Sakana AI's Self-Improving Model Boosts Performance but Struggles to Catch Its Own Blind Spots
Sakana AI's new MASS method enables language models to autonomously improve by solving tasks, judging results, and refining workflows, achieving a 70 percent win rate against base models. However, the system risks reinforcing errors when it fails to detect blind spots across all three operational roles during the self-training process.
Source: The Decoder