Sakana AI's self-improving model reaches 70% win rate but has blind spots

Sakana AI's MASS method lets a single language model improve itself by solving tasks, judging results, and refining workflows, reaching a nearly 70% win rate against the base model after two rounds. However, if the model makes the same errors across all three roles, it can mistake them for progress and reinforce them during training.

• Sakana AI Peer Review System Catches 73% of Core-Claim Errors2026-10-10
• Sakana AI Hires Deep Learning Pioneer Schmidhuber2026-09-24