Garry Tan: Own Your Intelligence

Garry Tan opens with the story of Baruch Spinoza, a 17th-century philosopher excommunicated for heretical ideas, yet who maintained intellectual freedom by grinding lenses by day and writing philosophy by night. Tan uses Spinoza's concept of 'conatus' (striving/the drive to increase one's power to act) as the foundation for understanding how AI agents can amplify human capability.

Tan reframes the AGI conversation, arguing that rather than arriving as a singular event, general intelligence is already diffusing through infrastructure as personal tools. He contrasts 'corporate AGI' (rented, reset when closed, generic) with 'personal AGI' (owned, running on personal infrastructure, improving daily with accumulated personal context). This distinction is central: one is a product consumed, the other an asset built.

Tan provides concrete evidence of this shift at Y Combinator, describing a 400x multiplier in his own coding output since 2013 through agent assistance, and noting that companies in the Winter 2025 batch with 95% AI-generated codebases are tracking as the fastest-growing, most profitable batch in YC history. He emphasizes that the leverage isn't in the model weights (which are commoditized) but in context relevance and application.

The technical architecture centers on three components: a frontier model (rented, getting cheaper), personal context (owned, unique), and a harness connecting them. Tan's personal system, 'G-Brain,' consists of 220,000 markdown pages of 25 years of life data, processed by agents for retrieval and synthesis. Skill files—markdown documents encoding processes and judgment—are the key innovation; they're executable, reusable pieces of cognition that an agent can deploy repeatedly.

Tan walks through his actual workflow: agents process his inbox while he sleeps, pull context before meetings, finish research overnight, and execute recurring tasks. He emphasizes the distinction between latent-space computation (judgment, taste, understanding intent) and deterministic computation (arithmetic, databases, scheduling), arguing that successful agents leverage both through markdown files calling code.

He provides a practical five-step implementation guide: pick a harness and run an agent locally, start a library with markdown files of personal history, write your first skill file for a weekly task you hate, wire it to run as a recurring job, and never throw away context—'skillify' completed work into reusable processes.

The critical second half addresses ownership and political implications. Tan argues that skill files represent externalized cognition—the same file creates two opposite futures depending on who controls the repository. He uses the example of a support engineer whose 40 skill files represent two years of judgment: if she owns them, they compound her career wherever she goes; if the company owns them, her expertise executes forever while she's forgotten. This mirrors the historical exploitation of craftspeople who lost leverage when factories owned the tools.

Tan positions this as a modern version of Spinoza's dilemma: in 1673, Heidelberg offered Spinoza a professorship with 'freedom to philosophize' provided he not disturb established religion—essentially offering comfort for compliance. Spinoza declined, protecting his power to act under his own terms. Similarly, founders should reject comfortable arrangements where their judgment compounds in someone else's repository, framing this as a choice between staying 'under your own power' or ceding it to an acquirer.

Addressing three objections: improving models don't make the harness obsolete (better models make personal context more valuable, not less); RAG is just a primitive, the real value is in what's worth retrieving; and consolidating personal context isn't a security risk—it's taking custody of data already scattered across ten corporate clouds.

Tan explains his decision to open-source everything: because he can (being at YC means he doesn't need to monetize infrastructure), but more importantly because powerful tools should be given away to prevent a priesthood and enable a renaissance. He frames the current moment as one where a private technology of leverage exists—the ability to operate agents at scale with personal context—creating widening gaps between those who have it and those who don't.

The talk concludes with Spinoza's final chapter: after visiting Spinoza secretly, Leibniz spent 40 years publicly denying it while privately obsessing over his notes. Tan notes he experiences a parallel pattern—announcing agent usage draws dunks by lunch, but the loudest critics are shipping agents 'all the way down.' Building in public reveals adoption curves disguised as criticism.

Tan offers a powerful example: a father built an 80,000-markdown-file brain for his son with rare epilepsy, creating a personal research system nobody was coming to build for him. This embodies personal AGI's purpose—not benchmarks or demos, but solving the problems you care about most.

He concludes that previous generations of founders needed to recruit believers before building anything; today's founders need only a laptop and years of personal history already accumulated. The machinery he described removes intermediaries between individual striving and actual work, and he questions whether the world understands what 7,000 people with that leverage will build when they leave the conference.