METR: Anthropic AI now drives 80% of its own research speedup

The METR findings land against a decade of dueling predictions. The author recounts siding for years with skeptics such as Ed Zitron, whose forecasts Dan Luu reviewed and found wrong in 25 of 26 cases, including a 2024 call that generative AI had "peaked" and a 2025 claim that Google's 500 million Gemini user target was unrealistic — Gemini reached 750 million users a year later. On the other side, Daniel Kokotajlo's 2021 essay "What 2026 Looks Like" anticipated AI agents browsing the web and hundreds of millions of people talking to chatbots regularly, and both came to pass.

The new data point sharpens that divide. METR's conclusion that Anthropic is improving its models about twice as fast as before, with roughly 80 percent of the speedup coming from the AI itself, is the kind of internal, checkable acceleration the author argues the wider economy has not felt. The body notes that AI improves fastest where answers can be checked — mathematics and code — and that the public models, even at a $200-a-month subscription, lag the much more powerful unreleased ones.

The safety picture around that acceleration is, by the author's account, deteriorating. Anthropic dropped its pause promise in February and, when outside investigators came to measure its acceleration, blacked out most of the evidence. The company has committed nearly 900 billion dollars to more data centers, and it plans to go public this month at a valuation of more than 1 trillion dollars — a prospectus that both promises to "bring forth an age of miracles" and warns the technology may doom humanity. Public opinion, meanwhile, is moving the other way: two-thirds of Americans say AI is moving too fast, and in last week's midterms voters picked the anti-data-center option in 40 of 41 places where data centers were on the ballot.