Garry Tan: Own Your Intelligence

Episode Summary

Executive Summary: The talk argues that AI’s real revolution is not a distant AGI event but “personal AGI”: owned, context-rich agents that compound with your life and work. Using Spinoza’s exile as a frame, the speaker claims skill files, memory libraries, and harnesses let individuals convert knowledge work into leverage, creating faster output, better memory, and startup-like compounding—while warning that ownership and custody of these systems are politically and economically decisive.

Main Topics: Spinoza as a model of intellectual independence: The speech opens with Spinoza’s excommunication, lens-grinding work, and posthumous influence to symbolize truth-seeking, resilience, and building despite social rejection. Personal AGI vs. corporate AI: The core thesis distinguishes rented chatbot products from owned systems that run on personal infrastructure, memory, and workflows, becoming an asset rather than a subscription. Agents as a labor multiplier: The speaker argues coding and broader knowledge work now scale dramatically through agents, turning one person into a far more productive operator across coding, design, PM, growth, and operations. Memory architecture: library, librarian, and skill files: The talk outlines a personal operating system built from markdown skill files, a curated knowledge base, and an agent that retrieves, processes, and acts using the owner’s context. Deterministic vs. latent computation: The speaker emphasizes that some tasks should be handled in the model’s latent space (judgment, interpretation), while others require deterministic systems like SQL and code. Ownership, custody, and power: A major warning is that the same skill-file system can either empower workers or extract their cognition, depending on who owns the repo and infrastructure. How to start building now: The talk ends with a practical playbook: choose a harness, create a starter library, write a skill file, schedule recurring jobs, and keep capturing reusable processes.

Key Arguments: AI’s true shift is not AGI as a single event, but personal AGI distributed through your own workflows and context. Corporate AI assistants are limited because they are rented, reset, and controlled by vendors; owned personal systems compound daily. Agents can massively increase output in coding and other knowledge work; the speaker claims his own output rose by orders of magnitude using them. The biggest performance gap is not model weights but the quality, relevance, and timing of context given to the agent. A curated personal knowledge library turns the agent from a generic assistant into a colleague that knows your history and can act with continuity. Skill files externalize human judgment into reusable executable instructions, making anyone who can write clear English effectively a programmer. Memory systems need hygiene, provenance, and contradiction checks; otherwise they become confident but unreliable garbage dumps. The political question is ownership: if skill files and cognitive workflows live in someone else’s repo, your judgment becomes extractable labor. Open-sourcing these tools is framed as democratizing leverage that would otherwise become a priesthood or private advantage. The speaker believes startups and entire organizations can now begin as one person plus agents, with dramatically higher revenue-per-employee potential.

Data Points: Spinoza excommunication age: 23 - He was excommunicated from the Sephardic Jewish community at age 23 for “evil opinions.” Years the ban remains in force: Still in force today - The excommunication is described as never lifted and technically still active. Offer to Spinoza: 1,000 guilders a year - The community tried to pay him to stay quiet and attend synagogue occasionally. Alternative offer mentioned: 10,000 - He is quoted as rejecting the idea that he could be bought even for 10,000. Age at death: 44 - Spinoza died relatively young after years of lens grinding. Lens-dust consequence: Lungs full of glass dust - His death is attributed to the physical toll of grinding optical instruments. Working memory limit: 7 plus or minus 2 - Used to illustrate the human limit compared with agent memory. Agent context window: 1 million tokens - Described as roughly a thousand pages open at once. Human effort benchmark (2013): 14 useful lines of code/day - The speaker says this was his approximate output while building Bookface at night. Productivity multiplier claim: 400x - He says his current output versus 2013 is about 400x, with caveats. Conservative productivity floor: 8x - After accounting for verbosity and scaffolding, he claims the gain is still at least 8x. YC batch AI-generated code share: 25% - In Winter 25, a quarter of companies reportedly had codebases 95% AI-generated. AI-generated code proportion: 95% - Describes the codebase composition of those companies in that batch. Revenue milestone: $15M annualized revenue - Emergent is cited as crossing this mark with 15 people. Headcount at milestone: 15 people - Emergent reached $15M annualized revenue with only 15 employees. Retail revenue: $60M annualized - Retail is cited as reaching this scale. Retail headcount: About 40 people - The company is cited as reaching $60M annualized with roughly 40 employees. Knowledge base size: 220,000 markdown pages - The speaker says his personal brain system contains this many pages. Personal history captured: 25 years - He claims the library includes 25 years of his life diarized. Open-source project size: 123,000 stars - His coding framework is said to have this many GitHub stars. GitHub ranking: Top 100 - He claims the framework is among the top 100 open source projects in GitHub history. Conference seating scale: 6,000 people - Used to illustrate a deterministic scheduling problem handled by code plus markdown. Biographies synthesized: 3 books / about 1,500 pages - His agent read three Spinoza biographies totaling roughly 1,500 pages. Support engineer skill files example: 40 skills - A fictional example of one worker externalizing two years of expertise into 40 files. Initial archive example: 80,000 markdown files - A friend built a medical brain for his son with this many files. Event attendance: About 7,000 people - Used to argue that many attendees could leave with this leverage and build immediately.

Pivotal Quotes: "AGI isn't arriving. As an event, it's arriving diffused as your agent, running on your context, doing your work." — Speaker: Core thesis redefining AGI as personal, distributed leverage rather than a singular breakthrough. "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." — Speaker: Explaining the split between latent-space judgment and deterministic computation. "Own your skills, because if you don't, your job becomes a skill file." — Speaker: Warning that externalized cognition can become worker extraction if owned by employers or platforms.

Implications: Listeners are urged to stop waiting for a future AGI and start building owned agent systems now. The likely winners will curate context, codify skills, and treat AI as infrastructure they control, creating major productivity and startup advantages.