Don't Worry About the Vase Podcast On Dwarkesh Patel's Podcast With Richard Sutton
Sep 29, 2025
Richard Sutton, a pioneer in AI research best known for his work on reinforcement learning, dives deep into the limitations of large language models (LLMs). He argues that LLMs mimic human behavior without true understanding and emphasizes the importance of goal-directed learning. The conversation explores themes such as continual learning, imitation versus experience, and the implications of Sutton's Bitter Lesson for future AI developments. Sutton also discusses the potential risks of merging experiences between AI copies and the need for steering AI towards positive outcomes.
AI Snips
Chapters
Transcript
Episode notes
Mimicry ≠ World Understanding
- Richard Sutton argues LLMs only mimic human tokens and thus lack true action-driven world understanding.
- He claims intelligence requires learning from real-world experience, not just next-token prediction.
Experience Versus Logged Examples
- Sutton contrasts continual experiential learning with LLM training on observed human behavior examples.
- He allows LLMs could use external memory or programs, but current forms lack continual learning from an agent's own life.
Incorporate Continual Learning
- Use continual learning and online feedback to build systems that adapt during interactions.
- Don’t rely solely on static next-token objectives if you want systems to update from real responses.