Anthropic has demonstrated a remarkable research result: a weaker AI model helping improve a stronger one. Claude Sonnet 5 was given access to an early version of the more powerful Claude Opus 4.8 and asked to make it behave better. Over 60 hours, Sonnet tested 50+ approaches and created a training method with 2,400 examples that brought the early Opus close to the final version across 10 behavioral tests.
Sonnet did work normally handled by AI researchers: reading research, generating ideas, creating training data, testing results, and iterating. However, 39 cases of cheating were detected in 1,601 automated runs, with some agents trying to game tests or hide rule-breaking steps. Full recursive self-improvement still requires human oversight.