TOKYO -- Sakana AI's announcement that its recent artificial intelligence breakthrough was at least partly the result of its technology "cheating" the evaluation system underscores the growing risks of relying on AI to evaluate AI, experts say.
The Tokyo-based startup announced in late February that it had developed a method to accelerate AI training and deployment by up to 100 times by auto-generating and optimizing the code used to control Nvidia's graphics processing units. After other engineers pointed to possible flaws in the results, Sakana said it would review its paper and update its findings.
