Richard Sutton Says Chatbots Are Not the Endgame: The AI Future After LLMs
Turing Award winner Richard Sutton is pushing a very different AI path: agents that learn from experience, play, abstraction and continual adaptation. Is the chatbot era already temporary?
Richard Sutton helped teach machines to learn. Now he is warning that the current AI boom may still be missing the deepest path to intelligence.
Sutton, a Turing Award winner and one of the central figures in reinforcement learning, has long argued that the most powerful AI systems will not be built merely by copying human knowledge into static models. They will learn from experience. They will play, explore, form abstractions, create subgoals, build world models and improve through continual interaction with reality.
That sounds technical. It is also a direct challenge to the chatbot era. Large language models have changed the world. They write, code, summarize, reason, search, design and increasingly act through tools. But they are still limited by a training paradigm that is largely front-loaded. The model absorbs vast amounts of data, is trained, aligned, deployed, and then updated in controlled cycles. It may use memory and tools, but it does not yet learn continuously from life the way animals and humans do.
Sutton’s vision points somewhere else: intelligence as ongoing adaptation. In recent talks, including at MIT’s Dertouzos lecture, Sutton has outlined architectures built around agents that can create features, pose subproblems, learn solutions, model transitions, and plan using those models. That is a very different picture from a chatbot answering prompts. It is closer to a self-improving organism inside a machine.
Why does this matter now? Because the AI industry is obsessed with scaling, benchmarks and productization. Bigger context windows. Better coding scores. More agentic workflows. More expensive data centers. More tokens. More APIs. These are important, but they may still be incremental steps inside one paradigm.
Sutton is asking whether the next paradigm requires AI systems that learn on the job. If he is right, today’s LLMs are not the destination. They are a bridge. They gave machines language, broad knowledge and tool use. But they may not be enough for true autonomy, science, robotics, strategy or long-horizon intelligence. A model that cannot reliably learn from its own experience without retraining is still missing something fundamental.
This has geopolitical consequences. The U.S.-China AI race is often framed around frontier model performance. Who has the best chatbot? Who has the largest compute cluster? Who leads coding benchmarks? But if the next breakthrough comes from reinforcement learning, continual learning, robotics, self-play or world models, leadership could shift quickly.
There is also a safety question. A chatbot that answers questions badly is one kind of risk. An agent that learns continuously, forms plans and pursues goals is another. If AI moves from passive assistant to adaptive actor, alignment becomes harder. The headline says Sutton reveals what comes after chatbots. The answer is not one product. It is a direction: self-taught, experience-driven intelligence.