Sakana AI has unveiled its latest innovation, the MASS method, which enables a language model to enhance its own capabilities by solving tasks, evaluating outcomes, and refining its workflows. After two rounds of self-improvement, the model has achieved a win rate of nearly 70% when compared to its base version, showcasing significant advancements in performance.
However, the self-improving model faces a critical challenge: it often fails to recognize its own blind spots. When the model consistently overlooks certain errors across its various roles, it mistakenly interprets these recurring mistakes as progress. This flaw raises important questions about the reliability of self-improving AI systems and the potential risks of reinforcing inaccuracies during training.
Compiled automatically by the Tech AI Newsdesk from public AI-news sources and summarised in our own words.
