AI R&D paper warns of an intelligence explosion

The Cambridge Programme on AI Science & Policy says AI systems completed 26% of Anthropic’s internal AI R&D with only high-level supervision in August, up from 1% five months earlier, in its intelligence-explosion working paper. The authors treat recursive acceleration as possible rather than established, and call for measurement of R&D automation, constraints on capability speed and preserved shutdown options. For labs and platform teams, OpenAI’s reported 3.1 agent-workday ratio is the useful adjacent baseline: count automated work and control surfaces before calling faster research productivity.