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

Sutton argues (0:55) that LLMs mimic what people say rather than building genuine world models. They predict what a person would say, not what will actually happen in the world. This distinction is fundamental to his critique—mimicking entities that have world models is categorically different from having a world model yourself.

Sutton defines intelligence as the computational ability to achieve goals (4:50), citing John McCarthy. He argues LLMs lack substantive goals—next token prediction doesn't change the world—and without goals there is no definition of right or wrong action, no ground truth, and therefore no foundation for genuine learning or prior knowledge.

Sutton lays out (10:45) a paradigm where intelligence is grounded in the stream of sensation, action, and reward. Knowledge is about that stream—predictions of what will follow what—and can be tested against the stream, enabling continual learning. This contrasts with LLMs, which learn from static training data unavailable during deployment.

While acknowledging LLMs leverage massive computation (7:15), Sutton suggests they also represent heavy use of human knowledge (the Internet corpus). He predicts systems learning purely from experience will eventually supersede LLMs, making this another instance of the bitter lesson where human-knowledge-dependent approaches get overtaken by scalable general methods.