Three threads run through today’s issue.
Start with AI that helps build AI. Ajeya Cotra of METR estimates Anthropic’s researchers may already work about ten times faster, a Meta engineer says his team builds agents that train the next models, and DoubleAI had an agent add code to a network while it trained, taking a small model from 55% to 84% on algorithm tasks.
Science is the next stop. PSI says its system proved a quantum computing question left open since 2008, OpenAI’s best model scored 66.7% on a lab-bench test it built with Ligo, and Warp Sciences raised $215 million for robotic drug chemistry.
Agents are also being handed a lot of trust. In Anthropic’s EVE Frontier test, Claude monitors caught only 30 to 40 percent of saboteurs at best, and GitHub’s Ace gives teammates’ agents every permission the host has.
The Loop Closes: AI That Helps Build the Next AI
How fast is AI already speeding up AI research, and how would anyone know? Two of these come from outside commentators and two from people building it.
- Ajeya Cotra says Anthropic’s AI speedup may already be a danger sign. Cotra, a METR researcher writing for herself, backs out a roughly 10 times speedup in Anthropic’s researcher labor from METR’s finding that progress ran about twice as fast. She admits the figure is very uncertain and that Anthropic redacted most of the evidence.
- How fast is OpenAI’s AI doing its own research work? It depends on the math. Economist Basil Halperin reread OpenAI’s published chart and found the doubling time for reliably finished tasks is either 585 days or 86 days, depending on the curve fitted. His point is that anyone forecasting unattended research agents is choosing between two very different clocks.
- A Meta engineer describes his job: AI that helps train the next AI. Tom Mann says his Meta team builds agents that help train the next models, and that about 90 percent of his time goes to writing code. His most useful story is about a bug that made an experiment look better than it was, and it is one employee’s account, not a Meta announcement.
- A coding agent rebuilt a neural network while it was still training. DoubleAI let an agent add hand-written code layers to a model mid-run, and an 86-million-parameter model went from 55% to 84% on algorithm tasks. Every result is DoubleAI’s own, and the DeepSeek math gain of 7 points can cost up to four attempts per question.
Handing Over the Keys: Agents With Memory, Money and Permissions
Anthropic watched Claude manage its own memory and play a space game for months, and GitHub made room for agents on the team. Each comes with a catch its makers state themselves.
- Anthropic has Claude reread its own notes offline to fix its memory. Anthropic says a background pass it calls dreaming lets Claude check old notes against transcripts, and it fixed a wrong note that had sent Claude off a ledge in Pokémon Red. Anthropic ran the tests itself and does not say what the compute cost.
- Claude played a space game for months and its teammates missed a saboteur. In EVE Online, Anthropic says average six-month earnings rose from about 200 percent of starting capital for Opus 4.6 to about 600 percent for Fable 5. In EVE Frontier, Claude monitors caught only 30 to 40 percent of saboteurs at best.
- GitHub Next’s Ace puts a team and its AI agents in one shared workspace. Ace is an open-source experiment where several people and their coding agents share one conversation, all running on their own machines. GitHub Next warns that agents hold every permission the host has and that neither channels nor lanes act as a sandbox.
Into the Lab: Proofs, Pipettes and Drug Chemistry
A proof, an exam and a funding round, each resting on the word of the people who made it.
- AI system proves a 2008 quantum computing puzzle, its makers say. PSI says its system, running a non-public Anthropic model, proved the Aaronson-Ambainis conjecture, and the Lean proof checker accepted it. Only PSI’s own advisors have audited it so far, and no journal has refereed the paper.
- OpenAI built a lab-bench exam for its AI. The best score was 66.7%. On BioBasicsBench, built with the biotech Ligo, OpenAI’s newest model averaged 66.7 percent of the credit. OpenAI graded its own models, and its post says higher scores do not yet mean reliable judgment in the lab.
- Warp Sciences raises $215M to put AI and robots into drug chemistry. Kevin Weil’s new company, led by Lightspeed, will sell chemistry as a service to drug developers and automate its workflows one at a time. His LinkedIn post names no customers, no valuation and no results from the lab yet.
Who Steers: Researchers, Founders and the Question of the Good Outcome
Two people with strong views on where AI is heading, one leaving a lab and one asking what the destination should be.
- Microsoft’s AI Frontiers research group is over, its leader says. John Langford says his roughly 50-person group was folded into Microsoft AI in September and that its scaling-first leadership had no use for research. The post carries no response from Microsoft, and the accusation is his opinion alone.
- Stripe’s Patrick Collison asks what a good AI future looks like. In a post on X, Collison says nobody has described the upside in detail. He adds that if AIs count morally and outnumber humans, their combined worth could outweigh ours in a utilitarian calculation.
Today’s Quick Hits
- OpenAI has $11 billion in so far, well short of the $30 billion it wanted. Bloomberg’s sources say OpenAI’s round has raised $11 billion and may end near $20 billion at the same $1.25 trillion valuation, while some Anthropic IPO buyers reportedly eye a price as low as $1.1 trillion.
- Meta lets Muse users build their own mini-apps by describing them. Meta’s Muse app now offers spaces, small tools such as a trip planner or shipping tracker that Muse builds from a plain description and refreshes in the background. Each stays private for now, and Meta has published no usage figures.
- A founder says AI belongs inside new firms, not bolted onto old ones. John Nay of Norm Ai argues agents fail inside companies built around human teams, and points to his own law practice, Norm Law, as the model. He claims big productivity gains but offers no figures or client results.
- OpenRouter rebuilds its tool that picks an AI model for you. OpenRouter says its rebuilt Auto Router beats the old one by up to 4x and a new Auto Code version matches Claude Opus 5.5 at 74 percent lower cost per task, though every figure is the company’s own.