Sakana AI's self-improving model gets better at solving and judging, but can't catch its own blind spots

via The Decoder · 10 views · 4d ago

What if AI could make itself smarter? Sakana AI is doing just that with its new MASS method. This approach lets a language model improve itself. It works by solving tasks, then judging its own results, and refining its internal workflows. After two rounds, the system achieved a nearly 70 percent win rate against its base model. That's a significant leap in capability, showing impressive self-correction. However, there's a critical catch: the model struggled to identify its own recurring errors. If it missed a problem initially, it often kept missing the same type of error in subsequent attempts. This highlights a fundamental challenge for truly autonomous AI. While self-improving models are powerful, they still need human oversight to spot and correct those stubborn blind spots, preventing them from repeating the same mistakes indefinitely.

← Back to the feed