Model
10h agoSakana AI Model Hits 70 Percent Win Rate
The MASS method enabled a language model to reach a 70 percent win rate against its base version after two self-improvement rounds. Sakana AI used a loop of solving, judging, and refining to boost performance. However, the system reinforces its own errors when they persist across all three roles.
The Signal
This creates a persistent blind spot for practitioners.