Garry Tan argues the leverage in AI engineering is not in the weights but in how you wire the work — the 2x and 100x people run the same Claude. An AI-native company encodes its org as skill files, resolver tables, and a curated company brain (the library plus the librarian) that compounds every day, and he urges founders to never do one-off work — skillify it instead.
Tan frames YC as a 20-year-old institution becoming AI-native, chasing companies where "one person does what used to take a thousand people" — mechanically, not as a metaphor. He grounds it in his own history: in 2013 as a YC partner and near-full-time engineer, he could write about 14 usable logical lines of code a day, roughly the median for that era. Running YC full-time today, he calculates his output at about 400x — and preemptively deflates his own claim, noting that even under the most pathological verbosity penalty it is still 8x at the floor.
The part he says he'd "tattoo on the inside of everyone's eyelids" is that this leverage is not in the model. The 2x people and the 100x people are using the exact same Claude — same weights, same context window, same API. The difference is entirely in how you wire the work.
With no slides, Tan lays out a direct mapping from software agents to org design. A skill file is an employee — one capability, one job written down clearly enough to execute. A resolver table (the thing you build when CLAUDE.md gets too big) is an org chart: a task comes in and the resolver decides who handles it. Filing rules are your internal process, and trigger evals — tests that check whether the right file actually loads — are performance reviews.
The upshot: the organization you used to hire a thousand people for is now markdown files (plus a little TypeScript). "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." He cites YC companies operating on this new physics — an AI app builder from the Summer 24 batch that went from public launch to nine figures of ARR in eight months (only 15 people at $15M ARR), and a Winter 24 company at $60M with about 40 people.
Tan uses cognitive psychology as the frame: humans hold about seven things in working memory (7±2), and every checklist, org chart, and filing cabinet is a prosthetic for that limit. An AI agent holds about a million tokens — roughly a thousand pages, or three Harry Potter books open at once, able to find a needle in any of them and synthesize across all three in seconds. But a company isn't three books; it's a library of every email, meeting, decision, and post-mortem. The question that decides whether your agents are geniuses or goldfish is who chooses which three books are open — that's context engineering, and that's a company brain.
He concedes the skeptic's point that this "is just RAG," replying that retrieval is the primitive the same way Postgres is just B-trees — the hard part is everything around it: what gets written down, how it's enriched and linked, what's promoted to hot memory, and who arbitrates when facts disagree. He's building his own in the open, GBrain: an MIT-licensed, harness-agnostic retrieval layer, a ~220,000-page warehouse written mostly by his agents from 20 years of email, meetings, and notes — his second brain.
Tan stress-tests his own pitch: company brains have failure modes. A brain nobody curates becomes "a garbage dump with great search," retrieval will surface a stale fact with total confidence, and a bad skill file encodes a bad process forever. So the primitive is not memory — it's memory plus hygiene: provenance on every fact, contradiction checks when new information collides with old, and a librarian (human plus agent) whose job is pruning. Treated like production infrastructure it compounds; treated like a dumping ground it yields a confident agent that's wrong in untraceable ways.
His signature discipline: never do one-off work. When you're happy with an agent's output, "skillify it" — turn it into a reusable skill file — because if you have to ask for something twice, you failed. "Model quality is rented, but if you build your brain, you own that brain."
Answering Theo's earlier question head-on, Tan says: build the AI-native company, not a company that merely uses AI — a thin team, skill files for everything, the founder still in the code, and a compounding company brain from day one. Tooling is secondary: OpenClaw is the Ferrari he'll always recommend, but Codex is a really good Honda that will do 90% of it. The concepts travel with you to any stack.
The greenfield he'd chase if he were 25: every company on earth is about to need a brain — the memory layer that means you never have to re-ask what you already knew — and he'd fund the team that builds the defining one. He closes on abundance as "shipped software," citing a friend with a rare form of epilepsy who built an 80,000-markdown-file company brain for his son's condition — a father, a laptop, and a library — as proof the architecture is real and available now.
It's not the model. The 2x people and the 100x people are using the exact same claude. Same weights, same context window, same API. So the leverage is not in the weights, it's in how you wire the work.
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.
Retrieval is easy. Being worth retrieving from is the product.
Model quality is rented, but if you build your brain, you own that brain.
| Time | Topic |
|---|---|
| 00:00 | Becoming AI-native: one person doing a thousand people's work |
| 01:34 | The 400x number — and why it isn't the model |
| 03:41 | YC batch data: 95% AI-generated codebases, 94 companies past $100M |
| 04:48 | Wiring the work: skill file = employee, resolver table = org chart |
| 07:57 | Non-engineers building skills; where the computation happens |
| 11:07 | Working memory, three Harry Potter books, and GBrain |
| 13:12 | The company brain: library plus librarian, RAG and hygiene |
| 15:16 | Never do one-off work; build the AI-native company; boil the ocean |