LLMs are stuck as mimics, says Sutton, but AlphaGo is real science
turing award winner richard sutton argues pure generative ai lacks the internal evaluation loop necessary for true scientific discovery.
reinforcement learning pioneer richard sutton critiques pure generative ai, stating it cannot perform true science because it lacks the ability to evaluate and select its own novel outputs. while models like alphago or alphafold succeed via objective feedback loops, llms rely on human or external verification, meaning their novelty is often just hallucination. sutton argues that combining generation with automated evaluation—formal proofs, code tests, or game wins—is the only path to genuine ai-driven discovery.