Tag
Rich Sutton argues that generative AI trained by supervised learning cannot achieve genuine novelty and quality simultaneously, and that true discovery requires a 'vary, evaluate, select' mechanism found in reinforcement learning rather than pure imitation.