Definition
Plain language
Rich Sutton's essay arguing that general learning methods that scale with compute always beat hand-crafted human cleverness.
As stated in the literature
Sutton's 2019 essay observing that across AI history, learning-based methods leveraging compute have repeatedly outperformed approaches relying on human-engineered priors and domain knowledge.
Why it matters: It's the philosophical backdrop behind the entire modern scaling movement and a reminder to bet on general methods rather than narrow tricks.
For example, Sutton points to chess and Go, where decades of clever handcrafted heuristics were eventually swept aside by search and learning with more compute.
Heard on the show
“Sutton called this the Bitter Lesson — durable progress comes from general learnable methods that scale with compute, not from elaborate human-authored priors.”Episode 060 — When Splitting One Model Across Three Agents Doubles Its Accuracy