Anthropic warns that it’s very close to RSI capability, aka self-building AI

Anthropic said in a new report that it is approaching recursive self-improvement (RSI) capability, the point at which AI systems can effectively train and improve themselves, triggering a rapid acceleration in AI progress. The company, which develops the Claude family of models, framed RSI as a threshold with major implications for AI safety and governance, and its warning comes as competition with rivals such as OpenAI and Google DeepMind intensifies. The announcement matters because RSI is viewed by AI-safety researchers, including prominent "doomers," as a potential point of no return, raising the stakes for frontier-lab commitments on oversight and control. Anthropic did not provide a specific timeline or technical benchmarks for when RSI would be reached. Investors and policymakers should watch for Anthropic's upcoming safety commitments, any follow-up research quantifying how close its models are to self-improvement, and whether competitors issue similar warnings or accelerate capability releases.