資料來源#
摘要#
Rich Sutton 在 2019 年的文章：善用運算（搜尋、學習）的通用方法，最終會勝過那些把人類知識與手工打造結構內建進去的方法——而且隨著運算量成長，差距會拉得很大。「苦澀」之處在於：這一再讓那些投入於精巧領域結構的研究者感到意外，因為那些結構成了天花板，而非地基。
這個頁面之所以存在，是因為這條原則在整個 wiki 中反覆作為承重的論據出現——被明確援引，用來證成把 harness 溶入模型的做法。
它在這裡被援引之處#
- Interaction Models——TML 直接引用「the bitter lesson」：手工打造的互動系統（VAD、turn-detection、dialog-management harness）「將被通用能力的進展所超越」，因此「若要讓互動性隨智能一起擴展，它必須成為模型本身的一部分」。見 Turn-Based Interface Bottleneck。
- Encoder-Free Early Fusion——在單一 transformer 中從零開始共同訓練所有模態元件，而不是把預訓練好的編碼器／解碼器拼接起來：手工打造的模組邊界更少。
- Time-Aligned Micro-Turns——移除人為的回合邊界，讓各種互動模式成為可擴展的模型行為，而不是每種模式各寫一套 harness 程式碼。
- Harness Shrinkage as Models Improve——把同樣的邏輯套用到 coding-agent 的 harness：prompt 鷹架彌補模型尚未能做到的部分，並應隨著模型進步而縮減。（那裡的但書：機械式驗證——測試、型別、linter——才是不會往內遷移的部分。）
- Agent Harness Engineering——「強制不變式，而非實作」：讓模型自己找出路徑；harness 只編碼那些必須為真的東西。
標準的但書#
The bitter lesson 講的是能力與結構往模型內部遷移，而不是「harness 沒有用」。有些東西理所當然地留在模型之外：機械式驗證（Harness Shrinkage as Models Improve 的綜論）、組織專屬的政策／風格、安全邊界，以及——根據 Claude Character as Product——刻意打造的性格／人格工作。對每個 harness 元件而言，懸而未決的問題都是：它落在那條線的哪一側。
相關連結#
- Evolutionary Proof Search——那套量身打造的演化裝置，正是 the bitter lesson 預測會被吸收掉的東西
- Interaction Models——近期最明確的一次援引
- Turn-Based Interface Bottleneck——「較不聰明的 harness 終究輸給規模擴展」
- Harness Shrinkage as Models Improve——coding-agent 版本，附帶機械式驗證的但書
- Agent Harness Engineering——以「不變式而非實作」作為一條意識到 bitter lesson 的設計準則
- Encoder-Free Early Fusion / Time-Aligned Micro-Turns——以它為依據的架構選擇
- Claude Character as Product——一個可能的反例：性格也許不會往內遷移
- Model Spec Midtraining (MSM)——把對齊從 harness 的 prompt 注入轉移到模型內化的價值觀，是在對齊這條軸線上的一次 bitter-lesson 式行動
- Compute Allocator——點明哪些東西留在那條線的人類這一側：分配決策、以及支撐它的、面向人類的鷹架不會往內遷移，即使面向模型的結構會
- HTML as the New Markdown——「為模型保留讓你驚喜的空間」是這條教訓在 prompt 層次的形式；但書是：面向人類的可讀性（HTML artifact）落在不會溶入模型的那一側
- MCP and Computer Use——Boris Cherny 的「對模型而言，這些都只是 token」把底層基質的選擇（MCP／API／computer use）變成模型的決定，而非 harness 的決定；是工具調度的 bitter-lesson 終點
- Agentic Loops Overtake Bespoke Systems——語料庫中最清楚的經驗性印證：隨著 LLM 進步，DeepMind 那套簡單的 agentic loop，在開放的數學問題上追平了它量身打造的訓練系統（AlphaProof ＋演化式搜尋）
- AI R&D Autonomy Evaluation (AECI)——如果 the bitter lesson 一路走到底，擴展後的通用方法最終會改進它們自己；AECI 就是 Anthropic 用來衡量那道門檻是否將近的方式
- Recursive Self-Improvement——這條原則最遠的外推：「研究進展主要是工具與資源的函數」，因此努力（那 99%）變得可自動化
- AI Accelerating AI Development——經驗性案例：kernel 最佳化迴圈從 3×→52× 的進步，就是擴展後的通用方法勝過手動調校，而且是實際量測到的
- Research Taste as the Human Bottleneck——對最後一塊頑抗陣地下的未定賭注：研究品味究竟是真正的天花板，還是 the bitter lesson 即將溶解的下一個結構？
- Build for the Next Model——產品策略上的推論：既然能力會隨著版本一路往內遷移，那就先把「差一點就能用的東西」做成原型，讓下一代模型去溶解那道落差，而不是繞著它做工程
- Task Time-Horizon Scaling——不斷上升的通用基準能力，正是那條持續削弱手工鷹架優勢的曲線
- The Verifiability Thesis——Karpathy 對為何擴展後的 RL 會勝過手工工程的說法：各實驗室把運算砸在可驗證獎勵的環境上
- Software 3.0——把神經網路當成宿主行程的外推，是 the bitter lesson 一路推到硬體層的結果
- Andrej Karpathy——這條原則的常客（可驗證性、ghosts、Software 3.0 全都以它為基礎）
資料來源#
- Interaction Models: A Scalable Approach to Human-AI Collaboration（明確引用了「the bitter lesson」）
Cited by 27
- Agent Harness Engineering
Patterns for scaffolding long-running LLM agents: environment design, progressive context disclosure, mechanical archit…
- Agentic Loops Overtake Bespoke Systems
DeepMind's *basic* Ralph-loop agent matched its bespoke evolutionary+AlphaProof system as the LLM improved; the bitter…
- AI Accelerating AI Development
The empirical core of *When AI builds itself*: measured evidence AI already speeds AI R&D at Anthropic — >80% of merged…
- AI R&D Autonomy Evaluation (AECI)
How Anthropic measures whether a model can automate or dramatically accelerate AI research — the capability that drives…
- Opinions on Using AI Tools & the Future of the Software Engineering Role
Debate map of four stances on using AI tools (bullish-insider / pragmatist-practitioner / skeptic-governance / architec…
- Andrej Karpathy
Co-founder OpenAI, ex-Tesla AI, Eureka Labs; coined "vibe coding," Software 1/2/3.0, "ghosts not animals," "agentic eng…
- Build for the Next Model
Prototype the thing that almost works, not the thing that already works: bet that the next concrete model release (not…
- Claude Character as Product
Personality as load-bearing product surface; Amanda's role at Anthropic; lunchtime vibe-checks as eval discipline; the…
- Compute Allocator
The human's evolving role: deciding what's worth spending compute on; ~1% of generated tokens ship, 99% is scaffolding…
- Encoder-Free Early Fusion
Multimodal design with minimal pre-processing instead of large standalone encoders: dMel audio embedding, 40×40-patch h…
- Evolutionary Proof Search
The full-featured agent's mechanism: population DB of proof sketches, Elo via Plackett–Luce/Gibbs, P-UCB selection, LLM…
- The Future of Agent Interfaces
Interface future is layered: native interaction models for human collaboration, MCP/APIs for structured action, app pro…
- Harness Shrinkage as Models Improve
Prompt scaffolding shrinks each model release; Cat Wu's pruning discipline; Boris Cherny "100 lines of code a year from…
- HTML as the New Markdown
Thariq Shihipar's thesis: as models improve, thousand-line markdown plans overwhelm the *human*; HTML artifacts (visual…
- Does the Human-Facing Harness (HTML Artifacts) Hit Its Own Bloat Ceiling?
Yes — HTML raises and reshapes the human-attention ceiling but can't remove it; bloat relocates from document-length to…
- Interaction Models
Thinking Machines Lab (May 2026): models that handle audio/video/text interaction natively in real time instead of via…
- MCP and Computer Use
Anthropic's two complementary connector mechanisms: MCP for structured programmatic access (Salesforce/Drive/Gmail/Slac…
- LLM Architecture, Training & Alignment
Map of Content for the llm-architecture domain — 19 concepts. Curated entry point; see Home for all domains.
- Model Spec Midtraining (MSM)
New training phase between pretrain and AFT: train base model on synthetic docs discussing the Model Spec; controls AFT…
- Recursive Self-Improvement
An AI system autonomously designing and developing its own successor; Anthropic Institute's *When AI builds itself* arg…
- Research Taste as the Human Bottleneck
The narrowing human role as AI absorbs execution: choosing which problems matter, which results to trust, and when an a…
- Software 3.0
Karpathy's taxonomy: 1.0 code, 2.0 weights, 3.0 prompting; LLM as programmable interpreter; MenuGen "shouldn't exist";…
- Task Time-Horizon Scaling
METR's measure of the task length AI can complete reliably on its own, doubling roughly every 4 months (up from every 7…
- Thinking Machines Lab
AI research lab behind interaction models (May 2026); harness-dissolves-into-model thesis; upstreamed streaming-session…
- Time-Aligned Micro-Turns
The core interaction-model move: input/output as continuous streams in ~200ms interleaved chunks, no turn boundaries; s…
- Turn-Based Interface Bottleneck
Why current AI interfaces limit collaboration: single-thread turn-taking is a bandwidth bottleneck; humans pushed out b…
- The Verifiability Thesis
LLMs automate what you can *verify* as computers automate what you can *specify*; RL verification rewards → jagged peak…
Related articles
- Harness Shrinkage as Models Improve
Prompt scaffolding shrinks each model release; Cat Wu's pruning discipline; Boris Cherny "100 lines of code a year from…
- Interaction Models
Thinking Machines Lab (May 2026): models that handle audio/video/text interaction natively in real time instead of via…
- Open Questions Backlog
_96 pages with open questions, as of 2026-06-14._
- Claude Code
Anthropic's agentic coding product; created by Boris Cherny late 2024; TypeScript/React; CLI/desktop/web/mobile/IDE sur…
- Recursive Self-Improvement
An AI system autonomously designing and developing its own successor; Anthropic Institute's *When AI builds itself* arg…