資料來源#
摘要#
Thariq Shihipar 對於當模型已經能完成大部分產出後，人類會變成什麼的 framing 是："we're all becoming compute allocators." 主要工作不再是生成作品，而是決定什麼值得花 compute，並投資在對齊和溝通上，確保花出去的 compute 花得值得。
這個 framing 最尖銳的統計是：Thariq 生成的 tokens 也許只有 1% 最後進了 production code。 另外 99% 花在豐富、通常可拋棄的 scaffolding 上：HTML plans、custom interfaces、狀態更新、design systems。那 99% 不是浪費；它是對齊與溝通的投資，讓真正 ship 的那 1%「正好就是它需要成為的樣子」。
1% / 99% 的分配#
"Thariq mentioned that maybe only 1% of the tokens he generates end up in production code. The other 99% are spent on this rich, beautiful, and sometimes disposable scaffolding."
仔細讀，這會反轉工程師平常的心智模型：輸出不再是你 ship 的成品；輸出是決策周圍的 scaffolding。ship 出去的 code，只是更大一片思考與權衡表面留下的小而高槓桿殘餘。把 compute 分配好，意思是選擇要資助哪些思考、以及要資助到多豐富：這裡一場 brainstorm，那裡一個可拋棄的編輯 UI，再加上一個 Living Design System，讓未來工作保持品牌一致。
「分配」在實務上代表什麼#
- 先決定什麼值得 build，再去 build：brainstorm-in-HTML 這一步存在，是為了讓 human 能在多個選項中做出好選擇，每個選項都有自己的風險評估。
- 花在可讀性上，不只是 production：為 human 真的會投入的計畫和介面支付 tokens（見 HTML as the New Markdown）。
- 花在可拋棄的 tooling 上：啟動一個 micro-app，讓單一決策變得更好，然後把它丟掉。
- 把其餘部分交給模型："I trust you here"；刻意少寫 spec，讓能力補上空白（一種 prompt 層級的 The Bitter Lesson）。
abundance mindset#
這種行為只有在 Thariq 所說的 abundance mindset 下才合理：生成已經便宜到你負擔得起製作可拋棄的 scaffolding，讓自己的流程更有效率也更愉快。稀缺思維（「不要浪費 tokens / 不要 build 你會丟掉的東西」）是錯的 framing；真正綁住你的限制是 human 注意力與判斷，不是生成成本。這正是 Printing Press Software Democratization 描述的同一種經濟轉移：當 production 成本崩塌，價值會移到做什麼與要不要做，而不是怎麼做。
與 wiki 其他部分的關係#
- "Coding is solved (for me)"（Boris Cherny）命名了前提；compute allocator 命名了當前提成立後留下來的角色。Boris 透過 Claude 寫 100% 的 code，並跑數百個 agents；剩下的 human 工作就是分配與指向。
- product taste 作為瓶頸技能（Engineer PM Convergence、Cat Wu："as code becomes much cheaper to write, the thing that becomes more valuable is deciding what to write"）是 org-design 側的同一個主張。Compute allocation 就是在單次模型呼叫尺度上行使的 product taste。
- Harness shrinkage，但帶著轉折：當 model-facing harness 收縮（Harness Shrinkage as Models Improve），human-facing harness（那 99%）會成長，因為 allocator 需要越來越豐富的 artifacts 才能做出好決策。見 HTML as the New Markdown 的 harness-tension 段落。
- 不會向內遷移的東西：The Bitter Lesson 會溶解 model-facing 結構；分配決策以及支撐它的 human-facing scaffolding，正是留在人類這側邊界上的東西。
相關連結#
- Thariq Shihipar：創造了這個 framing
- Claude Code：allocator 花費 compute 的工具；越來越服務於「決定要 build 什麼」，而不是「把 code 打出來」
- HTML as the New Markdown：allocator 花掉那 99% 的媒介
- Disposable Micro-Apps：abundance mindset 的 tooling，由分配資助
- Living Design System：一種持久分配：花一次，之後永遠作為 context 重用
- Boris Cherny："coding is solved (for me)" 是 allocator 角色的前提
- Engineer PM Convergence：「deciding what to write」作為瓶頸技能；分配就是呼叫尺度上的 taste
- Printing Press Software Democratization：讓 99% 可拋棄變得合理的 abundance economics
- Harness Shrinkage as Models Improve：model-facing harness 收縮，而 allocator 的 human-facing harness 成長
- The Bitter Lesson：分配決策就是不會遷移進模型的東西
- Outsource Your Thinking, Not Your Understanding：Karpathy 的雙生 framing：你可以 outsource 思考（以及 compute），但讓你能好好分配的理解仍留在人類身上
- Software 3.0：MenuGen（"that app shouldn't exist"）是正在發生的分配：決定應該讓 neural net 直接做這件事，而不是把 compute 花在 scaffolding 上
- Research Taste as the Human Bottleneck：把分配擴大到整個研究計畫：一旦 Claude 可以跑實驗，就要決定哪些實驗值得跑
- Recursive Self-Improvement：那篇 essay 的第三種未來，是人類把「大部分 effort 轉向 oversight、validation、verification」，pace 則由 compute 設定，也就是文明尺度上的 allocator 角色
開放問題#
- 1% 是 Thariq 專屬的數字，還是一種 regime？對更大型、更偏 code-heavy 的專案，production 殘餘比例想必更高；是什麼決定這個比例？
- 分配品質很難測量：什麼 feedback loop 會告訴 allocator 他們把 compute 花得很糟（相對於只是花了很多）？
- 把 human 視為 "compute allocators"，是否會冒著 oversight-fatigue / accountability 這些 HBR research 所標記的失敗模式：human 名義上做決定，但實際上只是 rubber-stamps？
衍生內容#
- Does the Human-Facing Harness (HTML Artifacts) Hit Its Own Bloat Ceiling?：有限的 human 注意力是 allocator 花費的守恆資源；rubber-stamping 是 human-facing harness 突破自身 bloat ceiling 的方式
- Human-in-the-Loop Boundaries：當 human 分配、理解並擁有風險，以及當他們變成手動 throughput 瓶頸時的操作邊界
資料來源#
Cited by 20
- AI Brain Fry
Kropp et al. 2026/03: mental fatigue from excessive AI oversight increases minor errors +11%, major errors +39%; cognit…
- Boris Cherny
Creator of Claude Code at Anthropic; phone-driven workflow with hundreds of agents; primary advocate of `/loop` primiti…
- Claire Vo
Host of the "How I AI" interview series (ChatPRD); interviewed Thariq Shihipar; runs a parallel component-visualization…
- Claude Code
Anthropic's agentic coding product; created by Boris Cherny late 2024; TypeScript/React; CLI/desktop/web/mobile/IDE sur…
- Disposable Micro-Apps
Throwaway custom UIs built per-task to edit a plan ("micro-software on top of micro-software"); copy-back-to-markdown;…
- Engineer PM Convergence
Generalists across disciplines; product taste as bottleneck skill; Anthropic Claude Code team as case study; "just do t…
- 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…
- Human-AI Accountability Redesign
HBR five-pillar prescription: span-of-control redesign, role redesign, performance management reset, decision-rights/es…
- 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…
- Human-in-the-Loop Boundaries
Humans belong at allocation, understanding, design-concept, risk, and accountability boundaries; they slow the system d…
- Living Design System
`design_system.html` extracted from repos as a portable, human- and machine-readable source of truth; component playgro…
- AI Engineering & Agent Tooling
Map of Content for the ai-engineering domain — 36 concepts. Curated entry point; see Home for all domains.
- Open Questions Backlog
_96 pages with open questions, as of 2026-06-14._
- Outsource Your Thinking, Not Your Understanding
"You can outsource your thinking but not your understanding"; understanding as the non-delegable human bottleneck; know…
- Printing Press Software Democratization
Boris Cherny's analogy: 1400s literacy expansion → AI software-writing expansion; domain knowledge displaces coding ski…
- 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";…
- Thariq Shihipar
Engineer on the Claude Code team at Anthropic; "HTML is the new markdown" and "compute allocator" framings; three HTML-…
- The Bitter Lesson
Sutton 2019: scaled general methods beat hand-engineered structure; recurring justification across the wiki for dissolv…
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…
- Disposable Micro-Apps
Throwaway custom UIs built per-task to edit a plan ("micro-software on top of micro-software"); copy-back-to-markdown;…
- Thariq Shihipar
Engineer on the Claude Code team at Anthropic; "HTML is the new markdown" and "compute allocator" framings; three HTML-…
- Claude Code
Anthropic's agentic coding product; created by Boris Cherny late 2024; TypeScript/React; CLI/desktop/web/mobile/IDE sur…