Cambridge report: AI-written code hits 80% at Anthropic as RSI could speed progress 10x
量子位 · wechat · 2026-10-03
A new report from the University of Cambridge's AI Science and Policy Programme (CASP), co-signed by over twenty researchers including OpenAI chief scientist Jakub Pachocki, Anthropic co-founder Jack Clark, Microsoft CSO Eric Horvitz, Geoffrey Hinton, Yoshua Bengio and Dawn Song, examines recursive self-improvement (RSI): whether progress accelerates as AI takes over AI R&D itself.
The first link of the loop is already documented: at Anthropic, AI-generated code that gets approved rose from single digits in early 2025 to over 80% by May 2026, and R&D work done autonomously by AI with only high-level human oversight jumped from 1% in March to 26% in August. OpenAI and Google confirm AI is deeply involved in training, evaluation and research for future models. Extrapolating current trends, by mid-2028 AI could handle projects requiring months of human expert effort.
The report frames RSI in three stages: assistant tool → part of the R&D pipeline (one person orchestrating hundreds of Agents, multiplying effective researcher headcount) → true recursion, where AI-built models re-enter the R&D loop. Under idealized assumptions, progress could speed up 10x within 18 months of full R&D automation; OpenAI's current inference compute alone could theoretically support millions of AI researchers. However, the report concedes evidence for an inevitable intelligence explosion is far from sufficient: duplicated Agent work, sequential bottlenecks, experiment compute, training data and real-world lab cycles all limit how much code can compress the loop.
More from AGI Musings
- If LLMs become the universal interface, who indexes the MCP services? — ThomasBuildLab · 2026-10-23
- Bindu Reddy: bearish on superintelligence until robotics breaks the digital loop — bindureddy · 2026-10-23
- AI layoffs arrive: Barclays cuts 5,000, eBay 9%, Amex 3,500, Netflix ~1,000 — MFordFuture · 2026-10-23
- 632 tests find no reliable link between AI search visibility and branded or direct traffic — lilyraynyc · 2026-10-23
- Random programs reliably evolve into self-replicators: why 'life is computation' may reshape how you think — paulnovosad · 2026-10-23
- Every generation feared its tech revolution would end the world — this one is no different — MoonL88537 · 2026-10-23