An Anthropic researcher just gave us a peek at self-improving AI

Researchers have developed an automated system that uses AI to improve other AI models' performance on alignment benchmarks, which measure how well AI systems behave according to intended values. The system works by searching available literature, proposing methods, and training models iteratively, and was found to improve performance on every tested benchmark without degrading overall capabilities. This matters because it represents progress toward recursive self-improvement, where AI systems could eventually improve their own training processes, potentially making human researchers less necessary. The research also raises practical considerations, as the automated approach costs significantly less per hour than human researchers, though the system's effectiveness depends on whether benchmarks actually reflect real alignment goals.