AI pioneer says generative AI can't make real scientific discoveries

Richard Sutton, a Turing Award winner, says regular generative AI can't make real scientific discoveries because it lacks self-evaluation and the ability to filter results. Sutton is a computer science professor at the University of Alberta and a key figure in reinforcement learning.

He argues that real discovery needs three steps: variation, evaluation, and selective retention. Generative AI is missing the evaluation part. Language and image models can create many variations, but without testing, they can't pick the best ones. He points to systems like AlphaGo and AlphaFold, which have evaluation loops beyond just generating text or images.

Sutton criticizes the AI industry for focusing too much on bigger language models. He prefers AI agents that can interact with environments over time, learn from experience, build world models, and plan strategies.