Bitter Lesson

The bitter lesson is the observation, set out by the computer scientist Richard Sutton in a short essay in March 2019, that over seventy years of AI research, broad techniques that make use of ever more computing power – above all search and learning – have beaten approaches that build human knowledge into systems. The lesson is bitter, Sutton argued, because researchers keep investing in clever, hand-designed shortcuts that help at first and are then overtaken.

The argument

Sutton’s examples come from the field’s history. In chess, IBM’s Deep Blue beat Garry Kasparov in 1997 mainly by brute-force search – a letdown for the many researchers who had tried to teach machines how grandmasters think, although Deep Blue also used a hand-tuned evaluation function and an opening book. In Go, AlphaGo and its successors won by combining large-scale search with learning, including self-play, rather than handcrafted strategy. Speech recognition moved from linguistic rules to statistical methods and then to deep learning; computer vision moved from hand-designed features such as edges to convolutional networks trained on data. Because computation keeps getting cheaper, Sutton concluded, methods that scale with it eventually win, and researchers should build methods that find structure on their own instead of hand-coding what they already know.

Influence

The essay became a touchstone of the scaling era. The success of large language models trained on vast data and compute, the scaling laws that describe their predictable improvement and the turn to test-time compute in reasoning models are often cited as vindication. In labs and startups the phrase now works as a warning: do not over-engineer what the next, bigger model may do on its own. In October 2026 the Wharton professor Ethan Mollick posted a music video about the essay that Anthropic’s Claude models and the AI music generator Suno made from a single request (see The Bitter Lesson — a song).

Objections

Critics say the lesson is overstated. Within days, the roboticist Rodney Brooks replied in “A Better Lesson” that human ingenuity still goes into the architectures, data and training setups that make learning work – convolutional networks themselves build in knowledge about images – and that brute-force computation has rising costs as chip improvements slow. Others point out that high-quality data, not only compute, has become a bottleneck; researchers at Epoch AI estimate that the stock of public human-written text could be used up for training within a few years. Sutton himself, who shared the 2024 Turing Award with Andrew Barto for reinforcement learning, argued in 2025 that today’s language models, trained mainly to imitate human text, do not yet fully follow his lesson, and that systems learning from their own experience will. He has also called the lesson an observation about one period of history rather than a law.

Why it matters

The bitter lesson is a lesson in humility – and an optimistic one. It suggests that progress depends less on designers knowing everything in advance than on open-ended systems that learn – an idea that echoes Friedrich Hayek’s argument that decentralized processes such as markets use knowledge no central planner has (see Knowledge Problem). It also explains why access to computation has become a central question of economics and policy.

Written with Claude, made by Anthropic, which trains Claude using the large-scale approach the essay describes; the music video linked above was generated by Claude models.

Related

Scaling Laws · Reinforcement Learning · Test-Time Compute · Emergent Abilities · Knowledge Problem · Compute Governance · AI Winter

Sources: Richard Sutton: The Bitter Lesson (March 13, 2019) · Rodney Brooks: A Better Lesson (March 19, 2019) · ACM: ACM A.M. Turing Award Honors Two Researchers Who Led the Development of Cornerstone AI Technology (March 5, 2025) · Epoch AI: Will we run out of data? Limits of LLM scaling based on human-generated data (2024) · Dwarkesh Patel: Richard Sutton – Father of RL thinks LLMs are a dead end (Dwarkesh Podcast) (September 26, 2025)

Leave a Reply

Your email address will not be published. Required fields are marked *