資料來源#
摘要#
Thariq Shihipar 對於當模型能承擔大部分產出後,人類將成為什麼角色的說法是:「我們都正在成為計算資源配置者。」 主要工作不再是產出成果,而是判斷哪些事情值得投入計算資源,並投入心力於對齊與溝通,確保花出去的計算資源用得其所。
這個說法最醒目的統計是:Thariq 產生的 token 可能只有 1% 最後進入正式程式碼。 另外 99% 用在豐富、往往用完即棄的腳手架上——HTML 計畫、自訂介面、狀態更新、設計系統。這 99% 並非浪費;它是對齊與溝通的投入,讓正式上線的那 1%「恰如其分」。
1%/99% 的分配#
「Thariq 提到,他產生的 token 可能只有 1% 最後進入正式程式碼。其餘 99% 都花在豐富、美觀,有時甚至用完即棄的腳手架上。」
仔細看,這顛覆了工程師一貫的思維模式:產出不再是你要交付的成果;產出是決策周遭的腳手架。正式交付的程式碼,只是大量審議過程留下的一小部分高槓桿成果。妥善配置計算資源,意味著選擇要支持哪些審議、投入多少資源——在這裡集思廣益,在那裡做個一次性的編輯介面,或建立一套持續演進的設計系統,讓未來的工作維持品牌風格。
「配置」在實務上的意義#
- 先決定值得打造什麼,再動手打造——先用 HTML 集思廣益,是為了讓人類能從各有風險評估的選項中做出好選擇。
- 投入資源讓事情易於理解,而不只投入生產——花 token 製作人類確實會使用的計畫與介面(參見 HTML as the New Markdown)。
- 投入資源打造一次性工具——建立一個微型應用程式,幫助自己更妥善地做出某個決定,然後用完就丟。
- 其餘交給模型——「這裡我信任你」;不要把規格寫得太細,讓模型的能力自行補足(提示層級的 The Bitter Lesson)。
豐裕心態#
只有在 Thariq 所說的豐裕心態下,這些行為才說得通:生成成本低到你能負擔製作一次性腳手架,讓自己的流程更有效率、更愉快。稀缺思維(「別浪費 token/別打造最後會丟掉的東西」)是錯誤的框架;真正受限的是人類的注意力與判斷力,而非生成成本。這正是 Printing Press Software Democratization 所描述的同一種經濟轉變——當生產成本大幅降低,價值便從「如何做」轉移到「做什麼」與「是否要做」。
與 wiki 其他內容的關係#
- 「對我來說,寫程式已經解決了」(Boris Cherny)指出前提;計算資源配置者則命名了此前提成立後仍然存在的角色。Boris 透過 Claude 撰寫 100% 的程式碼,也同時執行數百個代理程式;人類剩下的工作就是配置與指引。
- 產品品味是瓶頸技能(Engineer PM Convergence、Cat Wu:「程式碼寫起來便宜許多之後,更有價值的事就是決定要寫什麼」)從組織設計的角度提出相同主張。計算資源配置,就是在單次模型呼叫的層級發揮產品品味。
- 腳手架縮減,卻有轉折——隨著面向模型的腳手架縮小(Harness Shrinkage as Models Improve),面向人類的腳手架(那 99%)卻會擴大,因為配置者需要愈來愈豐富的成果來做出好決定。參見 HTML as the New Markdown 中討論腳手架張力的段落。
- 哪些東西不會移入模型——The Bitter Lesson 消解面向模型的結構;配置決策以及支援該決策、面向人類的腳手架,正是留在人類這一側的部分。
延伸關聯#
- Implementation Abundance Inverts Product Work — 從整個功能探索的層級來看,從 90 個低成本實作中篩選,就是資源配置
- Role Averaging, Not Role Elimination — 平均化後的 IC 是代理程式工作的配置者/引導者,而非逐字逐句的撰寫者
- Vibe Coding vs. Agentic Engineering — 「寫程式就是引導 AI」重述了配置者的角色:衡量引導次數,而非程式碼行數
- Thariq Shihipar — 提出這個說法的人
- Claude Code — 配置者投入計算資源所透過的工具;它愈來愈常用於「決定要打造什麼」,而非「敲出程式碼」
- HTML as the New Markdown — 配置者投入那 99% 資源時使用的媒介
- Disposable Micro-Apps — 由資源配置支持、體現豐裕心態的工具
- Living Design System — 一種持久的資源配置:投入一次,作為脈絡永久重用
- Boris Cherny — 「對我來說,寫程式已經解決了」是配置者角色的前提
- Engineer PM Convergence — 「決定要寫什麼」是瓶頸技能;資源配置就是在單次呼叫的尺度上運用品味
- Printing Press Software Democratization — 讓 99% 用完即棄也合理的豐裕經濟
- Harness Shrinkage as Models Improve — 面向模型的腳手架縮小,配置者面向人類的腳手架則擴大
- The Bitter Lesson — 配置決策正是不會移入模型的部分
- Outsource Your Thinking, Not Your Understanding — Karpathy 提出的雙重說法:你可以外包思考(以及計算資源),但讓你能妥善配置資源的理解力仍掌握在人類手中
- Software 3.0 — MenuGen(「那個應用程式根本不該存在」)就是配置的實例:決定直接讓神經網路完成工作,而不是把計算資源花在腳手架上
- Research Taste as the Human Bottleneck — 將資源配置擴大到整個研究計畫:Claude 能執行實驗後,決定哪些實驗值得進行
- Recursive Self-Improvement — 文章提出的第三種未來,是人類將「把大部分心力轉向監督、驗證與確認」,而進展速度由計算資源決定——也就是文明尺度的配置者角色
- AI as Primary Author — AI 撰寫大部分程式碼後,人類剩下的工作就是配置與指引,而非撰寫;這正是此頁命名的角色
- Planning / Execution Division of Labor — 對配置者角色的量化:Anthropic 對 400K 個工作階段的研究發現,人類做出約 70% 的規劃決策(要做什麼),而 Claude 負責約 80% 的執行;實證顯示,判斷什麼值得做是人類保留的工作範圍
- Unknowns as the Agentic Bottleneck — 未能上線的 99% token 都花在什麼事情上:Thariq Shihipar 的盲點探索、腦力激盪、訪談與參考資料,都是在釐清配置者自身未知之處
開放問題#
- 1% 是 Thariq 個人的數字,還是某種常態?規模更大、更偏重程式碼的專案,正式交付的部分想必更高;比例由什麼決定?
- 配置品質很難衡量——有什麼回饋機制能讓配置者知道自己把計算資源花得不好(而非只是花了很多)?
已解決問題#
- 把人類視為「計算資源配置者」,是否會引發 HBR 研究指出的監督疲勞/問責失效模式:人類名義上做決定,實際上卻只是照單全收?解答:Is Human Review of AI-Authored Code Still a Real Control, or Already Rubber-Stamping?——是的,這是此角色的核心失效模式,已有三層證據記錄(腦力疲勞造成的錯誤率增加 +11%/+39% 及投入不足的替代效應;Faros 的 31.3% 未審查遙測數據;CMU 理論中的 P1–P4 負荷與合理性機制),另有兩個配置者特有的惡化因素:配置品質沒有回饋機制,而且人類保留的 70% 規劃工作,正是照單全收在逐字稿中看不出來的部分。唯有採取結構性配套措施,這個角色才站得住腳——以理解力為門檻的合併機制(測驗關卡)、集中於高風險環節的抽樣深度審查,以及依風險分級的關卡;如此才能讓理解力,而不是簽名,成為合併的條件。
衍生文章#
- Does the Human-Facing Harness (HTML Artifacts) Hit Its Own Bloat Ceiling? — 有限的人類注意力是配置者所使用的固定資源;照單全收則是面向人類的腳手架突破膨脹上限的方式
- Human-in-the-Loop Boundaries — 描述人類何時負責配置、理解並承擔風險,何時又會成為人工處理速度的瓶頸
- Is Human Review of AI-Authored Code Still a Real Control, or Already Rubber-Stamping? — 照單全收的問題已有答案:三層證據皆顯示存在;唯有重新設計審查流程才能避免失控
資料來源#
Cited by 29
- Human-in-the-Loop Boundaries×5
Problem selection · Deciding what is worth building or investigating. This is the Compute Allocator…
- HTML as the New Markdown×3
The Bitter Lesson dissolves model-facing structure; it does not dissolve the human-facing structure…
- Does the Human-Facing Harness (HTML Artifacts) Hit Its Own Bloat Ceiling?×3
The binding constraint is "human attention and judgement, not generation cost" (Compute Allocator).…
- Thariq Shihipar×3
Engineer on the Claude Code team at Anthropic. Source of the "HTML is the new markdown" thesis (see…
- Disposable Micro-Apps×2
Spinning up a single-use application to make one editing decision would be absurd under scarcity…
- Is Human Review of AI-Authored Code Still a Real Control, or Already Rubber-Stamping?×2
Compute Allocator — does the allocator framing risk the oversight-fatigue / rubber-stamping failure…
- Living Design System×2
A conventional style guide drifts: the doc says one thing, the code does another. A…
- Outsource Your Thinking, Not Your Understanding×2
The human is "becoming a bottleneck of even knowing what we're trying to build, why is it worth…
- Planning / Execution Division of Labor×2
Anthropic's 400K-session study supplies the empirical shape of human–agent collaboration in agentic…
- Research Taste as the Human Bottleneck×2
Compute Allocator — taste exercised as allocation: deciding which experiments are worth the compute
- Vibe Coding vs. Agentic Engineering×2
Compute Allocator — "coding is steering the AI" is the allocator role: measure the steers, not the…
- AI as Primary Author
Compute Allocator — if the human is no longer the author, their residual role is deciding what the…
- AI Brain Fry
Oversight-quality risk: Compute Allocator — the "compute allocator" role assumes the human decides…
- Boris Cherny
Compute Allocator — "coding is solved (for me)" is the precondition; Thariq Shihipar's allocator…
- Claire Vo
Host of the "How I AI" interview series, published on ChatPRD (chatprd.ai), in which practitioners…
- Claude Code
Compute Allocator — the user role Claude Code increasingly serves: deciding what's worth compute,…
- Engineer PM Convergence
Compute Allocator — "deciding what to write" as the bottleneck skill, restated at the level of…
- Harness Shrinkage as Models Improve
Compute Allocator — names the human role that expands as the model-facing harness shrinks; ~99% of…
- The HTML Artifact Lifecycle: Where Plan History Lives, and When Disposable Becomes Durable
So "what happens to plan history" has a direct answer: it lives in the content layer or it…
- Human-AI Accountability Redesign
Role-evolution complement: Compute Allocator — names the individual-IC role ("decide what's worth…
- Implementation Abundance Inverts Product Work
Compute Allocator — curating 90 builds is allocation at the level of a whole feature exploration
- AI Coding Practice
Compute Allocator — The human's evolving role: deciding what's worth spending compute on; ~1% of…
- Open Questions Backlog
Compute Allocator ×2 (oldest 131d) — Is 1% a Thariq-specific number or a regime?
- Printing Press Software Democratization
Compute Allocator — the role the democratized builder plays: deciding what's worth producing, now…
- Role Averaging, Not Role Elimination
Compute Allocator — the averaged IC as allocator/steerer of agent work rather than…
- Software 3.0
Compute Allocator — deciding what's worth compute is the human role in a 3.0 stack
- The Bitter Lesson
Compute Allocator — names what stays on the human side of the line: the allocation decision and the…
- Unknowns as the Agentic Bottleneck
Compute Allocator — Thariq's other framing: if 99% of generated tokens are scaffolding, most of…
- What Scaffolding Survives Model Improvement — and How Do You Know When a Line Turns Harmful?
Human-facing legibility. The harness that keeps the human in the loop (HTML plans, review surfaces,…
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…
- HTML as the New Markdown
Thariq Shihipar's thesis: as models improve, thousand-line markdown plans overwhelm the *human*; HTML artifacts (visual…
- Verification as the New Bottleneck
Fiona Fung: coding is no longer the bottleneck — verification, review, maintenance are; shift-left; TDD loses its tax;…
- Claude Code
Anthropic's agentic coding product; created by Boris Cherny late 2024; TypeScript/React on Bun (itself Claude-rewritten…
- Returns to Expertise in Agentic Coding
Anthropic's 400K-session study: domain expertise (not coding skill) is what amplifies an agent — experts get 2× the act…
