Anthropic warns that it's close to RSI capability, a.k.a. self-building AI

Anthropic’s updated securities filing warns that using AI systems to speed up its own research could make model capabilities advance faster than the company can assess and secure them. The company says it has previously failed to recognize a significant acceleration in development until after it happened. Its internal projections suggest it may soon meet its Responsible Scaling Policy’s threshold for “dramatically accelerated progress attributable to AI,” a condition associated with recursive self-improvement: AI helping build more capable successors, potentially driving rapid advances.

Recursive self-improvement is seen both as a route to major gains in productivity and scientific discovery and as a safety risk if increasingly capable systems outpace human oversight. Anthropic stresses that no current model has crossed its threshold and says it does not expect to let Claude autonomously design and build successive generations of frontier models without meaningful human supervision or understanding. It is considering policy changes to preserve human control. The filing does not establish that self-building AI has arrived; rather, it signals that Anthropic believes the relevant acceleration may be near and acknowledges that it could be difficult to detect in time. Because the statements appear in an investor filing that also serves a promotional role, their framing warrants scrutiny, but the company’s reported concerns raise questions about how AI development should be monitored and governed.