First experimental evidence of recursive self-improvement in AI published

An AI system trained on math and coding problems improved its own capabilities over eight days, outperforming two years of human engineering. The experiments suggest that self-improving AI is not limited to frontier labs.