The Bitter Lesson

Richard Sutton's influential essay contends that the history of AI research reveals a consistent pattern: approaches that leverage general-purpose learning algorithms and massive computational resources ultimately outperform those relying on human-engineered features and domain-specific knowledge. Sutton reflects on decades of AI progress, from game-playing systems to language models, showing that the most successful breakthroughs came from scaling simple methods rather than incorporating expert knowledge.

The essay challenges the AI research community's tendency to invest in hand-crafted solutions and argues for a shift toward learning-based approaches that can exploit exponential growth in computation. Sutton emphasizes that this "bitter lesson" has been learned repeatedly—yet researchers often ignore it, returning to domain-specific engineering when general methods plateau temporarily.

Published in 2019, the essay became a touchstone in debates about AI methodology, influencing discussions around deep learning, reinforcement learning, and the role of inductive biases in modern AI systems. It resonates particularly with practitioners building large-scale models and has shaped thinking about research priorities in the field.