KI-Pioneer Sutton calls synthetic data a "big mistake" in the face of an infinitely complex world
decoder ·
AI researcher Richard Sutton, a Turing Award winner and pioneer of reinforcement learning, has criticized the use of synthetic data to overcome the limits of large language models, calling it a "big mistake." Sutton argues that the real world is infinitely complex and far larger than anything captured on the internet, making any simulation of it tiny and inadequate, particularly for modeling other agents' minds. He also highlights a human bottleneck, noting that expert judgment is required to distinguish good synthetic data from bad, which prevents the approach from scaling. As an alternative, Sutton advocates for agents learning from their own experience by building and continually correcting their own world models, rather than relying on static simulations built by humans. He further criticizes current language models for ceasing to learn after training, calling for true continual learning without catastrophic forgetting. Sutton acknowledges LLMs as an amazing breakthrough but considers them only about a quarter of true intelligence.
decoder