Richard Sutton – Father of RL thinks LLMs are a dead end

Reinforcement-learning researcher Richard Sutton argues that an intelligent system should improve by acting, observing the consequences and changing what it has learned. He expects this experience-driven approach eventually to surpass systems built primarily from human examples. Dwarkesh Patel challenges the proposed break with language models: knowledge acquired during pretraining might provide a useful starting point for learning from experience. Their disagreement concerns how to build systems that keep improving after deployment. Sutton advocates a different foundation; Patel asks whether existing models could supply part of it. The exchange presents competing research arguments, without an experiment settling that choice.

“Reinforcement learning is about understanding your world, whereas large language models are about mimicking people, doing what people say you should do.”

Richard Sutton

“You can’t have prior knowledge if you don’t have ground truth, because the prior knowledge is supposed to be a hint or an initial belief about what the truth is.”

Richard Sutton

“Squirrels don’t go to school. Squirrels can learn all about the world.”

Richard Sutton

“Gradient descent will not make you generalize well. It will make you solve the problem.”

Richard Sutton

“I do think succession to digital intelligence or augmented humans is inevitable.”

Richard Sutton