The Bitter Lesson

Source: http://incompleteideas.net/IncIdeas/BitterLesson.html Author: Rich Sutton Date: 2025-08-27 (saved), originally published 2019

Summary

Rich Sutton’s famous 2019 essay on the history of AI research. Core thesis: every time AI researchers have embedded human knowledge into systems, it has been eventually outperformed by general methods that leverage compute. The lesson is “bitter” because it invalidates decades of expert knowledge-encoding work.

Key Claims

• Across 70 years of AI research, the same pattern repeats: human-knowledge approaches plateau, then compute-based general methods overtake them.
• Examples: chess (domain-specific heuristics → search + compute), Go (pattern databases → Monte Carlo tree search + RL), speech recognition (phoneme models → deep learning on raw audio).
• The bitter truth: researchers’ instinct to incorporate domain knowledge is wrong — it hinders discovery of better general methods.
• General methods = search and learning, scaled by compute.
• The long-term answer is not to embed human knowledge but to discover how to meta-learn or meta-search effectively.

Connection to Other Sources

This is the ideological foundation for scaling laws. Dwarkesh references it. The “pretraining cliff” concern is implicitly about whether the bitter lesson has a limit.

Entities

• Rich Sutton — RL pioneer; wrote the RL textbook

Concepts

• Scaling & Compute — the bitter lesson predicts compute always wins
• RL Infrastructure — RL + search as the general method