Richard Sutton – Father of RL thinks LLMs are a dead end
Richard Sutton and Dwarkesh Patel debate goals, continual learning, world models and whether LLMs are a good starting point for experiential intelligence. Selected technical discussion, excluding the closing succession debate. Contested claims remain attributed to Sutton and challenged by Patel.
Executive Summary (TL;DR)
Richard Sutton and Dwarkesh Patel disagree about whether language models are a good foundation for intelligence that learns during use. This summary covers the technical debate, experiential learning, generalization and later knowledge-sharing risks, not the closing succession discussion. Sutton's critiques are his September 2025 views, challenged by Patel rather than treated as settled findings.
Key Takeaways
Prediction of words is not the same question as consequences
Sutton distinguishes predicting what people say from predicting what happens after an action. Patel counters that language models appear to contain useful world models. Their disagreement turns on what counts as prediction and whether unexpected outcomes cause durable changes, not simply whether outputs sound knowledgeable.