METR recorded a two-fold acceleration in AI development at Anthropic

The non-profit organization METR has completed an 11-week audit of internal research data from Anthropic. The audit showed that the pace of creating new models has approximately doubled compared to previous periods.

According to the report, about 80% of this speed increase is driven by research conducted with the help of AI systems themselves. Auditors estimate that the internal acceleration warning threshold was crossed back in July.

The audit faced limitations: the company itself redacted the most informative evidence in the documents. Because of this, independent observers have to rely heavily on the expert assessments of METR analysts, and skeptical views regarding these metrics persist within the company itself.

The findings demonstrate the practical effect of recursively applying AI to create subsequent generations of models. This changes the approach to assessing the pace of technological progress in leading AI labs.