RL Pioneer Sutton: Synthetic Data Is a ‘Big Mistake’

Richard Sutton — a Turing Award winner and the author of the best-known textbook on reinforcement learning — called synthetic data "a big mistake." The world is infinitely more complex than any simulation, he says, and expert control over the quality of synthetic data creates a new human ceiling — exactly what AI is trying to get away from. Sutton rates today's language models as "an amazing breakthrough," but only 20–25% of real intelligence.

Synthetic data became the industry's main answer to running out of human-written text in 2025–2026 — OpenAI, Anthropic, and Google are betting heavily on it. If Sutton is right, this whole scalable pipeline is built on a foundation that's limited by nature: the real breakthrough, in my view, would come from agents that learn from live experience continuously, not from frozen snapshots of reality.