Garry Tan Says He’s 400x More Productive Now. Here’s the ‘Brain’ Behind It.

Garry Tan’s real argument is about rebuilding how a company works, from the org chart down. Here’s his talk translated for people who don’t write code.

On July 17, Y Combinator posted a talk to YouTube called “Every company should have a Brain.” It’s Garry Tan, YC’s president and CEO, speaking at AI Engineer World’s Fair in San Francisco, no slides, twenty minutes, answering a question from host Theo Browne: what should founders actually build right now? The talk has since pulled 25,000+ views. It’s dense with specific numbers and a framework that’s worth unpacking slowly if you don’t spend your day writing code, because the ideas underneath it apply to any company, not just software startups.

The claim that started it: “400x”

Tan opens with a number he says got him “torn apart on the internet.” In 2013, as a YC partner also working as a near-full-time engineer, he could write about 14 usable lines of code a day, roughly the norm at the time. Today, running YC full-time, with fewer working hours because he has a 5pm daily commitment to pick up his kid, he did the math on his output.

I did the math on my output, and it’s about 400X.”

If that sounds like an exaggeration, he addresses it directly. He tells the audience to discount the number as harshly as they want, assume the AI-generated code is bloated, assume he’s flattering himself, and even then: “It’s still 8X at the floor and 80X in the middle. That number is large, no matter how you torture it.”

Tan says the gain isn’t coming from a better AI model, and that’s the part that matters most for anyone outside engineering. “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.” In plain terms: everyone has access to roughly the same AI tools right now. The gap between people getting a little value and people getting enormous value comes down to how they’ve organized their work around it.

Translating his framework into normal English

This is the core of the talk, and it’s the part most likely to sound like jargon if you’re not technical, so it’s worth slowing down here.

Tan’s argument is that the building blocks of AI systems map directly onto the building blocks of a company:

• A “skill file” is an employee. It’s a written instruction, in this case a simple text document, that describes one specific job clearly enough that someone (or something) can actually do it.
• A “resolver table” is an org chart. It’s the system that decides which employee (skill file) handles an incoming task.
• “Filing rules” are your internal processes, the ones that determine whether work is actually being done correctly and consistently.
• “Trigger evals” are performance reviews. They’re checks that confirm a process is still working the way it’s supposed to.

Put together, his point is this: “Literally every part of an organization, the organization that you used to have to hire a thousand people for… they’re markdown files.” Markdown is just a simple, plain-text way of writing documents, nothing exotic. His summary line for what this means day to day: “You’re not writing software, you’re hiring, training, and managing a workforce made of markdown.”

The proof: real companies, real numbers

Tan backs the framework with specific portfolio results.

In YC’s Winter 2025 batch, a quarter of the companies had codebases that were 95% AI-generated, a year before this talk. Tan says that batch became “the fastest-growing, most profitable batch in the history of YC.” He also states that 94 companies in YC’s history have now crossed $100 million in revenue starting from nothing more than a seed check, a figure he gave directly in the talk that we weren’t able to independently verify through other sources, so take it as his claim, not an outside audit.

Two company examples he cites by name check out against outside reporting. Emergent, an AI app-building tool from YC’s Summer 2024 batch, went from public launch to $100 million in annualized revenue in eight months, confirmed independently by TechCrunch, and had just 15 people when it crossed $15 million ARR. He also mentions a Winter 2024 company at “$60 million with about 40 people” (the audio transcript renders the name as “Retail,” which is almost certainly a mis-transcription; we couldn’t confirm the exact company or those specific figures independently, but a company called Retell AI is confirmed to be in that same YC batch). Tan’s point either way: “That kind of revenue per head did not exist before. Not in software, not in oil, not in railroads, never.”

What a “company brain” actually is

The second half of the talk introduces Tan’s central metaphor, and it’s a useful one for a non-technical audience.

Human short-term memory, he points out, holds about seven items at once, a well-known finding from cognitive psychology (it’s why local phone numbers are seven digits). “Every institution humanity has ever built, every checklist, every org chart, every filing cabinet is a prosthetic for that limit.” An AI system, by contrast, can hold roughly a million tokens (a “token” is just a small chunk of text) of information active at once, which Tan translates into something anyone can picture: “The AI agent can keep about three Harry Potter books sitting open in its head all at once. And it can find a needle in any of them and synthesize across all three in seconds.”

A “company brain,” in his framing, is the system that decides which three books get opened for any given task, what he calls “the library plus the librarian.” He’s built his own version, GBrain, and open-sourced it for free. It’s grown fast: as of this week the project has crossed 26,800 stars on GitHub, a rough measure of developer interest, and people are already building on top of it. One example we found: a developer named Ghiles Mssoui built something called a “Taste Index” directly on top of GBrain, a system meant to give an AI agent judgment, not just memory. Tan’s own version holds roughly 220,000 pages pulled from his email, meetings, and 20 years of notes. His example of how it’s useful in practice: when a founder emails him about a crisis, “before I even finish reading that email, my agent has already pulled every prior conversation with that founder, three portfolio companies that hit the same wall, and what actually worked for those people.”

He’s honest about where this breaks

To his credit, Tan doesn’t just sell the idea, he names the ways it fails. “A brain nobody curates becomes a garbage dump with great search.” Bad information can get pulled up with total confidence. “A bad skill file encodes a bad process forever.” His fix isn’t more technology, it’s discipline: someone (a person, an AI, or both) has to actively maintain the system, check facts against each other, and prune what’s stale. Skip that step, in his words, and “you get a very confident agent that is wrong in ways nobody can trace.”

That caution is backed up by what we found looking at how other people are actually implementing this. One founder running a six-person AI-native shop, a company called Swan AI aiming for $10 million in revenue per employee, described burning through six figures in AI costs to get there, a real reminder that “revenue per head” gains come with real spending. On the other side, someone helping a traditional financial-services company go AI-native independently arrived at a nearly identical framework to Tan’s, listing “establish reliable context and explicit ownership before agents” as design principle number one, evidence that this framework is converging with what practitioners elsewhere are learning on their own, not just something YC is evangelizing top-down.

The one instruction he wants you to actually remember

If the whole talk boils down to one habit, it’s this: never do one-off work. Every time you or your team solves something with an AI tool, the job isn’t done when the answer is good, it’s done when that solution is saved as something reusable. Tan calls this “skillify it.” His reasoning: “If you have to ask for something twice, you failed.” A company that captures what it learns gets smarter every day. One that doesn’t, in his words, “wakes up every morning with amnesia, no matter how good the model is.”

The story he ends on

Tan closes the talk with a story that has nothing to do with startups. A friend of his has a son with a rare form of epilepsy. With no lab, no grant, and no institutional backing, he built an 80,000-document personal knowledge base, a “company brain” for one child, documenting everything known about his son’s exact condition. Tan’s point in telling it: “That’s not a side story. That is the exact architecture I’ve been describing… pointed at the thing this man loves the most in the world.”

His closing line to the room was simple: “Build the AI native company, not a company that just uses AI.” The distinction he’s drawing, translated for anyone running a business rather than writing code, is this: adopting AI tools makes your existing company faster. Rebuilding your company’s structure, roles, and institutional memory around AI changes what your company is capable of in the first place.

---

By Anthony Batt — 20+ years building software and digital media products at scale. Podcasting host at Future-Proof Podcast by CO/AI.