Garry Tan: Own Your Intelligence

The video argues that the future of work belongs to "personal AGI"—agents built on your own context and infrastructure, not rented corporate subscriptions—drawing a parallel to Spinoza, who refused a salary for silence and was excommunicated for his dangerous ideas. The core driver is conatus, your innate striving to act and increase power, and AI tools should amplify that striving, not replace it with autocomplete. This leverage is already real: coding agents have multiplied output hundreds-fold, and in a recent YC batch, a quarter of companies shipped codebases that were 95% AI-generated. The key architecture is radically simple—markdown files as "skill" instructions that agents execute, turning non-programmers into managers of a workforce of agents. Garry Tan demonstrates this with a "G Brain": 25 years of diary entries compiled into a 220,000-page wiki, letting his agent read three full Spinoza biographies overnight and produce a cited compendium. In Spinoza's terms, this is joy—the feeling of your power of acting increasing—and it changes the economics of companies, with revenue per person reaching previously impossible levels. The closing advice: don't build the warehouse first, start with a single shelf of markdown files and let your library compound daily.

• ▶ 1:26 Spinoza was excommunicated by his Amsterdam Jewish community with the most violent curse they had; the ban forbade anyone from speaking to him, trading with him, or reading his writings, and uniquely had no repentance clause—never lifted, still in force today.
• ▶ 2:30 Before the curse, the community offered him a thousand guilders a year to show up at synagogue and stay silent—a "salary to stop building"—but Spinoza refused, keeping his torn cloak as a reminder of what ideas cost.
• ▶ 3:15 After being erased, Spinoza ground lenses by day and wrote a book so dangerous he couldn't publish it in his lifetime; he died at 44 from glass dust, and the manuscript was shipped to his publisher only after his death.
• ▶ 4:33 Spinoza's conatus — your innate striving to persist and increase your power to act — is the real driver, not resumes, titles, or jobs; tools should amplify that striving.
• ▶ 5:29 AGI won't arrive as a single god-like event; it is already diffused through infrastructure — a terminal, markdown files, background jobs — just as Spinoza saw God spread through nature.
• ▶ 6:09 "Personal AGI" means general intelligence for one person: you — not a $20 chatbot, not better autocomplete, but an agent running on your context and doing your work.
• ▶ 6:45 Corporate AGI is a rented subscription that resets, knows only public knowledge, and can be lobotomized; personal AGI runs on your own infrastructure, uses your memory and procedures, and compounds daily as an asset you own rather than consume.
• ▶ 8:06 Coding agents have amplified personal output roughly 400x (from ~14 useful lines per day in 2013), and even after extreme deflation for bloat, scaffolding, and self-flattery, the floor is still ~8x—making the multiplier real.
• ▶ 9:05 This leverage applies to all knowledge work, not just coding, and is visible at scale: in the Winter 25 YC batch, a quarter of companies had codebases that were 95% AI-generated and are now on track to be one of the fastest-growing, most profitable batches in YC history.
• ▶ 9:49 The fastest-growing founders treat AI as a workforce, not autocomplete: the leverage isn't in the model weights but in the context you give, its relevance, and the step at which it's applied.
• ▶ 10:26 In Spinoza's terms, joy is the feeling of your power of acting increasing—so when an agent does a week of your work in an afternoon, that's technically joy; sadness is your power of acting decreasing, like the heaviness of a Sunday night.
• ▶ 11:16 The equation for the next decade: frontier model (rented, commoditized) + your context (owned, unique) + a harness = an agent that acts like a very fast version of you; model quality is rented, but your brain is owned.
• ▶ 12:46 Human working memory holds about seven things at once, explain why phone numbers are seven digits and why we forget the eighth grocery item; every human institution (checklists, org charts, stand-ups) is a prosthetic for that seven-item limit.
• ▶ 13:27 An AI agent holds roughly a million tokens (about 1,000 pages or "three Harry Potter books open on its head"), can find and synthesize across all of them in seconds — a "different operating regime" from the seven-slot brain most people still run on.
• ▶ 14:40 G Brain is "the library plus the librarian": Garry has compiled 25 years of diarized life into a 220,000-page knowledge wiki, so his agent processes his inbox, preps meetings, and answers research overnight — acting as a colleague that "knows everything I know," not just an assistant.
• ▶ 16:45 G Stack's impressive scale (top 100 OSS projects) rests on a simple design: “Markdown, not magic. Fat skills, thin harness” — just a browser, English instructions, and a way to act on the world.
• ▶ 17:28 A skill file is literally a page of English instructions; the test is “if a smart intern could follow it, an agent can run it” — and this makes “markdown actually code,” letting non-programmers become managers of agents.
• ▶ 18:20 The most important design question is where computation happens: latent space for taste/judgment vs. deterministic space for arithmetic/SQL — agents must write code for large-scale tasks like custom schedules for 6,000 people, combining both spaces.
• ▶ 19:39 Garry Tan's entire AI workflow is built on a surprisingly simple foundation: markdown files that call databases and scripts — and this architecture is the real basis for everything else.
• ▶ 19:51 He demonstrates the power of this setup with a live example: his AI agent autonomously acquired three full biographies of Spinoza (~1,500 pages), read them all, and synthesized a compendium with a chronology, disagreements between authors, and verbatim quotes with citations.
• ▶ 20:43 Tan calls this output a "compendium skill" — a personal, deeper-than-deep research tool he uses daily, turning massive unstructured information into a stage-ready, editable narrative artifact.
• ▶ 20:58 The speaker’s system didn’t start as a massive archive—it began as a simple folder with a few Markdown files about the companies and people he worked with.
• ▶ 21:11 The library grew the same way anything gets big: a little every day, through compounding additions and automated agents handling the filing.
• ▶ 21:18 The core advice: don’t build the warehouse first—start with one shelf and let the library grow incrementally.
• ▶ 21:23 The work is not coding—you are managing a workforce made of markdown, directing agents as employees.
• ▶ 21:29 A skill file is an employee with one clearly written job, while a resolver acts as the org chart that routes tasks to the right markdown file.
• ▶ 21:47 Before any formal company structure, you can already run an organization of one plus your agents, with you as founder and the entire management layer.
• ▶ 22:09 A new company model is "breaking the old math": revenue per person now reaches levels that previously didn't exist in software, oil, or railroads.
• ▶ 22:13 Example results: Emergent hit nine figures revenue in 8 months with only 15 people at $15M annualized; Retail reached $60M annualized with ~40 people.
• ▶ 22:53 This is the current YC baseline: hundreds of founders each doing what used to be a person's entire year of work—if you're not built this way, your competitor is and will eat your lunch.
• ▶ 23:05 Software no longer has to be precious; the old constraints of cost, effort, and market validation no longer apply.
• ▶ 23:09 You can build exactly the tool you need for an audience of one in a single weekend, challenging traditional software project assumptions.
• ▶ 23:19 "Scratch your own itch" is now nearly free, so personal utility alone justifies building—and ▶ 23:27 you'll know a tool is becoming a company when other people start begging for it.
• ▶ 23:38 Without curation, a brain is just a garbage dump with great search — retrieval will surface stale facts with total confidence, and bad skill files encode bad processes forever.
• ▶ 23:46 The real primitive is memory plus hygiene: every fact needs provenance, new information must be checked against old for contradictions, and a librarian’s job is pruning.
• ▶ 24:03 Treat the brain like production infrastructure and it compounds; treat it like a dumping ground and you get a confident agent that is wrong in ways nobody can see.
• ▶ 24:16 Garry signals a shift from philosophy and evidence to practical, actionable how-to content.
• ▶ 24:21 He sets a clear bar: executing the instructions in the next 6 minutes will put you ahead of 99% of people who watched the talk, because most will only listen or agree.
• ▶ 24:26 This opening section ends as a bridge from theory to a short, dense, high-leverage action plan, with the step-by-step instructions immediately following.
• ▶ 24:26 Start by picking an AI agent "harness" and running it on your own machine — e.g., Open Claw, Hermes agent, or the free/open-source G brain (hosted at gbrain.io).
• ▶ 24:45 Tool choice matters less than you think: a "Ferrari" or a "Honda" both handle 99% of the work, and the simpler option means fewer breakdowns along the way.
• ▶ 25:01 The concepts and workflow matter more than any specific repo or product — don't get attached to a particular tool.
• ▶ 25:08 Start your library this weekend: create one folder of markdown files, export your notes and email into it, and write one page for each project and person you work with.
• ▶ 25:22 On each page, write what you actually know: what you’re building together, what the other person cares about, what you owe them, and what they said last time.
• ▶ 25:31 This matters because no model has this information — it lives only in your head, which holds about seven things at once — and the first time an agent answers from your context instead of the internet, it clicks.
• ▶ 26:08 Choose the recurring weekly task you dislike most (e.g., expense reports, meeting notes, status updates) to automate for immediate value.
• ▶ 26:21 Explain the task to your agent in plain English, as if teaching a smart friend on their first day — no need for technical precision.
• ▶ 26:28 Let the agent get it wrong first, then correct it by feeding every rule, exception, and “oh and also” detail back into the skill file so it becomes a living, accumulating document.
• ▶ 26:38 Treat the skill page as an employee: it's not a static document but an autonomous worker you can task and schedule.
• ▶ 26:50 The key psychological shift: waking up to finished work changes how you think, and "the day stops being the unit of work."
• ▶ 27:05 Recurring jobs should be driven by your goals and creative ambitions, not arbitrary busywork.
• ▶ 27:12 Never do one-off work: always turn each completed agent task into a reusable "skillify" file instead of discarding the context.
• ▶ 27:50 Follow the YC rule: "If you have to ask for something twice, you failed" — codify learnings so you never repeat the same work.
• ▶ 27:55 Compound your progress by capturing what you learn; otherwise you wake up each day with amnesia, and even a better model won't help you retain your own experience.
• ▶ 28:08 The entire system only works if you convert your work into real memory; without that, the 90-day progression will not happen.
• ▶ 28:15 Expect week one to feel like a toy—clumsy and time-consuming—but by week four the flywheel catches as your agent answers from your context and skill files prove useful.
• ▶ 28:35 By week twelve, the library answers before you finish asking, you have a dozen skill files, and custom tools others want to borrow—signaling a mature, externally valuable system.
• ▶ 28:48 The growth pattern behind a "life library" is the same compound curve found in startups: flat, flat, flat, then not.
• ▶ 28:56 The early phase shows almost no visible progress, which is why most people quit in week two.
• ▶ 29:02 Those who persist feel like they're "cheating" by week 12—the compounding payoff arrives suddenly for those who endured the unremarkable beginning.
• ▶ 29:10 The "part that isn't fun": everything taught so far "cuts both ways," marking a sharp pivot from the intuitive, natural feel of the methods to their political edge.
• ▶ 29:28 A skill file is not a document — it is a piece of your cognition (how you do a thing) extracted from your head, written down, and executable; every skill you teach an agent is "you externalized."
• ▶ 29:43 The exact same skill file represents two opposite futures, and the deciding variable is simple: who controls it — the same externalized cognition that empowers you can be used against you if control is lost.
• ▶ 29:51 A fictional support engineer, Maya, builds 40 distinct skills over two years — capturing her professional judgment in files.
• ▶ 30:09 If Maya keeps those files in her personal repo, her expertise compounds, gives her genuine ownership, and lets her start a company from them.
• ▶ 30:36 If the files live in the company repo, Maya leaves with nothing and the company keeps running on her judgment — “she didn't have a career, she had an extraction.”
• ▶ 30:52 Two otherwise identical situations diverge completely based on one variable: who controls the skill files.
• ▶ 30:56 The doctrine: skill files are yours — own your skills; this is an active responsibility, not passive.
• ▶ 31:05 If you don't own your skill files, your job becomes a skill file — you become replaceable.
• ▶ 31:09 Historically, freedom was tied to owning your tools; the factory broke that link, and knowledge workers wrongly assumed their internal skills were immune.
• ▶ 31:23 Skill files end that era of assumed safety: for the first time, cognition itself can be extracted, stored, and owned — leaving "by whom?" as the open battle.
• ▶ 31:38 The "thousand guilders" offer persists in modern form: any comfortable arrangement where your judgment compounds in someone else's repository is quiet absorption, not safety.
• ▶ 31:46 Starting a startup is the practical way to put your personal skill-file system to work and own your cognitive tools.
• ▶ 32:25 Spinoza declined Heidelberg’s professorship to stay “under your own power”; the real question is who commands your power of acting.
• ▶ 32:47 “Personal AGI is how you stay under your own power in the age of agents” – keep your brain and skills in a repo you control from day one.
• ▶ 33:33 The speaker concludes a previous point about personal ownership, using the phrase "like he owned his," serving as a thematic bridge into the next section.
• ▶ 33:36 The speaker pivots sharply by announcing "Now, three objections," acknowledging that the audience is mentally raising objections to his arguments.
• ▶ 33:38 He begins setting up his response ("so let's just do") to address the objections one by one, priming the audience for a structured rebuttal.
• ▶ 33:50 Better models don’t make the harness obsolete; instead, the differentiator shifts to context.
• ▶ 33:57 When everyone has the same engine, the race is won by the driver and the map — model weights are shared, but your library is yours.
• ▶ 34:07 Better models make your library more valuable, and every new release is a free upgrade to a workforce you already own.
• ▶ 34:29 Retrieval is a primitive, not the product: the speaker concedes "Sure, and Postgres is just B-trees" to show that pointing to RAG doesn't diminish the architecture.
• ▶ 34:33 The real value is the system around retrieval: deciding what gets written down, enriching/linking content, promoting hot memory vs cold reference, and arbitrating conflicting facts.
• ▶ 34:45 Core thesis: "Retrieval is easy. Being worth retrieving from is the product" — the product is curation, structure, and judgment, not the lookup mechanism.
• ▶ 34:51 The privacy objection is the most serious challenge: consolidating your entire life into one system raises the question of what happens when it leaks.
• ▶ 35:00 The answer is that it has to be yours — the solution is to own the infrastructure, repo, and keys, not to rely on third-party promises.
• ▶ 35:25 Taking custody of your data is the actual security model; trusting someone else's terms of service is weaker than trusting yourself with your own keys.
• ▶ 35:34 Garry open-sources his entire personal operating system, and his surface-level reason is "because I can" — being at Y Combinator removes financial pressure to keep it proprietary.
• ▶ 35:53 The real reason: tools of the powerful should be given away; literacy, capital, and now the harness/library/Markdown workforce have each been eras' private leverage, and the gap between those with and without them is widening monthly.
• ▶ 36:31 The conference's purpose is to let attendees build this power for themselves, because keeping such tools private creates a priesthood while giving them away creates a renaissance.
• ▶ 36:51 Garry Tan chooses a future he wants to live in, tying it to agency and conviction.
• ▶ 37:03 He recites a five-part creed: say, fund, build, write/give away, and leave behind what "other people won't."
• ▶ 37:14 The repeated "other people won't" frames the creed as a commitment to contrarian, high-leverage action.
• ▶ 37:17 Building in the open means you can see outcomes coming that others won’t yet recognize.
• ▶ 37:26 Spinoza’s story shows even the era’s greatest mind—Leibniz—secretly studied and borrowed from a publicly hated thinker, then lied about it for 40 years.
• ▶ 37:56 Garry Tan lives this weekly: agents write most of his code, the public criticism arrives fast, but the loudest critics are themselves quietly shipping with agents—so learn the pattern now.
• ▶ 38:12 Public response to new ideas follows two stages: first mockery (“quote tweet you”), then imitation (“clone you”)—being cloned is the real sign of validation.
• ▶ 38:14 The negative replies (“dunks”) are not genuine refutations but symptoms of the adoption curve announcing itself as the idea starts breaking through.
• ▶ 38:19 The speaker transitions to showing how this same architecture can be pointed at the one thing that truly matters, setting up the next section.
• ▶ 38:24 A father built a personal medical knowledge base for his son's rare epilepsy without waiting for institutional permission—no lab, grant, or authorization.
• ▶ 38:35 He created a repository of 80,000 indexed and cross-linked markdown files covering every specialist visit, paper, seizure log, and drug interaction—a "brain for one small boy."
• ▶ 39:04 This story defines personal AGI: not a benchmark or demo, but a real deployed system—the library, the librarian, and the right three books open—for one person's most important goal.
• ▶ 39:31 The real obstacle was never you—traditional requirements like teams, funding, and credentials were just workarounds for human limits.
• ▶ 39:48 Those limits have now expired: you can "fly" mechanically, not metaphorically, because the practical tools exist.
• ▶ 40:14 The thesis lands: "We can boil the ocean now"—the impossible scale of the past is no longer out of reach.
• ▶ 40:18 Garry's core creed: "It's all made up. But you get to make it up" — a source of agency, not cynicism, because invented structures can be reinvented.
• ▶ 40:28 Even imposing institutions (like the one that cursed a 23-year-old in 1656) were "made up by people no smarter than you," deflating authority and elevating the listener.
• ▶ 40:46 Today's founder doesn't need a crowd of believers — just a laptop and a few years of personal history. Your accumulated life experience is enough to begin building.
• ▶ 40:55 Garry highlights the scale of 7,000 attendees as "7,000 conatuses" and "7,000 strivings," contrasting them with historical striving that died waiting for funding, head count, permission, or belief.
• ▶ 41:18 He identifies the technology shown as the first that lets striving go straight to work—requiring one person, no intermediaries, and no permission—and asserts the world doesn't yet grasp what 7,000 people with such leverage will do.
• ▶ 41:41 Citing Spinoza's closing words—"All things excellent are as difficult as they are rare"—Garry declares the difficulty has collapsed and the rarity is now up to the audience, ending with "Go and build."

• ▶ 0:07 Spinoza’s Excommunication and the Dangerous Book (0:07 - 3:54) - - Spinoza is erased by his community, survives an attack, and keeps writing while grinding lenses.
• ▶ 3:54 From Conatus to Personal AGI (3:54 - 6:45) - - Spinoza’s heresy and conatus become the frame for a personal AGI that arrives diffused.
• ▶ 6:45 Corporate AGI vs. the Asset You Build (6:45 - 9:49) - - Rented corporate AGI is a product; your own AGI is an asset, and coding agents multiply knowledge work 400x.
• ▶ 9:49 The Equation and Spinoza’s Joy (9:49 - 12:29) - - Joy is power increasing, and a frontier model plus your own context is the equation for the next decade.
• ▶ 12:29 Spin Lenses, Working Memory, and G Brain (12:29 - 16:34) - - Mind-lenses extend seven-slot working memory to a life library, and G Brain works as a colleague.
• ▶ 16:34 G Stack, Skill Files, and Where Computation Happens (16:34 - 19:44) - - Skill files are plain-English instructions, and computation runs in latent or deterministic space.
• ▶ 19:44 The Spinoza Compendium and Your Life Library (19:44 - 42:06) - - Deep research builds the Spinoza Compendium, turning one shelf of a life into 220,000 pages.