![\"\"\/](\"https:\/\/www.europesays.com\/ai\/wp-content\/uploads\/2026\/08\/349ac2fb44d29d82_1786048789_inline_1.jpg\")
<\/p>\nIn 1656, Amsterdam’s Sephardic community excommunicated a 23-year-old lens grinder named Baruch Spinoza with the most violent curse in their arsenal \u2014 “Cursed be he by day and cursed be he by night” \u2014 and banned anyone from reading a word he wrote. A fanatic later attacked him with a knife. The community offered him a thousand guilders a year simply to show up at synagogue occasionally and keep quiet. He declined. By day, Spinoza ground lenses so precise that Europe’s finest scientists sought them out. By night, he wrote a book too dangerous to publish while he lived. When he died at 44, lungs full of glass dust, the manuscript sat locked in a writing desk, shipped by canal barge to a publisher who turned it into the foundational text of the European Enlightenment.<\/p>\n
Garry Tan \u2014 president of Y Combinator since 2013 \u2014 opened his talk on the YC Startup Podcast with that story, and it is not decoration. His argument, delivered across 42 minutes, is that Spinoza’s position \u2014 a man of compoundable intellectual leverage, surrounded by offers to stop building, threatened with deletion, and responding by owning his own tools \u2014 is the exact structural position of the knowledge worker in 2026. The technology that now makes this possible, Tan argues, is not waiting for a grand announcement. It is already in the terminal window, the folder of markdown files, and the overnight job that finishes while you sleep.<\/p>\n
Personal AGI versus corporate AGI: what you rent and what you own<\/p>\n
Tan’s central distinction is between two kinds of artificial intelligence. The first is what the industry markets: a $20-per-month chatbot, a better autocomplete, an assistant that resets when the tab closes. “That’s just a subscription you rent,” he said. “It’s a corporate AGI you don’t own. It knows what everyone else already knows. And when the company behind it pivots, your so-called assistant gets a lobotomy on someone else’s schedule.”<\/p>\n
The alternative \u2014 what Tan calls “personal AGI” \u2014 is built on three components: an agent that runs on infrastructure you control, reads from a memory you own, and executes procedures you wrote. The consequence is compounding. “Your personal AGI gets better every single day you use it because every day it knows more of your life.” The one-sentence summary: “One of these is a product you consume. The other is an asset you build.”<\/p>\n
The architecture rests on a simple equation. The frontier model \u2014 the large language model behind the agent \u2014 is rented, a commodity getting cheaper by the quarter. Your context \u2014 your emails, meeting notes, drafts, mistakes, and everything you know about the people you work with \u2014 is owned, and ideally unique. The harness, which wires the two together, can be built with tools like OpenClaw, Hermes, Claude Code, or Codex. “Model quality is rented, but your brain is owned,” Tan said. “AGI isn’t arriving as an event. It’s arriving diffused as your agent running on your context doing your work.”<\/p>\n
<\/p>\n
The library, the librarian, and a workforce made of markdown<\/p>\n
Tan’s architecture addresses a specific human limitation: working memory holds roughly seven items at once, a finding published by psychologist George Miller in 1956. An AI agent’s context window, by contrast, holds about a million tokens \u2014 roughly a thousand pages, “three Harry Potter books sitting open on its head all at once.” But a thousand pages is also very little; a life is a library, not three books. “The question that determines whether your agent is a genius or a goldfish,” Tan said, “is who decides \u2014 or what decides \u2014 which three books are open on the desk.”<\/p>\n
His answer is a system called GBrain: a personal knowledge library of roughly 220,000 markdown pages, encompassing 25 years of his life \u2014 every email, meeting, note, photo, draft, and mistake, “compiled mostly by agents, curated by agents, searched for by agents.” When a founder emails about a crisis, Tan’s agent has pulled every prior conversation with that founder plus three portfolio companies that hit the same wall and what actually worked before he finishes reading. “That’s the difference between an assistant and a colleague.”<\/p>\n
The atomic unit of the system is the skill file: a page of plain English instructions that an agent can execute. Tan showed an example from his own workflow \u2014 when a meeting recording arrives from the transcription tool Circle Back, the agent transcribes it with speaker labels, extracts commitments with deadlines, cross-checks every named person against the library and links their pages, files the summary and transcript, and flags contradictions without overriding them. “That’s it. That’s a skill. It’s a page of English,” Tan said. “If a smart intern could follow it, an agent can run it.” His test is deliberately deflationary, and his most provocative restatement of what coding has become is this: “Markdown is actually code. If you can write clear instructions in English, you’re a programmer. The compiler is a language model.”<\/p>\n
“You’re not coding,” he added. “You’re man managing a workforce made of markdown.”<\/p>\n
That workforce requires maintenance. “A brain nobody curates is a garbage dump with great search,” Tan warned. Every fact needs provenance. Every contradiction between old and new information needs arbitration. The librarian’s real 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.”<\/p>\n
![\"\"\/](\"https:\/\/www.europesays.com\/ai\/wp-content\/uploads\/2026\/08\/349ac2fb44d29d82_1786048877_inline_2.jpg\")
<\/p>\n
The productivity numbers that should not exist<\/p>\n
Tan measured his own output against a baseline: in 2013, building YC’s internal social network Bookface at night, he shipped roughly 14 useful lines of code a day \u2014 the median for a programmer at full effort. By 2026, running YC full-time with the same hours and a 5 p.m. kid pickup, he estimates his output at roughly 400x his 2013 rate. He pre-deflated the number for skeptics: apply the most aggressive verbosity penalty, assume half of it is scaffolding, assume he is flattering himself \u2014 the floor remains 8x, with the middle of the range around 80x. The multiplier, he said, applies to design, product management, and growth, not just code.<\/p>\n
The portfolio-level evidence comes from YC batch data.<\/p>\n
CompanyYC batchRevenueHeadcountRevenue per employeeEmergentSummer 2024$15M annualized15 people~$1MRetailWinter 2024~$60M annualized~40 people~$1.5M<\/p>\n
“That revenue per person did not exist before,” Tan said. “Not in software, not in oil, not in railroads.”<\/p>\n
In the Winter 2025 batch, roughly a year and a half before the talk, a quarter of companies had codebases that were 95 percent AI-generated, and that batch is on track to become one of the fastest-growing, most profitable in YC history. Tan is careful about causation \u2014 he cannot prove the AI code caused the growth \u2014 but the pattern is clear. The fastest-growing funded founders “are not treating AI as autocomplete. They are treating it as a workforce.” Two founders using the same model, same weights, same context window, and same API can differ by “2x people” versus “100x people.” The variable is not the engine. “The leverage is not in the weights,” Tan said. “It’s in what context you give it, how relevant it is, and does it happen at the right step.”<\/p>\n
Founders who fail to adopt agent-driven workflows, he predicted, will be outcompeted by those who do.<\/p>\n
The five steps that separate the 1 percent from the 99 percent<\/p>\n
Tan estimated that 99 percent of people who watch the talk will not follow the steps. Those who do, he said, enter a compounding curve with a distinct shape: flat, flat, flat, then not. The onboarding path is concrete.<\/p>\n
StepWhenAction1TonightPick a harness and run an agent on your own machine. Tan’s stack is OpenClaw plus Hermes with GBrain, hosted at gbrain.io; the brain itself is free and open source. Codex and Claude Code are fine alternatives \u2014 “I always recommend the Ferrari, but the Honda is really good, too.”2This weekendStart the library \u2014 not a grand archive, but one folder of markdown files. Export notes and email. Write one page per project and one page per person: what you are building, what they care about, what you owe them, what they said last time. “That content exists in no model on Earth.”3NextWrite the first skill file. Pick the weekly task you hate most \u2014 expense reports, meeting notes, status updates, competitor research. Explain it in plain English to the agent, let it get it wrong, correct it, encode every exception. “That page is now an employee.”4ThenWire it as a recurring job: every morning at 7, every Friday, summarize. “The first time you wake up to work that finished while you slept, something shifts in your head permanently.”5OngoingNever do one-off work. At the end of every task, ask the agent to “skillify” what it did \u2014 GBrain ships a skill called Skillify that extracts the procedure into a reusable markdown file. “If you have to ask for something twice, you failed.”<\/p>\n
The 90-day arc follows a classic compounding curve. Week one feels like a toy: the library is thin, skills are clumsy, you spend more time fixing than saving. Week four is when the flywheel catches \u2014 the agent answers with your context, the morning job produces something you actually read, and you write skills three and four because the first two worked. By week twelve, the library answers before you finish asking. A dozen skill files run the parts of your week you used to dread, plus one or two tools other people keep asking to borrow \u2014 “which in this room is called a startup,” Tan said. Most people quit in week two, precisely when the curve has not turned. Those who persist “feel like they’re cheating by week 12.”<\/p>\n
Tan’s demonstration of the loop was the talk itself. Five days before the event, he decided the argument needed Spinoza. His agent acquired three biographies \u2014 by Nadler, Goldstein, and Stewart, roughly 1,500 pages \u2014 read all three overnight, and produced a synthesis: a dated chronology, every place the biographers disagree, verbatim quotes with chapter citations, and the ten most tellable moments ranked with delivery notes. The opening beat of the talk came out of that overnight run. He called this a “compendium skill” \u2014 a personal mega-version of deep research.<\/p>\n
Whose repo gets the forty skill files?<\/p>\n
The pivot from how-to to politics arrives through a fictional but pointed case. Maya is a support engineer. Over two years, she teaches her agents forty skills: how to triage a critical incident at 2 a.m., how to de-escalate a customer about to cancel, how to write a postmortem that actually prevents the next incident. Those forty files are her judgment, externalized \u2014 “the thing that took her two years to build, sitting on a disk.”<\/p>\n
One variable forks everything.<\/p>\n
VariableOwnedExtractedWhere the 40 skill files liveMaya’s own repositoryThe company’s repository, under its IT policyWhat happens when Maya leavesSkills go with her; day one at a new company she operates at years-compounded judgmentThe company keeps running her judgment without her \u2014 40 files executing forever, her name absent from the commit historyWhat the career looks likeShe compounds yearly; the skill library is her expertise, and “entire startups these days will be markdown files””She didn’t have a career. She had an extraction.”<\/p>\n
Tan stated the doctrine directly: “Skill files are yours. Own your skills because if you don’t, your job becomes a skill file.” The historical analogy is explicit \u2014 craftsmen owned their tools, and that ownership made them free. The factory broke that arrangement: the loom belonged to the mill, not the weaver. Knowledge workers assumed they were safe because their tools lived in their heads, where nobody could confiscate them. Skill files end that protection. “For the first time in history, your cognition can be extracted, stored, versioned, and owned. The only question is by whom?”<\/p>\n
The thousand-guilder bribe, Tan insisted, never went away \u2014 it was 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.” In 1673, the University of Heidelberg offered the cursed heretic Spinoza a full professorship \u2014 salary, legitimacy, a chair, and freedom of philosophizing, provided he did not disturb the established religion. Spinoza declined. Tan’s gloss was pointed: “He read the terms of service and he declined the acquisition.” When Spinoza died, the room inventory listed 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.”<\/p>\n
The open-source bet and the Leibniz pattern<\/p>\n
Tan pre-empted three objections. To the argument that models are improving so fast that all harness work will be obsolete \u2014 what he called “the better bitter lesson crowd” \u2014 his response was that every model release moves the differentiator toward context. “When everyone’s engine is 1,000 horsepower, the race is won on the driver and the map. The weights are everyone’s. The library is yours.” A better model makes the library worth more: a smarter reader extracts more from the same books. Every release is a free upgrade to a workforce you already own.<\/p>\n
To the question of whether this is merely retrieval-augmented generation, Tan’s reply was terse: “Sure, and Postgres is just B-trees.” Retrieval is the primitive, not the product. The hard part is everything around it \u2014 what gets written down, how it gets enriched and linked, what gets promoted to hot memory versus filed as cold reference, who arbitrates when facts disagree. “Retrieval is easy. Being worth retrieving from is the product.”<\/p>\n
To the privacy objection \u2014 putting an entire life in one system \u2014 Tan argued the default is not privacy. The default is a life already scattered across ten clouds owned by companies whose incentives are not the user’s, “searchable by everyone except you.” Consolidating context is taking custody, not creating risk. “Custody is the security model.”<\/p>\n
ObjectionRebuttal”Just wait for the next model release \u2014 all this harness stuff will be obsolete.”Every model release moves the differentiator toward your context. A better engine makes your library more valuable. Every release is a free upgrade to a workforce you already own.”Isn’t this just RAG?”Retrieval is the primitive, not the product. The hard part is what gets written, enriched, linked, promoted, and pruned. “Being worth retrieving from is the product.””What happens when all your data in one place leaks?”The default is data already scattered across ten clouds, searchable by everyone except you. Consolidation is custody, not new risk. “Custody is the security model.”<\/p>\n
On why he open-sources the entire stack \u2014 the harness, the brain architecture, the skills, what he called “the whole personal operating system” \u2014 Tan gave a structural answer. “Tools of the powerful should be given away.” Every era has a private technology of leverage: literacy, then capital, and now this \u2014 the harness, the library, the workforce made of markdown. People who have it are operating at a different scale, and the gap widens monthly. “When something like that stays private, you get a priesthood. When it gets given away, you get a renaissance.”<\/p>\n
The final warning came through another Spinoza story. In November 1676, Gottfried Leibniz \u2014 the most celebrated intellect in Europe, silk stockings, a calculating machine in his luggage \u2014 traveled to The Hague and spent three days in an attic with the most hated man on the continent. He then spent the next forty years lying about it publicly, calling it “a few hours in passing,” while his private notes filled with obsessive commentary on Spinoza. Tan recognized the pattern. “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.” The sequence is predictable: “First, they quote-tweet you; then they clone you. The dunks are just the adoption curve announcing itself.”<\/p>\n
Tan closed on a single use case. A friend’s son has a rare form of epilepsy. With no lab, no grant, and no permission, the father built a repository of 80,000 markdown files \u2014 a brain for one small boy, indexed to the edge of what humanity knows about that exact condition. Every specialist visit, every paper, every seizure log, every drug interaction, cross-linked and ready. When a new doctor has an idea, the father knows in minutes whether it has been tried. “A father, a laptop, and a library. That is personal AGI. Not a benchmark, not a demo.” His closing address to the room: “Nobody is coming to build yours for you. That’s the good news.”<\/p>\n
Everything people were told they needed \u2014 the team, the funding, the permission, the credential \u2014 was a workaround for the fact that one person could hold seven things in their head and work sixteen hours a day. That fact, Tan argued, just expired. The two words that tie the entire architecture together \u2014 the library, the skill files, the harness, the open-source bet \u2014 are custody and compounding. Custody of memory, custody of procedure, custody of leverage. The same technology that enables one person to compound at four hundred times their previous rate also enables the cleanest extraction in the history of work: cognition that can be versioned, retained, and run after the person leaves. Tan’s bet is that giving away the tools \u2014 GBrain, GStack, the skills, the whole stack \u2014 is the counterweight, and that the outcome is a choice rather than a forecast. For investors, the signal is unambiguous: revenue-per-employee physics are being rewritten in real time, and the startups that win will not be the ones with the biggest headcount, but the ones whose founders treat markdown files as infrastructure and skill files as the compound interest on their own judgment.<\/p>\n","protected":false},"excerpt":{"rendered":"In 1656, Amsterdam’s Sephardic community excommunicated a 23-year-old lens grinder named Baruch Spinoza with the most violent curse…\n","protected":false},"author":2,"featured_media":132445,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[4],"tags":[6744,3013,2798,2317,5158,65940,65942,65943,43109,576,65946,1379,65945,65944,11376,65941],"class_list":["post-132444","post","type-post","status-publish","format-standard","has-post-thumbnail","category-agi","tag-agi","tag-artificial-general-intelligence","tag-claude-code","tag-codex","tag-emergent","tag-garry-tan","tag-gbrain","tag-gstack","tag-hermes","tag-openclaw","tag-personal-agi","tag-retail","tag-skill-files","tag-spinoza","tag-y-combinator","tag-yc-startup-podcast"],"_links":{"self":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/132444","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/comments?post=132444"}],"version-history":[{"count":0,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/132444\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media\/132445"}],"wp:attachment":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media?parent=132444"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/categories?post=132444"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/tags?post=132444"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}