Anthropic Put 6% of AI Research Compute Toward Safety Work

Anthropic’s Sept. 17 analysis put safety work at about 6% of the computing resources it used for artificial intelligence research and development during the week of July 13-20. For compute devoted to AI-driven AI research and development, its estimate was about 12%. Both estimates cover the same week.

Anthropic presented compute allocation as one possible measure of AI development’s pace. Such measures could include reserving some compute for safety work or restricting the share used by AI research agents.

These ideas address how resources are allocated rather than a cap on total company computing capacity. The Sept. 17 analysis did not announce a company-wide compute limit.