According to CoinMeta, Anthropic recently warned in its IPO prospectus that AI may soon cross the safety threshold for self-accelerated development. An independent evaluation agency, METR, published survey results indicating that this warning line could have been reached as early as July of this year. METR spent 11 weeks investigating Anthropic's internal development data, and researchers estimate that as of October of this year, Anthropic's model's ability to progress on tasks with clear scoring criteria has doubled compared to before. METR found that the growth rate of training computing power has not accelerated accordingly; the current acceleration is likely mainly due to Claude's participation in AI's development. METR also re-examined the security report released by Anthropic in August of this year, concluding that its public report was too conservative in describing the degree of acceleration and that it cannot be ruled out that the warning threshold was reached as early as July.