本文由 AI 分析生成
建立時間： 2024-12-21
Summary
EN: The “80/20 problem” is a reverse Pareto: AI gets to 80% of the desired output, but fixing the last 20% to reach production quality requires nearly as much effort as doing the whole thing manually. This problem is especially acute for: large tasks (complex, hard to decompose), non-decomposable tasks (can’t be broken into independently-acceptable parts), and all-or-nothing quality tasks (code, legal documents — either it’s correct or it’s wrong). The product implication: the wrapper that bridges the 80→100% gap creates 80% of the value.
ZH: 「80/20 問題」是反向柏拉圖法則：AI 能達到預期輸出的 80%，但將最後 20% 修正到生產品質所需的努力幾乎等同於完全手動完成。此問題在大型任務（複雜、難以分解）、不可分解任務，以及全有或全無品質任務（程式碼、法律文件）中尤為嚴峻。產品含義：彌合 80→100% 差距的包裝層創造了 80% 的價值。
Key Points
- Reverse Pareto: getting last 20% to production quality takes 80% of total effort when using AI
- Affected tasks: large + non-decomposable + all-or-nothing quality (code, legal, medical)
- Code is relatively more decomposable than prose → less affected; long-form prose is less decomposable → more affected
- Product opportunity: the wrapper/tool that takes AI output from 80% to 100% is where most value lies
- “All or nothing” quality: a contract with one wrong clause isn’t 95% good — it’s potentially harmful
- Decomposability is key: if you can independently accept each part, the 80% solution is actually 80% useful
Insights
- The “value is in the 80→100% bridge” insight reframes the AI product space: don’t compete with raw model capability, compete in the quality gap
- This directly explains why human-in-the-loop is valuable: the human’s role is specifically the last 20% verification/correction
- The code vs prose decomposability point is subtle but correct: reviewing and accepting individual functions is different from reviewing a legal contract holistically
Connections
- The 70% problem article covers the same territory from a coding/non-engineer angle; the 80/20 article provides the UX research framing
- Connects to Shreya Shankar’s DocETL: her tool is explicitly trying to solve the 80→100% gap for LLM-based document processing
- The PromptWizard and DSPy articles represent systematic attempts to close the 80→100% gap via automated optimization
Raw Excerpt
“Getting to 80% with AI is easy. The hard part is the last 20% — and for many tasks, that last 20% is not 20% of the work. It’s 80% of the work. And for some tasks — a legal contract, a safety-critical code path — 80% isn’t a useful deliverable. It’s a liability.”