{
  "video": {
    "id": "eBUyTS7SzV4",
    "title": "Every company should have a Brain — Garry Tan, Y Combinator",
    "duration": 1268,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": "In person"
  },
  "analysis": {
    "video_id": "eBUyTS7SzV4",
    "title": "Every company should have a Brain — Garry Tan, Y Combinator",
    "one_liner": "Garry Tan argues that the 400X leverage founders are getting from AI isn't in the model — it's in wiring the work like an organization: skill files as employees, resolver tables as org charts, and a curated \"company brain\" that decides which three books are open in the agent's head.",
    "summary": "Tan claims his own coding output went from ~14 usable logical lines of code a day in 2013 to roughly 400X that today, and insists the gap between 2X and 100X people has nothing to do with model weights — everyone is using the same Claude. His core move is mapping agent primitives onto an org chart: a skill file is an employee, a resolver table is the org chart, filing rules are internal process, trigger evals are performance reviews — \"a workforce made of markdown.\" He then argues the real scarce layer is memory: agents hold about a million tokens (three Harry Potter books) while your company is a library, so the product is a curated company brain — library plus librarian — with provenance, contradiction checks and pruning. He is building his own, G brain, as MIT open source, and urges everyone to never do one-off work: skillify it.",
    "key_points": [
      "In 2013, as a near-full-time YC engineer, he wrote about 14 usable logical lines of code a day — consistent with the literature's ~15 median for that era; his measured output now is about 400X, and even under the most punishing verbosity/scaffolding penalties it is \"8X at the floor and 80X in the middle.\"",
      "\"It's not the model\": the 2X people and the 100X people use the exact same Claude — same weights, same context window, same API. The leverage is in how you wire the work.",
      "In YC's Winter 25 batch, a quarter of companies had codebases that were 95% AI-generated; that batch became the fastest-growing and most profitable in YC history. 94 YC companies total have crossed $100M in revenue from a seed check.",
      "The org-chart mapping: a skill file is an employee (one capability, written down); a resolver table (load tests.md when altering a test) is the org chart; filing rules are internal process; trigger evals that check whether test.md actually got loaded are performance reviews.",
      "Revenue-per-head examples: Emergence (S24 AI app builder) went from public launch to nine figures of ARR in eight months and was 15 people at $15M ARR; Retail (W24) is at $60M with about 40 people.",
      "The transformation is not just engineers — YC's media, events and finance staff build skill files and cron jobs; one finance person who isn't a programmer collapsed ~100 Excel workbooks into a single app using the internal Open Claw and company brain.",
      "Know which side of the equation computation lives on: latent space (taste, judgment, understanding vague human intent, steered by markdown) vs deterministic space (code). Most AI engineering bugs are work happening on the wrong side — e.g. the 800-seat array for Startup School clustering must not live in the context window, letting a month of work become ~10 minutes and a couple hundred dollars of tokens.",
      "Humans hold 7±2 things in working memory; an agent holds a million tokens, about a thousand pages, \"three Harry Potter books open in its head at once\" — but your company is a library, so context engineering is deciding which three books are open. Failure modes of a company brain: a garbage dump with great search, stale facts surfaced with total confidence, and a bad skill file encoding a bad process forever — the fix is memory plus hygiene (provenance on every fact, contradiction checks, a human+agent librarian whose job is pruning)."
    ],
    "takeaways": [
      "Never do one-off work: when you're happy with an agent's output, skillify it into a reusable skill file (he has an X post you can find by searching \"skillify it\"). If you have to ask for something twice, you failed.",
      "Stop treating AI as autocomplete and treat it as a workforce — encode sales, support, ops and finance procedures as skill files, and hire engineers to maintain the skills and do what the skills can't do yet.",
      "Audit every agent bug by asking where the computation is happening: push storage and exact state into deterministic code, keep taste, judgment and ambiguous-human-intent work in latent space.",
      "Build the memory layer as production infrastructure, not a dumping ground: provenance on every fact, contradiction checks when new information collides with old, and someone whose actual job is pruning.",
      "Don't wait for the right tooling — the concepts travel to any stack. Open Claw is the Ferrari, Codex is \"a really good Honda\" that will do 90% of this."
    ],
    "topics": [
      "agents",
      "context-engineering",
      "skills",
      "memory",
      "rag",
      "ai-native-companies",
      "startups",
      "evals"
    ],
    "tools": [
      "Claude",
      "Claude Code",
      "Codex",
      "Open Claw",
      "Hermes agent",
      "G brain (MIT open source)",
      "Y Combinator",
      "Emergence",
      "Retail",
      "Postgres",
      "TypeScript",
      "Elixir/Erlang",
      "Excel",
      "X",
      "Startup School",
      "Linux"
    ],
    "quotes": [
      {
        "text": "The 2X people and the 100X people are using the exact same Claude. Same weights, same context window, same API.",
        "at": "03:07",
        "url": "https://www.youtube.com/watch?v=eBUyTS7SzV4&t=187s"
      },
      {
        "text": "When you sit down with Claude Code or Codex, you're not writing software, you're hiring, training, and managing a workforce made of markdown.",
        "at": "05:58",
        "url": "https://www.youtube.com/watch?v=eBUyTS7SzV4&t=358s"
      },
      {
        "text": "Retrieval is easy. Being worth retrieving from is the product.",
        "at": "13:20",
        "url": "https://www.youtube.com/watch?v=eBUyTS7SzV4&t=800s"
      },
      {
        "text": "Model quality is rented, but if you build your brain, you own that brain.",
        "at": "16:35",
        "url": "https://www.youtube.com/watch?v=eBUyTS7SzV4&t=995s"
      }
    ],
    "words": 4101
  },
  "summary_url": "/#eBUyTS7SzV4",
  "transcript": {
    "html": "/transcripts/eBUyTS7SzV4.html",
    "txt": "/transcripts/eBUyTS7SzV4.txt",
    "vtt": "/transcripts/eBUyTS7SzV4.vtt"
  }
}