Turing Award winner Richard Sutton says pure generative AI can't do real science

Sutton isn't just critiquing generative AI in the abstract. He's implicitly drawing a line between systems that can be evaluated against ground truth (math proofs, game outcomes) and those that cannot, which means his argument applies narrowly to domains with verifiable feedback and says much less about the vast majority of enterprise use cases where 'correct' is contested or contextual.

This connects directly to Hugging Face's recent piece 'Beyond LLMs: Why Scalable Enterprise AI Adoption Depends on Agent Logic,' which argued that the bottleneck in production AI has already shifted from raw model quality to reliable decision-making under uncertainty. Sutton's framing gives that argument a theoretical spine: agents matter not because they're architecturally fashionable, but because evaluation loops are what separate refinement from mere generation. The Nvidia GTC Taipei coverage also reinforces this, since Cosmos 3 and the robotics stack are explicitly built around world-model feedback rather than open-ended generation. Sutton's thesis, read alongside those stories, suggests the industry's serious capability bets are already quietly moving in the direction he's prescribing.

Watch whether AlphaProof's successor or any announced lab project explicitly cites evaluation architecture as a design requirement in the next six months. If that framing starts appearing in technical roadmaps from frontier labs, Sutton's argument is shaping investment priorities, not just conference talks.