AI Security Agency METR: The progress speed of the Anthropic model is about twice that of a year ago

According to The Verge, the AI safety research organization METR stated that after an 11-week investigation into Anthropic, the company's model improvement rate is about twice that of a year ago, with progress made in the past year equivalent to approximately two years prior. METR believes that as of October, most of the acceleration in research and development is due to AI, and its contribution is increasing.

Currently, the acceleration mainly comes from AI-assisted human researchers rather than autonomous work. Employees using Anthropic's strongest internal model reported in an informal survey that their work efficiency is on average about 12 times that of when not assisted by AI. Anthropic has heavily redacted the public report, with co-founder Sam McCandlish stating that the related information involves commercially sensitive content and could also help other companies accelerate their research and development.