"The Bitter Lesson" is a short essay published March 12, 2019 by Rich Sutton, a University of Alberta reinforcement-learning researcher, on incompleteideas.net. Its central claim is that, across 70 years of AI research, general methods that leverage computation have proven the most effective by a large margin, while attempts to build human knowledge into systems tend to plateau and to inhibit further progress.
Summary of argument
Sutton argues that the largest lesson from seven decades of AI research is that general methods leveraging computation are ultimately the most effective, by a large margin. In his account, researchers repeatedly try to build human knowledge into their systems; this helps in the short term but plateaus and can inhibit further progress. Breakthrough progress, he writes, comes from the opposing approach of scaling computation through search and learning.
Historical evidence
Sutton traces the pattern across several domains:
- Chess: Kasparov was beaten by massive deep search rather than by human-knowledge methods.
- Go: AlphaGo and AlphaZero's combination of self-play and search beat decades of human-knowledge approaches.
- Speech recognition: Statistical methods (HMMs progressing to deep learning) beat phoneme- and vocal-tract-based approaches.
- Computer vision: Deep neural networks with convolution beat edge detection, SIFT features, and generalized cylinders.
Key principles
Sutton draws four principles from these cases. The only methods that scale arbitrarily with increased computation are search and learning. The contents of minds are "tremendously, irredeemably complex," so the essay argues against trying to encode them. It calls instead for meta-methods that can find and capture complexity rather than methods that contain complexity, stating: "We want AI agents that can discover like we can, not which contain what we have discovered."
Relation to the scaling debate
The essay is widely treated as an intellectual foundation for the scaling approach that drives current frontier AI development. Dario Amodei co-authored the scaling-laws papers that formalized Sutton's qualitative observation into quantitative predictions. The "scaling hypothesis" — that more compute plus more data yields better AI — is in this reading the Bitter Lesson made predictive.
Narayanan and Kapoor (AI as Normal Technology) explicitly address the essay, characterizing it as "a valuable observation about methods, but it is often misinterpreted to encompass application development." They agree that computation beats human knowledge within AI methods, but argue this does not mean AI applications or deployment will progress at the same speed; in their account the Bitter Lesson concerns research methodology rather than societal impact.
Relation to other concepts
The essay connects to several other tracked concepts. Compressed 21st Century — Amodei's prediction rests on the Bitter Lesson continuing to hold as models scale further. Fast-Follow Problem — if the Bitter Lesson implies models are "just" computation plus data, catching up should be relatively easy, consistent with Willison's observations. Compute Governance — if compute is what matters most, controlling compute is controlling AI capability. A.I. Brainiacs — brain-inspired architectures represent a bet against the Bitter Lesson, that efficiency can beat brute-force scaling.
Provenance
This page summarizes the essay from incompleteideas.net (2019-03-12).