Richard Sutton, the Turing Award-winning reinforcement learning pioneer, has renewed a sharp critique of the current large-model boom: a system trained mainly to predict the next word cannot, by itself, produce real science.
His objection is not that large language models are useless. It is narrower and more fundamental. Scientific discovery requires causal understanding, active experimentation and the ability to test hypotheses against the world, not only recombine what humans have already written.
The data ceiling
Sutton's point echoes his earlier discussions with David Silver: human data has a ceiling. Models trained on existing papers, text and images can become powerful assistants, but they struggle to discover what no human has yet encoded in data.
He points to systems such as AlphaGo and AlphaProof as a different path. They include an internal evaluator and can improve through interaction, feedback and search rather than simply imitating human outputs. That loop looks closer to hypothesis, experiment and revision.
A critique of direction, not usefulness
The argument lands because most capital in AI is still pushing bigger models and more human data. Sutton, author of The Bitter Lesson, is not rejecting scale; he is questioning what should be scaled. For original science, he argues, AI needs its own experience, not only the accumulated record of ours.
The test will not be a debate panel. It will be whether the next genuinely new scientific result comes from a model that has read every paper, or from an agent that can act, measure, fail and revise inside an environment.
Sources: The Decoder, CocoLoop; checked Sutton's critique of next-word prediction, causal understanding, experimentation, human-data limits and the comparison with AlphaGo and AlphaProof.