AI now leads about a quarter of the work of building Anthropic’s AI. A report from the company sorts its development work by how involved AI is, from work AI barely touches, to work it does alongside researchers, to work it leads by itself, and charts how quickly that mix changed over the past year. Last August, AI was neither leading nor driving much of the development. This August, AI was collaborating with the researchers on it, and leading 26% of it on its own.
The work has a name across the industry: recursive self-improvement, the effort to build AI that can improve AI. AI companies are concentrating on it right now, and the pace is starting to increase rapidly. In practice it runs as a loop. A model helps build the next model, the next model is better at building, and each turn shortens the one after it. The part of Anthropic’s chart that grows as the loop speeds up is the share of development that AI leads by itself.
That share is what has unsettled people who follow the field. The more of its own development AI leads, the faster AI is expected to improve, so the size of that share says something about how fast every model that follows will arrive. At Anthropic it went from under 1% in February to 26% in August. The measure is a prototype, and a Claude model grades the work, which the company notes could repeat the errors of the model it is checking.
The company’s finances reflect the same change. Its filing to go public shows the company spending more and more on computing power. In the third quarter alone it bought $5 billion of computing power on short-term deals, which cost significantly more than multi-year contracts, and it has about $895 billion committed to computing and infrastructure in the years ahead.
Most of that computing does not serve customers. Over 70% of Anthropic’s computing is allocated to research and development, the work of building better models, rather than to running the models customers pay for. Jack Clark, one of the company’s co-founders, now thinks full self-improvement is more likely to become possible in 2027, a year earlier than he expected before. The computing budget is already weighted toward getting there.
