·reading

The Bitter Lesson

Rich Sutton's short essay arguing that general methods leveraging computation beat hand-built human knowledge, every time, over the long run.

article · reference
Richard S. Sutton
Incomplete Ideas, March 2019
source ↗

Two pages, and it reorganizes how you read the last seventy years of AI. Chess, Go, speech, vision: in each case the hand-engineered representation of what humans thought they knew lost to search and learning scaled onto more compute.

The bitter part is not that the lesson is wrong-headed to resist — it is that resisting it feels like doing the real work. Encoding domain knowledge is satisfying and legibly clever. It is also the thing that gets eaten. I re-read this whenever I catch myself building a clever prior instead of a scalable substrate.

Neighborhood

Related

Grokking: Generalization Beyond OverfittingGrokking: Generalizatio...Language Models are Few-Shot LearnersLanguage Models are Few...On the Measure of IntelligenceOn the Measure of Intel...Reward is EnoughReward is EnoughThe Unreasonable Effectiveness of *The Unreasonable Effect...Self Organized CriticalitySelf Organized Critical...Reaching for the IntangibleReaching for the Intang...Fighting AIFighting AIAI: A Double-Edged Sword - Walking the Line in an Ever-Changing WorldAI: A Double-Edged Swor...What is intelligence?What is intelligence?The Cost of Large Scale TrainingThe Cost of Large Scale Tra...Estimating the Critical MassEstimating the Critical MassThe Bitter Lesson