The New Physics of Business — Garry Tan, Y Combinator

Okay, great. Hey everyone, how's everyone doing? Allright, are we ready for the revolution? Okay, Theo just asked theright question: "What do we build now?" I'm going to answer it from the other side of the table. Uh, I'm a founder, I'm an investor, and I run a 20-year-old institution that is becoming AI nativeright now, which is a strange and wonderful thing to do to a 20-year-old institution.

Uh, and I'll spend about 20 minutes, um, talking about, you know, what YC is. We're trying to build companies where one person does what it took to, to—one person does what used to take 1,000 people. And I don't mean that as a metaphor, I mean that me-mechanically, this year, the people in this room will do this.

In about an hour, some of you will walk into the startup battlefield, and I want you to walk in knowing what's actually possibleright now, because what is possible now is much, much bigger than what people believe. So let me start with a number and, uh, you know, I got torn apart on the internet for this, but I'm going to say it again in front of all of you anyway.

Uh, this is the one room in the world that will stress test it, and I'd rather stress test it with you myself. In 2013, I was a YC partner, uh, building the internal social network, um, at, at YC.

Uh, I was also investing in companies, but, you know, I was also a near-full-time engineer. Um, and when I was doing that, I could maybe do about 14 usable, logical lines of code a day. Take out the comments, take out all the bullshit, and that's how many lines of code, uh, I was writing.

And if you look at the literature from that era, um, you know, that's kind of normal. Like, some people write 15, some people write 50. It was not, you know, the thousands of lines of code that I know a lot of you in this room are actually writing now, per day.

Um, that's about median 15. That was me at full effort at that time. This year I run, uh, YC full-time. Uh, same person, same hours, actually way less hours, weirdly. Uh, you know, but I have a 5:00 p.m.

kid pickup now, and I did the math on my output, and it's about 400x. Now, before the skeptic in the third rowright there deflates the number for me, let me deflate it myself. If you don't trust the raw code, well, fine, take the most pathological verbosity penalty you can stomach and assume the agent writes bloated code.

Assume half of it is scaffolding. Assume I'm flattering myself. It's still 8x at the floor and 80x in the middle. That number is large, no matter how you torture it. And here's the part that matters, the part that I'd tattoo on the inside of everyone's eyelids if I could.

It's not the model. The 2x people and the 100x people are using the exact same cloud, same weights, same context window, same API.

So the leverage is not in the weights, it's in how you wire the work. And it's not just me. At YC, we see this all the time. In the Winter '25 batch, a quarter of the companies had codebases that were 95% AI generated, and that was a year ago.

That batch has become the fastest-growing, most profitable batch in the history of YC. 94 companies, total, have now crossed 100 million in dollars in revenue from a seed check in the history of YC. So, uh, I think we know what we're talking about here.

And I can't prove that the AI-generated code caused the growth, but what I can tell you is the fastest-growing founders we fund are not treating AI as auto-complete. They're treating it as a workforce. The companies that wired the work differently are the ones that are bending the curve.

So what does wiring the work really mean?

Those slides.

This is the heart of the talk, and this is what I most want you to steal. Everything we've learned building with agents maps to an organization.

Can we still slide?

Those slides.

Sorry, there's no slides. I have no slides. I'm so sorry.

A skill file is an employee. It has one capability, one job, written down clearly enough that someone can execute it. Uh, a resolver table, uh, the thing that many of you, you know, when you run into Claude code and it says your context is too big in Claude.md, you run off and create a resolver table.

Well, you know, it—and if you don't know what that is, it's literally like whenever you need to, uh, alter a test, load tests.md, you have a whole table of these things. Um, that's an org chart. A task comes in, and the resolver decides who handles it and where it goes.

Uh, filing rules are your internal process. So this can be, um, you know, whether or not the resolver is actually working and, uh, is, you know, is there an—is it actually in compliance? And, uh, trigger evals. So going in and actually having a test that says, "When I need to alter a test file, does test.md actually get loaded?"

Those are performance reviews. So, you know, what have we done? Like, literally every part of an organization, the organization that you used to have to hire 1,000 people for, I just told you what those things are. They're Markdown files and other types of Markdown files.

And maybe there's some TypeScript in there too. We've been building organizations this whole time, but we didn't have a management layer. But now that's what we have. 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.

Uh, and, you know, that's—there are tons of companies at YC that are doing this. Uh, Emergence, an AI app builder out of Summer '24, they went from public launch to nine figures of ARR in eight months. Uh, when they crossed $15 million ARR, uh, they were only 15 people.

Retail, out of Winter '24, it's at $60 million with about 40 people. The kind of—that kind of revenue per head did not exist before. Not in software, not in oil, not in railroads, never. These are not freaks of nature.

They're just the first companies built natively on the new physics. And so how do companies like that actually run? Not by hiring hundreds of people for sales, support, ops, and finance. The AI-native companies that I see inside YC encode all of that as skills-written procedures that their agents execute, and they hire—they hire engineers whose job it is to maintain those skills, to do the work the skills can't do yet.

That is an AI-native company, and it's not a thought experiment. It'll actually file your taxes if you have a skill file for it. Now, picture YC's batch room. Like, it actually kind of looks like this. Um, 400 companies or 400 founders at long tables, and, uh, you know, I can imagine every single one of you, each of the laptop, every single day, you're doing, uh, a former person's entire year worth of work in a single day.

That's not the future. That's actually the barright now. And if you're not doing it, your competitor is, and they will eat your lunch politely and thank you for it. Here's the extension, uh, most engineering talks miss. It's not just the engineers.

Um, at YC, uh, as we make our transformation, it's our media people, our events staff, our finance team, people who have never opened a terminal in their lives are building skill files and cron jobs. And, uh, one of our finance folks just collapsed about 100 Excel workbooks into a single app she built with, uh, our internal OpenClaw and company brain.

She's not a programmer. She's a manager of agents now, and everyone at YC now is. So that's why YC can run at the scale it does with a staff that would look like a rounding error at any other comparable firm.

And that's not because we work harder, because it's because we have a different type of org. And that's whole—the whole game. It's not just 400x engineers. It's one company that operates at the level of 400x everyone else. And so, you know, if you remember, uh, only one thing about this, I mean, this is one of the things I had to discover along the way.

Like, you actually have to be really, really careful about where the compute computation is actually happening. Uh, it's happening almost always in two different places, and all of the bugs, all of the AI engineering that we run into that's a problem, it's usually because, uh, something is happening in one side of the equation that should be in the other.

Uh, the first area I would say is latent space, so the actual LLM. It's, you know, what do you use it for? Taste, judgment, understanding what a human actually wants when they say something vague, the nondeterministic calls, the computation that lives in the model, and you steer it with the Markdown file.

Uh, and then deterministic space is what, uh, engineers know. Like, your, your code agents go off and write TypeScript or, you know, maybe they're writing, um, you know, Erlang if you're using Elixir. Um, yeah. Deterministic space is the second place.

Um, let's say, you know, this is a real problem that we have for startup school coming up. We have 6,000 people, or, uh, we're going to try and, you know, one of the experiments we're going to try and do is can we seat 800 people at a time, um, perfectly clustered so the person sitting to the left and theright of you is the perfect person for you to meet at startup school.

Um, we have to do that in deterministic space combined with latent space. Uh, the computation, this computation, this actual storage of like where everyone is inside, like, you know, this multi-dimensional array of 800 seats, um, it actually must not live in the context window.

The LLM has to do the human part and seat people. Um, it's actually exactly what you would do if, um, you know, you were a human tasked with this thing. You would probably have to physically print out 800 pages and go in a, a big room and, like, say, like, "Well, where does this person go?"

Only now it can all happen, um, in your computer. And it can, instead of taking a month, it might be able to, you might be able to do it with, uh, you know, a couple hundred dollars' worth of tokens and probably 10 minutes.

Um, and so that's, you know, I would argue that's pretty remarkable. These are things that you couldn't do even, uh, I don't know, six months ago. Um, which brings me to working memory, and, uh, that's, you know, sort of my favorite way to understand it is, um, you and I, human beings, we only hold about seven things in our head at once.

Uh, seven plus or minus two. It's one of the most famous papers in cognitive psychology, and it's why, uh, local phone numbers are seven digits and why you forget the eighth item on your grocery list. Uh, that's the entire working memory, generally, of a human being.

And every institution humanity has ever built, every checklist, every org chart, every filing cabinet is a prosthetic for that limit. It's kind of a wild thing to think about. But an AI agent holds a million tokens. That's about 1,000 pages.

I was trying to explain to my 10-year-old what GBrain was recently. I said, "There's, you know, 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."

And that's quite magical, actually. Three Harry Potter books versus seven digits. I mean, that's pretty awesome. I mean, I don't know. Is that AGI? Maybe not, but it's already a very different operating regime. Almost every company on the earth is still running an org that's designed for the seven-digit brain.

But notice what that also tells you. Three books is a lot, but it's also very little. Your company is not three books. Your company is a library. Every email, every meeting, every decision, its reasoning, every customer conversation, every postmortem.

The question that determines whether your agents are geniuses or goldfish is who decides which three books are open on that desk. That's context engineering. And this is what a company brain is. It's the library plus the librarian. Now, some of you are already thinking, "This is just rag."

And you'reright that retrieval is the primitive, the same way Postgres is just B-trees. The hard part is everything around it. What gets written down in the first place into the knowledge wiki, how it gets enriched and linked, what gets promoted to hot memory versus filed as cold reference, who arbitrates when two facts disagree.

Retrieval is easy. Being worth retrieving from is the product. So I've been building mine in the open. It's called GBrain. It works with any harness, but it loves OpenClaw and Hermes Agent. It's effectively Postgres for agents, a retrieval layer whose job is to figure out for, uh, for any three, you know, for any task, what three books should be loaded into the agent's head.

My personal one started as a room's full of books or so. Now it's a warehouse, about 220,000 pages, written mostly by my agents, from my email, meetings, 20 years of notes, uh, and the lived experience of me. And that's the point.

It's my second brain. And when a founder emails me about a crisis, before I start reading this, 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.

Uh, when my agent does anything, it does, it does everything knowing what I already know. And that's the difference between an assistant and a colleague. So let me stress test my own pitch because you would anyway. Company brains do have failure modes.

Uh, a brain nobody curates becomes a garbage dump with great search. Retrieval will surface a stale fact with total confidence. Um, a bad skill file encodes a bad process forever. Uh, that's bad. So primitive, the primitive is not memory.

It's memory plus hygiene, provenance on every fact, contradiction, contradiction checks when new information collides with the old. And a librarian, human plus agent, whose actual job is pruning, treat the brain like a production infrastructure and it compounds. Treat it like a dumping ground and you get a very confident agent that is wrong in ways nobody can trace.

And, um, here's the discipline that I think, you know, personally, uh, makes our company brain and my personal AI compound. Um, that's my signature move. And, uh, you know, it's what I say to every YC company and every, uh, everyone inside YC, which is never do one-off work.

You can open OpenClaw, you can open your harness, you do some work, but then when you're happy, you know, and it'll come back, it's, you know, kind of a bad job. It's kind of like an intern that's not that good.

But the great thing is you can just say, "Hey, I didn't like that. Fix it." Right? I'm sure all of you do this. But don't stop there. You actually need to, at the end of that task, uh, skillify it.

And so I have a blog post on X about that. You can search for skillify it and, you know, go get that skill file and then just load it into your, your own harness and it'll just turn whatever you just did into a skill that you can reuse.

Because if you have to ask for something twice, you failed. Um, so yeah, if you remember only one thing, it's that. Like, when you, you know, use your AI agent and then when you're done with it and you're happy with the output, skillify it.

It's going to be awesome. Um, the organization that captures what it learns like this gets smarter every single day. The one that doesn't wakes up every morning with amnesia, no matter how good the model is. Model quality is rented, but if you build your brain, you're, you own that brain.

So Theo's question head-on. What do we build now? Build the AI-native company, not a company that just uses AI. A company that is shaped like what I just described from day one. A thin team, skill files for everything, the founder still in the code, library, this library, this company brain, your personal AI.

Uh, use GBrain if you'd like. Uh, it's open source and free, but you don't have to. There are a lot of really good ones. The, the library will compound from the first week and your whole org will be wired to run at about 400x.

And if you want the greenfield, the thing that I'd build if I were 25 and sitting where you're sitting, every company on this earth is about to need a brain. The memory layer that means that you never have to reask what you knew.

Personal AI that actually knows you. We're building GBrain in the open and MIT open source. Um, I'm not trying to make money from this because I think the layer should be open the way Linux is open. But the layer itself, company brains, personal context, the librarian that picks the three books, that's all wide open territory.

I hope somebody builds the defining company here. And I'd like to fund you at YC if you do. Now, let me be honest in a way that maybe undercuts my own pitch. You don't need my tools to start.

OpenClaw is the Ferrari. I will always recommend it. But Codex is a really good Honda. It will do 90% of this. Uh, it will not blow your face off, but it will get you there. Use whatever. The concepts are the point, not my repos.

You know, think about where the computation is. Use skill files as employees. The library, the librarian. Never do one-off work. Those travel with you to any stack. So let me land this. A lot of people in the worldright now are terrified about what happens to all the jobs.

And I understand the fear, but I want to say it plainly. That is a failure of imagination. And the people in this room are the answer to it. What I just described, you're going to take to your startup.

You will multiply yourself and every person in your company will multiply themselves. And you will go build the companies that become the beacon for how all of this works in society. Abundance is not a policy paper. It is shipped software.

I have a friend who has a rare form of epilepsy. He built a repo of 80,000 Markdown files, a company brain for one small boy. And he pushed himself to the absolute edge of human what humanity knows about his son's exact condition.

No lab, no grant, no permission, a father, a laptop, and a library. That's not a side story. That is the exact architecture I've been describing for the last 20 minutes. A library, the librarian, theright three books open at theright moment, pointed at the thing this man loves the most in the world.

You can do that now. Every problem where you thought, "I wish I had that person, but I can't get them." You can. Every code base you thought was too buggy to fix, you can fix all of it. Every archive too big to read, every data set too gnarly to clean, every ocean you were told not to boil, we can boil the ocean now.

And every single one of you can fly, not metaphorically, mechanically. And you need to, to survive, to thrive, to win. Theo asked what we should build now. And here's the whole answer. Build that AI-native company and build it, build the thing underneath it.

The brain, the memory, the compounding library. That makes every company after yours easier to build. Go boil the ocean. Go write that test. Go ship that skill. Some of the companies you're about to watch in the battlefield are already doing this.

Go build the one that does it best. Thank you.