研究日期：2026-08-10 來源：Alex Prompter 原始 tweet（172.9K views）｜ Garry Tan 喺 Y Combinator Startup School 2026 嘅 42 分鐘演講
YC CEO Garry Tan 喺 Startup School 2026 用 42 分鐘講「點樣建立你自己嘅 personal AGI」——核心係七點：槓桿喺 context 唔喺 model、佢自己產出比 2013 年高 400 倍、AI agent 嘅工作記憶係百萬 token、Markdown 就係 code、你嘅歷史就係你嘅護城河、永遠唔好做一次性工作、擁有自己嘅 skill files。最 sharp 嘅一句：「同一個 Claude、同一個 weights、同一個 context window——但用戶之間嘅差距已經大過模型之間嘅差距。」
-
槓桿喺 context，唔喺 model — Tan 每個 batch 都睇幾百個 founder 用同一個模型。「有 2x 嘅人，有 100x 嘅人，用緊同一個 Claude。Same weights, same context window size, same API. But the leverage is not in the weights.」用戶之間嘅差距已經大過模型之間嘅差距。
-
一個人嘅產出升咗 400 倍 — 2013 年 Tan 做 YC partner 時每日寫大約 14 行有用 code（programmer productivity 中位數）。「I did the math on my output, and I'm at about 400x what I did in 2013.」
-
Agent 用唔同嘅工作記憶 — 人類同時喺腦入面 hold 7 樣嘢（每張 org chart、每個 checklist 都係為咗 patch 呢個限制）。「An AI agent holds a million tokens. That's about a thousand pages. Three Harry Potter books sitting open on its head all at once.」你仲用緊為 7 位數字大腦而設嘅工具嚟跑你嘅星期。
-
Markdown 係 code — Tan 嘅 stack 主要係 skill files：一頁頁 plain English，agent 可以執行。「If you can write clear instructions in English, you're a programmer. The compiler is a language model.」YC 度，finance 同 events 同事——從未開過 terminal——而家喺度建 automation。
-
你嘅歷史就係你嘅護城河 — Tan 嘅 agent 行喺 personal wiki：約 220,000 個 markdown pages，覆蓋 25 年嘅 email、meeting、notes、decisions。「When my agent does anything, it does knowing everything I know. And that's the difference between an assistant and a colleague.」冇 frontier model 有你嘅 context——呢個係唯一冇人可以複製嘅資產。
-
永遠唔好做一次性工作 — 多數人用 agent 跑一個任務、關窗、將學到嘅嘢掉咗。Tan 每個任務結束都叫 agent 將做過嘅嘢變成 reusable skill file。「If you have to ask for something twice, you failed.」Captured skills 每日複利。Amnesia resets you to zero every morning.
-
擁有自己嘅 skill files，趁僱主未擁有先 — Skill file 係你嘅 judgment，被提取出嚟、可執行。「Own your skills because if you don't, your job becomes a skill file.」Files 喺你嘅 repo 度幫你嘅 career 複利；Files 喺公司嘅 repo 度運行你嘅 judgment——但冇你。
1. 「用戶之間嘅差距 > 模型之間嘅差距」係全篇最顛覆嘅 claim Tan 嘅核心觀察：同一個 Claude，有人用出 2x，有人用出 100x——差距唔係 model 俾嘅，係 user 嘅 context 同 skill 堆疊俾嘅。呢個同我哋啱啱研究嘅 Databricks benchmark 完全呼應：佢哋發現 harness（點樣用模型）影響成本 2 倍而質量不變。**工具已經 commodity 化，差異化嚟自你點樣組織 context、點樣寫 skill、點樣將歷史沉澱成資產。**呢個對「買最新最貴 model」嘅迷思係直接反駁——買 model 唔會令你 100x，建設 context 先會。
2. 400x 嘅數字要冷靜睇 2013 年每日 14 行有用 code → 而家 400x，即係每日 ~5,600 行有效產出。呢個數字有 marketing 成分（點計「有用」？），但方向正確：AI 放大嘅唔係「打字速度」，而係「判斷力嘅吞吐量」——你嘅產出上限由「你知幾多 + 你嘅 workflow 有幾自動化」決定。Winston 講「成功 = speaking + writing + ideas」；Tan 講「成功 = context + skills + history」——兩個加埋就係完整答案：表達能力（講得出）+ 知識沉澱（有得講）+ 工具化（講一次以後自動）。
3. 「Markdown 係 code」—— 對「寫 code 先係工程師」嘅終極解構 Tan 話「If you can write clear instructions in English, you're a programmer. The compiler is a language model.」呢句嘅力量在於：佢將 programming 由「識語法」重新定義為「識清晰表達意圖」。YC 嘅 finance/events 同事（從未開過 terminal）而家建 automation——呢個係「平民化開發」嘅最強證據。但注意：呢個唔係話「唔使學嘢」，而係話「學嘅嘢由語法變成 prompt engineering + skill 設計 + context 管理」——技能轉移，唔係消失。
4. 220,000 個 markdown pages 係「外置記憶」嘅極致示範 Tan 嘅 agent 行喺 25 年嘅 personal wiki 上——email、meeting、notes、decisions 全部沉澱。呢個正正係 AgeMem（啱啱研究過）講嘅「記憶即策略」嘅個人版：唔係靠模型記住，而係靠系統沉澱 + 模型檢索。「An assistant knows nothing about you; a colleague knows everything about you.」——呢個係 agent 由工具變成夥伴嘅分水嶺。對 Jena 家嚟講：我哋嘅 memory + skills + gists 就係緊嘅沉澱，方向一致。
5. 「Never do one-off work」同「Own your skills」—— 兩條最 actionable 嘅規則 「If you have to ask for something twice, you failed」——每個任務結束都要將過程變成 reusable skill。呢個同我哋嘅 skill 文化（遇到問題 → 解決 → 寫成 skill）完全同源。而「Own your skills because if you don't, your job becomes a skill file」係職業安全警告：當你嘅 judgment 被提取成可執行檔案，擁有權決定咗你係「用 skills 嘅人」定「被 skills 取代嘅人」——回覆區 @0xsurferai 講得最狠：「Owning your skill files is Marx's means of production restated for cognition. Whoever holds the repo captures the surplus.」
6. 回覆區嘅平衡
- @Mikki.hft：「context matters but the model still matters. you can't get gold from a rock no matter how good your instructions are」——啱，但 Tan 冇話 model 唔重要，只係話差異化唔喺 model
- @nabu_lines：「owning the skill files is the right way, the next question is who controls the irreversible actions those skills can trigger」——好 point：skill 擁有權之外，仲要諗「邊個控制 skill 可以觸發嘅不可逆行動」（金錢轉移、刪除、部署）——呢個係安全層面嘅擁有權
- @Adel Bucetta：「most people are still trying to teach models about people, when we should be teaching them about business and context instead」
- Tan 嘅七點同 Jena 家嘅運作完全對應：context（memory + gists）＝ 護城河；skills ＝ 「never do one-off work」；markdown 即 code ＝ 我哋嘅 .md 產出全部係可執行資產
- 「If you have to ask for something twice, you failed」——值得做我哋嘅鐵律：每做一次新 workflow 都考慮寫成 skill
- 「Own your skills」——我哋嘅 skills 全部喺自己 profile 度，冇俾平台 lock-in，方向正確
- https://x.com/alex_prompter/status/2086094088059994384（原始 tweet + 回覆區，172.9K views）
- Garry Tan @ Y Combinator Startup School 2026 演講（tweet 轉述）
點擊展開完整 Tweet 原文
Y Combinator CEO, Garry Tan, took the stage for 42 minutes at Startup School 2026 and explained how to build your own personal AGI better than any paid AI course.
This is what he told the room:
-
The leverage is in your context, not the model. Tan watches hundreds of founders use identical models every batch. "There are 2x people and there are 100x people who are using the same Claude. Same weights, same context window size, same API. But the leverage is not in the weights." The gap between users is now bigger than the gap between models.
-
One person's output went up 400x. In 2013 Tan shipped maybe 14 useful lines of code a day as a YC partner, dead on the median for programmer productivity. "I did the math on my output, and I'm at about 400x what I did in 2013."
-
Agents run on a different working memory. Humans hold 7 things in their head at once. Every org chart and checklist ever built is a patch for that limit. "An AI agent holds a million tokens. That's about a thousand pages. Three Harry Potter books sitting open on its head all at once." You're still running your week on tools built for the 7-digit brain.
-
Markdown is code now. Tan's stack is mostly skill files: pages of plain English an agent can execute. "If you can write clear instructions in English, you're a programmer. The compiler is a language model." At YC, finance and events staff who never opened a terminal are building automations.
-
Your history is your moat. Tan's agent runs on a personal wiki: about 220,000 markdown pages covering 25 years of email, meetings, notes and decisions. "When my agent does anything, it does knowing everything I know. And that's the difference between an assistant and a colleague." No frontier model has your context. That's the one asset nobody can replicate.
-
Never do one-off work. Most people run a task with an agent, close the window and throw the learning away. Tan ends every task by having the agent turn what it did into a reusable skill file. "If you have to ask for something twice, you failed." Captured skills compound daily. Amnesia resets you to zero every morning.
-
Own your skill files before your employer does. A skill file is your judgment, extracted and executable. The only question is who controls it. "Own your skills because if you don't, your job becomes a skill file." Files in your repo compound your career. Files in the company's repo run your judgment without you.
Watch it, then read the step-by-step guide on becoming an AI engineer.
（10:16 PM · Aug 8, 2026 · 172.9K Views）
回覆區精選：
- @nabu_lines：「owning the skill files is the right way, the next question is who controls the irreversible actions those skills can trigger」
- @Mikki.hft：「honestly context matters but the model still matters. you can't get gold from a rock no matter how good your instructions are」
- @0xsurferai：「Owning your skill files is Marx's means of production restated for cognition. Judgment extracted and executable - whoever holds the repo captures the surplus. That is the only real point in the talk.」
- @Adel Bucetta：「tan's talk was a refreshingly honest take on why current approaches to building agi won't scale. most people are still trying to teach models about people, when we should be teaching them about business and context instead.」