Y Combinator Startup Podcast
00:13:44 · Garry Tan
That's what g brain is meant to be, the library plus the librarian. I've been building g brain in the open. My personal Open Claw has a Karpathy style knowledge wiki with about 220,000 markdown pages. Twenty five years of my life diarized. Every email, every meeting, my notes, my photos, my drafts, the things I got wrong. Compiled mostly by agents, curated by agents, searched for by agents. But I never re ask a question I already answered. And the lived experience in the system is the point. A founder emails me about a crisis before I finish reading the email. My agent has already pulled every prior conversation I've had with that founder. Three portfolio companies that hit the same wall and what actually worked for them. When my agent does anything, it does knowing everything I know. And that's the difference between an assistant and a colleague. Let me walk you through an actual day because that matters more than an architecture diagram. While I slept last night, my agent processed my inbox. Not sorted it, processed it. It knows which emails are from founders in trouble, which are from people trying to sell me something, and which are from the 17 mailing lists I never quite unsubscribed from. The ones that matter are triaged with context pulled from the library, who this person is, my whole history with them, what they're really asking under what they wrote, and what that might mean for me. I wake up to a briefing, not a pile of emails. Before every meeting, a prep doc, who I'm meeting, what we said last time, what changed since, and what I should ask. Research I was curious about at midnight is finished by morning. And when something interesting happens in the world, my agent has usually read it cross referenced against what I care about and filed it before I've had coffee. On top of this library sits my agent coding framework g stack, a 123,000 stars now, which put it in the top 100 open source projects in the history of GitHub. And what's actually in the punchline of this full architecture? It's mostly skill files plus a browser that the agents can drive. Pages of English and a way to act on the world. Markdown, not magic. Fat skills, then harness. Let me show you what a skill file is because I keep saying this phrase, and I want you to see how unmagical it is. Here's a real one, lightly redact is.
00:16:27 · Garry Tan
It says, when a meeting recording lands from circle back, 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. File the summary here, full transcript there. If anything contradicts something we already believe, flag it. Don't overwrite it. That's it. That's a skill. It's a page of English. A smart intern, anyone really who could read could follow it. And that's the test actually. If a smart intern could follow it, an agent can run it, which means actually a kind of profound thing. I know I caught a lot of flack for talking about this, but I think it's more true than ever, especially now. Markdown is actually code. If you can write clear instructions in English, you're a programmer. The compiler is a language model. And that's why it's not just for engineers anymore. At YC, our media people, events staff, finance team, people who never open a terminal in their lives are building skill files and scheduled jobs. One of our finance folks compiled about a 100 Excel workbooks into a single app she built with an internal agent. She is not a programmer. She is a manager of agents now. Everyone is about to be. The most important question to ask here is where is the computation happening? And there are exactly two answers, and confusing them causes every agent failure I've ever seen. Some computation belongs in latent space. Taste, judgment, reading what a human actually wants from a vague request, that lives in the model and you steer it with a markdown file. And then some computation belongs in deterministic space, the arithmetic, the SQL query. For instance, the seating chart for what sessions you're going to go to today for your breakouts. All of that needs to be stored in a SQL database used by the markdown files. Being smart about this goes a long way. Ask an agent or human to seat five people around a table. That's easy. Do it in latent space. Ask it to make custom schedules for 6,000 people in an arena like we just did for you, and your latent space agent needs to write some code to keep track of it. Your experience at this conference had to be markdown files calling code in exactly this way. And you can't do it without the code. The model fails where we fail. The fix is having the model compute the way humans compute.
00:19:12 · Garry Tan
The latent and the deterministic, Markdown files calling databases and scripts. Simple, but it's what everything is actually built on. And I'll give you one more receipt, my favorite one because you're sitting inside it right now. Five days ago, I decided this talk needed Spinoza, one of my favorite philosophers, especially because of how canceled he got. So my agent went and acquired three of the best biographies about the man, books by Nadler, Goldstein, Stewart, about 1,500 pages. It read all three. It built me a synthesis, a dated chronology of his life, every place the three biographers disagree with each other, and the best verbatim quotes with chapter citations. And because it knows what I need, the 10 most tellable moments of his life ranked with delivery notes, The knife attack, the bribe, the desk. Every beat of our opening that might have given you some chills twenty minutes ago came out of that overnight run. 1,500 pages became a stage ready story that I could edit. I call it a compendium skill, and I use it daily. It's a personal skill that is a mega mega version of deep research, only deeper than anything the corporate AI products will give you. The spine of this talk you're watching was inspired by the machine we're describing now. And if you're wondering where my... Mine actually started, it was not 220,000 pages. It was a folder. It was a few markdown files about the companies I was working with and the people I kept emailing. And the library got big the same way anything gets big. A little everyday compounding with agents doing the filing. Nobody builds the warehouse first. First, you build one shelf. When you sit down with an agent tonight, you're not coding. You're managing a workforce made of markdown. A skill file is an employee. It has one capability, one job written down clearly enough that someone new could execute it. A resolver is an org chart. A task comes in and it decides which markdown file or who handles it. Which means that before you ever incorporate anything, before you have a co founder or a logo or a deck, you can already be running an organization. An organization of one plus your agents. You are the founder and the entire management layer of You Incorporated and the headcount under you is now whatever you decide it is. This already produces companies that break the old math.
00:21:56 · Garry Tan
Emergent out of our summer twenty four batch went from public launch to 9 figures of revenue in eight months. When they crossed $15,000,000 in annualized revenue, they were 15 people. Retail, winter twenty four, hit 60,000,000 annualized with about 40. That revenue per person did not exist before, not in software, not in oil, not in railroads. And these aren't freaks of nature. They're the first companies built natively on the new physics, and every one of them started as one or two people wired the way I just described. Now picture our batch room in the dog patch, Hundreds of founders every single day, each one of them doing what used to be a person's entire year of work. That is not the future. That is the bar right now with this batch. If you're not doing it, your competitor is, and they will eat your lunch politely and thank you for it. It also changes what software even is. Software doesn't have to be precious anymore. You can build exactly the tool you need for the audience of one in a weekend. The old advice was scratch your own itch and hope it's a market. The new version is much better. Scratch your own itch because scratching itches is nearly free, and some of your tools for one will turn out to be entire companies. You'll know because other people start begging for them. And one honest caveat before the how to because you catch me out in any way. A brain nobody curates is a garbage dump with great search. Retrieval will surface a stale fact with total confidence. A bad skill file encodes a bad process forever. So the primitive is memory plus hygiene, provenance on every fact, contradiction checks when new information collides with old, and a librarian whose actual job is pruning. Treat the brain like 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. Everything so far is philosophy and receipts. So let's get into some how to. If you do what I describe in the next six minutes, you'll be ahead of 99% of people who watched this talk and just nodded. Step one tonight, pick a harness and run an agent on your own machine. I use OpenClaw and Hermes agent with g brain. A hosted version of this is at gbrain.io. It's free. G brain itself is free and open source. I always recommend the Ferrari. But I'll be honest, the Honda is really good too.
00:24:42 · Garry Tan
Codex, Claude code, whatever. Any of them will do 99% of this. And the upside of not Ferrari is that it will also get you to your destination with a little less of... Less getting out to fix it on the side of the road. The concepts are the point, not any given repo or product. The intelligence is untapped, and there are many paths. Step two this weekend, start your library. Not a grand archive. One folder of markdown files, export your notes, export your email if you can. Write one page about each project you're working on and each person you work with. And on those pages, write the things you actually know, what you're building together, what they care about, what you owe them, what they said last time. That stuff no model on the earth... No model on earth has because it only exists in your head. And your head, as we established, only holds seven things. The first time an agent answers a question using your context instead of the Internet's, you'll feel the click and you won't go back. You're all sitting on... You are all sitting on five, ten years of your own history in one inbox or another. That's your moat just lying there, unindexed, doing nothing. The only gate between you and this entire architecture is probably twenty four hours. Step three, write your first skill file. Picking it is easy. You know, what's the task you do every single week that you hate the most? Might be expense reports, meeting notes, the weekly status update, competitor research. Explain it to your agent. What do you wanna do? In plain English, the way you'd explain to a smart friend on their first day of a job. And then let it get it wrong. If it gets it wrong, correct it. Every rule, every exception, every oh, and also, put it in there, and it'll fix it. That page is now an employee. Run it. Step four, wire it up to be a recurring job. Maybe it's the job you just created in step three. Every morning at seven, do this. Every Friday, summarize that. The first time you wake up to work that finished while you slept, something shifts in your head permanently. That's the day that the day stops being the unit of work for you. It becomes what you can imagine, and it should be driven by what your goals are and what you wanna create in the world. Step five, this is the discipline that separates the compounders from the dabblers. Never do one off work.
00:27:26 · Garry Tan
Most people run one operation with one agent and then throw the context away. They close the window. That's it. Don't. At the end of every task, ask the agent to Skillify what it did. Skillify is a special skill you can find in g brain. You can point it at that repo and say, extract Skillify, learn how to do it. Turn it into a markdown file you can use and reuse forever. I'll say it the way I say it at YC. If you have to ask for something twice, you failed. The person who captures what they learn gets smarter every single day. The person who wakes up every morning with amnesia, well, that... That's a waste of your time. And it sort of doesn't matter how good the model gets if you can't turn it into real memory. Do those five things and I can tell you what your next ninety days look like. Week one, honestly, it's a toy. The library's thin. The skills are clumsy. You're fixing more than you're saving. Week four, the flywheel catches. The agent starts answering with your context. The morning job produces something you actually read, and you write your third and fourth skill because the first two worked. Week twelve, you have a library that answers before you finish asking. A dozen skill files running the parts of your week you used to dread, and one or two tools that other people keep asking to borrow, which in this room is called a startup. The curve is the same curve as any compounding thing you've ever seen. Flat, flat, flat, then not. Most people who try this will quit this in week two, which is precisely why the ones who don't feel like they're cheating by week twelve. Now I need to tell you the part that isn't fun because everything I taught you just cuts both ways. I told you Spinoza's definition of sadness earlier, the feeling of your power acting... Power of acting decreasing. And I said it gets political. This is where. A skill file is not a document. It's a piece of your cognition, how you do the thing, extracted from your head, written down, and executable. Every skill you teach an agent is you, externalized. And the exact same file is two opposite futures depending on one variable, who controls it. Take a fictional example of a support engineer. Let's call her Maya. Over two years, Maya teaches her agents 40 skills. How to triage a p zero at two in the morning. How to deescalate the customer who's about to churn.