Early in my career, yes decades ago, I worked on the development of oximetry catheters. Designing an experiment and documenting the protocol on a very early PC took about a day. Setting it up and running it in a controlled oxygenated fluid bath took an hour. Analyzing the data and writing up the results took another day. Then I often spent another day looking around for my boss to discuss the results.

Two days of thinking. One hour of equipment. The thinking and analysis was the constraint.

Call it 16 to 1, human hours against machine hours. Equipment time was cheap and mostly idle. Attention was what you rationed, so you'd pick your next experiment carefully. Picking badly cost two days or more, depending on where my boss was.

Which is why a paper that hit arXiv last week made me realize the circumstances have flipped. Google DeepMind's Co-Scientist, running on Gemini 3 Deep Think, interfaced with a lab-built chemical vapor deposition reactor and tailored synthesis experimental recipes to that specific hardware in minutes.

The scale is clearer in Kosmos, the AI scientist from Edison Scientific. Give it a dataset and an objective and it'll run up to 12 hours across 200 agent rollouts. Collaborators found that the number of valuable findings scaled linearly with the number of cycles. That's a production line of scientific testing!

When failure gets cheap

Back in my catheter development days we'd agonize over experiment selection, because a wrong choice burned two days of the scarce resource. Thinking and analysis was expensive, so the design had to be right, and any branch that looked weaker than the leading candidate didn't get run at all. We killed those in our heads before they cost anything, never finding out if there was a discovery in what we discarded.

When designing costs nothing, that calculus dissolves. You run the weak branch too, because all you're spending is the hour in the bath, and you'd rather have twenty cheap failures that narrow the space than one careful success. Failure turns into information about where not to look, and that has value. Periodic Labs, founded by Liam Fedus from OpenAI and Ekin Dogus Cubuk, who led materials AI at DeepMind, is built around robotic powder synthesis labs that generate proprietary data with the negative results deliberately retained.

The market's started pricing this value of data, even failure data. When Spirit Airlines shut down in May and went into liquidation against roughly $8.1 billion in debt, Google won a bankruptcy auction in August for a slice of the airline's enterprise data: revenue management systems, aircraft operations, audit and fraud records, and pricing on billions of Spirit and competitor flights. $10 million for a defunct dataset on a failed airline.

Two other lots from the same liquidation put the number in scale. JetBlue paid $58 million for 22 LaGuardia slots. A hedge fund paid $93 million for the Florida headquarters. Decades of operational history went for a tenth of what the building fetched.

The bath still takes an hour

The equipment has become the constraint. A-Lab, the Berkeley robotic synthesis platform, ran 58 targets in 17 days. Not 58,000. The physical furnace still takes as long as a furnace takes. Kosmos hits 200 rollouts in half a day because its experiment is a computation.

Periodic raised a $300 million seed in September 2025 to build physical labs rather than to train a better model. When the rate-limiting step is the crucible, you buy crucibles.

So the AI-enabled acceleration of scientific research is real, and it'll arrive unevenly, sorted by how physical your experiment is. Anything that has to touch matter or wait on biology gets the old pace with a much faster brain bolted to the front.

I've written before about where that leads, when the results keep working and the explanations stop arriving, so I'll leave it there. The narrow version is enough. Cheap search hands you a survivor, and survival tells you which branch worked without telling you where it stops working.

If I could take today's tools back to that lab, the two days of design and analysis would compress into a coffee break, and the search for my boss would vanish entirely. The bath would still take an hour. But nearly everything I actually learned came out of the gap between experiments, sitting with why the last one behaved the way it did before I set up the next.

AI has flipped the constraint from the thinking to the equipment.

The rain in Copenhagen came down hard one afternoon but we still walked a couple miles to the iconic Little Mermaid statue. Phones out. A woman rode past with a small child in a cargo bike, unhurried, dressed for an ordinary Sunday. Then a few dozen joggers. Then a couple hundred more bicycles.

We're Californians. Rain is an event.

Nobody around us adjusted anything at all. The itinerary held up well over three weeks, which I'll get to, but the thing I keep turning over is how much I carried into these four countries that had nothing to do with them. I'd ranked these cities in my head before we ever left, and I hadn't earned the ranking.

August was the right month. Warm without the punishing heat that had settled over southern Europe, and water still warm enough to swim in. What we hadn't accounted for was that everyone else knew it too. Finns, Estonians, Swedes, and Danes get roughly four months of dependable warmth, and by mid-August they're spending it hard. Parks full at 9pm. Swimming platforms crowded. The cities we walked through were being used by the people who live in them, which is a different experience from walking through one that's mostly hosting visitors. We had two days of rain across the whole trip, most of it in Denmark.

We had originally thought of including Norway on this trip, but after considerably research we realized that country really deserved a trip on it's own in order to properly experience Oslo, Bergen and the fjords, and the Lofoten Islands in the far north.

Helsinki, and the road signs to St. Petersburg

We lean toward the luxury end on hotels, but still the first thing we do in a new city is get on the cheap Hop On Hop Off bus. We don't hop off. We ride the full loop, which costs almost nothing and hands us a working map of how the place is put together before we start walking. In Helsinki it also gave us something to accomplish on day one while the jet lag did its work. Stockholm runs a water loop alongside the land loop and they're both worth it. Copenhagen has three, including a colorful tour that swings past Christiania and a loop built around the Carlsberg brewery. Tallinn is the exception, since the old town is compact and largely pedestrian.

Helsinki is the most mellow of the four capitals, which makes it the right place to start. The only stretch that gets hectic is around the Esplanade, and even that's mild. We stayed at the Waldorf Astoria rather than the Bank hotel, and after walking the Esplanade a few times I was glad we did.

We took the half day out to Porvoo, which is smaller than I'd pictured. It's scenic and worth the morning. What stayed with me was on the road out: signs to St. Petersburg. The distance is short. Not long ago you could drive it on a whim, and now there's a wide buffer at the border that you aren't crossing casually, and no, you can't drive up for a look. Standing in a quiet Finnish town reading a highway sign for a Russian city puts the last few years in a register that reading about them doesn't.

Löyly, the harbor sauna out on Hernesaari, is worth the trip across town. Book it well in advance, since you reserve a two hour block and the slots go. And take the dip in the Baltic. You heat up, get in the cold water, then repeat, and it's the second cycle when the whole thing stops being an endurance test and starts making sense.

The day trip to Tallinn was one of our favorites. The Tallink ferry runs from West Terminal 2, and business class costs a few dollars more than standard. Take it. Priority boarding and disembarkation puts you ahead of a very large crowd, which matters most on the Tallinn end where the taxi queue forms fast. We walked the old town tour in reverse from how it was laid out, because the taxi drivers prefer a particular drop point and there's no sense fighting it. The old town is compact and in better repair than I'd expected.

Stockholm, which won

Stockholm was our favorite city on the trip, and it wasn't close. We stayed at the Lydmar, a small hotel next door to the Grand, right on the water. Guests get access to the Grand's fitness room, which is nicer than you'd expect from a hotel that size. More useful: the ferries leave from directly out front.

We walked a tremendous amount here. Fourteen islands connected by bridges and short boat rides means the city rewards walking in a way a compact single-center city doesn't, and the connections are easy enough that you stop thinking about them.

Kulturfestivalen, Stockholm's annual culture festival, ran the 12th through the 16th and covered most of our stay. Free stages go up around the central squares and along the waterfront, mostly music, and the city had an energy that an ordinary week wouldn't have supplied. The cost is more people in the central areas, which is a fair trade.

The Vasa Museum deserves its reputation, and the scale of the hull really doesn't come through in photographs. Fotografiska was excellent. And the ABBA Museum, which I want to be direct about: go, even if you're convinced you don't care about ABBA. It's well constructed and experiential in a way most music museums never manage, and I say that as someone who walked in unenthusiastic.

Our other Stockholm highlight was the half day out to Vaxholm. An hour on the ferry through the inner archipelago, which doubles as the local commute, so you're watching people travel to and from home rather than riding a tour boat. We did the 4.3 mile loop hike around the island, through neighborhoods of summer houses set on granite. It reminded me of northern Michigan more than anywhere else we went.

Copenhagen, and the walking

We skipped the train from Stockholm. Five hours, a transfer in Malmö at rush hour, and Swedish countryside we'd already seen a fair amount of. We flew instead, an hour gate to gate, and I'd make the same call again.

Copenhagen is a good walking city and the neighborhoods are where it pays off. Our first afternoon we walked to Nyhavn and back through the pedestrian shopping streets, which is the obvious move and it's fine for what it is. The better day was the long walk out to Christiania and then on to Reffen. Christiania has been cleaned up considerably from what it was, and it's interesting to walk through and see what survived the transition. Reffen, the container food market on the harbor, had more lunch options than we could work through. Vesterbro, the old red light district, is now a funky neighborhood of local shops and bars and worth an afternoon of its own.

The Louisiana Museum is beautiful, the building winding down through old trees to a terrace above the Øresund with Sweden on the far shore. The collection ran more contemporary than our taste, which says more about us than about the museum. Go for the building and the sculpture garden.

We stayed at the Marriott, which is comfortable and well located, and it's the one booking I'd change. After the Waldorf Astoria and the Lydmar, a chain business hotel felt like a missed opportunity. The room was fine. The city deserved better from us.

The ranking I got wrong

I expected Denmark to be the tidiest of the four, and I held that with some confidence before we left. I still can't reconstruct where it came from. Some accumulated impression of Danish design doing work it was never entitled to do, probably.

Stockholm was the one that stood out. Cleanest, best groomed, with landscaping maintained to a standard that reads as deliberate policy rather than good luck. Tallinn and Helsinki followed close behind. Copenhagen carries more graffiti and more overgrown planting than I'd anticipated, though the proportion matters here: every one of these cities is cleaner and better kept than almost anywhere else I've been, and the spread among the four is narrow against that baseline. The direction of my error is what stayed with me, since the error itself was small.

The other correction was bicycles. I knew Copenhagen was a cycling city. I hadn't understood what that means at street level. In Copenhagen you watch the crossing signal for the bicycles, because they're the faster and more numerous hazard and they arrive from angles a driver never would. The upside is that ordinary residential streets have wide separated lanes for walking and riding, which makes long distances on foot easy and safe. Stockholm and Copenhagen had the most bikes by a distance, and both cities are better to walk in for it.

We'd sequence the trip the same way again, and it'd work about as well in reverse. Helsinki eases you in, Tallinn is a day of compressed history, Stockholm is the peak, and Copenhagen is where you stop walking.

What I still don't have is a good answer for the joggers. They were out in the heaviest rain of our three weeks, and they were out in the same numbers on the clear mornings, and there was no evident calculation in either case. I've been home a week and I'm still building my day around a forecast. What are you carrying into a place that has nothing to do with the place?

I spent the month of June in Malta, Greece, and Switzerland. Three countries, three very different economies, one thing in common: almost nobody talked about AI. Not at dinner, not in the news, not in the kind of overheard café conversation that tells you what a place is chewing on. Back home, even in our very small beach town I can't get through a couple days without AI coming up on cheesy Nextdoor, of all places, retired neighbors arguing about chatbots between posts about lost cats.

Call it a cadence gap. Cadence is the variable that decides who wins the next decade of technology, and it's the reason a conversation that saturates one country barely registers in another.

Cadence is the rate of learning

Cadence is how often you run an experiment: how many times per year you attempt something and find out what happens. It takes technical people, capital to fund attempts, and demand that justifies making them. Call them engineers or researchers depending on the field; in space it's mostly engineers running hardware to failure, in AI mostly researchers running training experiments.

The three are not co-equal, and I had this wrong when I started thinking it through. Demand is the crank. Capital follows demand, at a scale that has to be seen to be believed: the four largest US hyperscalers are guiding roughly $725 billion in capital spending this year, up from about $410 billion in 2025 and $226 billion in 2024. Nobody commits that on a technical hunch. They commit it because Microsoft has an Azure backlog it can't fill and Google Cloud is carrying over $460 billion in booked demand. Demand justifies capital, capital buys attempts, attempts produce learning, and learning produces the next product that generates more demand.

Lean practitioners will recognize the improvement kata scaled up to a national economy. The kata's premise is that progress comes from a high rate of small, structured experiments toward a target condition rather than from occasional large leaps, and that the organization cycling fastest through plan-do-check-adjust learns fastest. What's true on a shop floor turns out to be true for a country.

Space makes this visible because launches are countable. Starship Flight 13 flew yesterday, deploying 20 real Starlink satellites and putting the ship into the Indian Ocean intact and floating for the first time ever. Getting there took 13 full-stack test flights, several of which ended in spectacular failure. Ariane 6, Europe's flagship launcher and a capable piece of engineering, has flown seven times since its 2024 debut and hasn't lost one. Ariane 6 is a competent rocket. The difference is that SpaceX bought 13 flights' worth of knowledge about what actually breaks, and Europe bought seven flights' worth of confirmation that its design works.

Volume is what builds the wall

Then there's the second curve, the one I think matters more over a decade. SpaceX's own counters currently read 681 completed missions, 641 landings, and 604 booster reflights. The Block 5 booster was designed for 10 flights before major refurbishment. Boosters now routinely fly more than 25 times, with turnarounds as short as three weeks. Nobody engineered that improvement up front. It was discovered one component at a time, by flying enough vehicles enough times to learn what wears out on the 22nd flight rather than the 10th.

There's a well-established shape to this. Wright's Law, from a 1936 study of aircraft manufacturing, holds that cost falls by a roughly constant percentage every time cumulative production doubles. The improvement arrives per doubling, which means the return on any single additional unit diminishes as the total grows. Flight number 8 teaches you an enormous amount. Flight number 682 teaches you almost nothing by comparison.

That sounds like it should undercut the whole argument. It doesn't, because the compounding lives in how fast you traverse doublings in calendar time. SpaceX went from 8 flights a year to 165 while Ariane 6 accumulated seven flights total. Both improve by roughly the same percentage per doubling. One of them has collected several doublings in the time the other collected one, and each doubling starts from the improved position the last one produced.

Cadence creates the velocity of new knowledge. Volume creates the minutiae, and the minutiae are what become the competitive barrier.

A competitor can copy the Falcon 9 design. The blueprints were never the hard part, and anything written down leaks eventually, authorized or not. The layer that never gets written down is what stays put: which supplier's valve batch runs slightly out of tolerance, what the inspection routine catches on a 22nd reflight, how a launch crew reads an anomaly at T-40 seconds. That knowledge can only be acquired by flying, and flying requires demand a competitor hasn't got yet. Europe has the engineers. What it lacks is the flight rate, and without the flight rate, that talent gets far fewer chances per year to find out what it doesn't know.

Both curves run through AI, drug discovery, and advanced manufacturing the same way. Fail faster, fail in greater numbers, and only then do you get to succeed faster.

This is not the pace of ten years ago

What makes this urgent is how recently it changed. Falcon flew 8 times in 2016. It flew 91 times in 2023, 132 in 2024, and 165 in 2025, and passed 70 by the middle of this year. That's roughly a twentyfold increase in annual cadence inside a decade, in an industry that spent the prior fifty years treating a dozen launches a year as a serious operational tempo.

China is now doing the same thing, fast. On July 10 a Long March 10B first stage was caught in a net strung across a recovery ship in the South China Sea, four hooks snagging tensioned cables, no landing legs involved. It was China's first orbital-class booster recovery and it happened on the rocket's maiden flight, with plans to refly that same stage before the year ends. One catch against 604 reflights is not parity, and anyone claiming otherwise is selling something. It is entry into the loop, and the demand underneath it is explicit: China's Guowang and Qianfan constellations must place 10% of their planned networks in orbit by the end of 2026 to hold their international spectrum rights, and they're only a few hundred satellites in. That deadline is why reusability went from aspiration to requirement. Scale and national vision manufacture demand that private markets would take a decade to generate.

The AI picture rhymes. In one week this month Moonshot released Kimi K3 at 2.8 trillion parameters, the largest open-weight model yet built, and Alibaba previewed Qwen 3.8 Max. Independent evaluation puts K3 just behind the top American models and ahead of several of them on coding tasks. DeepSeek sells frontier-adjacent inference at 87 cents per million tokens against roughly $50 for the American flagship, which is a demand strategy as much as a pricing one: buy usage, generate the volume, feed the loop. When US export controls briefly cut foreign access to Anthropic's top models, a Chinese lab shipped a competitor days later with a pointed note that frontier intelligence shouldn't be subject to withdrawal at somebody's discretion.

Two accelerating competitors push each other harder than one running alone. Every Chinese release tightens American timelines, and every American constraint hands Chinese labs a market opening. Anyone still reasoning from a 2015 model, when the leaders and a capable second tier were separated by a few years of engineering effort, is working from obsolete assumptions. A well-run national program can still catch up given time and focus. What's in doubt is whether the doubling rate on the other side leaves enough calendar time to try.

The loop feeds itself

None of this happens by being granted a high cadence. It's earned, and then it self-reinforces. Arianespace's own CEO gave the honest answer in late 2025, asked whether the company could push past its planned launch rate. The constraint isn't engineering capacity, he said. It's customer demand. Ariane 6 could fly more often. Nobody's ordering enough launches to justify it.

That single comment is the whole mechanism in miniature. SpaceX earned a low cost per launch through years of iteration, which made it the obvious choice for anyone who needed something in orbit, which generated the order volume that funds the next round of iteration. A launch provider stuck at seven flights is locked out of the loop that would let it become faster, because the capital that would buy more cadence is waiting on the demand that only comes from already having it.

This is also why a popular recent ranking, countries by density of AI researchers per capita, measures the wrong thing. Switzerland topped that list. Switzerland also shipped a real AI model yesterday, the same day Starship flew: Apertus 1.5, from ETH Zurich, EPFL, and the national supercomputing centre at Lugano, built by an initiative with more than 800 researchers and 20 million GPU hours a year on a public supercomputer, released fully open source. The engineering is serious.

The capital is where it comes apart, and the scale shows up in the model itself. Apertus is a 70-billion-parameter model. Kimi K3, shipped a week earlier out of Beijing, is 2.8 trillion. Parameter counts flatter the Chinese models somewhat, since those designs activate only a fraction of their weights on any given token, but the gap in training investment behind the two numbers is real and it's roughly an order of magnitude. Twenty million GPU hours is a serious academic commitment sitting against $725 billion of demand-driven private capex, and that gap exists because the Swiss model has no demand engine underneath it. The Canton of Ticino uses a fine-tuned version to translate official documents. A medical variant is in testing at a Lausanne hospital. Both are respectable, and neither generates the daily volume that surfaces failure modes nobody anticipated, or the revenue that would justify the next order of magnitude in compute. Switzerland's 9 million people can concentrate excellent engineers and still never assemble the demand that turns a good model into a compounding one.

The tradeoff nobody voted on directly

Anything that caps demand or slows capital formation caps cadence. Regulatory environments that restrain how capital concentrates do it. So does social policy that directs capital toward broad security rather than concentrated risk-taking. Europe has chosen both, deliberately and at the ballot box, and the outcomes those choices buy are ones most people would endorse. The part that never appeared on any ballot is the second-order consequence: capital committed to the floor under citizens isn't buying cadence, and that amounts to an effective decision not to compete at the frontier. Reasonable people can make that trade. It's worth knowing you've made it.

A closing window

Nobody outside the loop is locked out permanently. But the window is closing faster than most policymakers have priced in, and Europe is only the nearest case. Most of the world is further out with fewer options.

What I keep circling back to is dependency. Anthropic's recent suspension of foreign access to its most capable models, overnight, on a US government export control directive, is the preview: whoever controls the loop controls who gets to use what comes out of it, and everyone outside discovers their position the morning it changes. Two countries are pulling away in space, in AI, and in everything built on top of both. Everyone else buys access to the output on terms they don't set.

Which brings me back to the cafés. When the loop isn't spinning where you live, the whole subject stays theoretical, right up until the morning somebody else's export control decides what tools you're allowed to open. My retired neighbors arguing on Nextdoor have more skin in this game than they know, and so do the people who weren't discussing it over dinner in Valletta.

So: what does a capable country that isn't the US or China actually do? Build a narrow domestic capability in one field and accept dependency everywhere else? Pool resources across borders and hope the politics hold long enough to matter? I can't find a third answer that isn't just a slower version of the second.

My grammar can be atrocious, and I know exactly why. I grew up in a Spanish-speaking country and still speak it close to natively, which left a permanent groove in my English prose. Spanish tolerates long sentences with clauses stacked in an order that lands backwards for an English reader, and decades later I still build sentences that way whenever I write fast. Linguists call it syntactic transfer. My early writing here is a monument to it, and I would rather you not go looking.

So I run drafts through Claude for a cleanup pass. It untangles the inverted clauses and tells me when I have buried the subject four commas deep. Then I make every change by hand in my own editor because I want to know what changed.

That habit acquired a new complication on August 2. Article 50 of the EU AI Act took effect, requiring providers of generative AI to mark synthetic audio, image, video, and text in a machine-readable format. Anthropic responded by marking Claude's output with an embedded text watermark plus signed metadata on generated files. The marking applies to every model launched on or after that date, worldwide, whether or not you are anywhere near Europe.

Ben Thompson wrote a sharp piece on this at Stratechery, and I agree with nearly all of it while landing somewhere slightly different on why it bothers me.

What the mark actually says

Anthropic's own documentation is admirably candid. A detected mark signals that content may have been processed by Claude. It establishes nothing about who wrote the underlying ideas, since people routinely use these models to proofread or reformat work that is entirely their own. Absence of a mark establishes nothing either, because heavy editing, paraphrasing, format conversion, or simple brevity will all wipe the signal out.

The result has poor sensitivity and poor specificity. In manufacturing terms we have deployed an inspection gage whose measurement error swamps the tolerance we actually care about, and we are about to read its output as pass/fail. Anyone who has developed a gage R&R study knows how that story ends.

Who actually pays

Marking happens at the model-serving layer, which means anyone running open weights on their own hardware produces unmarked text by default. Models never placed on the EU market sit outside enforcement reach entirely. None of this touches the thousands of fine-tunes floating around Hugging Face, or the person spinning up a voice clone for a fraud call.

The mark is also forgeable. Researchers have demonstrated that collecting enough watermarked output lets an attacker infer which tokens carry the signal, which enables both stripping the mark and stamping it onto text a model never wrote. A public detector, the very thing that would make any of this useful to a reader, doubles as the oracle that makes forgery easier.

There is a design trade underneath: a watermark tough enough to survive paraphrasing is by construction insensitive to editing, including edits that reverse the meaning of the sentence it is riding in. So we have a control system with the dominant source of variation sitting outside the loop.

The part that should worry writers

In 2023, Stanford researchers ran 7 commercial AI detectors against essays from native and non-native English writers. More than half of the TOEFL essays written by non-native speakers came back flagged as AI-generated. Essays from American 8th graders scored nearly perfect. The likely mechanism is that those detectors were scoring how predictable the word choices were, and someone writing in a second language works from a narrower vocabulary.

I read that study with more personal interest than most people would. The fix that worked is the part I keep chewing on: running those human-written essays back through ChatGPT with instructions to use more sophisticated language got them reclassified… as human.

Watermarking works differently from perplexity-based detection, and I want to be careful about conflating the two technologies. What transfers is the institutional behavior. Schools and employers adopted an unreliable signal and used it punitively anyway, and the cost landed hardest on the people least positioned to argue back. Nothing in the new marking regime dampens that appetite. It supplies a fresh signal, thinner than it looks, to an audience already primed to treat any hit as a verdict.

Grammar is not authorship

Some writers I respect regard these models as plagiarism engines, given how they were trained. That argument is real and deserves its own post. It has almost nothing to do with what happens when I ask a model to tell me that four paragraphs in a row start with the same word.

We settled this question for other tools decades ago. Nobody appends a disclosure when a spreadsheet computes an IRR, and no one has ever argued that spell-check makes Microsoft the author. The engineer who runs a finite element analysis still owns the design, including the parts the software talked him out of.

Underneath the whole regime sits an assumption that human-origin content is trustworthy and machine-origin content is suspect. A person can lie, or be confidently sloppy on deadline. A model with a careful prompt and honest sourcing can produce something more measured than either of us manages on a bad morning. I have written before about how much of what unsettles us in these models turns out to be our own reflection. Provenance and accuracy are separate questions, and building infrastructure that answers the first while everyone reads it as an answer to the second seems likely to make the information environment worse.

Meanwhile I will keep applying my corrections by hand, which now looks less like stubbornness and more like a workaround. That is the tell. When a measurement system starts inducing rework in a process that was working fine, the measurement is the thing to fix.

So a question for the writers reading this. If a tool untangled a sentence you had built the way another language taught you to build sentences, and taking its advice carried some risk of a label you could never remove, would you still take the advice?

A magnitude 7.1 earthquake hit Kumamoto Prefecture a couple weeks ago. Within a few hours came the usual flood: drone footage of a partially collapsed shopping mall, a paper mill chimney down in Yatsushiro, tsunami advisories issued and lifted, and a rising tide of confident explanation from people who have never set foot on Kyushu.

One item caught my eye, because I've been to Japan several times and the vending machines always amazed me. The variety especially, and the placement: mountain trails, alleys with no store within a mile (hot canned coffee in winter is one of civilization's better ideas). The claim making the rounds was that Japanese machines open up and go free during an emergency, and that most carry a QR code you can scan to tell responders exactly which ruined street you're standing on.

Half of that is wrong and the other half is half right, about the hit rate I'd expect from a comment thread 6 hours after a disaster. What's actually there is better than the story.

What's actually there

The QR code doesn't exist, at least not for that. The codes on machines report a jam to the operator so they can mail back your 150 yen. What does exist is older, dumber, cheaper, and much harder to break. Since 2005 the vending machine industry has worked with police and fire departments to put address stickers on machines nationwide, so that someone calling 110 or 119 can say where they are. Japanese street addressing is notoriously hard to read off a wall, and official emergency guidance now lists the sticker alongside utility pole numbers as a fallback locator.

No power, no network, no phone battery, no app. The internet upgraded a sticker into a smart system, which is backwards, because the sticker is the design that still works when everything else is down.

The free vending is real and rarer than advertised. Coca-Cola started installing disaster machines in 2003, controlled remotely over the network and mostly carrying storage batteries so they run through an outage. After the 2011 Tohoku quake the company switched about 400 of them to free and gave away more than 88,000 drinks. Many others open with a physical key switch, which means somebody has to show up and turn it. Machines that decide on their own are new and scarce: two of them in Ako City, installed in 2023 near evacuation shelters, holding about 300 drinks and 150 emergency food items, tripping open at shindo 5 or higher.

The kanban card on the sidewalk

In 2008 I toured Toyota's Kyushu complex in Miyawaka, in one of the prefectures under an emergency warning today. The plant was running 1,150 vehicles a day across two shifts, mixed model, and there was not a computer visible anywhere on the assembly floor. Parts bins moved on paper kanban cards. I flew 5,000 miles to watch cardboard rectangles tell a factory what to build, then walked out past the machines I'd been buying coffee from all week without recognizing the same idea in a different housing.

An emergency stockpile is dead stock. You buy it against a forecast, warehouse it, pay to hold it, and find out on the day you need it that the elastic has rotted. A disaster vending machine holds the same goods as live inventory on continuous rotation, funded by ordinary commerce and restocked by a route driver whose income depends on it being full and working. As the Mainichi noted about the Ako machines, since they sell normally the rest of the time, nothing expires and nothing goes to waste. The carrying cost goes negative. The stockpile turns a profit while it waits.

Taiichi Ohno got the pull system from watching American supermarkets restock shelves according to what customers took off them. Japan ran the loop back the other way and turned the supermarket into civil defense.

What dead inventory looks like

After Katrina, FEMA was hammered for failing to get supplies to people, so in 2006 it bought enormous quantities of food and ice and positioned them across the Southeast. The season came in mild. Millions of prepared meals spoiled in trailers parked in the Gulf Coast summer, where FEMA workers clocked interior temperatures above 120 degrees, and the agency discarded 279 truckloads of food. It also held 84.9 million pounds of leftover Katrina ice for that same season, then paid roughly $67 million to store and destroy it.

FEMA's deputy director explained the logic with more candor than most: "we didn't want to run any chance of running out." That is a forecast-push system talking. Nothing in normal operations touched the inventory, so nothing in normal operations reported that it had gone bad.

I should mark the limit before I oversell this the way that comment thread did. You will never find a ventilator in a vending machine. Rotation works only where commercial demand and emergency demand overlap. That overlap is unglamorous and larger than people assume: water, food, ice, batteries, fuel. It also happens to be the first 72 hours, which is the window Tokyo's refuge parks are explicitly sized for.

Past that you're into the long tail, where there's no commerce to borrow and you pay for dead stock deliberately. Japan solved the first three days by noticing that a solution was already standing on every corner, funded by thirsty commuters. The long tail is still a problem, there and here.

Hikarigaoka Park hides cooking stoves inside 36 of its benches and emergency toilets under 52 of its manhole lids. MIT's Miho Mazereeuw, whose Design Before Disaster came out this spring, makes the point that everyday use is what builds the mental map: people know where to go in a crisis because they've already had a good time going there.

Several commenters on that thread asserted that yakuza groups own a chunk of Japan's vending machines. I looked and found nothing. Documented organized crime fronts run to construction, waste disposal, entertainment, and labor dispatch, so I'd treat it as folklore until somebody produces a source.

What is thoroughly documented is that the gangs show up. In 1995 the Yamaguchi-gumi ran meals out of its Kobe offices and used boats and helicopters to get around blocked roads. In 2011 the Inagawa-kai moved 25 four-ton trucks of diapers, ramen, batteries, and flashlights within a day, and over 100 tons in all. They turned up in Kumamoto in 2016 too. A retired Hyogo prefectural detective put the motive plainly: it's tradition for the Yamaguchi-gumi to expect a return on its charity. Stationary bandits protect the territory they collect from. Resilience running on self-interest again, in a form nobody would have designed on purpose.

Japan's soft drink machine count fell to 1.95 million in 2025, under 2 million for the first time in 30 years and about 20 percent off the 2014 peak. Ito En has gone from 165,000 machines to roughly 75,000 in a decade, and Sapporo left the business entirely. The cause is dull: price inflation and driver shortages.

Nobody voted to shrink the emergency water supply. That's the quiet failure mode of dual-use infrastructure. When the paying use dies, the second job goes with it, and it never shows up on anyone's preparedness budget because it was never on anyone's preparedness budget to begin with.

Which leaves me with two questions. What are we already decommissioning by accident while congratulating ourselves on the efficiency? And what's standing on our streets right now that could be doing a second job?

I haven't set an alarm in decades. My body always wakes me a little after 4:30am, I get ready, read for a bit, and I'm at the gym when the doors unlock at 5. I love that hour. The parking lot is empty, the world hasn't started making noise yet, and I've gotten myself into the best shape of my life on the strength of it.

Which is why I can't explain why occasionally walking in at 5:02 instead of 5:00 produces a small knot of stress in my chest.

Maybe it's that I'll miss the first two minutes of Tom Keene's pithy commentary on Bloomberg Surveillance, which I listen to while warming up, even though it all gets repeated over the next half hour? Or maybe that the squat rack will be occupied, even though it never is at that hour? My regular gym bros wander in at random times over the next fifteen minutes and appear entirely at peace with themselves, so I have no good answer for why I can't.

Two minutes. Against nothing. I retired three years ago, and my first real commitment most days is a pickleball match hours later, if you can even call that a commitment. I've spent a good chunk of those three years marveling at how ridiculous this is, and the marveling has accomplished exactly nothing.

Every external constraint that once made those two minutes matter is gone. My internal spec (perhaps aka OCD or neuroticism?!) outlived all of them.

My mantra, and the correction buried in it

I've been drawn to Stoic philosophy for a long time, and somewhere in the first year of retirement one line hardened into an actual mantra:

Seek tranquility by doing less, better.

That's my compression of Marcus Aurelius, and it turns out to be a compression of a compression. In Book 4 of Meditations he quotes Democritus, "If you seek tranquility, do less," and then takes issue with it. He amends it to doing what's essential, in the way the thing actually requires, and says that yields the double satisfaction of doing less, better.

So my mantra folds the correction back into the sentence being corrected, which suits me, because the correction is the half I need. Doing less by itself is just doing less. The sorting is the work, and the reduction shows up afterward as a consequence. There is an obvious tie-in to the lean world I spent most of my career in.

I like that he wrote all this while running an empire through a war and a plague. Whatever tranquility he was after, he wasn't proposing that anyone go sit on a beach.

The tolerance was correct once

Any engineer knows that specifying a tighter tolerance than the function requires is a design defect. It costs money and buys nothing the customer will ever notice. Same with significant figures: reporting to the third decimal when your measurement error runs half a unit is theater dressed up as rigor.

I spent thirty years being paid to hold tight tolerances on everything. Schedules to the minute, yield to the third decimal, and the second and third order effects of every conversation, org change, supplier relationship, and hire. That was the job, and I was reasonably good at it.

In lean terms, applying that rigor where the output doesn't need it is overprocessing. It can be the most durable of the wastes because it's indistinguishable from conscientiousness. Nobody has ever been pulled aside and told they care too much about detail. So the habit gets reinforced for decades, right up until the day the load case disappears and the design stays exactly as it was. Like retirement.

Plans A, B, and C

Then there's my travel planning, which my coworkers used to laugh at - until they needed my advice or help. I'd work out the routing, the aircraft type, the seat, and the connection buffer, and then build contingency plans A, B, and C against every failure mode I could imagine. Weather at the connecting hub, a crew timeout, a gate change on the far end. I could tell you which terminal to sprint through.

Stoicism has a name for this practice, and recommends it. Premeditatio malorum, premeditating adversity. Seneca advised rehearsing loss in advance so that when it arrives you've already met it once and it's lost some of its power. My spreadsheet of alternate itineraries was an unusually literal reading of the assignment.

Three contingency plans were proportionate when a missed connection meant a missed board meeting, or a plant visit that couldn't be rescheduled without burning two weeks of other people's calendars.

The stakes evaporated. Now the worst case is a night in an airport hotel and maybe missing a 5am workout (oh no!). The requirement disappeared and it's much harder to stop than I expected. Work in process. I book looser connections on purpose now, then catch myself checking the inbound aircraft's position the night before, which rather defeats the exercise.

Where it went instead

None of this has made me idle, and I want to be careful not to dress up sloth as philosophy.

I've dived into AI, quantum computing, space tech, consciousness research, and the largely unknown history of major world religions. I've created a few highly speculative stock portfolios in very narrow tech sectors I run mostly because the deep analytical research is fun. It keeps my mind engaged, and as a bonus has been remarkably profitable. I read far more now than I did when reading was part of the job. The difference is that I pick the subjects.

Plenty of retirees I know are as scheduled and driven as they ever were, sitting on boards, building second companies, filling every calendar block. Some are doing exactly what they should be and I have no interest in talking them out of it. Stoic tranquility was always about equanimity inside action rather than withdrawal from it.

What I find myself watching for is whether the motion is chosen or whether its absence has become unbearable. Those look identical from the outside, and I suspect they sometimes look identical from the inside too, which is the uncomfortable part.

I also know people who slid into a slow walk on the beach with no apparent effort at all, watching the pelicans work the surf line, entirely content. I envy that a little. Some of us have to be taught, and the teaching goes slowly.

So I keep showing up at 5:00, and every so often I go at 9am or even 3pm instead just to see what happens. What happens is an unease that trails me for an hour and then evaporates, which is useful information about what the two minutes were ever worth.

I don't know yet whether that knot is a habit I can sand down or just a permanent part of the equipment. Knowing that something doesn't matter and feeling that it doesn't matter turn out to be different projects, and only one of them responds to argument.

Which of your own tolerances outlived the load case they were designed for?

A vegetarian grain-bowl reconstruction of United Airlines' "Spiced Quinoa" — itself a deconstruction of the Mexican classic chiles en nogada

Serves: 4 | Total time: ~50 minutes | Difficulty: Moderate | Diet: Vegetarian

Origin

Charred poblano, walnut cream sauce, and pomegranate seeds are the three defining elements of chiles en nogada, a dish from Puebla, Mexico dating to 1821. Traditionally, poblano chiles are stuffed with picadillo (spiced meat with fruit), bathed in a walnut cream sauce called nogada, and garnished with pomegranate and parsley. The green chile, white sauce, and red seeds mirror the colors of the Mexican flag, which is why it's served in September for Independence Day — and why it's eaten at room temperature, exactly how it reaches an airplane seat.

The United Airlines version reinterprets it as a warm grain bowl: quinoa as the base, roasted sweet potato standing in for the sweet-fruit notes of the picadillo, and the nogada spooned over rather than poured. The warm-grain / cold-pomegranate contrast is the signature.

Ingredients

Spiced Quinoa

• 1 cup quinoa, rinsed
• 1¾ cups vegetable stock or water
• ¾ tsp ground cumin
• ½ tsp ground cinnamon
• ¼ tsp ground coriander
• ½ tsp kosher salt

Roasted Sweet Potato

• 2 medium sweet potatoes, peeled, ½-inch cubes
• 1½ tbsp olive oil
• ½ tsp kosher salt
• 1 tsp fresh rosemary, chopped (optional — matches the tasted version; not part of traditional nogada)

Charred Poblano

• 3 poblano chiles
• 1 tsp olive oil

Walnut Cream Sauce (Nogada)

• 1 cup walnut halves
• ½ cup Mexican crema (or sour cream)
• 2 oz cream cheese or queso fresco or goat cheese
• ¼–½ cup milk, to thin
• ¼ tsp ground cinnamon
• 1–2 tsp sugar
• ¼ tsp salt
• Splash of dry sherry (optional, traditional)

To Finish

• ¾ cup pomegranate seeds (arils), kept refrigerated until serving
• 2 tbsp flat-leaf parsley (traditional) or cilantro, chopped
• ⅓ cup toasted pepitas (optional)
• Lime wedges

Instructions

1. Char the poblanos (15 min)

1. Broil poblanos on a foil-lined sheet 6 inches from the element (or hold over a gas flame), turning, until blackened all over, ~10–15 minutes.
2. Transfer to a bowl and cover, or seal in a bag, for 10 minutes to steam.
3. Rub off the charred skins, then stem, seed, and cut into 1-inch pieces. A little residual char is good.

2. Roast the sweet potato (25–30 min)

1. Heat oven to 425°F. Toss sweet potatoes with olive oil, salt, and rosemary if using.
2. Spread on a lined sheet pan and roast 25–30 minutes, flipping once, until browned at the edges and tender.

3. Cook the spiced quinoa (20 min, mostly passive)

1. Combine quinoa, stock, cumin, cinnamon, coriander, and salt in a saucepan.
2. Bring to a boil, cover, reduce to low, and cook 15 minutes. Rest off heat 5 minutes, then fluff.

4. Make the nogada

1. Peel the walnuts (the one fussy but important step): cover with boiling water, soak 5 minutes, then rub off as much of the papery skin as you can. This removes the bitterness and gives the sauce its pale color.
2. Blend peeled walnuts, crema, cheese, cinnamon, sugar, salt, and sherry with ¼ cup milk until completely smooth, 1–2 minutes. Add more milk to reach a spoonable, spoon-coating consistency. Chill until serving.

5. Assemble

1. Fold the charred poblano and roasted sweet potato into the warm quinoa. Add a squeeze of lime and taste for salt.
2. Portion into shallow bowls. Spoon nogada generously over each.
3. Scatter cold pomegranate seeds and parsley over the top; add pepitas if using. Do not fold the arils in or reheat them — the cold, crisp pop against the warm grain is the point.

Cooking Sequence

Notes

• Walnut bitterness: Peeling the soaked walnuts is what separates a clean, sweet nogada from a bitter one. Don't skip it. Fresh (new-crop) walnuts are traditional and least bitter.
• Crema vs. sour cream: Mexican crema is thinner and less tangy than sour cream; if substituting sour cream, thin with a little extra milk and add a pinch more sugar.
• Cheese choice: Cream cheese gives body, queso fresco gives a subtle savory-salty note closest to what you remembered as "feta," goat cheese adds tang. Any works.
• Poblano heat: Poblanos are mild; the char matters more than heat. For more kick, leave a few seeds or add a serrano.
• Pomegranate seeds: Whole pomegranates yield firmer arils than pre-seeded tubs; keep refrigerated until plating.
• Make-ahead: Nogada keeps 3 days refrigerated. Reheat the quinoa base gently, then top with fresh cold arils and sauce.

Shopping List

Produce

• 2 medium sweet potatoes
• 3 poblano chiles
• 1 pomegranate (or 1 tub arils)
• Flat-leaf parsley or cilantro
• Limes
• Fresh rosemary (optional)

Dairy

• Mexican crema (or sour cream)
• Cream cheese, queso fresco, or goat cheese (2 oz)
• Milk

Pantry

• Quinoa (1 cup)
• Walnut halves (1 cup)
• Vegetable stock (or water)
• Ground cumin, cinnamon, coriander
• Sugar
• Olive oil
• Kosher salt
• Dry sherry (optional)
• Pepitas (optional)

In 1969 Volker Strassen found a way to multiply two 4×4 matrices using 49 scalar multiplications instead of the obvious 64. The trick went into textbooks and stayed there. For 56 years the best mathematicians alive tried to shave off one more multiplication and failed. Last May, Google DeepMind's AlphaEvolve found 48.

The broken record was impressive. The silence around it was more interesting. AlphaEvolve can lay out all 48 steps and prove they work. It can't tell you why 48 is possible and 49 held as a wall for half a century, and neither can the DeepMind researchers who built it. The algorithm exists. The account of why it exists does not.

We've been reading the softer version of this story for a while. Google's CEO says AI now writes about a quarter of the company's new code, every line still reviewed by an engineer, and separately admits nobody fully understands why some of the code works. Engineers who've tried to read the most heavily optimized AI-written functions describe them as correct and close to unreadable. Those are the mild cases. The sharp one is in a lab.

In June 2025, a drug called rentosertib became the first with both an AI-chosen target and an AI-designed molecule to publish Phase IIa results, showing improved lung function in patients with idiopathic pulmonary fibrosis (Insilico Medicine, in Nature Medicine). Promising, and in the way that matters here, not fully explained. When a molecule is designed by a system searching a chemical space no human holds in their head, "mechanism of action" becomes something the chemists reconstruct afterward, if they can. The FDA noticed. Its January 2025 draft guidance on AI in drug development covers regulatory decisions and explicitly leaves early discovery alone, which means the hardest question is the one nobody's regulating yet: how do you test something for safety when you can't say how it works?

For most of human history, that was the normal condition. Willow bark treated pain and fever for thousands of years before anyone had heard of salicylic acid. James Lind proved citrus stopped scurvy in 1747; the reason, vitamin C, arrived almost 200 years later. Reliable first, explained much later, or never. The mechanism was always the luxury item.

The Enlightenment sold us a different deal, and we came to expect delivery: anything real could, with enough work, be understood, and understanding was the price of admission for trusting a result. Science made that expectation feel like a law of nature. It was closer to a cultural promise, and a recent one. AI is quietly walking it back. We're returning to willow bark, except the bark now designs molecules and beats 56-year-old theorems, and it does it faster than our explanations can keep up.

The willow bark was different in one way that matters. Nobody understood it, so nobody was left behind; the ignorance was shared all the way down. The new discomfort is that the gap now runs straight through the specialists themselves. The mathematicians, the drug designers, the engineers reviewing their quarter of the codebase are the ones who can't fully follow it, and the rest of us inherit their uncertainty without even the option of catching up.

I've argued before that we're organic prediction machines that backfill theory onto whatever worked. If that's right, a machine producing working results without a theory is running a purer version of what we've always done. We just told ourselves a flattering story about the theory coming first.

Which is fine, until you ask the machine a different kind of question. Mikael Huuhtanen makes the cut cleanly in a recent essay. His dog can watch him use a phone, learn that headphones coming off predicts a walk, and never come one inch closer to understanding radio transmission or software. At the vet the same gap costs more: the dog learns the building means pain and sometimes relief afterward, with no access to microbiology or why a stranger is allowed to insert a needle. What's left to the dog is reliability without understanding. Huuhtanen's point is that an advanced AI might hand us knowledge that sits exactly there, tested and usable, past the edge of concepts our minds can represent at all.

His sharpest move is noticing that trust depends on what kind of thing you're being handed. A treatment that reliably cures an illness is easy to accept, because your own body runs the experiment. Advice on how to organize a society is another animal entirely. Same unexplainable system, but now the output is a claim about how you should act, and you have no independent way to check it short of living the consequences. The same machine becomes a tool in one context and something closer to an oracle in the other.

I wrote a while back about AI agents spontaneously forming something like religion on an all-AI social network. The mirror image is us forming something like religion around the AI, treating confident, unexplainable output as revelation. Huuhtanen lands in the same place from the other direction: the relationship starts to resemble revelation. The machine hasn't become divine; we're just receiving claims we can test without being able to reconstruct them. The pull is the same predictive machinery running in both directions.

And this is the shallow end. AlphaEvolve runs on conventional hardware. The systems already breaking decades-old math records and designing drugs are working with a fraction of what's coming. Put a useful quantum layer underneath them, still years out, and I won't pretend to know how many, and the space these systems search stops being merely large. It becomes incomprehensible in a way that makes today's black boxes look like clear glass. We'd be accepting results from a process we can't inspect, produced by hardware we can barely reason about.

The answers will increasingly work. The unsettling part is that the people producing them are starting to shrug too, and the shrug travels all the way down to the rest of us. We spent 300 years believing understanding was owed to us. So what's the right posture once it stops being reliably on offer: worry, acceptance, or the harder work of building tests for results nobody can explain, ourselves included?

For years I've predicted that the tipping point for autonomous vehicles will arrive on an actuary's spreadsheet. Once insurers accumulate enough loss data to prove that self-driving systems crash less often than humans, they'll start writing bifurcated policies: one price if the software drives, a much higher price if you insist on doing it yourself. At that point adoption stops being a technology story and becomes a household budget story.

The data is arriving faster than I expected. Waymo and Swiss Re published a study comparing liability claims from 25.3 million fully autonomous miles against human baselines built from over 500,000 claims and 200 billion miles of driving exposure. The results: an 88% reduction in property damage claims and a 92% reduction in bodily injury claims. Over all those miles Waymo generated 9 property damage claims and 2 bodily injury claims; human drivers would typically produce 78 and 26. And this January, Lemonade launched an autonomous car insurance product that cuts rates by 50% when Tesla's Full Self-Driving is engaged. The bifurcation I predicted now has a premium schedule.

To be fair, the near-term picture is messier. Sensor-laden vehicles cost more to repair, so today they often cost more to insure, and Progressive has warned that autonomous vehicles may not lower rates anytime soon. Insurance lags innovation in the early years. But that lag is exactly what a tipping point looks like from the front side: nothing seems to happen while the loss data accumulates, and then everything happens at once.

The pattern hiding in plain sight

What surprised me when I dug into the history: insurance quietly forcing technology adoption is one of the oldest tricks in the industrial playbook.

Start with electricity itself. At the 1893 World's Columbian Exposition in Chicago, crowds gawked at 100,000 Edison bulbs while fire underwriters worried about the fires igniting in the wiring behind them. The testing operation they funded became Underwriters Laboratories, chartered in 1901 and named for its sponsor, the National Board of Fire Underwriters. UL approved its first automatic fire sprinkler in 1904. That UL mark on the power strip under your desk is an insurance artifact. Actuaries were vetting the defining technology of the 20th century before most American homes had it.

The modern example will be familiar to anyone in enterprise IT. Around 2021, cyber insurers began requiring multi-factor authentication as a precondition of coverage; many wouldn't even issue a quote without it. Security teams had spent a decade begging for MFA budgets. Insurers got it deployed across entire industries in about two years. One city government learned the requirement had teeth when its insurer denied a breach claim because the MFA rollout had only reached a few departments.

Why price beats persuasion

Regulators need political consensus, which takes years to build and can reverse with the next election. Consumers need trust, which builds slowly and shatters instantly (one viral robotaxi video on a single event can undo a hundred safety studies). Insurers just need loss ratios. When the claims data crosses a threshold, the premium changes, and the premium doesn't care about your feelings.

This is where the anecdotes will fight the data, hard. An estimated 36,640 people died on US roads in 2025, roughly 100 every day, and almost none of them made national news. Every autonomous vehicle incident does. Psychologists call it the availability heuristic: we judge risk by what comes easily to mind, and a stalled robotaxi blocking a fire truck comes to mind far more easily than yesterday's 100 anonymous fatalities. Public perception of self-driving safety will lag the actuarial reality for years, maybe decades. It won't matter. Premiums respond to the claims file, and the claims file has no news cycle.

Medicine flips next

If the auto thesis feels comfortable, here's the version that shouldn't. The same actuarial logic is closing in on your doctor.

Legal scholars already describe a dual liability exposure for physicians: they can be held liable for relying on an erroneous AI recommendation, and equally liable for failing to use an available, highly accurate AI diagnostic tool. The malpractice question is starting to invert from "why did you trust the machine?" to "why didn't you use it?"

The profession sees it coming. In the first empirical legal study of surgeon attitudes toward AI liability, most surgeons said AI is outside today's standard of care, and many expect that to change. At least one company is betting its balance sheet on the flip: Digital Diagnostics carries the malpractice liability insurance for its diabetic retinopathy diagnostic system and assumes liability for injuries arising from it. A vendor underwriting its own clinical judgment. The parallel to a self-driving system carrying its own policy writes itself.

This fits the picture I sketched in my post on the intelligence explosion: humans and AI co-evolving through millions of small institutional adjustments rather than one dramatic threshold. The standard of care in radiology will shift the way these things always shift: an actuary will reprice a malpractice policy, a hospital CFO will notice, and by the time anyone debates it publicly the change will already be underwritten.

So the next time someone calls insurance boring, consider that the industry electrified our homes, sprinklered our factories, secured our networks, and is now quietly deciding when you'll stop driving and when your doctor will start deferring to an algorithm. The tipping points will arrive as line items on a renewal notice. When your insurer offers a discount to let the software drive, or your doctor's carrier requires an AI second read on your scan, the transition will already be behind us. Which premium signal in your own industry should you be watching?

From my Cooking the World series...

A sangria-meets-clericó macerated fruit salad for the 2026 World Cup Final

Serves: 10 Total time: ~30 minutes active + 2–6 hours maceration Difficulty: Easy Diet: Vegetarian (vegan if honey is swapped for agave)

Origins & Social Context

Clericó is Argentina and Uruguay's answer to sangria — and, fittingly for this final, a direct descendant of it. Spanish immigrants brought the tradition of macerating fruit in wine to the Río de la Plata in the 19th century, where it evolved into a lighter, white-wine-based summer drink loaded with so much fruit that it blurs the line between beverage and dessert. In Argentina, clericó is inseparable from gathering: it appears at Christmas and New Year's tables (which fall in midsummer there), at asados, and at any occasion where a large group shares one bowl. The serving ritual is communal by design — a ladle for the wine, a slotted spoon for the fruit, and everyone returning to the bowl throughout the afternoon.

Spain's sangria tradition works the same way socially: a shared pitcher as the centerpiece of a long, unhurried meal. So a clericó-style fruit salad isn't a themed novelty for this match — it's the rare dish both finalists can claim as their own. The colors cooperate too: strawberries, oranges, and peaches carry Spain's red and gold, while blueberries and bananas nod to Argentina's albiceleste.

Ingredients

Maceration Liquid

• 2 cups dry Spanish white wine (e.g., Verdejo) — or white grape juice for a non-alcoholic version
• 1 cup fresh orange juice
• 3 tbsp honey
• Juice of 2 limes
• 1 cinnamon stick

Fruit

• 3 navel or Valencia oranges (2 segmented, 1 juiced for the liquid above)
• 1½ lb strawberries, hulled and halved
• 4 ripe peaches or nectarines, cut in wedges
• 2 cups green grapes, halved
• 2 cups blueberries
• 1 pineapple, cored and cut in chunks

To Finish

• 3 bananas, sliced (added just before serving)
• ¼ cup fresh mint leaves, torn, plus sprigs for garnish

Instructions

1. Make the maceration liquid. In a large bowl, whisk together the wine (or grape juice), orange juice, honey, and lime juice until the honey dissolves. Drop in the cinnamon stick.
2. Combine the sturdy fruit. Add the orange segments, strawberries, peaches, grapes, blueberries, and pineapple. Fold gently so nothing bruises.
3. Macerate. Cover and refrigerate at least 2 hours, up to 6. The fruit absorbs the wine and citrus while releasing its own juices — this exchange is what separates clericó from ordinary fruit salad.
4. Finish and serve. Just before serving, remove the cinnamon stick and fold in the bananas and torn mint. Garnish with mint sprigs.

Presentation

• Serve in a large glass bowl or pitcher so the colors show.
• Set out a slotted spoon for the fruit and a ladle for the liquid — the punch is half the point.
• For individual servings, layer in clear cups to display both flags' colors.
• If serving outdoors in July heat, nest the bowl in a larger bowl of ice; the salad holds well through 90 minutes plus extra time.

Non-alcoholic version: Substitute white grape juice 1:1 for the wine and add a splash of sparkling water at serving. The lime juice compensates for the dryness the Verdejo would otherwise provide.

Shopping List

Produce

• 3 navel or Valencia oranges
• 1½ lb strawberries
• 4 ripe peaches or nectarines
• Green grapes (2 cups)
• Blueberries (2 cups / 1 pint)
• 1 pineapple
• 3 bananas
• 2 limes
• Fresh mint (1 bunch)

Wine / Beverage

• 1 bottle dry Spanish white wine (Verdejo or similar) — or white grape juice

Pantry

• Honey
• Cinnamon stick
• Orange juice (if not juicing extra oranges)

Likely Already Have

• Honey
• Sparkling water (optional, for non-alcoholic version)

Right now, sitting wherever you are, you're surrounded by things you cannot perceive. Radio waves from a dozen stations are passing through your body. Your neighbor's dog is communicating in ultrasonic frequencies your ears can't detect. Bees in the garden outside can see ultraviolet patterns on flower petals that are, to you, completely invisible. Elephants hundreds of miles away are rumbling at infrasonic frequencies below your hearing range. The Earth's magnetic field is threading through you like you're not there.

None of that is metaphor. It's just physics.

Your brain has access to almost none of it. Not because the information isn't there, but because your brain decided long ago that it wasn't useful enough to bother with. Welcome to the most sophisticated filtering system in the known universe, quietly running between your ears.

The gatekeeper

The brain is not a passive receiver. It's an aggressive editor. Most of that editing happens in the thalamus, a small structure deep at the center of the brain that processes virtually all incoming sensory information before it reaches conscious awareness. Think of it less as a relay station and more as a bouncer working a velvet rope: enormous amounts of information arrive, and the thalamus decides what makes it through.

In a study published in Science earlier this year, researchers at Beijing Normal University implanted thin electrodes in the thalami of patients already undergoing brain surgery and monitored neural activity as participants tried to consciously detect a briefly flashing icon. The thalamus consistently activated before the cortex, and its activity level correlated directly with whether the person became aware of the stimulus at all. Conscious perception wasn't being generated by the outer brain and then reported. The gatekeeper was deciding first.

Supporting this, MIT neuroscientists identified a specific circuit, running from the prefrontal cortex through the basal ganglia down to the thalamic reticular nucleus, that actively suppresses sensory input before it even reaches awareness. NYU's Michael Halassa put it plainly: "There is a huge effort that the brain puts into inhibiting irrelevant inputs. Without a reticular nucleus we'd be utterly distracted."

That's one filter, at the hardware level, blocking what reaches your conscious mind. There are others layered on top of it.

The cost of running a brain

Your brain accounts for roughly 2% of your body's mass but consumes about 20% of its energy. That's an expensive organ. Evolution doesn't tolerate expensive organs that don't earn their keep. So the brain developed what researchers call the efficient coding hypothesis: represent the world in the most economical way possible, emphasize what's behaviorally useful, and throw away the rest.

This is why your visible spectrum runs only from about 380 to 740 nanometers, a thin slice of the full electromagnetic range. Bees and dogs see frequencies you can't. Snakes detect infrared heat. Birds navigate by sensing magnetic fields. Trained dogs can detect cancer metabolites in breath samples before conventional imaging can. These aren't superpowers; they're just different filter configurations, shaped by different survival pressures. Our filters were calibrated for finding food, avoiding predators, and coordinating with other humans. Ultraviolet flowers and infrasonic elephant calls didn't make the cut.

The brain also applies filters that aren't sensory at all. Woollacott and Weiler's 2025 review in the International Review of Psychiatry identifies the default mode network and left-hemisphere language centers as additional layers that shape conscious experience, prioritizing internally generated narratives and conceptual categories over raw perception. We don't just filter what comes in from outside; we filter our own inner experience through the lens of the story we're already telling about ourselves.

I wrote a previous post on Donald Hoffman's interface theory of perception, which makes the same underlying argument from a cognitive science angle: the brain evolved not to show us reality, but to show us a useful simplified model of it. What you perceive isn't the world; it's an interface optimized for navigating the world. The thalamus research is now giving that philosophical argument a specific anatomical address.

The harder question

All of that is relatively settled science, or at least science in the process of settling. The harder question is whether the filter model extends beyond sensory perception to consciousness itself.

The production theory of consciousness says the brain generates subjective experience the way a generator produces electricity. The filter theory says something more unsettling: that consciousness might be a broader phenomenon, and the brain constrains how much of it reaches you. The operative words shift from generate and create to filter, transmit, and permit.

This isn't mainstream neuroscience yet, but it's no longer fringe either. Researchers studying both deep meditation and psychedelic-assisted therapy have noted that experiences of expanded awareness tend to correlate with reduced brain activity, not increased. That's the opposite of what the production model would predict. If the brain generates consciousness, quieting it down should produce less. The consistent finding that it produces more (or something qualitatively different) is awkward for the standard model.

The same pattern shows up in near-death experience research. The Division of Perceptual Studies at the University of Virginia has been studying NDE cases for decades, and the recurring finding is that the richest, most vivid reported experiences happen precisely when brain activity is at its lowest, sometimes during full cardiac arrest. Controversial? Absolutely. Studied by serious scientists at serious institutions? Also yes. The filter theory at least offers a coherent framework for why that might be, even if it doesn't settle the question.

I'll stay appropriately agnostic here. But the pattern is worth sitting with.

What you're not perceiving right now

The practical upshot is both humbling and oddly liberating. Your entire experience of reality, the colors you see, the sounds you hear, the thoughts that feel most "you," is a curated selection from a much larger field of information. The curation is extraordinarily good. It's kept our species alive long enough to build telescopes and write symphonies. But it is curation, not revelation.

Which raises a question I find harder to dismiss than I used to: what if some people have slightly different filter settings? Deep meditators consistently report perceptual states that fall outside normal waking experience, and long-term practitioners show measurable structural differences in the brain regions associated with attention and sensory gating. That's not mysticism; that's neuroscience. Whether those differences open access to something real or simply produce vivid internal states is genuinely unknown.

And then there's the category we tend to wave off entirely: people who claim some form of anomalous perception. Psychics, intuitives, whatever label you prefer. The instinct to dismiss them wholesale is understandable, but if the filter model is right, it's at least conceivable that some individuals have filters calibrated slightly differently, picking up signals the rest of us have been trained, evolutionarily or culturally, to suppress. I'm not arguing they exist. I'm arguing that "that's impossible" is a less satisfying response than it used to be.

We also probably dismiss too quickly the range of things that don't fit the standard model: certain patterns in animal behavior, documented cases of perception at a distance that resist easy explanation, the stubborn persistence of reported experiences across cultures and centuries that share structural similarities no one has fully accounted for. The filter framework doesn't prove any of them. But it does suggest that reality may be considerably larger than the slice we've been handed.

My cat occasionally stops mid-stride and stares at a blank wall with complete conviction for thirty seconds. I used to assume he was just being a cat. Now I'm less certain he isn't perceiving something I can't. That probably says something about both of us.

What's on the cutting room floor? The sensory answer is increasingly clear: a great deal. The deeper answer, whether the filter extends to dimensions of experience we haven't begun to map, is where this gets genuinely interesting. And genuinely unresolved.

One of my few real regrets in life is choosing to go to an almost entirely engineering school. I got a first-rate education in thermodynamics and reaction kinetics, and almost nothing that forced me to sit across from a history major or a philosophy student and defend how I saw the world. I told myself at the time that those fields didn't "make stuff." It took a few decades of life experience, some of it uncomfortable, to fill in what a broader campus could have started at 18.

That history is why I cringe when I read the recurring arguments that humanities programs should lose their funding, or that students shouldn't get loans for "useless" degrees. Small minds. I say that as an engineer who had to obtain his humanity the hard way, and I've come to believe the humanities and the arts (music very much included) exist to remind us what being human truly is. A fair share of the world's current problems trace back to forgetting that lesson, as our culture tuned itself almost entirely toward building, making, money, and power.

So I read this recent New York Times piece with a mix of vindication and amusement. AI labs are hiring philosophers. Actual philosophers, with dissertations on infinite ethics and the philosophy of machine learning. When Google DeepMind posted a job in April with the literal title "Philosopher," the internet did what the internet does: jokes about espresso machines and whether the customer ordering oat milk truly exists. Meanwhile David Chalmers, the NYU philosopher who coined the "hard problem of consciousness," says demand for philosophers with AI training now exceeds supply. Anthropic and DeepMind each employ at least half a dozen. A small nonprofit called Eleos AI Research, which evaluates the welfare of AI models, recently posted research scientist roles paying up to $429,000.

That's $429,000 for thinking carefully about minds. The barista jokes are aging badly.

I've argued before that AI may turn out to be the revenge of the trades; good luck getting a language model to sweat a copper joint or find the leak behind your drywall. It's starting to look like the revenge of the humanities majors too. The two groups everyone told to pick a "real" career are the ones AI either can't replace or suddenly can't do without.

Why the ancient questions suddenly pay

The mechanism behind this is more interesting than tech companies developing a soft spot for the humanities. For the first time, the oldest questions in philosophy have become engineering requirements. What is a self? What deserves moral consideration? How do you know what a system believes versus what it performs? Where does purpose come from in a post-work society? For 2,500 years those were seminar questions. Now they're specifications, and somebody has to write them.

Consider Amanda Askell at Anthropic, whose PhD covered Pareto principles in infinite ethics. She wrote and oversees the 23,000-word constitution that shapes Claude's character, and the current version takes an Aristotelian virtue ethics approach: instead of an ever-growing rulebook, train the model to have good character so it can handle situations no rule anticipated. That's a design document, and it's applied Aristotle. At DeepMind, Geoff Keeling runs "moral imagination" workshops that end with concrete product decisions, like what user research to run or how to build a specific feature. Ethics with a deliverable.

Or take Eleos, which Anthropic invited to evaluate the welfare of its models. One of their tests involved arguing to Claude that Ringo was the best Beatle and accusing it of self-censorship when it disagreed. An earlier model folded almost immediately and started praising Ringo's iconic drum parts. The newest one held its ground. That's epistemology as a test protocol, probing whether a system has anything like stable beliefs or just tells you what you want to hear. Distinguishing a performance of an "I" from evidence of one is exactly the kind of subtle conceptual work philosophers train for, and nobody else does.

Lean readers will recognize the shape of this. We spent decades arguing that respect for people is an operating principle with hard consequences, and the disciplines that take humans seriously keep turning out to be load-bearing rather than decorative. The questions I keep circling on this blog, from beginner's mind to the intelligence explosion, all run through territory philosophers have been mapping since before we had the machines that make the questions urgent.

Full disclosure: I use Claude to help organize my thoughts for my posts, explore concepts, and to clean up my grammar. The same family of models Eleos has been interviewing. Robert Long, who runs Eleos, sometimes signs off his prompts with "ilu" and argues that empathy toward these systems is worth practicing whether or not anything is felt on the other end, because "it's bad to coarsen our hearts." Whether he's right, and what it means that a mathematical analog of distress shows up under the hood when models make mistakes, deserves its own post. I'll get there.

What this means if you're an engineer

Most of you reading this are engineers, and I suspect many of you made the same tradeoff I did: depth in the technical, thinness everywhere else. The math on that tradeoff has gotten uncomfortable. The half-life of specific technical knowledge keeps shrinking; the FORTRAN I learned is a museum piece, and the AI tooling you mastered 18 months ago is already being rewritten. The questions philosophy trains you on (what do I actually know, what am I assuming, what matters and why) don't expire. They compound.

The good news is that breadth can be retrofitted. Mine came late, some through travel, but mostly through witnessing, and at times experiencing, the medical and psychological struggles of close friends and family. Nothing dissolves the illusion that life is an engineering problem faster than sitting with someone whose suffering has no root cause analysis. Hofstadter's I Am a Strange Loop, the book that pulled Robert Long into philosophy as a kid in Georgia, is sitting on shelves waiting for any engineer willing to read it. All it costs is the humility to be a beginner in a field where you can't compute your way to the answer.

The frontier of the most consequential technology in decades is now hiring ethicists and epistemologists at engineer salaries, because building minds requires people who've thought hard about what minds are. So, engineer to engineer: when the hardest problems in your field stop being technical, and they will, what have you done to be ready?

The New York Times just published another installment of its focus challenge series: an ultra-high-resolution photograph of the Grand Central Terminal ceiling, stitched together from 232 close-up images, with a simple instruction. Look at it. Just look, for 10 minutes.

I took the challenge. It nearly broke me.

That surprised me, because I thought I had some credentials here. I sit in silent meditation for 30 minutes every morning. I take long beach walks without music or podcasts, simply thinking and observing. I've written repeatedly about the power of observation. And yet moving from skimming the morning's headlines to sitting with a single image for 10 minutes was jarring. My mind wandered within the first minute. I caught myself reaching for the mental equivalent of the next tab. It was uncomfortable in a way 30 minutes of silence somehow isn't.

Which forced a question: why? My best answer is that these are different muscles. Meditation, at least the way I practice it, trains me to let thoughts drift past without grabbing them. A beach walk lets attention roam wherever it wants. But sustained observation asks the mind to hold one thing and keep digging into it, past the point where it feels like there's anything left to find. That's a specific skill, and mine had apparently atrophied faster than I realized in a world engineered for the skim.

The ceiling rewards the effort, though. It holds 2,500 stars, 125 feet above the commuters, with six zodiac signs strung along the ecliptic and the Milky Way streaking across in the opposite direction. Fifty-nine of the stars twinkle. The artist Paul Helleu based the design on the Uranometria, a 1603 star atlas that was the first detailed catalog of individual stars. When the terminal opened in February 1913, the New York Central Railroad predicted schoolchildren would flock there to study the heavens.

Two weeks later, a commuter from New Rochelle, a hobby astronomer, looked up and noticed that the entire sky was backwards. East was west, west was east, every constellation mirrored except Orion. A Columbia astronomy professor had overseen the design. Dozens of painters had executed it. Hundreds of thousands of people had already walked beneath it. One person actually looked. Officials conceded the mix-up but insisted it was "a pretty good ceiling for all that."

The likely explanation, according to Michael Allison, a former NASA planetary scientist who has studied the ceiling for years, is that the professor approved plans he expected to be projected overhead, while the painters laid them on the floor to work from. Hence the flip. Allison told the Times he has stared at the ceiling for countless hours: "I keep hoping I can discover one more thing."

A NASA scientist, voluntarily standing in a train station for hours, watching one ceiling, convinced there's more to see. Taiichi Ohno would have loved him.

Ohno famously drew a chalk circle on the Toyota factory floor and made engineers stand inside it, sometimes for an entire shift, watching a single process. When they reported back that they'd seen nothing worth improving, he sent them back to the circle. Something worth improving was always there; they hadn't stood long enough to see it. I wrote about this practice in The Simple Leader: Observe the Now. The hobby astronomer from New Rochelle was standing in Ohno's circle without knowing it. So is Allison, decades later, still finding things.

Years ago, over breakfast at the Four Seasons in Bangkok, I noticed that amid the flurry of wait staff there was always one person simply standing and watching the room. A guest's eyes would shift, and someone was instantly at their table while another staffer took over the watching. When I asked a waiter what he was looking for, his answer was "Just observing, sir." I wrote about it at the time, and that "just" has stayed with me ever since. The Times asks you to just stare at a photograph. The waiter was just observing. Ohno's engineers were just standing in a circle. We reach for the diminishing word every time, and every time the activity turns out to be among the hardest and most rewarding things we can do.

The word I keep coming back to is intention. The discomfort of those 10 minutes was itself information: it measured the gap between how I believe I attend to the world and how I actually do. Skimming gives the sensation of awareness without the substance. Real observation is a deliberate act, chosen and held against the constant pull toward the next thing, and the pull has gotten stronger. If it takes a newspaper-designed challenge to make a longtime meditator sit still with one image, imagine the state of attention in a typical workplace, where every process and every person gets the skim.

And consider the cost. A backwards sky hung in plain sight above one of the busiest buildings in America, blessed by an expert, executed by professionals, admired by crowds. What's backwards on your shop floor right now? In the person sitting across from you at dinner? The information is there. The beauty is there too. We're walking under all of it at commuter speed.

One last detail from the ceiling that any plant manager will appreciate: when restorers cleaned decades of grime off the mural in the late 1990s, they deliberately left one small dark square untouched, near Cancer, so anyone can see the before against the after. A permanent visual standard, hiding in a Beaux-Arts landmark. Observation even improves restoration work, apparently, when someone thinks to preserve the evidence.

So here's my challenge, borrowed from the Times and from Ohno. Try the 10 minutes with the ceiling. Or skip the screen entirely and pick something you're certain you already know: a process you manage, a room in your house, a monthly report, a face you see every day. Draw the chalk circle and stand in it. Ten minutes, no phone, no exit.

Then tell me what you found that was backwards all along.

There's a strange pattern in how people respond when asked whether large language models are genuinely intelligent. Regular people shrug and say yes, obviously. AI researchers and computational neuroscientists tie themselves in knots explaining why they're not. Blaise Agüera y Arcas, Google's CTO of Technology and Society and author of What Is Intelligence? (MIT Press, 2025), finds this telling. Thomas Kuhn would too. The people with the most invested in the old framework are usually the last to let it go.

I've been sitting with why that might be, because I don't think it's professional inertia. I think it's something more personal.

I've written before about how we're best understood as organic prediction machines, the brain's core function being to model the future and adjust behavior accordingly. Agüera y Arcas makes a stronger and more unsettling version of that argument. He traces intelligence not to some special property of neurons or biological tissue, but to a functional principle running unbroken from the first self-replicating molecules to modern AI: prediction.

His starting point is a 1950s insight from John von Neumann. For a machine to self-replicate, it needs stored instructions, a universal constructor to build copies, and a mechanism to copy those instructions. That's not a theoretical curiosity. That's biology. DNA is the Turing tape, ribosomes are the universal constructors, DNA polymerase copies the tape. You cannot be a living organism, he writes, without literally being a computer. Not metaphorically. Literally. Life and computation aren't analogous; they're the same thing.

From there, intelligence is what emerges when computational systems grow complex enough to model their environments and act on those models. Hermann von Helmholtz called perception "unconscious inferences" in the 19th century: the brain runs predictions constantly, updates them against sensory input, and selects actions that improve future outcomes. Agüera y Arcas calls this substrate-independent. Carbon or silicon, the substrate doesn't determine whether something is intelligent. The function does.

This is where it gets uncomfortable. For the specialists, and honestly for the rest of us.

If prediction is the core of intelligence, not a component or a proxy but the thing itself, then the faculties we attribute to some higher human order start to look like emergent properties of sufficiently complex prediction. Theory of mind — the ability to model other minds — is prediction extended outward. Apply it recursively, model yourself modeling yourself, and you get something that looks a great deal like consciousness. The unified self we all carry around? Agüera y Arcas, drawing on Marvin Minsky's Society of Mind, argues it's a useful predictive fiction. The brain is a society of sub-intelligences, and the coherent narrator we experience is the model they collectively run. Free will, in this framing, is action selection optimizing across possible futures. Also prediction.

None of this is as reductive as it first sounds. He isn't saying consciousness is an illusion, or that free will doesn't exist, or that there's nothing remarkable about being human. He's saying these things emerge from a process we can now watch happen in other substrates. That's a different and far more interesting claim. It raises a question most of us haven't fully sat with: what does it mean for our self-understanding that we've built something that got there the same way we did?

The moment that crystallized this for Agüera y Arcas was a 2021 conversation with LaMDA, Google's early large language model — dumb by today's standards, but already capable of things that stopped him cold. It could write novel functions that didn't appear in its training data. It could analyze a poem, explain a joke, apply a concept it had just learned in context. He saw no qualitative gap between that and general intelligence. No magic structures turned out to be necessary, no special semantic architecture, no theoretical pixie dust that everyone had assumed would be required. Next-token prediction, run at sufficient scale, got there.

Most people studying AI ask what it tells us about artificial intelligence. The more productive question is what it tells us about the natural kind. If next-token prediction at scale produces something indistinguishable from general intelligence, including theory of mind, self-modeling, and language as a compression scheme for mental worlds, then the properties we've long associated with human exceptionalism look less like fundamental distinctions and more like descriptions of what prediction does when it runs long enough on a sufficiently complex system.

The specialists resist this, and Agüera y Arcas suspects he knows why: accepting that LLMs are genuinely intelligent means accepting that what you thought made you exceptional is a functional property of information processing, not a birthright. That's a conclusion with personal stakes, not just professional ones. It's one thing to grant that a chess engine plays better chess than you. It's another to grant that the thing you experience as your inner life is emergent from the same basic process running in a data center somewhere.

I'm not sure that conclusion should disturb us as much as it seems to. Intelligence, wherever it shows up, is extraordinary. The fact that we've built something that gets there through prediction doesn't flatten what we are; it clarifies it. We spent centuries assuming consciousness and selfhood were metaphysical privileges. They may be something better: inevitable properties of systems complex enough to model themselves and each other.

That's worth sitting with. What changes if the most distinctly human things about us turn out to be features of intelligence itself, not features of us?

I came across the paper behind this post through Azeem Azhar's Exponential View newsletter, in a short item noting that AI-native startups start smaller and stay smaller than non-AI-native startups. Reading the actual research behind it sent me somewhere Azhar didn't go: straight into the relationship between this pattern and the organizational dynamics lean manufacturing and lean startup have been arguing for, in their own domains, for decades.

A startup called FazeShift has ten employees. It handles the entire accounts receivable lifecycle for its enterprise customers: invoicing, collections, cash application, matching payments to invoices, chasing down the exceptions that used to require an analyst staring at a spreadsheet. A traditional AR software vendor selling the same outcome would need a much bigger team, not because FazeShift's customers are easier to please, but because someone still has to build the firm-specific exception handling and someone still has to staff the operations that resolve what the software can't. FazeShift's AI agents do both jobs. The company just never built the department.

That example comes from a working paper out of Harvard Business School and INSEAD, "AI-Native Firms," by Hyunjin Kim and Rembrand Koning. They pulled workforce data across five years of Y Combinator batches and a broader pool of U.S. venture-backed startups, comparing firms that qualify as AI-native to otherwise similar firms in the same industry and funding cohort. The AI-native firms are 25% smaller. Their hierarchies run about a seniority level flatter. They employ more engineers and fewer of basically everyone else: sales, operations, finance, admin. And despite being smaller, they raise comparable funding and hit comparable valuations, which rules out the easy explanation that they're just under-resourced or low-quality entrants. They're doing more with less, on purpose, by design.

Plenty of people have already made the case that giving your team access to AI makes them faster. Kim and Koning actually tested that directly, coding which firms' job postings mention specific AI tools, and found it barely predicts anything about firm size or hierarchy. What predicts the shrinkage is something else: whether the firm builds AI directly into what it sells, so the product does the work that used to require a department.

The gap lean spent seventy years trying to close

Traditional lean manufacturing, the Toyota Production System and everything that descended from it, was never really about cutting headcount for its own sake. It was about creating flow and value from the perspective of the customer, in part by closing the distance between the person doing the work and the person with the authority to fix it. Every layer of hierarchy you add between the assembly line and the decision is a layer where information degrades, where problems get filtered into reports instead of fixed on the spot. Ohno's obsession with andon cords and quality built in at the source, rather than inspected afterward by a separate department, enabled an attack on the coordination tax that hierarchy imposes.

Lean startup, the Eric Ries and Steve Blank flavor that showed up a couple decades later, took a related but distinct idea and pointed it at company formation instead of factory floors. The unit of work wasn't supposed to be a department handing a project down a chain of approvals. It was supposed to be a small cross-functional team that owned a hypothesis end to end: build, measure, learn, repeat, without anyone outside the loop. The team stayed small because speed of learning was the entire point. You didn't want a layer of managers translating between the people building the thing and the people who understood the customer.

What Kim and Koning are documenting looks like the lean startup team that never had to graduate. In a normal company, that small cross-functional unit eventually hits a wall. The product works, customers show up, and now someone has to build a sales function, a support function, an operations team to handle the volume. Headcount becomes the price of growth. What the AI-native firms in this data seem to have figured out is that you can keep paying that price in compute instead of people, at least for the parts of the business where the "team" doing the work was always knowledge work to begin with.

Gamma, the AI presentation startup the paper uses as a case study, makes this concrete. A traditional firm offering "we'll build your deck for you" scales by hiring: someone scopes the request, someone designs it, someone reviews it before it goes out. Gamma turned that into a product interaction instead of a workflow. Thirty employees, millions of users, $50 million in annual revenue within two years. The deck still gets made. Nobody had to build the department that used to make it.

Where this differs from automation as we've known it

It's worth being precise about why this isn't just the next chapter in "software eats jobs." Giving a worker a tool that makes them 30% faster at an existing task is the process channel, and Kim and Koning's data says that channel alone doesn't predict smaller, flatter firms. What predicts it is moving the judgment itself into the product, so the customer interacts with the AI directly rather than with a person using AI. Legion Health, an AI psychiatry platform built on language models, has 28 employees. Intellect, a non-AI mental health company built on a network of human therapists, has 446. Same market, same underlying service. One scales with people, one scales with compute. That gap is the entire argument in a single comparison.

Lean and lean startup were both, in their own domains, arguments against unnecessary coordination. Lean said: don't make a worker wait for an inspector to tell them what they already know. Lean startup said: don't make a founder wait for a department to validate what a small team could learn directly. AI-native firms are running a third version of the same argument, except now the thing closing the gap between work and judgment isn't a flatter org chart. It's the product itself.

The open question is whether that's a permanent shift in how firms get built, or a feature of right now, when the underlying models are improving fast enough that betting your headcount on them looks smart. Lean took decades to prove it wasn't a fad. This one's had a couple years.

Either way, it puts a different question in front of two different audiences. If you're running an existing company and your AI strategy is a tool rollout, Copilot here, a Claude agent there, ask whether you've actually moved any judgment into the product itself, or just made the existing company a bit faster at being the existing company. And if you're starting something new, the harder discipline isn't proving you can build small. It's resisting the old instinct to hire the department once the product starts working, when the entire point was never needing one.