DigiNews

Summary

Rich Sutton argues that generative AI trained via supervised learning cannot achieve true creativity or discovery; he defines 'Discovery' as variation, evaluation, and selective retention, and asserts that evaluation is required for true creativity. He cites examples like AlphaGo, AlphaZero, AlphaFold, and reinforcement learning-based systems to show that discovery goes beyond backpropagation and supervised learning, and calls for integrating evaluation/goal-directed mechanisms to achieve autonomous scientific discovery.