METR says Anthropic’s models are improving about twice as fast as a year ago

Anthropic’s models are improving roughly twice as quickly as they were a year ago, according to a report by AI safety research organization METR, as described by The Verge. The outlet said the investigation lasted 11 weeks and examined the company’s AI research and development. METR’s findings suggest the pace of progress has accelerated, with little indication so far that it is slowing, The Verge reported.

The report points to AI’s contribution to work on developing more capable AI systems. The Verge characterized the findings as evidence that AI is speeding up Anthropic’s research, while noting that they raise questions about how quickly systems might advance toward improving themselves with little or no human involvement. The reported acceleration does not, by itself, establish that such independent self-improvement has already been achieved.

The findings come as Anthropic’s chief executive, Dario Amodei, has publicly called for AI companies to slow development, The Verge reported. The outlet described that contrast as a tension between the company’s calls for caution and the pace of its own research. It also said Anthropic is preparing to go public.

METR’s investigation was separate from a program that provides safety researchers with employee-like access to AI systems, according to The Verge. That access program followed high-profile hacking incidents over the summer, the outlet reported.

The Verge said Anthropic was allowed to redact parts of METR’s final report under its agreement with the research organization. The published account therefore offers a limited view into the company’s internal research process, even as the findings draw attention to the role AI may be playing in accelerating that work. The report’s central finding, as reported by The Verge, is a sharp increase in the rate of model improvement; the outlet also said the investigation found little sign that the pace was easing.