Why Sutton Says Generative AI Cannot Make Novel Discoveries

Richard Sutton, co‑inventor of temporal‑difference learning and the man behind the bitter lesson, just put a pin in the hype balloon: generative AI built on supervised learning is constitutionally incapable of making a discovery that is both novel and good at the same time.

Sutton frames the argument with an old researcher joke — the review that says “the parts that are good are not novel, and the parts that are novel are not good.” That quip, he argues, describes the current state of generative AI exactly.

The Joke That Describes Generative AI

Generative AI — LLMs, image models, video generators, even world‑model learners — works by ingesting billions of examples and outputting a statistical mimic of that data. When you ask it for a fact, you want fidelity to the training corpus; any deviation is a hallucination, and you don’t want that. When you ask for fiction or a bedtime story, you rely on stochastic sampling to produce novelty, but the quality comes from how closely the output hews to the source material. Sutton’s conclusion: the output trajectory is either novel or good, based on randomness or data, but never both simultaneously.

That’s fine for entertainment, summarization, or code completion. Sutton is clear: “Generative AI is meant to be a mimic. … It is okay if Generative AI cannot be both novel and good at the same time. It is still a transformative technology.”

Why Mimicry Fails for Science

Science and mathematics don’t get that pass. For a discovery to matter, it must be both new and correct — novel and good in the same output. A model that can only interpolate within its training distribution will never propose a theorem or a chemical synthesis that genuinely extends human knowledge. Sutton calls this a “devastating” limitation for anyone trying to use vanilla generative AI for research.

Sutton contrasts these systems with the ones that do produce real novelty: AlphaGo’s move 37 in the game against Lee Sedol, AlphaZero’s emergent chess style, GT‑Sophie’s racing lines, AlphaFold, AlphaProof, Claude‑Code, and RL‑Lyft for ride‑hail optimization. All of those combine a generative component with a second mechanism — reinforcement learning, search, or a reward signal — that lets them try many candidates and select for both quality and novelty.

Systems That Actually Discover

“Discovery is just the idea of trying many things and seeing which ones are both novel and good,” Sutton says. That’s what supervised learning alone cannot do. AlphaGo didn’t learn move 37 from human games; it played millions of self‑play games and found an unexpected move that beat the best human player. That’s the pattern: generate diverse candidates, score them against an objective, and keep what works.

Sutton notes that some language models have been augmented with these extra mechanisms — tool use, search, self‑play — and that those are the ones showing hints of real creativity. The distinction matters because researchers funding the next generation of AI for science need to stop expecting breakthroughs from pure supervised learning and start building systems that explicitly separate generation from evaluation.

The hard takeaway: generative AI as currently deployed is a brilliant mimic, but a mimic cannot discover. The path to AI‑driven science runs through architecture that includes the ability to try, fail, and learn from the outcome — exactly the loop Sutton has championed for decades.

Source: Rich Sutton on AI creativity and discovery
Domain: twitter.com