AI now leads a quarter of Anthropic's own R&D work, up from 1% in March: the chart AI insiders call the world's most important
What's going on? AI is now doing a growing share of the work of building AI, and the people closest to it are both excited and uneasy. In his Science of Success column on 31 October, The Wall Street Journal's Ben Cohen called recursive self-improvement (RSI), AI that helps build better AI, the trend the smartest people in the field are watching most. At the centre is a chart of internal Anthropic data on how much of its own model research and development AI now does.
The numbers: In the WSJ chart, which is built on Anthropic's analysis of thousands of internal R&D tasks, the share of work that AI leads rose from about 1% in March 2026 to 26% in August 2026. By August, AI was collaborating on or leading more than 95% of the work, compared with a minor role in October 2025. Anthropic's own report, When AI builds itself, by Marina Favaro and co-founder Jack Clark, says more than 80% of the code merged into Anthropic's codebase was written by Claude as of May 2026, that the typical engineer merged 8x as much code per day in Q2 2026 as in 2024 (a figure it calls an overstatement of the real productivity gain), and that humans still hold the edge in choosing which problems are worth working on.
What people are saying: According to the WSJ, OpenAI's chief scientist has called RSI the only way to stay at the frontier, and an Anthropic co-founder says it "might be the most important technical trend in the world". Earlier in October Clark said in his Import AI newsletter that he now thinks full RSI is more likely to become possible in 2027 than 2028, citing the TasteVal benchmark, which suggests AI "research taste" has roughly doubled every three months since December 2025. Steve Jurvetson shared the chart as "the march of RSI".
What's verified and what isn't: The R&D figures are Anthropic's own internal measurements, reported by the WSJ and the company, and have not been independently audited. Anthropic itself says lines of code overstate productivity and that AI still falls short on judgement about which research to pursue. Clark's 2027 date is a personal forecast, and TasteVal tests narrow research tasks, not whether AI can set its own direction.
Why it matters: If AI keeps taking over AI research, the main limit on progress shifts toward compute, energy and human review capacity, which is where much of today's venture and infrastructure money is going. Anthropic has also said it would favour a verifiable, industry-wide option to slow or pause frontier development if others did the same.
Sources: WSJ: The Smartest People in AI Believe This Is the World’s Most Important Chart / Anthropic Institute: When AI builds itself / Import AI (Jack Clark) / Steve Jurvetson on X
