Zvi Mowshowitz, who writes the Don’t Worry About the Vase newsletter on Substack, reads METR’s Anthropic report as proof that the company is running several months ahead of the “AI 2027” scenario. His argument is that Anthropic’s AI is now speeding up the work of building better AI, and that nobody outside is ready for where that leads.
The idea at issue is recursive self-improvement, or RSI: AI systems doing the research that makes the next AI systems better, which then do the research even faster. METR, the evaluation group, estimated that Anthropic’s capabilities progress now runs about twice as fast as in the stretch between early 2024 and late 2025, with most of the gain probably coming from AI doing the work. AI Insiders covered that report on 6 November and Ajeya Cotra’s reading of it yesterday. Mowshowitz adds a different question: if this is already happening, what should anyone do about it?
His own take on the numbers is blunt. He says he made only small updates because he had already made big ones, and he finds it odd that METR’s authors seemed surprised. Ryan Greenblatt of METR told readers he had expected a speedup of only 1.15 to 1.3 times this year. He now estimates, from public modeling, that AI is speeding up Anthropic’s software research by about 4.5 times, and he notes that the AI 2027 scenario reaches that level around March 2027. Mowshowitz also points to the system card for Opus 5.5, published about six weeks ago, which already put the figure at 1.5 times or more.
He does not treat the result as settled. Toby Ord argues that allowing for a one-time jump from the Mythos pretraining run drops the slope change to about 1.4 times. Peter Wildeford suggests Anthropic may simply have been catching up on reinforcement learning, though he still believes AI accounts for at least a 1.4 times speedup and trusts METR’s access to private data. Mowshowitz answers that the same trick can be played against any future jump, so a forecast that deletes outliers is not much of a forecast. He also cites Buck Shlegeris, who says the public graph was only the part of METR’s evidence that Anthropic allowed out.
That redaction is where his sharpest complaints land. METR says Anthropic blacked out most of the evidence it found most informative, and a planned review of how Anthropic splits compute between capabilities and safety was postponed. Mowshowitz notes that Anthropic itself has called compute one of the most verifiable inputs to AI research and a candidate lever for pacing the frontier. If the study never resumes, he writes, it is hard not to read that as Anthropic not wanting compute limits taken seriously. Anthropic cofounder Sam McCandlish defends the redactions as protection against handing competitors a roadmap, and says a report with Anthropic’s own analysis is due in the coming weeks.
Mowshowitz’s list of preparations is short and political. He backs Kelsey Piper’s call for lawmakers to get the full unredacted report and to bring in both believers and skeptics of RSI to check it. He backs the push, from Representative Greg Casar and Alex Bores, to ban RSI, but warns that a ban needs a legal definition that survives clever workarounds. A narrow rule against fully handing research to AI with no way to monitor it, he says, would just produce a process that is 99 percent automated and only looks stoppable. He also notes that neither OpenAI nor Anthropic mentions internal-deployment RSI risk in the compliance frameworks they filed under California’s SB 53 and New York’s RAISE Act.
On security, he quotes Cotra: very strong security, for the weights and for the algorithmic secrets alike, is a minimum before an intelligence explosion next year. His own estimate is that full RSI is not the favorite for 2027, though he would no longer be surprised by it.
McCandlish’s promised report will put Anthropic’s own numbers beside METR’s 1.9 times figure for the first time, and a gap between the two would say more about the redactions than any analyst’s thread.
Reported by Zvi Mowshowitz in Don’t Worry About the Vase (thezvi.substack.com), whose post carries no publication date.
