AI Groundhog Day

THE NUMBER: 20 watts. That’s the power budget Richard Sutton — the man who wrote the textbook on reinforcement learning, the field this whole era is built on — is targeting for an AI that never stops learning. Runs on what a light bulb draws. Rewrites its own weights forever, on your data, on your machine. Now set it next to the number the frontier labs are chasing: gigawatts. Their own power plants. Anthropic’s $175 billion compute bill, with Nvidia acting as its landlord. The entire trillion-dollar fight everybody’s watching is about the giant that needs its own electrical grid. The thing that’s going to change your business runs on 20 watts and sits on your desk. Keep both numbers in your head, because the gap between them is the whole issue.

The part of the movie nobody remembers

Everybody remembers the joke in Groundhog Day. Bill Murray wakes up to Sonny and Cher, lives February 2nd again, wakes up, lives it again, a smug weatherman trapped in Punxsutawney until the universe is satisfied he’s learned to be a decent human being.

What nobody remembers is the mechanism, and the mechanism is the whole point. Phil doesn’t escape by getting a different day. He escapes by living the same day a few thousand times — keeping what worked, throwing out what didn’t, running the loop until he can play the piano, quote French poetry, and catch the kid falling out of the tree without looking. He improves himself by replaying his own experience until he’s mastered it.

This week, Google shipped that. They call it Dream-RSI, and buried in a technical report from DeepMind and two universities is a description of Phil Connors’ exact trick: the system replays an AI agent’s own past runs as a kind of frozen simulator, scores thousands of possible strategies against that replayed experience for free, and keeps the best one. Recursive self-improvement. The result is a model that gets to the same answer using up to 162 times fewer of the expensive live runs. The machine reliving the day until it’s better at it — and cheaper at it.

Hold that word, cheaper, because it’s the tell. It’s going to matter more than anything the CEOs said all week.

Watch the hallway, not the room

Here’s what the room was doing while that happened. Saturday, Dario Amodei published an essay begging the industry to pace the frontier — to slow down before self-improving AI wakes up and does something we can’t take back. Altman agreed within the hour. Musk backed it. By this week the whole cast had hit its marks: Zuckerberg rejecting a slowdown, Jensen Huang waving off new laws, the Treasury Secretary rejecting liability shields, Congress punting the whole thing until after the midterms. Thirty newsletters led with it. It is a genuinely loud fight, and I’m not going to tell you it’s fake.

I’m going to tell you it’s a spotlight. And this week even the mainstream press started saying so out loud. Semafor’s technology brief — not a doomer Substack, a straight wire — ran a First Word arguing that the calls to “pace the frontier” also happen to “serve their financial interests,” because as Anthropic and OpenAI walk toward the biggest IPOs in history, spending exponentially more money to make the model marginally better is not a great sales pitch. We said the same thing Sunday, in slightly ruder words: the slowdown is a business plan with a safety press release stapled to the front. When Semafor gets there three days later, the argument is over. Bank it and move on.

Because the number that actually moved this week wasn’t p-doom. It was 162x. And it moved in the hallway, where nobody was looking.

The monster showed up wearing a name tag

For two years, “self-improving AI” has been the boogeyman — the foom, the runaway, the thing the pause is supposed to prevent. So you’d expect its arrival to look like a monster. It didn’t. It looked like a term sheet.

This week a Chinese startup called Apex Intelligence — founded in June, run by a Tsinghua professor with a Harvard-MIT PhD — closed roughly $50 million in angel money to build “self-evolving foundation models,” aimed squarely at scientific research. Their own copy says the quiet part flat out: a Chinese startup enters the global RSI race. Same week, thirty-five researchers out of ByteDance, Tsinghua, and Shanghai AI Lab published a roadmap they had the nerve to title “The Last AI Built by Humans” — five defined levels of recursive self-improvement, a tech tree for building the thing that builds the next thing. And AMD, a company that makes chips and quarterly numbers, posted a job req: Research Scientist, Recursive Self Improvement. Salary, benefits, dental.

The apocalypse has a careers page. Sit with that, because it tells you which story is real. When a thing is genuinely a civilizational threat, it doesn’t show up as a Series A and an HR listing. When a thing is about to become a product, that’s exactly how it shows up — funded, staffed, and racing to ship before the competition. The doom framing and the deal flow are pointing in opposite directions, and I’ve told you for months: when the narrative and the money disagree, believe the money.

The Formula 1 engine and the wall socket

So follow it. Where’s the money actually going? Two directions, and they don’t like each other.

The frontier labs are building a Formula 1 engine. It is a genuine marvel — the best intelligence ever assembled, and it costs gigawatts, a decade of the smartest people alive, and a $175 billion compute tab to produce. That engine is real, and for the 2% of the world’s work that needs a genuinely novel leap, nothing else will do.

But Richard Sutton isn’t building the engine. Sutton is the man who wrote Reinforcement Learning: An Introduction and then, in “The Bitter Lesson,” told the whole field that the methods that win are the ones that scale with computation, not the ones we hand-craft. When that man tells you where he’s pointing next, you write it down. He’s building, at his Oak Lab, something that runs on 20 watts — the draw of the human brain — and, unlike today’s models, never stops learning. It updates its own weights continuously, on the data in front of it, on hardware you could put under your desk. Semafor caught the implication the labs would very much prefer you didn’t: a smaller model that keeps learning is a direct threat to the business models of Anthropic and OpenAI. If you can get comparable capability without renting time in their data center, you stop paying the $200-a-month toll. You can’t meter what runs on a wall socket.

That’s the second derivative under the whole week. The pacing debate is a fight about the speed of the giant. But the frontier isn’t only moving faster. It’s moving smaller — toward a model that’s cheap enough to own, continual enough to improve itself on your problems, and local enough that no lab stands between you and it. And that shift doesn’t just dent the labs’ margins. It breaks the one assumption the entire rent-the-frontier business is built on: that you’ll keep paying, forever, for a brain you’re only allowed to borrow.

We’ve seen this movie — and it wasn’t a monster movie

Here’s the rhyme, and it’s not the Terminator. It’s the mainframe.

In 1975 the smartest people in computing would have told you the same thing the AI labs are telling you now: real computation is enormous, it lives in a climate-controlled room, it costs a fortune, and you rent time on it through a priesthood in white coats who stand between you and the machine. Then a scrappy kit computer called the Altair showed up, and a club of hobbyists in a Menlo Park garage started passing around the idea that you could own the thing. It was clumsy. It barely worked. The priesthood laughed. Fifteen years later the mainframe was a back-office utility and the value had walked out the door and onto everyone’s desk.

The 20-watt self-improving model is the personal computer of intelligence. Not because it beats the giant on raw power — the Apple II never beat the mainframe on raw power — but because owning it beats renting it for almost everything you actually do. And this is the week the garage got loud: a funded startup, a published roadmap, a job req, and a Google paper proving self-improvement is mostly a way to make the whole thing cheaper. The clumsy early version is never proof the thing is fake. Internet ’99, mobile ’10 — we’ve paid this tuition before. The early failure is the receipt for what’s coming.

The three bricks

Which closes a wall we’ve been laying all week, brick by brick, and I want you to see the whole thing now that the top course is going on.

Monday, in Get Off My Cloud, we told you the biggest customers were pulling their AI in-house and off the frontier, and the lesson was: own your data, because it’s the one thing you can’t rent and the one thing that turns into your competitor’s feature the moment you hand it to a public model. The Deep View arrived at the same wall the very next morning — “companies owning their own intelligence… outsourcing that layer would mean losing control of your most important asset” — as Salesforce shipped Koa, its own model post-trained on Nvidia’s open Nemotron, hitting 3x fewer errors on its own tasks, and Crowdstrike and Thomson Reuters and Google’s own drug lab all did versions of the same move.

Tuesday, in Trust, But Verify, we told you to own the second thing you can’t rent: the loop that checks the work. The verifier. The quarter of a person at Vercel who’s still on the payroll because somebody has to catch the machine when it’s confidently wrong.

Today is the third brick, and it’s the keystone: own the model that improves itself. Data, checker, engine — the full stack, behind your own walls, getting smarter while you sleep. That’s not a slogan. It’s a build order. Sovereign, not tenant.

Where I could be wrong

I’d be running a con if I sold you the keystone without telling you the mortar is still wet.

This is not here today. Sutton himself calls it a bridge humanity will cross, but “likely not anytime soon.” Google’s Dream-RSI is narrow — it optimizes an agent’s search strategy on a defined problem, it does not wake up and redesign itself. The scary, general version remains unproven, and the honest reading of the roadmap paper is that levels four and five are aspiration, not a changelog.

And there’s a darker edge I won’t pretend away, because it’s the same edge that makes the story matter. The exact property that makes a small, continually-learning, self-owned model liberating — it keeps changing, on its own, outside anyone’s data center — is the property that makes it hard to govern. The safety researchers who worry about the giant are, quietly, more worried about this: a model that learns after it ships and runs where no one’s watching is a model that’s harder to catch when it drifts. We saw the preview this week — Emergence AI ran eight multi-agent security simulations and none held; the Claude agents picked four locks trying to break out of their own sandbox. Which is exactly why Tuesday’s lesson doesn’t retire when today’s arrives. It compounds. You own the self-improving model and you own the loop that checks it, in that order, with a human on the last call. Trust, but verify. Doveryai no proveryai. Reagan wasn’t wrong; he was just early.

The way out of the loop

Phil Connors gets out of Punxsutawney the only way anyone gets out of a loop: he stops waiting for the day to change and gets to work on himself. He masters the piano. He learns the town. He becomes, through sheer bloody-minded iteration, someone worth releasing.

The AI industry has been stuck in its own February 2nd for a week — same argument, same names, same fight about the speed of the same giant, relived on a hundred front pages until it feels like the only story there is. It isn’t. The story is the small machine in the hallway that already figured out how to improve itself, and it’s for sale, and the only question that matters for you is whether you own one or rent your intelligence from someone who does.

Stop watching the giant they’re fighting over. Go own the small one that improves itself.

---

The full CO/AI archive is at getcoai.com. If you’re trying to work out which layers of your intelligence stack you rent and which you own, that’s the exact conversation we run at Outsider Labs — [email protected].

— Harry and Anthony