Jack Clark, co-founder of Anthropic, has updated his forecast for the arrival of full recursive self-improvement (RSI) in AI systems. He now anticipates this capability will emerge in 2027, shifting his previous estimate from late 2028.
Data from August 2026 indicates that AI agents based on the Claude model led 26% of core AI research and development tasks. Over 90% of these tasks achieved a collaborative level of involvement, designated as AL3 or higher.
Currently, approximately 30 000 AI agents operate concurrently at Anthropic with low override rates reported by safety monitors. The Vals AI RSI Index projects that full automation of machine learning engineering at the company may occur between July and August 2027.
In September 2026, Anthropic CEO Dario Amodei emphasized the need for paced development of frontier AI models. He highlighted that the acceleration of risks associated with RSI necessitates coordinated industry and government efforts.
