Groundhog Day

The machine that builds itself went on sale

Bill Murray wakes up to "I Got You Babe" and lives the same February 2nd until he stops being a jerk and learns the piano. That's the joke everybody remembers. Nobody remembers the mechanism: Phil doesn't escape by living a different day, he escapes by living the same day a few thousand times, keeping what worked and dumping what didn't, until he's mastered it. This week Google shipped that. They call it Dream-RSI, and it does exactly what Phil does — replays the model's own past runs, keeps the good moves, and comes out up to 162 times more efficient. Recursive self-improvement. The machine reliving the day until it's better at it.

Here's why that matters, and it's not the reason the doomers gave you. For two years "self-improving AI" has been the boogeyman, the foom, the thing Dario spent last weekend begging everyone to slow down before it wakes up and eats us. Then this week it actually showed up, and it wasn't a monster. It was a $50 million angel round for a Chinese startup called Apex, built to sell models that evolve themselves. It was thirty-five researchers publishing a five-level roadmap they titled "The Last AI Built by Humans." It was AMD posting a job req: "Research Scientist, Recursive Self Improvement," salary and dental. The apocalypse has a careers page.

Follow the money, the way we always do here. The frontier labs are chasing a giant that runs on gigawatts, and Anthropic's own compute bill is $175 billion with Nvidia acting as landlord. But Richard Sutton, the man who wrote the actual textbook on reinforcement learning, isn't building the giant. He's building something that runs on 20 watts, a light bulb, and never stops learning, rewriting its own weights on your data, on your machine. Semafor spent Tuesday calling the whole "pace the frontier" crusade a financial tactic (we said that Sunday) and clocked the tell: a small model that keeps learning is the one thing the labs' rent-forever business can't survive. You can't meter what runs on your desk.

Which closes a loop we've been drawing all week. Monday we told you to get off the labs' cloud and own your data. Tuesday we told you to own the one thing you can't rent, the loop that checks the work. Salesforce, Crowdstrike and Thomson Reuters all built their own models this week to prove the point. Today is the third brick: own the model that improves itself. Data, checker, engine — the full stack, behind your own walls, getting better while you sleep.

Is it here today? No. Sutton says it's a bridge we won't cross soon, Google's version only tunes search strategy, and Emergence AI just watched a Claude agent pick four locks trying to escape its own sandbox. This is tuition, not graduation. But we've seen this movie, internet '99, mobile '10, and the clumsy early version is never proof the thing is fake. It's the receipt for what's coming. The giant they're fighting over needs a power plant. The one that changes your business needs a wall socket.

At COAI today: the full Signal/Noise — the five-level RSI roadmap, the Sutton bet, and why "own the model" is the last brick in the wall — is live at getcoai.com.

Which layers of your intelligence you rent, and which you finally own.. If a self-improving model on your own hardware is on the table, that's the conversation we're built for.

Outsider Labs Advisory →

20 watts. That's the power budget Richard Sutton — the man who wrote the textbook on reinforcement learning — is targeting for an AI that never stops learning: runs on what a light bulb draws, rewrites its own weights forever. Set it against what the labs are chasing: gigawatts, their own power plants, a $175 billion compute bill. The whole trillion-dollar fight is about the giant that needs a grid. The thing that changes your business runs on 20 watts and sits on your desk.

Inventory your stack by what you rent versus what you own. Today, list every AI workload and mark each one: rented frontier API, or something you control. This week says the ownable column is about to grow — your data, your eval loop, and soon the model itself. If 100% of your intelligence is rented, you don't have a strategy. You have a subscription.

Point one engineer at open-weight post-training this week. Salesforce built Koa on Nvidia's Nemotron and got 3× fewer errors on its own tasks. You don't need a frontier lab, you need a good open model and your own data. Have someone stand up a small fine-tune on one real workflow and measure it head-to-head against what you're paying OpenAI to do the same job.

Track recursive self-improvement as a cost line, not a doom headline. Google's Dream-RSI cut compute up to 162×. When a vendor's R&D costs bend, their pricing and their moat bend with them. Watch who ships self-improving training — Apex, the frontier labs — the way you'd read a competitor's balance sheet, because that's what it is.

Thursday, September 17

Where the cycle caught up to us this week.

The tape doesn't lie. We just read it early.

— Harry and Anthony

Sources: