METR Finds Evidence of Recursive Improvement

On November 5, METR, a nonprofit that evaluates frontier AI for catastrophic risk, published an 11-week review of internal Anthropic data on whether AI is accelerating Anthropic’s internal AI development. The review concluded that Anthropic is likely improving its models on well-specified tasks at about 2x the rate it did from March 2024 to October 2025, meaning it doubled its pace of progress in the past year relative to that period. METR attributed roughly 80% of this speedup to Anthropic’s internal use of AI, although Anthropic redacted most of the informative evidence that METR used. However, Anthropic co-founder Sam McCandlish affirmed that “many of their [METR’s] conclusions match my own.”

Recursive self-improvement (RSI) refers to the concept of AI contributing to its own improvement, creating a feedback loop and accelerating progress at an increasing rate. METR report suggests that loop has started, although it remains human-in-the-loop, since AI mostly assists Anthropic’s researchers rather than working on its own. METR found that no compute, data, or algorithmic bottleneck is likely to stop the speedup from compounding over the next 12 months, and said it “cannot rule out a substantial further acceleration” within the next six months. Anthropic’s amended IPO filing, made three days before the report, says it “may imminently meet” its safety policy’s threshold for “dramatically accelerated progress attributable to AI.”