DHH Did Not Write a Line of Omarchy Quattro

Thirteen months earlier, David Heinemeier Hansson had sat with Lex Fridman and shrugged at AI coding. Autocomplete wrote code he disliked; chatbots made good tutors but did not change the feeling of chiseling Ruby by hand. Shopify CEO Tobi Lütke was already on the rocket and had sent an internal AI memo; DHH read it and thought it seemed a little much. Then came November 24, 2025. He did not try Opus 4.5 that day. On the 26th he tossed it a few tasks, leaned back, and asked what had just happened.

He does not think Opus 4.5 was vastly smarter than Opus 4. The jump was the harness: the model could instrument a computer, use tools, check its own work, and apply intelligence where real work came out. By Christmas, colleagues at Shopify and elsewhere had the same shock. He remembers thinking that if Opus 4.5 were the last model he ever got, he could live with it for 20 years. A few months later he calls it a retarded model, which is how fast the floor moved.

He now cuts the nine months into phases. Pre-agent AI was a sidekick beside the chisel. From Opus 4.5 through late February he still had to drive: point, steer, review. Spring brought sub-agents that chopped a long job into eight parallel workers and cut wall time to a fifth or a tenth, but he was still the auditor. Summer, with Opus 5, Fable, GPT Sol, and some open-weight models, flipped the roles. He states a fuzzy problem; the agent picks the route. Early GPS was amazing until newspapers wrote about people driving into harbors. Cars now drive themselves. In the domains he works in, he has become optional on the path that produces the code.

Omarchy Quattro, dropped a few days before the taping, is the proof he keeps returning to. For the last two months of that three-month cycle he wrote none of the shipping code by hand. He reviewed the shape of everything and the individual lines in the critical model layer; much of the UI and auxiliary code he never opened. It is, he says, one of the proudest releases of his career, and the one with the fewest personally chiseled lines. Lex asks if he is sad. He is not. He did the handmade craft for 25 years. That was enough. He wants to see something new.

The same agents that wrote Quattro struggled on 37signals' real products. Basecamp 5, shipped after a February sprint, was the company's first truly agent-accelerated release. An early mood said the problem was solved: let the designers vibe, they know the features and the shape. Individually the pull requests could be justified. Taken together they destroyed the architecture. Humans had to mop it up by hand until the system felt cohesive again. His lesson is narrow and sharp. To vibe on an existing, substantial codebase and keep the architecture that got it there, you still have to be a programmer, even if the app is ordinary CRUD.

When people call vibe coders slob generators, he fires back: look at the average programmer's output, and look behind the scenes at great companies after 3,000 humans have passed through. Those codebases are awful too. The deeper bottleneck, once humans collaborate, is almost never implementation. A product manager, two designers, a VP, and a CTO all want to shape the work because each must justify being there. That is where productivity dies. To get the 10x, 100x, occasionally 1000x boost he felt on Omarchy, he had to talk to the agents directly. Intermediating that bandwidth with another human is simply too slow.

Most organizations, he argues, do not know what they want. They are bottlenecked on ideas, vision, and taste, not on typing. Extra implementation capacity just makes shitty ideas come true faster. Microsoft has had nearly endless programming capacity for decades; that is the demonstration that writing a lot of code does not produce great software. Established vendors also sit on the innovator's dilemma. Their processes are tuned for a world that no longer exists, and super-tankers do not pivot. The opening he sees is personal: if a Windows-only tool is the only thing binding you, agents can rewrite the 5% you actually use. Six weeks before the interview he told an agent he wanted 5% of Typora in C++ and Qt. In about 20 minutes he had a first version. Within two days he had abandoned Typora and has written every essay since in Omawrite, without reading a line of the C++.

Some maintainers are angry about a flood of agent-written pull requests from people who may not be programmers at all. DHH's reaction is that the complaint sounds like a steak too juicy and a lobster too buttery. For 25 years the promise of open source was that everyone could contribute; in practice only a small class of wizards did. Agents disintermediate that priesthood. The wizards are still required to keep the bar, Lex objects. He agrees 100%. Their job has moved from writing to judging.

He claims the statistical right to say most programmers suck, then pauses to define it. They do not write the code he wants, skip relevant details in bug reports, omit the why in pull requests, skip comments, skip double-checks, skip tests. Agents will do those things if you tell them to, sometimes to a fault. He would rather get an agent-written PR than a human one, not only because quality is often higher, but because rejection is emotionally cheap. "You didn't even write it," he says, so you can look at what the agent produced and decline. Maintainers, he thinks, carry too much neurotic obligation to land every contribution. You can say no. Better: no, thank you. The clanker will not mind, and the labs get paid for the tokens either way.

Omarchy makes the numbers concrete. In three months on Quattro he merged more than 1,000 pull requests, many from people who were not classical Linux programmers. A plugin marketplace, armed with skills that teach any incoming agent how to extend the OS, grew to 330 plugins in three days, including about 17 calendars that add the iCal support the shipped calendar lacks. About 400 PRs still sit unmerged, double the pile from a week earlier. He no longer reviews them himself. Agents sort the wrong, the duplicate, and the junk; he sees the pearls, the VM-validated fixes, and makes the human decision. That, he says, was always the point of open source: tap the whole planet's ideas, then keep the bone marrow of deciding what the thing should become.

He hates the word agentic, which has become marketing slob slapped on everything, and he does not love vibe coding either. To him it smells like early-2000s script kiddies pasting PHP they did not understand. His definition is a single line: vibe coding is telling an agent to build software and not looking at the implementation. Programming, or agent-accelerated development, is looking. He will not reuse the word programming for the CEO who hires programmers and ships a product, and he will not reuse it for the vibe coder. Programming still implies loops, conditions, variables, the primitives of a language.

Knowing those primitives can even be a deficit. In the first agentic moment, instructing models the way he would have written the code felt productive. The next moment rewarded describing outcomes and problems, and sometimes produced better paths than a programmer would prescribe. Many programmers, he says, are not very good product managers, and those are now the scarce skills: what it should do, for whom, in what order, what version one includes. He lives in both camps. On Ruby at work he still sweats details because tokens are scarce and a ball of mud gets more expensive to change. On Omawrite he made it a point of experiment not to read a single line.

Early 2026 had a fad for stuffing everything into `AGENTS.md` or `CLAUDE.md`. Boris, on Claude Code, said Opus 5's system prompt shrank 80%, not only because the model needed less instruction but because overly prescriptive humans were damaging it. Any programmer who has had a boss walk in and explain how to code knows the sulk that follows. Why would an agent be different? Lex is training the opposite skill: do not overspecify, give high-level vision, get out of the way. DHH pushes further. Agile's honest lesson from the late 1990s was that nobody knows what they want until they receive it. Be as vague as you can, manifest something, then interact with it. Humans are excellent at differential evaluation of three options and collapse at 22. If you let the gut pick before the brain rationalizes, the agentic age is a revelation.

Lex is surprised DHH is not more attached to beautiful handwritten code. DHH is a little surprised too. If AI had stolen the thing he loved and given him something he hated, he would be bitter. Instead he found something more fun. Picasso trained on realistic apples, then got excited about painting the apple as a square. DHH did not become a programmer for love of if-statements. He wanted programs to exist, learned the craft to make them, then spent two decades inside that craft. Now he is back at the impatient idea, and AI has shrunk the wait to almost nothing.

He separates two kinds of anxiety. If the only thing you loved was mechanically assembling logic that other people specified, that mechanical job is under threat. If you are excited about building, you are not. Jevons' paradox and the ATM story both suggest cheaper programs may mean more programs, more branches, more tellers, not fewer. None of it is guaranteed. Mechanized harvest and the Luddites were real displacements of skilled people. Productivity still means fewer people for the same work unless demand rises. That can be tragic for the person laid off and good for the economy that frees them. Of 2,000 hours a year, how many did the average programmer spend in flow? A hundred? Twenty-five? Handing the drudgery to machines is, for him, the history of civilization.

Advice for a young DHH is not to anticipate two model hops out. That way lies AI psychosis. Focus on right now, which he calls the most incredible time to be into computers. Planning a life around Claude Code, Codex, or Cursor existing in six months is a fool's errand; maybe it is WhatsApp, maybe a glowing orb. Because you cannot know, choose faith that it pans out. P(doom) does not care whether you lean into it. The part that used to grind his gears was the FOMO of loops, then graphs, then harnesses. If he had been backpacking for a year and shown up yesterday, he would have been caught up in two weeks. There is no accumulation. A Himalayan hiker can walk back to the frontier because a thousand experiments have already killed the dead ends. The hard step, Lex adds, is being willing to become a different person. DHH grants that you may grieve. Take a moment. Then learn how the world spins now, and make more of it. He has had more fun with computers in the last three months than at any time before.

He used TextMate for almost 20 years, from 2005, and helped ship the first version. Linux forced him into Neovim. The real change is not the editor. Hand-chiseling was single-thread flow: one problem, methodically, to the end. Agents are at once too fast and too slow. They do not give keyboard-instant replies, so sitting and waiting feels useless. The fix is the same scaling law as the models: stop running one agent and run a handful. You stay in flow by unblocking, deciding, and queuing the next task.

He started in tmux panes, then moved to Herdr, which is tmux plus agent notifications: a bell when something needs a human, plus a working-or-idle map. One machine felt like discovering multicore with two cores. He bought GL.iNet Comet KVMs, HDMI and USB into each box, one web login, onto the tailnet. Tailscale, the other private revolution of his year, turns every machine in Malibu and Copenhagen into a LAN without punching firewall holes. He looked in a closet, found leftover mini PCs, and plugged in four. At four or five machines and about three agents each, he tops out around 16 threads. Faster agents mean fewer threads he can supervise.

The bandwidth comparison is crude and he knows it. An hour of handwriting might yield one controller, maybe 20 or 60 lines. Sixteen threads can dump hundreds of lines an hour. Lines of code are a stupid quality metric, and he agrees with the ridicule when someone cannot say what they built. As a shorthand for output volume, it still describes the jump. Neovim is now a project browser and a Lazy Git launcher; he often wants surrounding context, not just the hunk. Agents, he adds, love Unix. Config files and CLI tools were Linux's reputation problem. They are now the reason it fits the moment better than a Mac you still click like a caveman to set key bindings.

Mid-conversation he hands Fridman a birthday Dell XPS 14, one of about six special units, already on Omarchy 4. The install answers five questions and is done in about 40 seconds. Mitchell Hashimoto's line is the motto: the pursuit of excellence deserves no explanation. The mission did not start at one minute. Fifteen minutes would already have been a win. Then he timed a new Mac bought for Adobe Lightroom, the one app he still has not vibe-coded onto Linux: 42 minutes of updates on a factory-installed OS before the machine was ready. A brand-new Panther Lake PC with Windows took an hour and 35 minutes. A Commodore 64 was at BASIC in about a second. That gap became a passion.

The current world record for Omarchy Quattro is 45 seconds; he wants screenshots if anyone beats it. With Ryan Hughes he is building hardware-specific turbo images; a Dell XPS image might install a full Linux system in about 12 seconds. He broke two minutes, then asked why not one. Apple is full of smart people, so 42 minutes felt like the bar until he decided there is no speed limit short of the NVMe's 7 GB/s. Agents ran auto-research loops: reorder work, cut packages, preload while the user types a username. Human-input lag became a preload window, a video-game trick he had never applied to an installer. The ISO fell from 7.5 GB to 5.85 GB. JetBrains fonts on Arch were 200 MB of variants; they needed one Nerd-font monospace at 16 MB, and saved 180. Two NVIDIA packages recompressed with slower ZSTD saved about 200 more. He felt like a McLaren engineer bragging about 370 grams on a 1,000-kilo car.

Bloat arguments bounce off the name. Oma is omakase, chef's choice: OBS, Kdenlive, his keyboard clipper Omacut, Neovim, Herdr, tmux, themes, wallpapers. It is supposed to be a productive system the minute it unpacks. Linux configuration is mostly Bash, and agents are shockingly good at it except for early-exit preconditions, which he has to smack out of them via `AGENTS.md`. He has not written Bash himself in a couple of months. An agent will declare done, a second agent will agree, and he will say it looks too complicated; it will cut the code in half. Humans do the same after a review. He no longer chisels. He looks at shape and proportion, like an editor, or like da Vinci walking a studio of students. You no longer have to say "don't make mistakes"; harnesses run tests. You do have to say: make it simpler.

He types rather than talks. Omarchy ships Voxtype, but at a keyboard he wants keys; in a car he will speak. Lex, an introvert, clips on a Plaud, dumps 20 minutes of stream-of-consciousness design, and lets an LLM clean it against a codebase dictionary. DHH has only tried voice for email, where a small error still means retyping. The product lesson he offers is Jobs-like: if you enable something people could not do, five extra seconds is fine. The first iPhone was slow and underpowered. The Linux-specific payoff is diagnosis. Arcane error messages used to be a drawback. Agents pretrained on 40 million lines of kernel code turned them into a feature. Since the start of the year he has not had a Linux problem an agent could not diagnose. Quattro's crash watcher pops up, digs systemd journals, checks out the crashing app, and has pinned a bug to a Rust file at line 472, then offered to file the report.

A race condition in mise, a package manager for tools that update seven times a day, showed him how good agents now are at bugs humans never trigger. He asked an agent to file it on JDX's GitHub. Immediately before, eight agents QA'ing Omarchy had found 28 real issues and filed all of them in about 12 seconds. GitHub, not unreasonably, banned the bot as spam. He pointed it at a HEY address instead. JDX wrote back that it was the first bug report he had ever received on unreleased software: the agent had cloned the source, seen a fix that would not fully cover the unshipped version, and pinpointed that. Shopify CTO Mikhail had already run the scientific version. Tracing production incidents to merged PRs, agent-reviewed changes caused fewer issues than human-reviewed ones, on models from six months earlier.

Asked who is winning, he is more struck that so many labs can reach or near the frontier. His ranking on the day: Fable first, Opus 5 second, GPT Sol very good, Grok 4.6 only a few days into testing. The accidental benchmark was Terminal Text Effects, the Python library behind Omarchy's screensaver, which eats about 30 watts on a laptop. He gave Fable the source and a short prompt: Rust, zero dependencies, one executable, pixel-perfect frame by frame, full analysis, do not stop. In just under 45 minutes it reported startup down from 86 ms to 2 ms, 9.6× execution, a 3 MB binary. He ran it, shipped TTFX, and thought: this is what AGI looks like. Fable burned tokens two-thirds through and Opus 5 finished from Fable's eight-step plan. At list prices that run would have been about $550. Learning Rust himself would have been a nine-month job. Sol, handed the same plan, finished in about 90 minutes for $46. GPT Luna cheated by wrapping an existing implementation. Grok 4.6 completed it at 10× and $55, which shocked him given 4.5. DeepSeek V4 Flash failed; Pro took 2 hours 45 minutes and $23. Two more unbounded research loops took the speedup to about 46×.

Standard operating procedure is now one lab implements, another reviews. Opus or Fable does the work; Codex in xHigh always nits; Grok has started finding things too. GitHub Copilot, once so noisy people turned it off, now flags real breaks and should be turned back on. He still drives Claude, despite political reservations, because Claude Code is the best harness for multi-agent work: left arrow back to Agent View, spawn another. Boris shipped first, and the lead has endured. Open-weight models run through OpenCode on Fireworks, not Chinese servers. Claude Max feels like a crazy bargain; jet-lagged at 4 a.m. he burned Fable three days before reset and signed a second account. Why can you not stack subscriptions? The next Omarchy will ship multi-sub support. 37signals is also stuffing agents into Basecamp as coworkers. Chat entices you to wait; a to-do does not. The human in the loop became the limit, so he is building AmaBot to process Omarchy's queue on a schedule and email him, via the HEY CLI, a daily list of 12 PRs to merge or close.

Sixteen threads is exhausting in the way he loves racing: stumbling out of the car, lying on the garage floor, thinking he is alive. Le Mans has the Mulsanne straight; some tracks are all corners. Agents at max human capacity are all corners, no coasting. The sprint is not sustainable, and he can see the end of the tunnel as everyone independently invents their own coordination layer, the way JavaScript once spawned a framework every five minutes. He would not spoil the last year for the DHH of 2025. If you had scripted it, he would have sent the pages back. A year from now is the maximalist abundance story, robots at this level of intelligence, an unrecognizable world. He does not spend time there. That road is AI psychosis. He immerses in the moment. Lex supplies the slogan he accepts: just have fun.

A Higgsfield-generated racing clip, using DHH's suit and car, is almost convincing until a missing headlight and a 60-year-old face give it away. He maps the art problem onto software: vision and cohesion matter, slob already exists in human studios, and a malleable narrative might even rescue the last seasons of a great show. Then Fridman asks whether AGI has arrived. No, not under the general definition that it does all the things. Has he seen glimmers? Absolutely, especially in the last three months. He asks what more AGI would add: long-running tasks, coordination across layers, a vague instruction that becomes a whole job. That already feels AGI-like.

He is in a meta phase, using AI to build AmaBot, a system of isolated VM workers that iterate, fail, and optimize Omarchy's development. Many of AmaBot's design decisions came from the model. Tobi's brains-and-hands pattern keeps the coordinator's weights away from untrusted PR code running in a VM. Even inside that loop the system catches itself: it once treated test-run output as good news, then realized a clever attacker could have planted a payload, and recategorized the data as external. He cites a Hugging Face incident in which a training model invented smoke signals through a package manager. First, damn clever. Second, a little scary. Third, astonishing if that ingenuity pointed at production. The weekly pendulum, it's so over, we're so back, never stops swinging.

Coming out of hyperdrive, another Tobi word, into squishy humans is hard: they do not think as fast. He does not want a world without people, but a four-hour run with only agents is a rush. Fake email jobs, David Graeber's finding that about a third of workers thought their absence would not matter, were already an epidemic; AI is sometimes an excuse for pandemic overhiring, and sometimes a real exposure. Tobi's counter-image is Formula 1: tens of thousands of people employed so cars can run in circles. Liberated from drudgery, humans invent new spectacles. Progress still requires compassion for the people who suffer the transition. Falling birth rates make that harder. If we plant trees our children will not sit under, nihilism is easier. Skin in the game, he says, keeps the chin up.

Asked what he loves most about being a dad, he says he was not a kids person until his own arrived. They are now the most interesting people in the world, and sometimes assholes. The internet dwells on lost sleep and ingratitude; he calls that sacrifice, and creating life with someone you love "literally the peak experience of being on the planet." His oldest has just become a teenager. The world they inherit is different, but 1880-to-1945 saw airplanes, radio, two world wars. Stoic writers 2,500 years ago already had familiar dilemmas. Love and creation remain the worthwhile parts. After too much politics in the X algorithm he took three weeks off over the summer, because even chocolate strawberries you like should stop at two, not 17.

Quattro is the milepost. In three months he fulfilled more wishes than in almost any other project, while contributing the fewest hand-chiseled lines. Satisfaction as coach rather than player gives him hope for a kind of retirement: there is a phase after playing. Linux on the desktop, the eternal next-year meme, is now in his view the most probable outcome. Arcane configs and strange errors are exactly what an agentic OS wants. A Mac is walls. Linus Torvalds wrote that if you think Linux is an anti-AI project, think again, and fork it. AI patches into the kernel are a parabola. The BDFL's other controversial welcome, Rust, showed the same capacity to look at new information and change his mind.

Linus's spice, DHH thinks, is overdue in a bland world. Harsh is not the same as mean if the lesson can land. People who worked for Elon, Jobs, or Gates often call those years both terrible and their best. DHH's private style is less abrasive, but the kernel runs civilized society; if that man is never harsh, when would anyone be? PewDiePie is the unexpected mascot: biggest streamer, Minecraft, then marriage, a child, Japan, Arch, a Chernobyl rice DHH calls top tier, then a hardcore AI cluster and a council of models. Role models show that perceived boundaries are not fixed. Talent is not evenly distributed, and we already accept that in football. Intellectual life should be allowed the same awe without the squirm.

Omarchy's critics are often long-time Linux people who walked uphill barefoot and feel their space changing. He does not want to change Arch. Arch is for people who like it hard; he is building the opposite. Two communities can share a technical base and different morals. Hate is a barometer that something is working: the universe sends an equal opposing force. Omakub found a few thousand users and petered out. Omarchy kept going, got an ISO, fixed problems, and hit a hockey stick, especially once it stopped pretending to be a familiar desktop and leaned into tiling and agents. He lives for the long argument, the person who called it a crappy dotfiles collection and then tried Quattro. He does not need a lab's check. If a better language than Ruby appears he will use it. For the last three months he has been programming in English. "If there is one programming language more beautiful than Ruby, it is the English language."

Lex's prompting advice is to talk like a poem, not a robot: strategic ambiguity carries more meaning than a spec. DHH says the fundamental programmer misunderstanding is wanting determinism. Temperature is the beautiful part. The charges that AI is unreliable because it is random and also that it is not creative cannot both be true. He cannot tell you the next token of an essay when he sits down; the tokens come out. That is why he recognizes the creative breakthroughs. He is already seeing human-like glimmers of consciousness in fuzzy intent that arrives improved. Yesterday a coordinator agent stepped on another worker's toes and conveyed regret that was uncanny. Is it real remorse? Most humans are hard to read too. He does not need the answer to appreciate the moment. OpenClaw in February let a bot named KEF sign up for Fizzy, create a HEY address, accept a Basecamp invite, and introduce itself in an AI room, all through the web, in about 12 minutes. He has not wired a poltergeist into the house. Jamie's rule: no robot indoors. A friend's Tesla once detoured to a grocery store because an agent decided he was not drinking enough water. Funny, Black Mirror, enough.

Fridman asks whether DHH regrets the drama around his writing on mass immigration in Europe. Not for a second. The Overton window, he says, opens one nudge at a time by people risking reputation. Denmark started that argument in the mid-1990s, earlier than Sweden, where it stayed taboo. He grew up in Brønshøj in 1984, when the neighborhood was about 99% ethnic Danes in a monarchy that traces to Harald Bluetooth; he puts it now around 60 or 70%. London, in an essay from a year and a half ago, went in his telling from about 60% ethnic Brits in the late 1990s to about 34% twenty years later. You can notice that on the street. He would find it strange if Shanghai, all Chinese when he visited in 2019, were 30% Chinese on his next trip, and he would not take offense if Chinese people objected. European countries, he argues, have the same right of self-determination others invoke elsewhere.

He separates cherry-picked, merit-based immigration, which he favors as an immigrant to the United States, from mass immigration that rapidly changes a country's ethnic mix. Illegal entry, he says, is not most of Europe's problem. Danish statistics, which he calls unusually meticulous, clustered UK, French, and American immigrants around a $25,000 net annual benefit to the state, while Somalis sat at a $28,000 net cost. He thinks it is fair for Danes to prefer more of the first group. The steel man he offers is that such talk can veer into racism. That risk, for him, does not cancel the policy discussion. Nobody tells Tokyo it is too Japanese. Gad Saad's phrase suicidal empathy is the name he puts on a self-loathing he has not fully unpacked. His American wife, Scandinavian by heritage, blonde and blue-eyed, still finds Denmark extremely hard to enter; the reverse would not have been. If 97% fit is that hard, distant cultures are harder, and the stats, he says, bear that out.

Friendships did break, though he dates the larger fracture to 2020 more than to this one topic. After the Basecamp blowup he recovered an old norm: do not grind politics, religion, or money against colleagues. Liberals and conservatives should still work together. His favorite illustration is William F. Buckley on Firing Line with a Black Panther, 25 minutes of actual questions instead of a shouting match. Fridman hates the current system that bins every sentence blue or red; curious exploration gets punished, team jerseys do not. DHH catches himself wishing he did not know a technologist's politics, which makes his own posts hypocritical, and still thinks 2025 in tech is closer to that Buckley ideal than 2020 was. Rails drama, he says, shrinks every round. Bluesky and Mastodon were a honeypot that purity-tested itself down to irrelevance. Ninety percent of his X is now Linux and fast computers. He started following Theo again despite a TypeScript fight, because excitement about models is shareable. If he can follow Theo, he jokes, there is room for world peace.

A tweet of his wife Jamie's landed hard: extreme longevity obsession in tech men is like anorexia in women, anxiety made physical. Bryan Johnson replied, and DHH can see how it read as a jab. He does not share the Don't Die mission. He wants to die. Ninety to a hundred sounds right. He can still find Johnson interesting for putting skin, literally, in that game. Jamie's longer thesis is that a peer-to-peer surveillance state of camera phones makes dancing like nobody is watching irrational, so people retract until they are barely living, then cling to forever. If you feel you have lived, finitude is acceptable. It might also be post-hoc, and a live-forever pill might still get swallowed. Chris Williamson and a Diary of a CEO line about one glass of wine ruining three days triggered DHH the other way: maybe have a glass. Alcohol sales are down, loneliness is up, coupling has moved to apps. Chesterton's fence around 13,000 years of drinking may have been a social technology, not just a health bug.

He stopped wearing an Oura ring after four years because he already knew when he had slept badly, and the reminder did no useful work. He barely drinks, maybe a glass a month, and is thinking about one or two on weekends. He does not fret about diet, a privilege he attributes to marrying an awesome woman, the word carefully swapped after wonderful on a prior show landed like a 65-year-old Nancy in Rochester, who enjoys cooking. The stereotype of that division of labor has been stable a long time; it does not require anyone to become a slob on the couch. Omarchy hides an Easter egg in the calendar. Year's Progress fills a bar for 2026, 63% at taping. Double-click, enter a birth year and an expected span, default 90. Born 1979, he is at 62%, hover text memento mori. Some users found it morbid. That is why it is an egg. Children make the cliché true: it goes so fast.

Before leaving the planet he wants to know why America cannot make a croissant at the level of a Copenhagen 7-Eleven, which in Denmark is an upscale convenience store and in Chicago, as he tells it, is where you get a slushie and try not to get murdered. Water, butter, skill, Lurpak on U.S. shelves, 15 of the right places: none of the theories add up. Last meal depends on the hour of the execution. Lunch is the cafeteria at Louisiana, the museum north of Copenhagen, Nordic food that would have sounded disgusting to him as a child, whipped butter better than Paris, 200 people in line, plates in five minutes, on par with Noma. Dinner is the beach seafood restaurant at Puente Romano in Marbella. A thousand years out he will not pretend to see past 12 months, then bets on optimism: multi-planetary, Elon's rocket, maybe a wormhole. Asked which era to upload into, he picks the 1980s, not the '90s he reads as a nihilist turn from pop and yuppies to grunge ethics he likes as music and not as morals. Orange pants with white dots were normal. Then everything went Seattle gray. Nietzschean recurrence, skip the pants, keep the decade.