Anthropic warns it is close to recursive self-improvement capability

TechCrunch reports that Anthropic's revised S-1/A filing, submitted on a Monday, contains language suggesting the company may be near building a model capable of recursive self-improvement, or RSI. The outlet defines RSI as the point where AI becomes so good at training itself that capabilities accelerate dramatically. Coverage had focused on updated financials, but TechCrunch calls the risk-factor wording a further notable disclosure.

According to the article, the filing says Anthropic uses AI in its own research and development, and warns that doing so could pose a considerable safety risk if progress speeds up sharply. It points to the company's Responsible Scaling Policy, which sets out plans to monitor and mitigate such risk. New statements added to the Risk Factors section say that increasing use of AI in research may let model capabilities advance faster than safety, alignment and security measures can keep up. Anthropic acknowledges it has previously failed to notice a significant acceleration until afterward.

The filing adds that internal estimates and expectations for future models, though not current ones, indicate the company may imminently reach the policy's own threshold for dramatically accelerated progress that is attributable to AI. It states that no current model has been designated as exceeding that threshold and that the underlying data is still being assessed.

TechCrunch also notes the company's statement that it does not expect to safely use Claude to design successive frontier models without meaningful human supervision, and that it is weighing changes to that policy so that people stay in charge. The article observes that an S-1 serves as both disclosure and marketing, yet concludes that rapid capability gains this year, especially in coding, math and cybersecurity, mean the claim deserves to be taken seriously.