AI 安全
METR Reviews Evidence of Anthropic R&D Acceleration, Says Twofold Warning Threshold Remains Open to Interpretation
In an external review published November 5, METR said Anthropic’s public conclusion that R&D acceleration had not reached its warning threshold was too definitive. Disputes concern the acceleration metric, counterfactual baseline, and evaluation saturation; some key evidence remains unpublished.
On November 5, METR published a review of Anthropic’s August risk report, concluding that as of July 15, AI-driven R&D acceleration may have reached the company’s twofold warning threshold. Most relevant researchers at METR and Anthropic still lean toward the view that the threshold had not been crossed at the time, but the review said the public report did not adequately convey the uncertainty behind that judgment. Both sides still agree that the degree to which models directly posed a catastrophic risk by automating R&D was low at the time. METR review
How the threshold is defined matters. Anthropic’s policy measures the rate of overall AI capability progress: achieving in one year the progress that would otherwise have taken two, with a substantial portion of the acceleration reasonably attributable to research or engineering automation. Saving engineers half their coding time cannot be directly translated into a doubling of the rate of capability progress; changes in headcount, compute, and algorithmic efficiency must also be accounted for separately. Anthropic’s original report, p. 13
Differences in measurement are at the heart of the dispute. Using figures from a private version of the report, METR estimated that one leading indicator had a central estimate of about 1.97 times the historical rate. Under different reasonable definitions and time periods, estimates ranged from 1.3 to 2.8 times. These figures alone cannot establish AI’s causal contribution, but they suggest that describing the rate as “below twofold” may obscure how close it was to the threshold. Anthropic, meanwhile, acknowledged that its most concrete task evaluations had saturated, making it difficult to keep detecting capability growth. METR review, original report
For research teams, the review underscores the need to track task efficiency, model capability, and end-to-end R&D outcomes together, while clearly stating the comparison baseline for a scenario without AI assistance. This is a practical implication drawn from the report. Going forward, attention should turn to whether unsaturated evaluations and measurement methods can be made public. In an overall assessment published the same day, METR said that most of the most informative evidence had been redacted, making it difficult for outsiders to independently reconstruct its conclusions. METR overall assessment