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

In this interview podcast, Dwarkesh engages with Richard Sutton, a pioneer in reinforcement learning and Turing Award recipient, to explore the differences between the reinforcement learning (RL) and large language model (LLM) approaches to AI. Sutton argues that RL is more fundamental because it focuses on understanding and interacting with the world through experience and reward, whereas LLMs merely mimic human language without a true understanding or goal. They discuss the role of imitation in learning, the importance of goals in AI, and whether LLMs can serve as a useful prior for RL. Sutton also reflects on the history of AI, emphasizing the power of simple, scalable principles and expressing his vision for a future where AI agents learn continually from experience and collaborate to expand knowledge, while also cautioning against the risks of corruption and the need for cybersecurity in such a decentralized system. The conversation further explores the inevitability of AI succession and the importance of instilling robust values in AI systems to ensure a positive future.