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

Reinforcement learning pioneer Richard Sutton argues that Large Language Models (LLMs) represent a limited approach to AI because they are designed to mimic human output rather than truly understand the world. He contrasts this with reinforcement learning, where an agent learns from direct experience and feedback to achieve a defined goal, providing a "ground truth" for what constitutes a correct action. Sutton contends that LLMs lack this mechanism for continual, on-the-job learning from real-world interaction, as there is no inherent goal or reward signal in their framework. Ultimately, he suggests that a new architecture enabling continual learning from experience will be necessary, potentially rendering the current LLM paradigm obsolete.