Garry Tan: "Personal AGI Is How You Stay Under Your Own Power"
In a Nutshell
Garry Tan argues that personal AGI—your own agent running on your infrastructure with your context, memory, and skill files—is how individuals maintain ownership of their cognitive abilities in an age of agents. The core mechanism is building a personal library of markdown files (220k pages in his case) plus skill files that capture your processes, then wiring them to agents so they compound daily while you sleep. This setup produces 400x productivity gains, lets one founder run companies that break old revenue-per-person math, and ensures your judgment stays yours rather than being extracted into corporate repos.
These notes were generated by AI and may contain inaccuracies.
The internet calls Garry Tan one of the most AI psychotic people online. He begins his talk with the story of Baruch Spinoza, one of the most cancelled men in history. In 1929, a New York rabbi challenged Einstein by telegram asking if he believed in God, requesting an answer in 50 words. Einstein answered in 25 words: "I believe in Spinoza's God who reveals himself in the lawful harmony of the world, not in a God who concerns himself with the fate and doings of mankind."
On July 27, 1656, in Amsterdam, Spinoza was 23 years old and a member of a tight-knit Sephardic Jewish community. The elders excommunicated him with the most violent curse the community ever produced, stating he was cursed by day and by night, when he lies down and when he rises up. Nobody could speak to him, trade with him, come within four cubits of him, or read anything he writes. This ban uniquely had no repentance clause among the roughly 40 bans the community issued that century, and it remains technically in force today.
Spinoza's crime was evil opinions and expressing forbidden thoughts. His punishment was complete deletion from the community. Before the excommunication, the community tried to buy him off with a thousand guilders a year, serious money at the time, requiring only that he show up at synagogue occasionally and keep his mouth shut. In founder terms, they offered him a salary to stop building. He refused, saying he wanted truth, not comfort, and would not accept it for 10,000 guilders.
Shortly before his excommunication, a fanatic came at him with a knife. The blade tore through his cloak and missed him. Spinoza kept that cloak, scar unmended, for the rest of his life, wanting to remember what ideas cost.
After being cancelled, Spinoza ground lenses by day, making optical instruments so well that the best scientists in Europe sought them out. By night, he wrote a book so dangerous he could not publish it while alive. He died at 44 with lungs full of glass dust from making other people's lenses. His dying instruction was to ship the manuscript by canal barge to his publisher in Amsterdam. That manuscript became his posthumous works, which attracted immediate attention across Europe and inspired some of the most important philosophers of the Enlightenment.
Spinoza had a name for the engine that kept him going: conatus. It means your striving, the drive in every living thing to keep going and to increase its power to act. Not your resume, title, or job, but the striving itself.
Spinoza's heresy was that God is not a king on a throne but has spread through everything that exists, described as "God or nature." Four hundred years later, we make a similar mistake about intelligence. Everyone waits for AGI as a singular event, a god in a data center, some announcement, threshold, or day when the sky changes color.
The updated heresy states that everyone watches the sky while the thing they are watching for is already in the room. It does not look like a god but like infrastructure: a terminal window, a folder of markdown files, a job that finishes while you sleep, spread through everything. AGI is not arriving as an event but diffused as your agent running on your context doing your work.
Personal AGI is not a chatbot you pay $20 a month for, not a slightly better autocomplete, and not an assistant that knows your calendar and nothing else. That is a subscription you rent, corporate AGI you do not own. It resets when you close the tab, knows what everyone else already knows, and when the company pivots, your assistant gets a lobotomy on someone else's schedule.
Personal AGI is an agent that runs on your infrastructure, reads from a memory you own, executes procedures you wrote, and compounds. The corporate AGI you do not own gets better only when the company ships something. Your personal AGI gets better every single 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.
Intelligence of this kind should be owned by you, not rented. If you build this for yourself, 2034 does not have to be like 1984.
In 2013, Tan was a YC partner building Bookface, YC's internal social network at night, shipping about 14 useful lines of code a day, which is dead on median for programmer productivity. In the current year, running YC full-time with the same brain and hours plus a 5:00 kid pickup, his output is about 400x what it was in 2013.
Even applying the most pathological verbosity penalty and assuming half the code is scaffolding, the output is still 8x at the absolute floor and 10 times that in the middle of the range. This multiplier applies not just to coding but to design, product management, growth, and every piece of knowledge work.
At YC, a year and a half ago in the winter 25 batch, a quarter of the companies had codebases that were 95% AI generated. Those companies use AI agents for everything now, not just code, and that batch is on track to becoming one of the fastest growing, most profitable batches in YC history.
The fastest growing founders treat AI as a workforce rather than autocomplete. There are 2x people and 100x people using the same Claude with the same weights, same context window size, and same API. The leverage is not in the weights but in what context you give it, how relevant it is, and whether it happens at the right step.
In Ethics, Spinoza defines joy as the feeling of your power of acting increasing. The first time an agent does a week of your work in an afternoon, it does not feel like a convenience but like joy. The technical term is that your power of acting increased and your conatus got bigger. Spinoza defined the opposite as sadness, the feeling of your power of acting decreasing.
The equation for the next decade is: frontier model (rented commodity getting cheaper by the quarter) plus your context (owned by you and unique) plus a harness that wires them together (such as OpenClaw, Hermes agent, Claude Code, or Codex). This gives you an agent that acts like a very fast version of you. Model quality is rented but your brain is owned by you.
Marshall McLuhan said technology is an extension of man. Steve Jobs called a computer a bicycle for the mind. With this setup, you have a self-driving rocket. Paul Graham taught every founder to make something people want and do things that don't scale. Agents are how one founder now does unscalable things at scale.
Human working memory holds about seven things at once, seven plus or minus two. This is why local phone numbers are seven digits and why you forget the eighth item on a grocery list. Every institution humanity has built, every checklist, org chart, filing cabinet, and standup meeting is a prosthetic for that limit.
An AI agent holds a million tokens, about a thousand pages or three Harry Potter books sitting open on its head all at once, and can find a needle in any of them and synthesize across all three in seconds.
Your life is not three books but a library: every email you ever sent, every meeting, every decision and every reason behind it, every conversation with every person you know. The question determining whether your agent is a genius or a goldfish is who or what decides which three books are open on the desk.
Tan has been building GBrain in the open. His personal OpenClaw has a Carpentry-style knowledge wiki with about 220,000 markdown pages representing 25 years of his life diarized. Every email, meeting, notes, photos, drafts, and things he got wrong are compiled, curated, and searched by agents, so he never re-asks a question he already answered.
When a founder emails him about a crisis, before he finishes reading the email, his agent has already pulled every prior conversation with that founder, three portfolio companies that hit the same wall, and what actually worked for them. When the agent does anything, it does so knowing everything Tan knows.
While Tan sleeps, his agent processes his inbox, not sorting it but processing it. It knows which emails are from founders in trouble, which are from people trying to sell something, and which are from the 17 mailing lists he never quite unsubscribed from. The ones that matter are triaged with context pulled from the library.
He wakes up to a briefing, not a pile of emails. Before every meeting, there is a prep doc with who he is meeting, what was said last time, what changed since, and what he should ask. Research he was curious about at midnight is finished by morning. When something interesting happens in the world, his agent has usually read it, cross-referenced it against what he cares about, and filed it before he has had coffee.
On top of this library sits his agent coding framework, GStack, with 123,000 stars, putting it in the top 100 open source projects in GitHub history. The architecture consists mostly of skill files plus a browser that agents can drive. Pages of English and a way to act on the world. Markdown, not magic. Fat skills, thin harness.
A skill file example states: 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 already believed, flag it but do not override it.
If a smart intern could follow the instructions, an agent can run them. Markdown is actually code. If you can write clear instructions in English, you are a programmer. The compiler is a language model. This is why it is not just for engineers anymore. At YC, media people, event staff, and finance team members who never open a terminal are building skill files and scheduled jobs.
One finance person compiled about a hundred Excel workbooks into a single app she built with an internal agent. She is not a programmer but a manager of agents.
There are exactly two answers to where computation happens, and confusing them causes every agent failure. 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.
Some computation belongs in deterministic space: the arithmetic, the SQL query, the seating chart for sessions. All of that needs to be stored in a SQL database used by the markdown files.
Asking an agent or human to seat five people around a table is easy to do in latent space. Making custom schedules for 6,000 people in an arena requires the latent space agent to write code to keep track of it. The experience at this conference had to be markdown files calling code in exactly this way.
Five days before the talk, Tan decided it needed Spinoza. His agent acquired three of the best biographies about the man by Nadler, Goldstein, and Stewart, about 1,500 pages. It read all three, built a synthesis, a dated chronology of his life, every place the three biographers disagree, and the best verbatim quotes with chapter citations. Because it knows what Tan needs, it extracted the 10 most tellable moments of Spinoza's life ranked with delivery notes: the knife attack, the bribe, the desk. Every beat of the opening came out of that overnight run.
Tan calls it a compendium skill and uses it daily. It is a personal skill that is a mega version of deep research, only deeper than anything the corporate AI products will give you.
The library did not start with 220,000 pages but with a folder, a few markdown files about the companies Tan was working with and the people he kept emailing. The library got big the same way anything gets big: a little every day compounding with agents doing the filing. Nobody builds the warehouse first. First, you build one shelf.
When you sit down with an agent, you are not coding but managing a workforce made of markdown. A skill file is an employee with 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.
Before you ever incorporate anything, before you have a co-founder or a logo or a deck, you can already be running 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 whatever you decide it is.
This produces companies that break the old math. Emergent out of the summer 24 batch went from public launch to nine figures of revenue in eight months. When they crossed $15 million in annualized revenue, they were 15 people. Retail winter 24 hit 60 million annualized with about 40 people. That revenue per person did not exist before, not in software, not in oil, not in railroads.
These are not freaks of nature but the first companies built natively on the new physics. Every one of them started as one or two people wired this way.
Software does not have to be precious anymore. You can build exactly the tool you need for an audience of one in a weekend. The old advice was to scratch your own itch and hope it is the market. The new version is to scratch your own itch because scratching itches is nearly free. Some of your tools for one will turn out to be entire companies. You will know because other people start begging for them.
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. 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.
Step one tonight is to pick a harness and run an agent on your own machine. Tan uses OpenClaw and Hermes agent with GBrain. A hosted version is at gbrain.io, which is free. GBrain itself is free and open source. Any harness will do 99% of this work. The concepts are the point, not any given repo or product. The intelligence is on tap and there are many paths.
Step two this weekend is to start your library, not a grand archive but one folder of markdown files. Export your notes. Export your email if you can. Write one page about each project you are working on and each person you work with. On those pages, write the things you actually know: what you are building together, what they care about, what you owe them, what they said last time. That is stuff no model on earth has because it only exists in your head, and your head only holds seven things.
The first time an agent answers a question using your context instead of the internet's, you will feel the click and you will not go back. Everyone is sitting on five to ten years of their own history in one inbox or another. That is your moat just lying there unindexed doing nothing. The only gate between you and this entire architecture is probably 24 hours.
Step three is to write your first skill file. Pick the task you do every single week that you hate the most: expense reports, meeting notes, the weekly status update, competitor research. Explain it to your agent in plain English, the way you would explain to a smart friend on their first day of a job, then let it get it wrong. If it gets it wrong, correct it. Every rule, every exception, every "oh, and also" goes in there, and it will fix it. That page is now an employee. Run it.
Step four is to wire it up to be a recurring job. Maybe it is the job you just created in step three. Every morning at 7, 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 is the day the day stops being the unit of work for you. It becomes what you can imagine, driven by what your goals are and what you want to create in the world.
Step five is the discipline that separates the compounders from the dabblers. Never do one-off work. Most people run one operation with one agent and then throw the context away. They close the window. Do not do that. At the end of every task, ask the agent to skillify what it did.
Skillify is a special skill you can find in GBrain. 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. 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 wastes their time. It does not matter how good the model gets if you cannot turn it into real memory.
If you do these five things, here is what the next 90 days look like. Week one, it is honestly a toy. The library is thin. The skills are clumsy. You are fixing more than you are 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 12, 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 have ever seen: flat, flat, flat, then not. Most people who try this will quit in week two, which is precisely why
Most people who try this will quit in week two, which is precisely why the ones who don't feel like they're cheating by week 12.
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.
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 P0 at 2 in the morning. How to deescalate the customer who's about to churn. How to write a postmortem that actually prevents the next incident. 40 files. That's her judgment. The thing that took her two years to build sitting on a disc.
Those files live in Maya's repo. She changes jobs. They go with her. Day one at a new company, she's operating with years of compounded judgment on tap. Every year she works, she compounds. That's ownership. And if she wanted to start a company that does this, it's her expertise.
Those 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. 40 files executing forever and her name isn't even in the commit history. She didn't have a career. She had an extraction.
Skill files are yours. Own your skills because if you don't, your job becomes a skill file.
This happened before. Craftsmen own their tools. That's what made them free. The factory broke that. The loom belonged to the mill. The knowledge workers assumed we were safe because our tools lived in our heads where nobody could confiscate them. Skill files end that. For the first time in history, your cognition can be extracted, stored, versioned, and owned. The only question is by whom?
Remember the thousand gilders? That offer never went away. It got rebranded. Every comfortable arrangement where your judgment compounds in someone else's repo is a thousand gilders a year. To show up, keep quiet, and stop building your own thing.
Spinoza faced the upgraded version two. In 1673, Highleberg offered the cursed heretic a full professorship, salary, legitimacy, a chair, and quote, freedom of philosophizing, provided he not disturbed the established religion. His answer was, I do not know what the limits of that freedom of philosophizing might have to be. He read the terms of service and he declined the acquisition.
He had a phrase for what he was protecting. Under your own power as opposed to under someone else's. Your power of acting exists either way. The political question in 1673 and in 2026 is who commands it? Personal AI is about controlling your own cognitive abilities and protecting yourself.
Personal AGI is how you stay under your own power in the age of agents. So keep your brain and your skills in a repo you control from day one before any platform or any acquirer has an opinion about it.
When Spinoza died, they inventoried the room. Two pairs of pants, seven shirts, a lens lathe, 160 books, and the ethics. Locked in a desk, he owned almost nothing, and nobody ever controlled his skill files. The desk drawer was his repo. Own yours like he owned his.
The models are improving so fast that all this harness stuff will be obsolete. Just wait for the next release. This is the better bitter lesson crowd and I love them. But notice what actually happens in every model release. The better the models get, the more the differentiator moves to context. When everyone's engine is a 1000 horsepower, the race is won on the driver and the map. The weights are everyone's. The library is yours. At least I hope it is.
A better model makes your library worth more because a smarter reader extracts more from the same books. I'm rooting for the labs as hard as anyone in this building, but every release they ship is a free upgrade to a workforce I already own and a workforce I want you to own.
Is this just RAG? Sure, and Postgres is just B trees. Retrieval is the primitive, not the product. The hard part is everything around it. What gets written down in the first place? 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.
You put your entire life in one system, your email, your meetings, your kids schedules. What happens when it leaks? My answer is the same answer as the whole talk. That's exactly why it has to be yours. My brain runs on my own infra in my own repo under my own keys.
Compare that to the default, which is not privacy. The default is your life is already scattered across 10 clouds owned by companies whose incentives are not yours, searchable by everyone except you. I didn't create the risk by consolidating my context. I took custody of it. Custody is the security model. And if you don't trust yourself to hold the keys, I promise you the answer isn't trusting someone else's terms of service more.
Why did I open source all of it? The harness, the brain architecture, the skills, the whole personal operating system. People ask me this because they seem like they think there must be a catch. Well, the answer is because I can. Because being at YC for me means I don't have to monetize my own infrastructure. But because I can is also the answer to the wrong question.
The real question is why anyone should. And the answer is that I believe tools of the powerful should be given away. Every era has a private technology of leverage, a thing the powerful have and everyone else doesn't. For a long time it was literacy. Then it was capital. Right now today it's this. The harness, the library, the workforce made of markdown. The people who have it are quietly operating at a different scale than the people who don't. And the gap is widening every month.
When something like that powerful stays private, you get a priesthood. When it gets given away, you get a renaissance. I know which one I want to live in, which means I get to do the thing that I actually believe in.
Say the things other people won't. Fund the people other people won't. Build the buildings other people won't. Write and give away the code that other people won't. Leave behind the institutions that other people won't.
When you build in the open, you should know what's coming because Spinoza's story has one more chapter. November 1676, Gottfried Leibniz, the most glittering genius in Europe, silk stockings, a calculating machine in his luggage, travels to the Hague to spend three days in an attic with the most hated man on the continent. And then he spends the next 40 years lying about it publicly. The visit was a few hours in passing. Privately, his notes are crammed with obsessive commentary on Spinoza.
I live a small version of this weekly. I say, Agents write most of my code now, and the dunks arrive by lunch. Then I look at what the loudest dunkers are actually shipping, and it's agents all the way down. So, learn the pattern now because building in public guarantees you'll meet it. First, they quote tweet you, then they get clone you. The dunks are just the adoption curve announcing itself.
I have a friend whose son has a rare form of epilepsy. No lab, no grant, no permission. He just went and you can just do things. He built a repo of 80,000 markdown files, a brain for one small boy, and pushed himself to the absolute edge of what humanity knows about his son's exact condition. Every specialist visit, every paper, every seizure log, every drug interaction indexed and cross-linked and ready so that when a new doctor has an idea, he knows in minutes whether it's already been tried. A father, a laptop, and a library. That is personal AGI. Not a benchmark, not a demo. The entire architecture I've described tonight, the library, the librarian, the right three books, open at the right moment, aimed at the one thing, one man loves the most in the world. Nobody was coming to build that for him. So, he built it. And nobody is coming to build yours for you.
Everything you were told you needed, the team, the funding, the permission, the credential was a workaround for the fact that one person could hold seven things in their head and work 16 hours a day. That fact just expired. You can fly now. Not metaphorically, mechanically. Every problem where you thought, I wish I had this person. I wish I could hire this person, but I can't get them. You can. 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.
I have a sentence I live by and I want to leave it with you. It's all made up, but you get to make it up. Every institution in the world, including the one that read a curse over a 23-year-old in 1656, was made up by people no smarter than you.
The difference between you and every generation of founders before you is that they had to recruit dozens of believers before they could build anything at all. You need a laptop and a few years of your own history you're already sitting on. There were about 7,000 people at this whole event. 7,000 kannons, 7,000 strivings. For most of history, almost all of that striving never got an audience. It died waiting for funding, waiting for headcount, waiting for permission, waiting for someone else to believe first. The machinery I showed you tonight is the first technology I've ever seen that lets the striving go straight to work. One person, no intermediaries, no permission. I genuinely do not think the world understands yet what 7,000 people with that kind of leverage walk out of a building and do.
Spinoza closed the ethics, the book that had to be smuggled out in a desk with nine words. All things excellent are as difficult as they are rare. The difficulty just collapsed. The rarity is now up to you. Go and build.
Keep Y Combinator in your library
Save the videos and channels worth coming back to, and find them again in one place.
