資料來源#
- Gemma 4 Technical Report
- Interaction Models: A Scalable Approach to Human-AI Collaboration
- Single-Rollout Asynchronous Optimization for Agentic Reinforcement Learning
摘要#
Rich Sutton 在 2019 年的文章指出:運用計算能力的通用方法(搜尋、學習),最終會勝過將人類知識與手工設計結構內建其中的方法,而且隨著計算能力增加,差距會大幅擴大。「苦澀」之處在於,這總是讓那些投入心力打造精巧領域結構的研究者感到意外,因為結構最後成了天花板,而不是地基。
這個頁面存在的原因,是這項原則在本 wiki 中反覆作為關鍵論據出現——明確用來支持將 harness 溶入模型。
本頁的引用位置#
- Interaction Models — TML 直接引用「苦澀的教訓」:手工打造的互動系統(VAD、turn-detection、dialog-management harnesses)「會被通用能力的進展超越」,因此「若要讓互動性隨智慧一同擴展,就必須將其納入模型本身」。請參閱 Turn-Based Interface Bottleneck。
- Encoder-Free Early Fusion — 在單一 transformer 中從頭共同訓練所有模態元件,而不是拼接預訓練的編碼器/解碼器:減少手工設計的模組邊界。
- Time-Aligned Micro-Turns — 移除人為的輪次邊界,讓互動模式成為可擴展的模型行為,而不是各模式專用的 harness 程式碼。
- Harness Shrinkage as Models Improve — 將同樣的邏輯套用到 coding-agent harness:提示 scaffolding 彌補模型尚做不到的部分,並應隨模型進步而縮減。(但該文的例外是:機械式驗證——測試、型別、linters——不會向內遷移。)
- Agent Harness Engineering — 「強制執行不變量,而非實作」:讓模型尋找路徑;harness 只編碼必須為真的條件。
標準例外#
苦澀的教訓談的是能力與結構向模型內遷移,而不是「harness 沒用」。合理地留在模型外的事物包括:機械式驗證(Harness Shrinkage as Models Improve 的綜合結論)、組織特定的政策/風格、安全邊界,以及——依據 Claude Character as Product——刻意塑造角色/個性的工作。每個 harness 元件都必須面對一個開放問題:它究竟位於這條界線的哪一側?
部署豁免#
Gemma 4 同時朝兩個方向違反這條界線,讓界線更加清晰。同一份報告移除手工設計的結構——550M 視覺編碼器變成 35M matmul,305M 音訊 conformer 被完全刪除(Encoder-Free Early Fusion)——同時也加入大量結構:5:1 的區域至全域注意力比例、p = 0.25 的 p-RoPE、全域層中的 values = keys、每層叢集的量化位元寬度,以及 drafter head 中對 token-clusters 的 top-k 投影。
這並不矛盾,說明原因很有用。苦澀的教訓針對的是編碼人類對任務之先驗的結構——模態編碼器主張,在模型看見音訊之前,音訊應以某種特定方式被消化,而這項主張會成為天花板。KV-cache 與量化技巧沒有編碼任何事物的先驗;它們只是讓模型已學會的網路得以運行的算術。部署工程不受此限,Inference Efficiency as Capability 因此主張它會累積而非縮減:它沒有可供遷移的「內部位置」。
對 Harness Shrinkage as Models Improve 而言,推論路徑本身是留在外部事物清單上的第二項,與機械式驗證並列。
同樣的豁免也適用於訓練迴圈。SAO 在 RL 端執行完全相同的雙向移動——它移除機制(舊政策模型 π_θ_old、GRPO 的群組基準線),同時加入大量機制(frozen-attention critic、skip-observation GAE、length-adaptive λ、每任務的 clip 不對稱性)。而它手工打造的 critic 在推論時會被丟棄;就像 KV-cache 技巧一樣,這是沒有可供遷移之「內部位置」的結構。界線仍然成立:模型學到的任務知識會向內遷移;產生或服務模型的 scaffolding 不會。
相關連結#
- Why AI Lags at Design — 「這些模型將會擅長設計」是將苦澀教訓押注於設計落差的應用
- Evolutionary Proof Search — 苦澀教訓預測會被吸收的,正是這套特製的演化裝置
- Interaction Models — 最近最明確的引用
- Turn-Based Interface Bottleneck — 「較不智慧的 harness 會敗給擴展」
- Harness Shrinkage as Models Improve — coding-agent 版本,附帶機械式驗證的例外
- Agent Harness Engineering — 以苦澀教訓為意識的不變量而非實作設計規則
- Encoder-Free Early Fusion / Time-Aligned Micro-Turns — 由此原則支持的架構選擇
- Claude Character as Product — 一個可能的反例:角色也許不會向內遷移
- Model Spec Midtraining (MSM) — 對齊從 harness-prompt-injection 移向模型內化的價值觀,是對齊軸上的苦澀教訓式移動
- Compute Allocator — 說明哪些事物留在人類一側:分配決策與支援人類的 scaffolding 不會向內遷移,即使面向模型的結構會遷移
- HTML as the New Markdown — 「為模型留下讓它帶來驚喜的空間」是這項教訓在提示層級的形式;但例外是,面向人類的可讀性(HTML artifacts)位於不會溶入模型的一側
- MCP and Computer Use — Boris Cherny 的「對模型而言,那就只是 tokens」表示 substrate 的選擇(MCP/API/computer use)是模型決策,而不是 harness 決策;這是工具派送的苦澀教訓終點
- Agentic Loops Overtake Bespoke Systems — 語料庫中最清楚的實證確認:隨著 LLM 進步,DeepMind 的簡單 agentic loop 在開放數學問題上,追上了其特製訓練系統(AlphaProof + evolutionary search)
- AI R&D Autonomy Evaluation (AECI) — 如果苦澀教訓一路貫徹,擴展通用方法最終會改進自身;AECI 是 Anthropic 用來衡量該門檻是否接近的方法
- Recursive Self-Improvement — 這項原則最遠的外推:「研究進展主要是工具與資源的函數」,因此勞力(99%)將可自動化
- AI Accelerating AI Development — 實證案例:核心最佳化迴圈從 3×→52×,是擴展通用方法擊敗手工調校的量化結果
- Research Taste as the Human Bottleneck — 對最後一個尚未被突破之處的開放押注:研究品味是真正的天花板,還是苦澀教訓將溶解的下一個結構?
- Build for the Next Model — 產品策略推論:既然能力會在各版本發布時向內遷移,就製作「差一點就能運作」的東西,讓下一個模型消除落差,而不是繞著它進行工程設計
- Task Time-Horizon Scaling — 通用基準能力上升的曲線,正是持續縮小手工 scaffolding 優勢的曲線
- The Verifiability Thesis — Karpathy 對擴展 RL 為何跑贏手工工程的解釋:實驗室將計算能力投入可驗證獎勵的環境
- Software 3.0 — 將神經網路作為主機程序的外推,是一路把苦澀教訓推到硬體層的結果
- Universal AI (AIXI) — 形式化版本:「智慧是穿越假設/政策空間的搜尋」,其中交錯執行的 AIXI 近似保證會隨計算能力增加而改進(但代價是暴力搜尋不可行)
- Effective Compute Scaling — 「擴展是否足夠?」是以預測問題形式提出的苦澀教訓;DeepMind 對樸素暴力搜尋在玩具領域之外失效的警告,呼應 Sutton 的「搜尋需要良好先驗」
- Andrej Karpathy — 經常引用這項原則的人(可驗證性、幽靈、Software 3.0 都建立在其上)
- Repository Exploration Subagent — 一個活生生的測試案例:FastContext 訓練專門的探索器(手工打造的結構),但它自身的「同模型探索」基準線顯示,持久的收益來自架構分離,而不是訓練出的模型——苦澀教訓預測,隨著基礎模型變得更便宜,這項收益將會縮小
- Inference Efficiency as Capability — 例外:不編碼任務先驗的結構(KV cache、量化、drafters)永遠不會向內遷移,反而會累積
- Gemma 4 — 在同一版本中移除編碼器並加入推論路徑結構,清楚劃出界線
- Single-Rollout Optimization — 同樣雙向移動在訓練迴圈中的案例;手工打造的 critic 是 scaffolding,在推論時被丟棄,因此永遠不會向內遷移
- Asynchronous RL for LLMs — DIS 是「透過進一步移除而變得更簡單」(
π_θ_old、checkpoint history);RL 管線與推論路徑一樣,皆不受此限 - Group Relative Policy Optimization (GRPO) — 移除 critic 本身就是苦澀教訓式的移動;SAO 是一項衡量這種移除何時走得太遠的論證
資料來源#
- Interaction Models: A Scalable Approach to Human-AI Collaboration(明確引用「苦澀的教訓」)
- Gemma 4 Technical Report — §2(手工設計的推論路徑)以及 §2.3(移除編碼器)(
empirical) - Single-Rollout Asynchronous Optimization for Agentic Reinforcement Learning — §3(移除
π_θ_old與群組基準線;加入 frozen-attention critic + skip-observation GAE):訓練迴圈雙向移動的案例(empirical)
Cited by 40
- Agentic Loops Overtake Bespoke Systems×3
The headline empirical finding of DeepMind's Ai Driven Formal Proof Search paper, and its clearest…
- Opinions on Using AI Tools & the Future of the Software Engineering Role×3
The bitter lesson recurs. The Bitter Lesson: scaled general methods beat hand-engineered structure.…
- Compute Allocator×3
What doesn't migrate inward — The Bitter Lesson dissolves model-facing structure; the allocation…
- HTML as the New Markdown×3
The Bitter Lesson dissolves model-facing structure; it does not dissolve the human-facing structure…
- Single General Agent vs. Multi-Agent Coding Architecture×3
The Bitter Lesson: scaled general methods beat hand-engineered structure over time; the structure…
- Agent Harness Engineering×2
Does a single general-purpose coding agent outperform a multi-agent architecture with specialized…
- Build for the Next Model×2
This is the product-side expression of The Bitter Lesson and Harness Shrinkage As Models Improve:…
- Effective Compute Scaling×2
The Bitter Lesson — "is scaling enough?" is the bitter lesson as a forecasting question; search…
- Encoder-Free Early Fusion×2
The Bitter Lesson — "co-train from scratch, drop the modular encoders" is a bitter-lesson move
- Does the Human-Facing Harness (HTML Artifacts) Hit Its Own Bloat Ceiling?×2
The model-facing harness can shrink toward zero as capability migrates inward (Harness Shrinkage As…
- Inference Efficiency as Capability×2
The report simultaneously removes hand-engineered structure (encoders, per Sutton's logic) and adds…
- Interaction Models×2
The central bet: interactivity should scale alongside intelligence. If interaction is part of the…
- MCP and Computer Use×2
This connects to The Bitter Lesson: as models improve, the boundary between "use an MCP" and "use…
- Recursive Self-Improvement×2
Perspiration is becoming automated. AI advances rarely come from "eureka" moments; paradigm shifts…
- Research Taste as the Human Bottleneck×2
The Bitter Lesson — "research progress is mostly tools and resources" is the bitter lesson aimed at…
- Software 3.0×2
Pushed to the limit: a "completely neural computer" — raw video/audio in, diffusion rendering a UI…
- Thinking Machines Lab×2
Position: interactivity should scale with intelligence → it must be in the model, citing The Bitter…
- Turn-Based Interface Bottleneck×2
The Bitter Lesson says these hand-crafted systems get outpaced by general capability growth → the…
- Universal AI (AIXI)×2
The Bitter Lesson — "intelligence as search through hypothesis/policy space" is the shared premise;…
- What Scaffolding Survives Model Improvement — and How Do You Know When a Line Turns Harmful?×2
The question's examples (org style, security rules, brand voice) all survive, and the sorting rule…
- AI Accelerating AI Development
The Bitter Lesson — "research progress is mostly a function of tools and resources" is the bitter…
- AI R&D Autonomy Evaluation (AECI)
The Bitter Lesson — the acceleration AECI tracks is what makes "scaled general methods improve…
- Andrej Karpathy
The Bitter Lesson — the Sutton principle his neural-net-as-host-process extrapolation rests on
- Asynchronous RL for LLMs
The Bitter Lesson — DIS removes hand-built machinery (π_θ_old, checkpoint history) — "simpler by…
- Authority and Audit Survive Abundance
The shared shape: both ask whether abundance retires a layer — model capability retiring authority…
- Claude Character as Product
The Bitter Lesson — character is a candidate counterexample: a deliberately hand-crafted asset that…
- Evolutionary Proof Search
The Bitter Lesson — Elo/P-UCB/evolution is exactly the hand-engineered structure the bitter lesson…
- The Future of Agent Interfaces
Interaction Models is the strongest claim in the covered pages. It says interactivity should be…
- Gemma 4
The Bitter Lesson — the report both obeys it (drop the encoders) and defies it (hand-engineer the…
- Group Relative Policy Optimization (GRPO)
The Bitter Lesson — GRPO's critic-free design was itself a "remove hand-built structure" move…
- Harness Shrinkage as Models Improve
The Bitter Lesson — the underlying principle: hand-crafted scaffolding gets outpaced by scaled…
- Jeff Dean
The Bitter Lesson — his TPU design rule (specialize to the arithmetic, not to the architecture) is…
- Model Capability & Training
The Bitter Lesson — Sutton 2019: scaled general methods beat hand-engineered structure; recurring…
- Model Spec Midtraining (MSM)
Underlying principle: The Bitter Lesson — moving alignment from harness-prompt-injection of values…
- Repository Exploration Subagent
The Bitter Lesson — a live tension: training a specialized explorer adds hand-built structure,…
- Single-Rollout Optimization
The Bitter Lesson — SAO both removes structure (drops the group baseline, drops π_θ_old) and adds a…
- Task Time-Horizon Scaling
The Bitter Lesson — rising capability on general benchmarks is what makes hand-built scaffolding a…
- Time-Aligned Micro-Turns
The Bitter Lesson — "no turn boundaries → interaction modes become scalable model behavior" is a…
- The Verifiability Thesis
The Bitter Lesson — RL-at-scale in verifiable environments is the general method outrunning…
- Why AI Lags at Design
The Bitter Lesson / Build For The Next Model — "these models will get good at design" is the…
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…
- Open Questions Backlog
_456 actionable open questions across 205 pages · 107 predictions · 9 notes · 147 in progress · 69 watching (entities),…
- Interaction Models
Thinking Machines Lab (May 2026): models that handle audio/video/text interaction natively in real time instead of via…
- Claude Code
Anthropic's agentic coding product; created by Boris Cherny late 2024; TypeScript/React on Bun (itself Claude-rewritten…
- Jagged Intelligence (Ghosts, Not Animals)
"Ghosts not animals": jagged statistical circuits, no intrinsic motivation; car-wash/strawberry failures; stay in the l…
