Richard Sutton, the Turing Award winner who literally wrote the book on reinforcement learning, isn’t buying the synthetic data hype. In a recent conversation introducing his new company Oak Lab — founded with former student Khurram Javeed — Sutton dismissed the approach outright: “No, that’s just a big mistake.”
His reasoning hinges on what the Alberta group calls the “Big World Hypothesis.” The argument is straightforward: the real world is infinitely complex, and any simulation of it is microscopic by comparison. A small program can only produce a small world, one with wrong friction values or an inaccurate model of a robot’s motor behavior. Sutton also points to a problem that’s easy to overlook — the world contains other agents whose inner workings simply can’t be generated as synthetic data. “There’s no way we can have synthetic data for other people’s minds.”
The second objection cuts deeper. Even if synthetic data were useful, deciding which synthetic data is good requires human experts. Javeed’s argument: you’d need domain specialists to evaluate every dataset, which means the approach is bottlenecked by humans and doesn’t scale. Want to train a drone that moves like a bat using echolocation? Hire bat experts first. Even self-driving car companies, which have poured years into simulation, still need humans to fix the gap between simulation and reality.
Sutton’s alternative is characteristically blunt: take humans out of the loop entirely. Agents should learn their own world models and keep correcting them — “simulators they make themselves” — rather than relying on frozen models built by people. He’s equally critical of how today’s language models stop learning after training. “Their weights never change.” Real learning, he argues, is continual by definition, and catastrophic forgetting is a solvable problem — his team published a method called Continual Backprop in Nature. His verdict on LLMs? An amazing scientific breakthrough, but only “like 20% or a quarter of intelligence.” Coming from the author of “The Bitter Lesson,” the message is consistent with a career-long argument: anything that relies on built-in human knowledge eventually hits a ceiling.
