Sutton says synthetic data is a 'big mistake' for scaling AI
reinforcement learning pioneer richard sutton claims synthetic data bottlenecks ai scaling because simulations are 'microscopic' compared to the infinite complexity of the real world.
richard sutton, author of the influential 'bitter lesson' essay, criticized the reliance on synthetic data to solve llm scaling bottlenecks, calling it a 'big mistake' that re-introduces human bottlenecks. citing the 'big world hypothesis,' sutton explained that because the world is infinitely complex, any simulation remains 'microscopic' and fails to accurately represent reality or the minds of other agents. instead of relying on human-curated simulations, sutton advocates for agents that learn continually from their own experience to build dynamic world models. he concludes that while large language models are a scientific breakthrough, they currently represent only '20% or a quarter of intelligence' because their weights remain frozen after training.