AI-Wiki

Garry Tan: Own Your Intelligence

16 min read

Garry Tan: Own Your Intelligence

▶ Watch on YouTube · Y Combinator · 42:08

The next generation of startups will be built by smaller teams than ever before.

At Startup School 2026, YC President & CEO Garry Tan explains why we’re entering the era of personal AGI: AI agents that run on your own infrastructure, compound your knowledge over time, and dramatically increase your ability to build. He shares the tools and workflows he uses every day, why every founder should own their intelligence instead of renting it, and what it means to build under your own power.

— Channel description, Y Combinator

A ~42-minute Startup School keynote by Garry Tan built around a single thesis — AGI is arriving diffused rather than as an event, as “your agent running on your context doing your work” — and structured in three movements: a Spinoza frame, a concrete how-to for building personal AI infrastructure, and a political argument about who ends up owning externalised cognition.

Its value to the wiki is that it is the most explicit first-party account of the LLM-wiki pattern operating at scale in production — GBrain is a ~220,000-page markdown knowledge base with agents as its compilers, curators and searchers — and that it supplies the corpus’s sharpest statement of the ownership question at the level of the individual worker, which open-source-ai has so far held only at organisational scale.

TL;DR

  • The thesis, and the deliberate rejection of the event model. “Everyone is watching the sky and the thing they’re watching for is already in the room. It doesn’t look like a god. It looks like infrastructure, a terminal window, a folder of markdown files, a job that finishes while you sleep.” He calls it personal AGI — “not artificial general intelligence for everyone all at once; general intelligence for one person” — and separates it explicitly from what marketing has captured: not a chatbot subscription, not better autocomplete, not a calendar assistant. “That’s just a subscription you rent. It’s a corporate AGI you don’t own. It resets when you close the tab… and when the company behind it pivots, your so-called assistant gets a lobotomy on someone else’s schedule.”
  • The equation. A frontier model (rented, commoditising, cheaper each quarter) + your context (owned, ideally unique to you) + a harness that wires them together (OpenClaw, Hermes, Claude Code, Codex) = an agent that acts like a fast version of you. “Model quality is rented but your brain is owned.”
  • The productivity claim, and his own discount of it. ~14 useful lines of code a day in 2013 — “dead on median” by the programmer-productivity literature — against roughly 400× today. He then discounts it himself: apply the most pathological verbosity penalty, assume half is scaffolding, assume self-flattery, “it’s still 8x at the absolute floor.” The generalisation he actually wants: the multiplier “is not just for coding. It’s for every piece of knowledge work.”
  • The leverage is not in the weights. “There are 2x people and there are 100x people who are using the same Claude. Same weights, same context window size, same API. But the leverage is not in the weights. It’s in what context you give it, how relevant it is, and does it happen at the right step.”
  • Working memory as the framing device. Humans hold about seven things at once; “every institution humanity has ever built — every checklist, every org chart, every filing cabinet, every standup meeting — is a prosthetic for that limit.” An agent holds a million tokens, about a thousand pages. But “a thousand pages is a lot, and it is also very little. Your life is not three books. Your life is a library.” Hence the operative question: who or what decides which three books are open on the desk. “That’s what a brain is… the library plus the librarian.”
  • GBrain, described concretely. ~220,000 markdown pages, 25 years diarized — email, meetings, notes, photos, drafts, the things he got wrong — “compiled mostly by agents, curated by agents, searched for by agents.” The lived test: a founder emails about a crisis and before he finishes reading, the agent has pulled every prior conversation with that founder and three portfolio companies that hit the same wall and what worked. “When my agent does anything, it does [it] knowing everything I know. And that’s the difference between an assistant and a colleague.”
  • “Fat skills, thin harness.” GStack, his agent-coding framework, is “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.” A skill file is read out in full on stage and it is a page of plain instructions. The test he gives: “if a smart intern could follow it, an agent can run it.” Hence “markdown is actually code… the compiler is a language model” — and hence YC’s media, events and finance staff, who never open a terminal, write skill files. One finance colleague compiled ~100 Excel workbooks into a single app. “She is not a programmer. She is a manager of agents.”
  • The single most useful diagnostic: latent versus deterministic space. “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.” Taste, judgment and reading what a human means from a vague request live in latent space and are steered with a markdown file. Arithmetic, SQL, a seating chart live in deterministic space and must be written as code against a database. Seat five people at a table: latent. Seat 6,000 people in an arena: the latent agent must write code. “The model fails where we fail. The fix is having the model compute the way humans compute.”
  • The compendium skill, as a worked example. Five days before the talk he decided it needed Spinoza. The agent acquired three biographies (~1,500 pages), read all three, and produced a synthesis: a dated chronology, every place the three biographers disagree with each other, verbatim quotes with chapter citations, and the ten most tellable moments ranked with delivery notes. “1,500 pages became a stage-ready story that I could edit.”
  • The five steps. (1) Tonight: pick a harness, run an agent on your own machine. (2) This weekend: start the library — “not a grand archive. One folder of markdown files” — a page per project and per person, containing “stuff no model on earth has because it only exists in your head.” (3) Write one skill file for the weekly task you hate most, let it get things wrong, correct it. (4) Wire it to a recurring job. (5) Never do one-off work — at the end of every task, ask the agent to skillify what it did. “If you have to ask for something twice, you failed.”
  • The 90-day curve, stated including the bad part. Week one “honestly, it’s a toy” — thin library, clumsy skills, fixing more than saving. Week four the flywheel catches. Week twelve, a library that answers before you finish asking. “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.”
  • The curation caveat, unprompted and load-bearing. “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 stated primitives: 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.”
  • The political turn. 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.” The worked case: Maya, a support engineer, teaches her agents 40 skills over two years. If they live in her repo, she carries two years of compounded judgment to her next job. If they live in the company’s repo, “she leaves with nothing. The company keeps running her judgment without her… She didn’t have a career. She had an extraction.” Hence the doctrine: “Own your skills, because if you don’t, your job becomes a skill file.” The historical frame: craftsmen owned their tools and that made them free; the factory broke that; knowledge workers assumed they were exempt because their tools lived in their heads. “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?”
  • Three objections, answered. (1) Models will obsolete all this. “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… A better model makes your library worth more because a smarter reader extracts more from the same books.” (2) Is this just RAG? “Sure, and Postgres is just B-trees. Retrieval is the primitive, not the product… Retrieval is easy. Being worth retrieving from is the product.” (3) What happens when it leaks? “That’s exactly why it has to be yours… 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.”
  • Revenue-per-person examples. Emergent (S24) public launch to nine figures of revenue in eight months, and 15 people at $15M annualised; another portfolio company at $60M annualised with about 40. “That revenue per person did not exist before. Not in software, not in oil, not in railroads.”
  • The closing case. A friend whose son has a rare epilepsy built a repo of 80,000 markdown files — every specialist visit, paper, seizure log and drug interaction, indexed and cross-linked — so that when a new doctor proposes something he 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.”

Why this matters to the wiki

1. It is a first-party account of the pattern this repository implements. llm-wiki describes a persistent, interlinked markdown knowledge base that an LLM incrementally builds and maintains, compounding rather than being re-derived per query. GBrain is that pattern at 220,000 pages with a quarter-century of source material, and Tan independently states the same three disciplines this wiki’s schema encodes: provenance on every fact (the wiki’s citations beat assertions), contradiction checks when new information collides with old (the ## Debates and supersession contract), and a librarian whose job is pruning (the maintenance argument for why the pattern needs an agent at all). The convergence is worth recording precisely because it was reached independently.

2. It relocates the own-vs-rent thesis from the firm to the worker — and adds a politics. open-source-ai carries own-vs-rent as an economics of production: cost at scale, sovereignty, concentration of power. Tan runs the identical structure one level down and reaches a labour conclusion the concept page has no analogue for: skill files make cognition extractable for the first time, so the question of who holds the repo becomes a question about careers. That connects to ai-deskilling from an unexpected direction — the risk is not that the worker’s skill atrophies but that it is successfully captured while remaining intact.

3. The latent-versus-deterministic diagnostic is a genuinely portable failure taxonomy. agent-harness accumulates failure patterns from many sources; this is a single axis claimed to explain “every agent failure I’ve ever seen”, with a crisp decision rule (does the task require holding exact state at a scale beyond judgment?) and a worked example at both ends. It is also a reframing of why tool use exists at all, and it converges with the same week’s Gemini video-understanding source, where the whole design move is refusing to do in latent space what a tool can do exactly.

4. It supplies the corpus’s clearest statement of why context beats weights and why better models don’t erode that. The wiki has the first half from many angles. The second half — that model progress increases the value of an owned corpus rather than commoditising it — is an argument agent-harness’s “the model is rented, the harness is owned” motif implies but never states.

Dynamic-capabilities reading

  • digital-sensing/digital-mindset-crafting — Most of the talk is mindset work aimed at an audience of founders: the AGI-as-diffusion reframe, the 90-day curve with its explicit warning that week two is where people quit, and the closing insistence that “it’s all made up, but you get to make it up.” The keynote’s function is to change what the room believes is possible before it changes what they build.
  • digital-seizing/rapid-prototyping — The five steps are a prototyping protocol: run something tonight, one folder this weekend, write a skill and let it get things wrong, correct it, wire it to a cron. Explicitly anti-planning — “nobody builds the warehouse first. First, you build one shelf.”
  • digital-seizing/strategic-agility — The claim that software need no longer be precious: “you can build exactly the tool you need for the audience of one in a weekend”, and the revised Paul Graham advice — scratch your own itch because scratching itches is nearly free, and some of the tools for one turn out to be companies.
  • digital-transforming/improving-digital-maturity — The non-engineer story is a maturity claim about a workforce, not a person: YC’s media, events and finance staff writing skill files and scheduled jobs, and the finance colleague who replaced ~100 workbooks with an app. “She is not a programmer. She is a manager of agents. Now, everyone is about to be.”
  • strategic-renewal/business-model — The revenue-per-person examples and the “company of one plus your agents” framing are a claim that the unit economics of a software company have changed, not merely its tooling: “before you ever incorporate anything… you can already be running an organization.”
  • contextual/internal-enablers — Open-sourcing the harness, the brain architecture and the skills is presented as the enabling move, with an explicit rationale: “tools of the powerful should be given away… When something that powerful stays private, you get a priesthood. When it gets given away, you get a renaissance.”

Linked entities and concepts

Debates and supersession

  • The 400× figure is not a measurement and he half-says so. It compares lines of code in 2013 against agent output in 2026 — a metric he simultaneously invokes and disowns (“you don’t trust the raw lines of code. Fine.”). The 8× floor is offered as robust but is derived by discounting the same unmeasured quantity. The wiki should carry the mechanism claims — context ownership, skillify, latent-versus-deterministic — as the transferable content, and treat the multipliers as illustration. See ai-coding-productivity-evidence, where measured studies land far below practitioner self-report.
  • Selection is doing work in the portfolio evidence. The W25 batch statistic (a quarter with 95% AI-generated codebases, on track to be among YC’s fastest-growing) is presented with a correlation caveat he states himself — “I cannot prove that the AI generated code and everything else caused the growth” — which is the right caveat, and worth preserving alongside the claim. The revenue-per-person examples are the accelerator’s outliers by construction.
  • The tension with the aggregate-productivity cluster is unaddressed. Frey and the wider return-gap literature hold that vivid individual gains coexist with absent aggregate growth. Tan’s talk is a maximal statement of the individual side and does not engage the aggregate side at all. This is recorded as a contradicts edge in the frontmatter not because the facts conflict but because the two sources cannot both be a complete account, and the wiki should not let the keynote’s vividness settle a question the empirical cluster holds open. See micro-productivity-trap.
  • The ownership doctrine has an unexamined asymmetry. “Own your skills” is advice a YC audience can act on — founders own their repos by default. The Maya case is about an employee, and an employee’s ability to keep skill files developed on company time and company systems is a legal question about work-for-hire that the talk does not raise. The doctrine may be sound and still not be available to the worker it is addressed to, which is the more uncomfortable version of the argument.
  • Open question — does the curation discipline scale to 220,000 pages, and how is it verified? He names the failure mode precisely (confident retrieval of stale facts) and names three remedies, but gives no account of how contradiction checks are run at that volume, what fraction of the corpus is agent-generated versus source material, or how pruning decisions are audited. For a wiki built on the same pattern this is the operationally interesting question, and it is left open.

What was actually ingested

Full ~42-minute keynote transcript (auto-generated English captions). Proper nouns corrected against the channel description and chapter list; several — notably Spinoza, conatus, Leibniz and Vannevar Bush — are badly mangled in the ASR and are rendered here in their standard forms. Slides are referenced throughout (the skill-file screenshot, the architecture diagram, the stack slide) and were not ingested; all descriptions of the system are as narrated. One portfolio company’s name in the revenue-per-person examples is unrecoverable from the ASR and is referred to descriptively rather than guessed. The Spinoza biographical material is the speaker’s framing device and is reported as such, not as a claim the wiki vouches for.

Relationships

Supports

  • How Companies Are Building Their Own Intelligence | Sonya Huang, Sequoia Capital — The same slogan and the same month, one level of aggregation apart — and the coincidence is not one, since both events are literally titled Own Your Intelligence. Huang argues a firm should own its weights because intelligence is too core to outsource; Tan argues an individual should own their context and skill files for the same reason. They also agree on what stays rented: Huang is explicit that companies should not get off Opus or GPT, and Tan's equation keeps the frontier model as the rented, commoditising term. The interesting divergence is what the owned asset IS — weights and post-training for Huang, accumulated personal context and procedures for Tan.
  • Hugging Face's CEO on why companies are done renting their AI | Equity Podcast — Both make ownership the thesis, but Tan answers an objection Delangue's cost-driven argument does not reach: what happens when models improve. Tan's reply is that a better model makes an owned library worth MORE, because a smarter reader extracts more from the same books — which turns the frontier labs' progress from a threat to the owned layer into a free upgrade to it. That is a different defence of ownership than cost-at-scale.
  • I built a Claude Cowork system that does a week of PM work in a day — The same personal-infrastructure architecture reached independently from opposite starting points — a PM inside an enterprise's fixed stack, and an accelerator CEO with unlimited latitude — and converging on the same primitives: one harness holding centralised context, markdown skill files as the unit of procedure, scheduled jobs producing a morning briefing, and a loop that turns observed repetition into new skills. Tan's skillify discipline is Blum's self-improvement loop stated as a rule; Blum's J-curve is Tan's week-one-is-a-toy, week-twelve-it-answers-before-you-ask timeline measured from inside one job.
  • Stanford CS153 Frontier Systems | The AI Native Company: How One Founder Becomes a 1000x Engineer — Same speaker, three months on, with the argument moved from the firm to the person. CS153 makes the 1000x-engineer case about how an AI-native company is built; this keynote asks who owns the leverage once it exists, and adds the ownership politics — skill files as extractable cognition — that the earlier talk does not raise.

Contradicts

  • Why AI Isn't Actually Boosting Productivity | Trumponomics — Not a factual conflict so much as an unresolved one about what individual gains imply. Tan reports ~400x his own 2013 output and floors it at 8x under maximal discounting, and cites a YC batch where a quarter of companies ran 95%-AI-generated codebases; Frey's position is that such gains are real at the task level and yet absent from aggregate productivity. Tan does not engage the aggregate question at all, and his revenue-per-person examples are selected from an accelerator's winners — precisely the survivorship channel Frey's argument predicts would produce vivid anecdotes without moving the statistics.

Graph View

Backlinks