Borrowed intelligence does not compound — what Garry Tan said on stage, and what he left out

It isn't where everyone is looking

In August 2026, on stage at Startup School 2026, Y Combinator president and CEO Garry Tan opened with Spinoza.

Excommunicated from his community at 23, never absolved. The substance of the heresy: God is not a king on a throne but something diffused through everything that exists. Tan skips four hundred years and lands on his own claim. We are making the same mistake about intelligence.

Everyone is looking at the sky, waiting for AGI to arrive. Some day a threshold gets crossed, an announcement goes out, the color of the sky changes. But the thing being waited for is already in the room, and it doesn't look like a god. It looks like infrastructure. A terminal window, a folder of markdown files, a job that finished while you slept. It arrives diffused through everything.

AGI does not arrive as an event. It arrives distributed — as agents that run on your context and do your work. He calls it personal AGI: not general intelligence for all of humanity, but general intelligence for one person.

He then warns that the phrase has already been seized by marketing departments. This is not a $20-a-month chatbot. It is not slightly smarter autocomplete. It is not an assistant that knows only your calendar. That is a subscription you rent, not something you own.

1. The library and the librarian

At the center of his argument is the most famous number in cognitive psychology.

Human working memory is 7±2. About seven things at once — which is why local phone numbers were seven digits and why you forget the eighth item on a shopping list. And Tan says every institution humans have built — checklists, org charts, filing cabinets, standups — is a prosthetic for that limit.

An agent holds a million tokens. Roughly a thousand pages: three Harry Potter books open on the desk simultaneously. And it can find the needle inside them and synthesize across all three.

Seven digits against three books. Whether that counts as AGI is arguable, but it is already a different operating regime, he says. And nearly everyone on earth is still running their life on org charts and working habits designed for a seven-digit brain.

Then he runs the same number in reverse. A thousand pages is a lot, and it is also almost nothing. Your life is not three books. It is a library. Every email you sent, every meeting, every decision and the reason behind it, every conversation with everyone you know.

Which produces the most important question in the talk.

What decides whether your agent is a genius or a goldfish is who — or what — decides which three books are open on the desk.

The library, and the librarian. His system is those two things. His own wiki runs to roughly 220,000 pages of markdown, twenty-five years of a life kept as a journal. When a crisis email arrives from a founder, before he has finished reading it the agent has pulled his entire history with that person plus three portfolio companies that hit the same wall and what they did about it. He called that the difference between an assistant and a colleague.

2. Borrowed intelligence, owned memory

The equation he put on stage is simple.

A frontier model (rented, commoditized, cheaper every quarter) + your context (owned, specific, held by nobody else on earth) + a harness that connects the two.

Then he draws the line between an AGI you don't own and one you do. The company's AGI gets better only when that company ships. Close the tab and it resets. It knows what everyone else already knows. And when the company changes direction, it gets lobotomized on someone else's schedule.

Your own AGI gets better with every day you use it, because every day it knows more of your life.

One is a product you consume. The other is an asset you build.

He puts numbers on it. In 2013, writing internal tools at night as a YC partner, his output was 14 lines a day — the actual median in the programmer productivity literature. Today, while running YC full time, same brain, same hours, roughly 400× that. Then he discounts his own number as hard as he can: apply the most pathological verbosity penalty, assume half of it is scaffolding, assume he is grading himself generously, and the floor is still 8×. In the middle of the range it's ten times that.

The same thing is happening across YC. In the Winter '25 batch, a quarter of the companies had 95% of their codebase AI-generated. That batch is becoming one of the fastest-growing and most profitable in YC's history. Careful about correlation and causation, he restates it: the fastest-growing founders treat AI as labor, not autocomplete.

Same Claude, same weights, same context window — and some people get 2× while others get 100×. The difference isn't in the weights. It's in what context you hand over, how relevant it is, and in what order.

3. The parable of Maya

Late in the talk he says the next part isn't fun, and tells a fictional story. It is the heaviest passage in the whole thing.

A support engineer named Maya. Over two years she teaches her agent forty skills. How to triage a P0 at 2 a.m. How to calm a customer on the edge of churning. How to write a postmortem that actually prevents the next incident. Forty files. They are the judgment she built over two years, laid down on disk.

Version one: the files live in Maya's repo. When she changes jobs they go with her. From day one at the new company she operates with years of compounded judgment. It compounds for as long as she works. That is ownership.

Version two: the files live in the company's repo, under the company's IT policy. Maya leaves with nothing. The company keeps running her judgment without her. Forty files execute forever and her name isn't even in the commit history.

Same files, same Maya. One variable — who holds them.

His closing line: what she had wasn't a career. It was extraction.

Then he draws the lineage. Craftsmen owned their tools, and that is what made them free. The factory broke it — the loom belonged to the factory. Knowledge workers assumed they were safe, because their tools were in their heads and nobody could confiscate them. Skill files end that. For the first time your cognition can be extracted, stored, versioned, and owned. The only question is by whom.

4. Three objections and his answers

He pre-empts three objections. They work as a map of the arguments in this space.

"Once models get good enough, the harness becomes unnecessary. Just wait for the next release." → Look at what actually happens with each release, he says. The better models get, the more the differentiator moves to context. If everyone's engine makes 1,000 horsepower, the race is decided by the driver and the map. The weights belong to everyone; the library is yours. A sharper reader gets more out of the same book, so model progress raises the value of your library.

"Isn't this just RAG?" → Yes, and Postgres is just a B-tree. Retrieval is a primitive, not a product. The hard part is everything around it: what gets written down at all, how it gets reinforced and linked, what gets promoted to hot memory versus filed as a cold reference, and who adjudicates when two facts disagree. Retrieval is easy. Being in a state worth retrieving is the product.

"What if you put your whole life in one place and it leaks?" → He calls this the objection that deserves the most respect, and gives the same answer as the rest of the talk: that is exactly why it has to be yours. The default is not privacy. The default is your life already scattered across ten clouds, searchable by everyone except you. Aggregating it didn't create the danger. It accepted custody. Custody is the security model.

5. "A brain nobody cleans is a highly searchable garbage dump"

He leaves one honest weakness in his own prescription.

A brain nobody tends becomes a highly searchable garbage dump. Retrieval will hand you a stale fact with total confidence, and a bad skill file locks a bad process in permanently. So what's needed isn't just memory but hygiene: provenance on every fact, contradiction checks when new information collides with old, and a librarian whose job is pruning.

Treat it as production infrastructure and it compounds. Treat it as a dump and you get an agent that is confidently wrong in ways nobody can trace.

6. And then he said most people quit in week two

That is the argument. Here is why we wrote this.

Tan gives the prescription too. Tonight, pick a harness and run an agent on your own machine. This weekend, make one folder of markdown and write a page each about the projects and people you're involved with. Write your first skill file. Wire it to a scheduled job. And the last one — what he calls the discipline that separates the people who compound from the people who dabble — never do one-time work. Every time you finish something, convert what you did into a skill file.

Then he forecasts the next ninety days. Week one is honestly a toy and costs more in fixes than it returns. Week four is when the flywheel catches. By week twelve you have a library that answers before you finish asking, a dozen or so skill files running the weekly work you used to hate, and one or two tools other people start asking to borrow.

And then he adds this.

Most people who try this quit in week two. Which is exactly why the people who don't feel like they're cheating by week twelve.

The president of YC stood on stage and named the dropout rate of his own prescription. That is the gap.

There is nothing dishonest in the talk. It may be the most honest presentation in this space. But list what the prescription assumes and the outline gets sharp. A terminal. A GitHub repo. Markdown you write yourself. Choosing a harness. Configuring scheduled jobs. Running contradiction checks and pruning as an ongoing practice. In short: it is for people who can assemble it themselves.

He says this is an asset everyone should have, and in the same talk that most people won't keep it up. Those two statements don't contradict each other. The need is universal. The implementability isn't.

The diagnosis went mainstream. The prescription arrived. But the prescription is still a personal skill.

And a field that stays a personal skill does not stay open for long.

Coda — authority is borrowed, memory is built

Put Tan's argument next to what Sam Altman said on another podcast shortly before, and the industry's position gets fairly precise.

Altman gave the diagnosis: what separates a good decision from a worse one is not intelligence but the context in hand at that moment — and he forgets it. Tan gave the prescription: so own your context, don't rent it.

That the two of them are looking at the same place is no longer arguable. And both stop in front of the same question — where the memory goes. Altman withholds an answer. Tan answers "put it in your own repo," and the set of people who can execute that answer is small.

Model intelligence will be fully commoditized within a few years. Every product will pull the same grade of intelligence from the same APIs. What's left between products at that point is not intelligence. It's whose context that intelligence runs on.

Intelligence is borrowed. Tomorrow it gets swapped for something stronger. What accumulates is memory: it survives the swap and it thickens with time. And — this is the part that matters — as long as it stays on your machine, it stays yours.

Tan built his inside a terminal. We are building ours somewhere you don't have to open one.

ShogunAI — Your AI has memory. Now it acts.

Source: Y Combinator, Garry Tan: Own Your Intelligence (Startup School 2026, published August 6, 2026).