What to know

• The AI industry has promised recursive self-improvement—where AI systems autonomously enhance their own training—but evidence suggests the timeline is more distant than marketed.
• AI systems can write code and optimize hardware, but lack the creativity for open-ended AI research needed to drive genuine recursive improvement.
• New research benchmarks test whether AI agents can design training algorithms, the fundamental capability on which recursive self-improvement depends.

“Recursive self-improvement (RSI) asks whether an AI system can improve the process that produces AI systems, so that the next system inherits the improvement.”

Yizhe Chi et al., AI researchers · arXiv · Aug 20

Michelle Kim MIT Technology Review reporterThomas Macaulay MIT Technology Review writerYizhe Chi et al. AI researchers