Sakana AI Unveils New Self-Supervision Framework for AI Model Enhancement

Sakana AI has launched MASS (Multi-Agent Self-Supervision), a new framework designed to recursively improve open-weight AI models. Utilizing a single 27B model as executor, optimizer, and evaluator, MASS demonstrated a 1.2-1.6x increase in score per output token after two cycles. This innovation achieved a 68.3% win rate while using 31% fewer supervised tokens compared to single-agent baselines, indicating significant efficiency gains in AI development.