Garry Tan Runs YC at 400x His 2013…

1. Key Themes

Theme 1: Context Is the New Moat — Rent the Model, Own the Memory

The central investment and operating thesis is that frontier AI models are rapidly commoditizing, but the context you feed them compounds over time into a durable, proprietary asset.

"Personal AGI is a different animal. An agent that runs on your infrastructure, reads from a memory you own, executes procedures you wrote, and compounds."

"Model quality is rented, and it gets cheaper every quarter. Your context is the only part of this stack you actually own, and the only part that compounds."

Theme 2: AI Is Structurally Breaking the Revenue-Per-Headcount Ceiling

YC portfolio data is surfacing a new economic regime where tiny teams generate revenues that have no historical precedent across any industry.

"Emergent, out of our summer 24 batch, went from public launch to 9 figures of revenue in 8 months."

"That revenue per person did not exist before. Not in software, not in oil, not in railroads."

Retell AI hit $60M annualized with ~40 people. Emergent hit $1M per head at $15M ARR. Both started with 1–2 people running AI-native workflows from day one.

Theme 3: The "Skill File" as the New Primitive Unit of Work

Plain-English instruction files — not code, not headcount — are becoming the atomic unit of organizational leverage. Non-technical employees are now effectively programmers.

"Markdown is the code now. The compiler is a language model."

"When a meeting recording lands, transcribe it with speaker labels, pull out the commitment made, who made it, and the deadline. Cross-check every person named against the library and link their pages. If anything contradicts something we already believe, flag it. Don't overwrite it."

A YC finance team member consolidated 100 Excel workbooks into a single internal tool using only plain-English instructions — no code written.

Theme 4: The Productivity Multiplier Is Already Significant — Even at Its Floor

Tan attempts to stress-test his own 400x productivity claim and finds a defensible floor that still has major implications for how individuals and organizations should budget capacity.

"It's still 8x at the absolute floor, and 10 times that in the middle of the range."

In 2013, Tan shipped ~14 useful lines of code per day — median for a working programmer. He now runs YC full-time and still hits this output while making school pickup most nights. The effect extends across design, product, and growth work — any domain where an agent runs the full loop.

Theme 5: Institutional vs. Personal AI — Who Holds the Repo Holds the Power

Tan frames skill file ownership as the defining labor question of the AI era, with direct implications for employment, IP, and organizational leverage.

"I believe skill files are yours. Own your skills because if you don't, your job becomes a skill file."

"The company keeps running her judgment without her. 40 files executing forever, and her name isn't even in the commit history. She didn't have a career. She had an extraction."

2. Contrarian Perspectives

Contrarian 1: AGI Has Already Arrived — It Just Looks Like a Folder of Text Files

The mainstream narrative is still waiting for a dramatic AGI announcement. Tan's take: it's already here and most people are looking in the wrong direction.

"Everyone is watching the sky, and the thing they're watching for is already in the room."

His evidence: a 220,000-page markdown knowledge base running for years, a father building an 80,000-page medical brain for his epileptic son with no lab, no grant, and no permission. The capability threshold has been crossed; what's missing is the practice of using it systematically.

Contrarian 2: Which AI Lab Built the Model Is Nearly Irrelevant

The popular obsession with model rankings and lab horse-racing misses what actually drives the 2x-to-100x performance gap between founders using identical tools.

"Same model, same context window, same API. Some founders get 2x and some get 100x on identical tools, and Tan says the gap comes down to what context the agent gets, and at what step it gets pulled in. Which lab built the weights underneath barely registers."

The implication for investors: model access is not a durable competitive signal. Workflow discipline and context architecture are.

Contrarian 3: AI-Generated Code Correlates With — Not Against — Startup Performance

The instinct in engineering culture is to distrust or discount heavily AI-generated codebases. YC's Winter 2025 batch data challenges this directly.

"A quarter of the companies had code bases that were 95% AI generated. That batch is on track to becoming one of the fastest growing, most profitable batches in the history of YC."

Tan is careful to note correlation, not causation — but the direction of the signal runs opposite to the conventional quality concern.

3. Companies Identified

Y Combinator

Description: Leading startup accelerator behind Airbnb, Stripe, and Coinbase
Why mentioned: Primary case study; Tan is its president and uses his Personal AGI system to run it at claimed 400x output
Quote: "He now runs YC full time, does a 5 o'clock pickup most nights, and still lands at roughly 400x that pace."

Emergent (YC Summer 2024)

Description: AI-native startup from YC's Summer 2024 batch
Why mentioned: Case study for new revenue-per-headcount economics; went from launch to 9-figure revenue in 8 months; hit ~$1M revenue per employee at $15M ARR
Quote: "Emergent, out of our summer 24 batch, went from public launch to 9 figures of revenue in 8 months."

Retell AI (YC Winter 2024)

Description: AI-native startup from YC's Winter 2024 batch
Why mentioned: Second case study for the new unit economics; $60M annualized revenue with ~40 people
Quote: "Retell AI, from the Winter 2024 batch, hit $60 million annualized with about 40 people, a similar ratio at more than triple the scale."

4. People Identified

Garry Tan

Description: President of Y Combinator; formerly a founder and partner at YC; built "G-Brain," a 220,000-page markdown-based personal knowledge system
Why mentioned: Central subject of the article; primary practitioner and advocate for the Personal AGI methodology
Quote: "He has run it in the open for years, 220,000 pages deep, and still makes kid pickup most nights."

5. Operating Insights

Insight 1: Never Let a Task Die Without Extracting a Reusable Skill File

The compounding effect of Personal AGI comes from converting every task into a persistent, rerunnable procedure — not from any single AI interaction.

"Never do one-off work. At the end of every task, ask the agent to skillify what it did. Turn it into a markdown file you can use and reuse forever. If you have to ask for something twice, you failed."

Tactical application: After completing any recurring task with an AI, prompt the model to document the procedure as a plain-English skill file. Wire it into a scheduled job. The standard is zero repeated manual inputs.

Insight 2: Use the "Smart Intern Test" to Validate Whether an Agent Can Execute

Tan's practical calibration for skill file quality is whether a human could follow the same instructions without ambiguity — if so, the model can run it reliably.

"Anyone who can read could follow it, and that's Tan's actual test for whether an agent can run it: if a smart intern could execute the instructions, so can the model."

Tactical application: Before deploying a skill file, read it aloud and ask whether an intelligent, motivated non-expert could execute each step without clarification. Vagueness at that test will produce failures in the agent run.

Insight 3: For Investors — Audit AI Workflow Discipline, Not Model Access

The due diligence signal that actually predicts the output gap is how a founder has structured their agent context and workflow, not which API they're calling.

"You get more signal from how a founder talks about AI in due diligence than from asking about model access. The 2x-to-100x gap Tan describes comes from context and workflow discipline, and it shows up in burn multiples before it ever shows up in a pitch."

6. Overlooked Insights

Overlooked Insight 1: The Context Window Is Already a Ceiling — And Closing It Requires a Library Architecture, Not a Better Model

Tan's point about the 1M-token context window sounds large until measured against an actual working life — and this gap is where the structural value of a curated personal library sits.

"Your life is not three books. Your life is a library. The question that determines whether your agent is a genius or a goldfish is this: who decides which three books are open on the desk?"

The 1,000-page (~1M token) window is the operating constraint today. The durable advantage goes to whoever has already done the curatorial work to decide which pages get surfaced — not to whoever has the largest context window.

Overlooked Insight 2: Compendium Skills Enable Asynchronous Expert-Level Research at Near-Zero Cost

Tan's Spinoza example reveals a specific, underutilized technique — the "compendium skill" — that runs deep synthesis across multiple long-form sources overnight, with citations and disagreement flagged automatically.

"My agent acquired 3 of the best biographies about the man, about 1,500 pages, read all three, and built me a synthesis: a dated chronology, every place the biographers disagree, the best verbatim quotes with chapter citations, and the 10 most tellable moments of his life ranked with delivery notes."

This technique is applicable well beyond talk preparation — to competitive intelligence, investment due diligence, technical literature review, and domain onboarding — and requires no specialized tooling beyond a capable model and a well-written skill file.