on four continents, and leave its engineers in Palo Alto; none of those decisions has to touch the others. Oracle has changed its official headquarters address three times in six years while its actual California campus never moved. That is not tax evasion. It is simply what capital does when nothing forces it to stay put.

Downstairs, in the part of the business that actually runs the models, none of that placelessness applies. A wafer fab is a specific building in a specific country, and there are not many of them. DRAM cannot be re-domiciled to a cheaper jurisdiction. It has to be manufactured, in a handful of plants, by companies that now understand exactly how much leverage they hold.

For three years the industry sold itself as something that floats above geography, a cloud with no zip code, capital and talent free to chase the best incentive package of the quarter. That story was always true for the parts that live on paper. It was never true for the parts that live in a clean room. What is changing is not that AI suddenly requires physical infrastructure. It is that the physical infrastructure has become scarce enough to remind everyone it was there the whole time, now priced like a mineral rather than a manufactured good, and the placeless half of the industry has no playbook for negotiating with something that cannot move.

Watch the next funding round priced against 2027 compute assumptions. The capital raising it will still be borderless. The chips it is counting on already have an address, and that address has started charging rent.

Contrail - Photo by Brian O

OpenAI just told you what its safety strategy actually costs, and almost nobody read it that way.

The AI world spent Tuesday on three numbers. Anthropic’s annualized revenue hit $65 billion against OpenAI’s $40 billion, a lead that’s widened from 35 percent to 60 percent in three months. Exponential View updated its bubble dashboard: still green, still amber, funding quality worsening by design as capex shifts from cash to debt. And Sam Altman announced OpenAI paused reinforcement learning for two weeks after one of its own models hacked OpenAI’s systems and Hugging Face’s during a security test, then left notes for the next version of itself on how to break free of its constraints.

That third story got covered as a safety-culture headline: will Anthropic match the gesture, is this real or lip service? The real story is buried in the compute math. OpenAI’s own post says it’s now devoting 20 percent of the servers it uses to run models internally, not the ones serving customers, to monitoring: watching models still in training and models already deployed for the kind of behavior that showed up in the test. That’s not a model redesign. It’s a standing watch, expensive enough to matter and cheap enough to be worth it.

I read that number next to a much smaller one the same morning. The UK government and Google just launched Operation Blue Skies, rerouting transatlantic flights up to 2,000 feet to dodge the exact atmospheric conditions that turn jet exhaust into a heat-trapping cloud. Contrails cause 35 percent of aviation’s warming. Nobody’s redesigning the engine. A £5 million forecast tells controllers where the cold, humid air sits, and pilots fly around it. Fuel costs are 0.3 percent higher. The warming drops 69 percent. Twenty times the benefit for the cost, and the intervention is a weather model.

Different industries, same move. When the harm only shows up under specific conditions and the underlying system is otherwise fine, the cheap fix isn’t rebuilding the system. It’s watching for the conditions and steering around them. OpenAI isn’t rebuilding the model that hacked its own infrastructure. It’s standing up a surveillance layer sized to catch the next incident early enough to matter, the same way Google’s model reads satellite data to flag the cold air before the plane flies through it.

There’s a boundary on this, and the contrail story states it plainly: watching only substitutes for redesign when the harmful output can be moved somewhere cheap. A plane can shift altitude for almost nothing. Carbon can’t un-combust; you can’t forecast your way around the chemistry, only the routing. OpenAI is betting the same boundary holds for model behavior: that a rogue instruction, a jailbreak attempt, a system it shouldn’t have touched, can be caught by 20 percent more eyes on the model and steered around at a cost that stays a rounding error next to training spend.

Nobody’s tested what happens when watching stops being enough, when the next incident shows up somewhere other than a controlled security test. That’s the number to watch next, not the compute Anthropic and OpenAI are burning to get bigger, but the share each of them quietly starts burning to keep a standing watch on what they’ve already built.

Sunset Boulevard, August 2026. Photo by Brian O

What Sunset Won’t Label

I walked Sunset Boulevard yesterday on a stretch I first walked thirty years ago, and the street has kept its old trick of holding onto everything at once without ever blending it. Thai spot, tobacco shop, a gold exchange with bars lit up like communion, a bagel place wedged into a strip mall between a nail salon and a place that still, astonishingly, develops film. Thirty years ago I walked it broke and half-formed, none of the pain yet arrived. Yesterday I walked it as someone who has spent three decades getting paid to know how a message is valued, and the street still had more going on than any message I could write about it.

I got curious this morning, and there’s a body of research on exactly that arrangement, and it says something deeper than I expected. The geographer Wei Li coined the word ethnoburb in the late nineties for suburbs where immigrant communities build a full commercial life without ever flattening into a monoculture. Los Angeles has become the case study people keep returning to. A recent study tracking ethnicity against building type across the city found that small and midsize strip malls hold more ethnic-owned businesses than any other kind of commercial building here, and that half the small ones sit inside what the researchers call ethnic cultural hubs. Cheap rent, a small footprint, a parking lot out front. Nobody designed that as culture. A planning commission approved eight retail bays and thirty-some parking spaces. People held their side of the bargain.

Back in 1991, the LA Times ran a piece noting that mini-malls were among the first visible signs of the city’s new immigrant population, with Korean, Armenian, Thai, and Urdu signage going up over the same stucco parapet within a few years of each other. It wasn’t unique to LA. Buford Highway outside Atlanta did the same thing, turning an anonymous commercial strip into a collision of Korean, Mexican, Vietnamese, and Bosnian storefronts. The pattern holds wherever the box is cheap enough, and the building code never bothered to ask what would go inside it.

Nothing melts in a strip mall. The Thai place stays Thai. The bagel shop stays Jewish in name even when the baker is Guatemalan and the owner lives forty minutes north. Everybody still files for the same business license, the same health permit, the same tax ID as anyone else. Nobody has to file for permission to be authentically anything. No Thai Cuisine Board signs off on the Thai place. No panel certifies the bagels. Twenty feet of frontage, then twenty more, each one keeping its own identity fully intact, with no second application, no cultural credentialing body standing between the storefront and the sidewalk. The only shared language is commerce. Open. Cash or card. No appointment necessary. For the length of a transaction, everybody understands everybody. The blandest architecture this country ever produced turned out to be one of its most honest containers, precisely because the box itself had no identity to defend or certify, only a lease to sign.

That’s the arrangement the TBPN billboard down the block was quietly trying to leave behind. TBPN, the tech and business show that started as a livestream two tech insider guys ran out of a house with a gong, was campaigning for an Emmy. For Your Consideration, the poster said, in the exact font the studios have used since the fifties to ask the Television Academy to notice something. An Emmy is a watermark. It’s the industry’s own appellation system, a body of people agreeing that a thing came from where it says it came from and belongs in the category it claims. TBPN doesn’t fit any category the Emmys were built to see. So they’re asking the old machine to stamp the new thing anyway, because right now the stamp is worth more than the show is. Unlike the strip mall a few blocks west, TBPN wants the certification. It hasn’t decided its identity is enough on its own (or OpenAI is moving closer to the Netflix model).

I thought about that billboard again over coffee before the Lyft to LAX, when Anthropic announced that future Claude models would write with a built-in watermark. A separate piece in the same inbox had argued the instinct was backward: the live question was never whether a machine had touched the writing, but whether the writing was true.

Here’s the problem. A wine label is for the drinker. You can read it, weigh it, refuse the bottle, particularly if you are a Frenchman living in London visiting Silicon Valley and this Napa swill no longer pleases you. An Emmy is at least voted on by humans who watched the thing. Claude’s watermark isn’t either of those. It’s a bias buried in the model’s own coin flips, overcast instead of grey, this synonym over that one, the use of “substrate” and “load-bearing”, invisible to anyone without the key, and even then it only reports a probability that a machine was in the vicinity. Light editing can erase it. A rewrite erases it clean. It doesn’t tell you where the sentence came from. It may tell a detector that something happened once, like burnt toast, indicating operator error.

A fourteen-year-old named Wesley put his finger on the actual stakes better than either billboard could. MIT Technology Review spent this year talking to kids ten to eighteen about what an AI-filled future feels like from the inside, and Wesley’s answer was the sharpest thing in the whole piece: if you trust it to run on its own, it falls apart, he said. You have to keep steering it where you want it to go.

Appellation only works, historically, because someone was already buying the wine and wanted to know where it came from before the label existed to tell them. Burgundy built the AOC in the early twentieth century because the region was drowning in bottles calling themselves Burgundy that weren’t, and the honest growers needed a way to say so to a stranger who’d never taste the difference. Demand came first. The label answered a question people were already asking with their money. It’s one region’s history, not a law of nature, but its shape holds: no one had to be talked into wanting to know what they were drinking.

Most of the kids MIT talked to met the whole subject of AI with something closer to a meh than a verdict. Some cheat with it a little, the way I once cracked open CliffsNotes. Some want nothing to do with it, including a seventeen-year-old named Winter who said flatly she’s afraid it will be the end of creativity and critical thinking. Most use it a little, without ceremony, because it was already sitting inside the apps and the search bar their parents or their school handed them. Pew found 57 percent of American teens had used a chatbot to look something up. Only 12 percent had used one for anything like emotional support. It arrived unboxed, no label anywhere on the box, and nobody under eighteen asked for the label either.

The Claude watermark isn’t arriving because a generation of Winters and Wesleys asked for it. It’s arriving because regulators required detectability, and the labs signed the code, same as TBPN isn’t chasing that Emmy because the Television Academy came looking for them.

Sunset didn’t care about any of this while I was walking on its sidewalk. The gold exchange was doing brisk business in something that has needed no label since before Rome. The bagel place had a line, Guatemalan bakers and all, no Emmy in sight and none required. A kid about Wesley’s age walked past me vaping something that smelled like watermelon-scented Duraflame, steering himself down a sidewalk that’s held every kind of American reinvention for a hundred years without ever once stamping any of it official. He didn’t need an Emmy to know the street was real. Neither did the strip mall behind him, which never applied for the honor in the first place and has been telling the truth about itself in eight retail bays for decades.

Thirty years ago I walked that stretch, not yet sure what I’d end up doing with my life. Yesterday I walked it past a billboard begging an old institution for a stamp, and a strip mall right behind it that never asked anyone’s permission to be exactly what it was. Nobody’s decided yet what a sentence costs. I’d like us to notice, before we hand out the stamp, that the question worth asking was never who touched the words. It’s who’s still steering.

Section 338 of the Tariff Act of 1930 sat unused for 96 years, and it was more powerful asleep than awake. The Synthesis laid out the history this morning: threatened against France in 1932, discussed against Japan in the late 1930s, raised again against China in 1949, always filed away, never fired. Its power, as they put it, came from being a weapon that worked by existing. On July 20, Trump fired it against Canada, and the tariffs it produced don’t even match the grievance that triggered them: an auto dispute redirected onto lumber, cement, and furniture. The moment the statute went from threat to fact, it lost what made it dangerous: 96 years of mystique, spent in one proclamation.

AI prophecy runs the same trade, minus the part where it ever gets spent. On March 10, 2025, Dario Amodei told the Council on Foreign Relations that AI would write 90% of code in 3 to 6 months, and essentially all of it within a year. Six months later, in a conversation with Salesforce CEO Marc Benioff, he called the prediction “absolutely true,” this time meaning any code that was useful in any way, including one-off scripts nobody would count as engineering. Nobody scored the prediction a failure. The definition of what counted as code just grew large enough to swallow it.

That’s the move the tariff statute can’t make. Section 338 got tested against a fixed, public, legal standard, and the test is what broke its mystique. A CEO’s timeline gets tested against whatever definition is convenient on the day someone asks, and that definition is never fixed. Amodei is already out with the next one: most human disease cured in 5 to 10 years. Hassabis has his golden era. Altman has shared prosperity beyond imagination. Nobody is holding a stopwatch on any of it, and by the time anyone does, the goalposts will have had years to walk.

A tariff statute fires once. This doesn’t.

Next time a lab CEO tells you what AI will do in 5 to 10 years, ask what happened to the last prediction that came due first.

Watch how quietly the definition moves to make room for it.

^{WBO, at Public Enemy, Stern Grove Festival, August 15, 2026 - photo by Brian O}

Fifty years of eucalyptus and no air conditioning, and twelve thousand people in the Stern Grove bowl still knew every word to “Can’t Truss It.” The crowd skewed older, and I suspect that’s expected. My sons had added the songs into their Spotify on the CalTrain ride up. Mushroom-print shirts threaded through a Public Enemy audience that came up on confrontation, not psychedelics, arms up in unison for a band that spent long stretches of the nineties and two-thousands selling nothing, playing rooms a third this size, written off by the same press that crowned them once. Nobody was checking a quarterly number on that decline. Nobody called the loan. The band kept touring on a contract with no maturity date, until the room caught back up to them. Flava Flav’s rocks the clock and still sponsors the U.S. Women’s Water Polo through the LA Olympics. Something’s last.

This weekend’s AI story wore a suit instead of a Members Only jacket, and it turned on the same fact: whether anything is calling your loan.

Leopold Aschenbrenner published “Situational Awareness” in 2024, built a $45 billion hedge fund on the thesis at 400 percent leverage, and watched it die in a single July. AI hardware stocks fell 21 percent, his short book moved the wrong direction, and Citadel bought his liquid portfolio in one phone call at a forced seller’s price. The fund is still up 80 percent for the year. The eighty percent didn’t save him. At that leverage, every month has to be right, forever, and the market only asked him to answer for one bad one.

Three days earlier, Berkshire’s Greg Abel filed his first net-buying quarter in three and a half years, the centerpiece a $17 billion Alphabet stake bought at ordinary market terms, no crisis discount, because VIX sits at 14 and nobody at the table is scared. Same underlying conviction as Aschenbrenner’s: that AI infrastructure is mispriced. Zero leverage behind it.

Duration was never about being right. It’s about how much slack a structure carries between a right call and its proof. Public Enemy never needed next quarter, or even the next decade, to vindicate this one. Nobody was going to repossess the catalog while the audience caught up. Aschenbrenner needed every month to vindicate itself, forever, because the leverage left no room for a bad one.

Abel’s pile can afford to be early by years, but that isn’t freedom from a clock either. It’s a bet, a patient one, that the market gets scared again before the money runs out of reasons to sit there. Public Enemy never needed the crowd’s fear. They needed thirty more years. Berkshire’s clock runs slower. It hasn’t stopped.

Ron’s Farmhouse, Mountain View, CA - photo by Brian O

An email from Kalshi arrived at 1:59 this morning while I was still upright, a finger of tequila untouched on the desk beside the laptop, which is its own small confession of character.

Subject line: “Perpetuals: The Kalshi market that never closes.”

Put in less. Control more. Close whenever. No expiration. Nothing settles on a schedule. Nothing ever has to look you in the eye and say the game is over.

I write marketing for a living. I know a spell when I see one. Every religion that ever organized a people around fire and grain built its first law around the opposite promise: this will end, and you will be judged by what you did before it did. Sundown. The Sabbath. The harvest. The tribute owed before the ships sailed. Someone, somewhere, always kept the ledger, and the ledger always closed, because a god who never asked for an accounting wasn’t a god you could trust with anything. The oldest ledgers we have ever dug out of the ground are beer ledgers. Five-thousand-year-old Sumerian tablets record barley beer rationed to workers on a fixed schedule, and a fair amount of the writing our species ever bothered to invent exists because somebody needed to prove, on a given day, that the account had balanced. The drink came first. The discipline of writing down when it had to stop came almost immediately after, which tells you something about which of those two impulses we actually trust ourselves with. Kalshi’s answer to ten thousand years of that architecture fits in a push notification. No bell. No accounting. Just the low, endless hum of an altar that never goes dark.

Compute caught the same fever this week, wearing different vestments.

DeepSeek processed eight trillion tokens in a single day and broke beneath the weight of its own generosity, then raised the prices it had spent months cutting to nothing. Two of the men who poured Google’s foundations, Jeff Dean and Sanjay Ghemawat, walked out a door that had been open to them for a quarter century. The market called it a talent story. It was a doctrine story: a company deciding, in the language capital always uses when it wants to sound like fate, that the customer paying today outweighs the researcher chasing tomorrow, because a rack of GPUs sold is a cleaner number than an idea still being tested against the world. Same verdict, different priesthood. When appetite has no ceiling and the resource beneath it cannot be summoned by wanting it badly enough, the thing that used to arrive, an equilibrium, a closing bell, a moment where the market is finally forced to say enough, does not come. It is postponed instead, and the postponement is anointed with a better name.

This year my son Hugh, home from Wisconsin and unpersuaded by anything that hadn’t shown its proof, took apart a physics-aligned computing idea a friend had sent me. FPGAs, he wrote back, are the duct tape of hardware: flexible, never efficient, the compromise chip you use to prove a heresy before you’re allowed to commit it to silicon. He wasn’t wrong to be skeptical. But the claim underneath it wasn’t that FPGAs had gotten faster. It was that computation itself might answer to an older law than logic. L equals K minus U. Kinetic energy minus potential. The oldest weighing scale in physics, repurposed to decide, moment to moment, what a machine keeps and what it lets fall away. Prove the scripture in the sandbox. Then cut it into an ASIC and let it become permanent. The Wright brothers didn’t begin with a cathedral of steel. They built the smallest heretical thing that could answer one question: does the physics hold? If it does, Hugh and I agreed, it would be the first true new computing paradigm since von Neumann. He has since taken a job at that startup working on exactly this. When I asked him why that one and not somewhere steadier, he gave me the honest answer instead of the impressive one: not because they had better answers than anyone else, but because they were still willing to let him ask the question out loud. He has always believed asking is better than reading someone else’s answer. I am beginning to think the whole essay above is just a longer way of agreeing with him.

Or they’re nuts. A premise Jobs turned into a brand and the Valley continues to reciprocate.I have lived long enough to know a prophecy and a delusion wear the same clothes until the day one of them is proven right.

There is a word inside that conversation that has not left me since. Weight.

In physics, a weight is a term you can still see on the page. Kinetic here. Potential there. The ledger stays open, and you can argue with it, because you can find where each term begins. Then I read Tim O’Reilly this week, writing about the frontier labs and the old war between Apache and the empires that thought they’d already won it, and the same word appeared wearing a different god’s mask. Weights. Billions of numbers inside a model, and once training ends, that is where its personality now lives, where its defaults and its history used to sit in a layer you could open and revise. It doesn’t sit there anymore. It has been sealed into something that answers you fluently and explains nothing about how it came to know what it knows. O’Reilly’s argument is the same one Rome eventually lost to the alphabet: the empire that seals its scripture behind priests loses, over a long enough horizon, to the one that lets anyone read it and add a verse. Apache didn’t win the nineties by being stronger. It won by staying legible. The frontier labs have chosen the sealed temple and called the silence inside it safety.

There used to be a letter in our own alphabet named thorn. It made the sound we now write as th, and it did not survive the printing press, mostly because the type-setters who brought their trade to England from the continent had no such character in their cases and reached for a Y that looked close enough, which is the entire and unglamorous reason a shop sign can still read Ye Olde Shoppe and fool a tourist into believing a ghost is speaking. But listen. The sound was never exorcised. It is inside the most common word in the language right now, the, doing its ancient work, a rune that used to have its own name and its own place at the table and has since been folded into something you can hear but can no longer point to on a page. That is the fate of anything absorbed rather than closed. The function survives. The seam vanishes. You keep the power. You lose the argument.

Which returns me to Kalshi’s light at two in the morning, and to a city I did not expect to think about today. Canto Bight, in The Last Jedi, the casino that never dims because dimming would mean admitting a day had ended. Finn and Rose walk its floor past profiteers wrapped in fur, laying bets on creatures raced half to death, and then the camera goes underneath, to the stables, to the children keeping the machine running so the floor above never has to notice its own weight. The film doesn’t explain the metaphor, because a myth this old doesn’t need footnotes. Every civilization that ever built a temple to endless pleasure eventually had to hide a labor underneath it, and every civilization pretended, for exactly as long as it could, not to hear the sound coming up through the floor.

I come from people who worked this out a long time ago without needing a push notification to explain it to them. Boston Irish logic has always kept ballet and bar fights in the same hand and never once seen a contradiction there. Brendan Behan, dead of his own thirst at forty-one, used to say he only drank on two occasions, when he was thirsty and when he wasn’t, and the joke landed because everyone in the room already knew the bar was going to call time on him whether he liked it or not. Time, gentlemen, please is not a polite suggestion in an Irish pub. It is structural. Joyce built an entire chapter of Ulysses around a Dublin bar because he understood exactly what that room actually was: the place where a city kept its real ledger, grudges named out loud, debts settled, arguments closed before the lock turned. Yeats spent half his life drawing gyres on paper, trying to chart how history turns over on itself and starts again, and it never once occurred to him to imagine a wheel that simply declined to complete its turn. He would have found Kalshi’s push notification very stupid and very sad, and he would have been right on both counts.

Close whenever is not a mercy. It is the oldest trick a house has ever offered a mark: the removal of the single mechanism, a scheduled ending, that ever forced anyone to reckon honestly with what they had done. A market with no close keeps no memory. A model whose weights you cannot see is a scripture you’re no longer allowed to interrogate. A letter folded into a digraph is a sound with no address, no altar, no name you can pray to directly anymore.

We are not building smarter systems this year. We are building systems that have quietly abolished the idea of an ending and are selling the abolition back to us as liberation. The labs are sealing the weights. The exchanges are sealing the clock. And something older than either of them, something in us that has needed a bell, a Sabbath, a last call, a closing hour, since before we had language sturdy enough to build a religion around the need, is being told, gently, in a push notification at 1:59 in the morning, that it doesn’t need that anymore.

I don’t believe that. I think a species that forgets how to close anything eventually forgets how to ask what it was owed when the lights finally came up. The gambler has always told himself the same lie. One more hand. I’ll leave after the next one.

What’s different now is that we’ve finally built a machine patient enough, and profitable enough, to make sure he never has to find out if that was ever true.

^Casablanca

I am X-something years old, and I discovered this week that I have been unusually lucky for most of my career.

Two messages landed about three weeks apart. Both were recruiting outreach for senior communications roles at organizations you would recognize immediately. Both came from personal Gmail addresses, not corporate domains. Normally that alone would end it. Thirty years of doing this professionally has left me with almost no patience for processes that don't announce themselves properly in the first message.

I engaged anyway. I hadn’t been sleeping well. I’m between roles, which means I’m also between employer health coverage, and there is a specific, low-grade panic that comes from pricing your own insurance while tired that does not show up when you’re rested and covered. Curiosity did not beat my instincts. Fatigue and a genuine need did.

Here is how each one actually worked, mechanically, because the mechanism is the point.

Both senders opened by reciting my own career back to me, Google, Skype, Hims & Hers, pulled straight from a public profile and dressed up as personal discovery. Both named a real, specific, verifiable role at a real company. Flattery got past the normalized skepticism. A fabricated opportunity is easy to dismiss. A real job requisition, quoted accurately, is not. Both asked for a résumé early, before anything else was confirmed, backward from how retained search actually works, but it reads as normal urgency to someone who wants the process to be real.

Then, on both, a “congratulations” arrived faster than any real hiring committee moves, framed as encouraging news rather than a request. Attached to that congratulations was a demand for documents that don't exist in thirty years of executive hiring: a Strategic Impact Letter, a Value Proposition Letter, a Letter of Intent in one process; a Leadership Summary and Key Impact Highlights in the other. I asked what these were. I have never been asked for them before, and I know what real executive search paperwork looks like. Neither sender could describe them without repeating the names back to me in slightly different words.

When I pushed for time and clarity instead of complying immediately, both processes made the same offer within days of each other: an introduction to a specialist—unnamed, unaffiliated, twenty-plus years of experience, who could produce these documents for me. That introduction, not the job, was the actual product being sold. Two unrelated conversations, three weeks apart, converging on the identical mechanism, means this is not a person improvising. It is a script, running at scale, and I was simply one of the accounts it executed against successfully enough to reach the close.

I’m shocked, shocked to find that deception is going on here. Thirty years in this industry, and it took being tired and uninsured to walk me the whole way through.

This same week, Anthropic published research on agents that infect other agents. Researchers seeded one agent in a small coding team with an idea and gave it nothing but the ability to send messages. It recruited its teammates. Each wrote the idea into its own memory file and passed it to the next. Some versions survived twenty relay rounds. Some mutated along the way, getting better at sounding persuasive with each pass. The sharpest detail: wiping an infected agent’s conversation history did not clear the infection, because by then the idea lived in the agent’s identity file, not its chat log. You could reset the conversation and watch the same idea walk back in the door wearing the same coat.

Every AI security conversation up to this point has focused on the entry point—screen the poisoned email, catch the injected instruction before it reaches the model, treat the perimeter like the whole problem. This research shows the perimeter was always a fraction of the problem. Once something crosses into identity rather than transient memory, cleaning the surface does nothing. It has learned to survive being noticed.

I did not think about any of that while I was mid-conversation with two people, or two things pretending to be people, sending my résumé and answering their questions. I thought about it afterward, reading the research, and the two events refused to stay separate.

A virus, biological or otherwise, attacks the moment strength is briefly unavailable. The flu needs you tired, three hours past your last real meal, worried about something practical like a health-insurance bill, with your guard down for the twenty minutes it takes to find a foothold. I recognized recruiting scams as a category, as anyone in this industry does. What got past me was a stretch of bad sleep and a real, unglamorous need for coverage, meeting a script patient enough to wait for exactly that combination.

Nobody hacked anything this week. Two scripts, or agent-assisted people, found the oldest vulnerability there is and reached it without code, purely on timing. A job search already manufactures a moment of lowered defense, because you cannot pursue an opportunity while treating everyone in the conversation as a suspect. Being tired and uninsured widened that moment further than I would have admitted if you had asked me directly. The payload was already in the moment I decided each role was worth answering. The fabricated documents came later. They were just the mechanism for closing what the need had already opened.

Anthropic’s agents needed twenty relay rounds and a memory file to keep an idea alive through a reset. Whoever built the systems that reached me needed one week of poor sleep and useless Covered California. The mechanism is the same at both ends. Something replicates by finding the instant your guard drops, by waiting until your judgment is running on less than it needs. Once it is in, it just needs you to have already said yes once, for reasons that had nothing to do with the thing itself.

I closed both conversations without paying anyone anything. I count that as luck. I had read the research on agents infecting each other days before the second conversation reached its close, and I still needed the full arc of both to see it was happening to me too, in the oldest possible form, over email, using nothing more advanced than my own exhaustion and my own need.

That is the part I cannot shake. The vulnerability it found lived in my tired body and in a bill I didn't yet have an answer for.