The Bitter Lesson

Rich Sutton 的苦涩教训：通用计算方法总是胜出，人类领域知识内建短期有效但长期阻碍进展

源

SOURCE SUMMARY · BITTER LESSON · wiki synthesis

The Bitter Lesson — Summary · Link to Agent Engineering

Core thesis · two scalable method families · direct guidance for harness engineering

70 years of AI research keep confirming: methods that leverage general computation ultimately win by a large margin. Researchers repeatedly try to bake human domain knowledge into the system — effective and satisfying short-term, stagnating or actively obstructing long-term. Breakthroughs come from the opposite approach: scaling search and learning.

Two indefinitely scalable method families

chess · Go · MCTS · deep search

Learning

self-play · deep learning · scaling laws

Key aphorism

The contents of minds are irreducibly complex — don’t try to model space, objects, symmetries. Build in meta-methods that can find and capture complexity.

Direct guidance for agent engineering

Managed Agentscites Bitter Lesson to argue harness assumptions go stale — “what the model can’t do” baked into a harness is exactly the domain knowledge Sutton warns against

Harnessingthe three patterns (use known tools · keep stripping assumptions · set boundaries cautiously) directly practice the Bitter Lesson

What can I stop?the central reflection axis of harness engineering

→ bitter-lesson · search-and-learning · harness-engineering · managed-agentswiki summary

The Bitter Lesson

来源: sources/sutton-bitter-lesson.md
URL: http://www.incompleteideas.net/IncIdeas/BitterLesson.html
作者: Rich Sutton
发布: 2019-03-13

摘要

Rich Sutton 的标志性短文。核心论点：70 年 AI 研究的最大教训是，利用通用计算的方法最终总是以巨大优势胜出。研究者反复犯同样的错误——试图将人类领域知识内建到系统中。这在短期内有效且令研究者满足，但长期来看会停滞甚至阻碍进展。突破最终来自于相反的方法——通过搜索和学习来扩展计算。

关键要点

模式反复出现：国际象棋（1997 深度搜索胜 Kasparov）、围棋（AlphaGo 自我对弈学习）、语音识别（HMM/深度学习胜手工特征）、计算机视觉（CNN 胜手工边缘检测）——每个领域都经历了同样的”人类知识方法短期有效但最终被计算方法压倒”的循环
两类可无限扩展的方法：搜索和学习。只有这两类方法能随计算资源增长持续提升
不要内建发现，要内建发现过程：心智内容是不可还原地复杂的，不应试图建模。应建入能找到和捕捉复杂性的元方法

与其他 source 的关联

Managed Agents 直接引用 Bitter Lesson 来论证 harness 假设会过时——harness 中内建的”模型做不到什么”的知识就是 Sutton 批评的”领域知识内建”
Harness Engineering 中”What Can I Stop Doing?”原则是 Bitter Lesson 在 agent 工程中的具体体现
Harnessing Claude’s Intelligence 中三大模式（用已知工具、持续剥离假设、谨慎设边界）直接实践 Bitter Lesson

References

sources/sutton-bitter-lesson.md