本文由 AI 分析生成
建立時間： 2025-12-16 來源： https://x.com/sagacity/status/2000752125970981265
Summary
Reflecting on Andrej Karpathy’s blog, the author distills two ideas. First, Karpathy’s experiment auto-grading a full month of 2015 Hacker News threads with hindsight — scraping 31 days x 30 posts and having a model output structured “after-the-fact analysis” — shows AI now lets us cheaply run thought experiments that were previously impossible. Second, and more pointed: verifiability is a ruler for which jobs get transformed or replaced first. Math, Go, programming, and translation are verifiable; creative writing, negotiation, strategy, leadership, and research taste are not.
作者讀 Andrej Karpathy 的部落格後提煉出兩個想法。其一，Karpathy 用「事後諸葛」視角自動為 2015 年整整一個月的 Hacker News 帖子打分——抓下 31 天 × 每天 30 帖，讓模型輸出結構化的事後分析——顯示 AI 讓過去做不到的思想實驗變得廉價可行。其二更尖銳：可驗證性是判斷哪些職業會先被改造或替換的一把尺。數學、圍棋、編程、翻譯是可驗證的；創意寫作、談判、策略、領導與研究品味則不是。
Key Points
- Karpathy scraped a full month of 2015 Hacker News front pages and had a model retroactively grade posts and commenters.
- AI now makes previously-impossible thought experiments cheap to actually run and verify.
- Verifiable work (math, Go, coding, translation) is most exposed to AI replacement.
- Hard-to-verify work (creative writing, negotiation, strategy, leadership, research taste) resists replacement, partly because it can’t be environment-reset.
- “Be good — the future LLM is watching you” — your public traces may be graded later.
Insights
The verifiability lens reframes job-displacement risk in terms of RL-style training feasibility: tasks with cheap, resettable reward signals get optimized away first, while tasks lacking a ground-truth grader stay human longer. The half-joking “future LLM is watching” point has teeth — Karpathy’s HN experiment demonstrates that today’s models can retroactively score decade-old public reasoning, so the quality of one’s recorded judgments becomes a durable, gradeable asset.
Connections
Raw Excerpt
可验证的工作最快被替换，be good，未来的 LLM 正在看着你