Conference Presentation, Keynote
Garry Tan: The Future of AGI Is Personal
- Core Thesis: The talk argues for "Personal AGI"—an agent that runs on user-owned infrastructure, utilizes unique personal context, and executes user-defined procedures, distinguishing it from rented corporate AI subscriptions.
- Spinoza as a Metaphor: The speaker uses Baruch Spinoza's 1656 excommunication (the "Herem") as a historical parallel for founders who refuse to trade truth for institutional comfort; Spinoza declined a salary to silence his ideas and was nearly killed for his work.
- The $1,000 Guilders Offer: A specific historical detail notes that Spinoza was offered 1,000 guilders annually to attend synagogue occasionally and remain silent, which he refused to preserve his intellectual independence.
- Productivity Multipliers: The speaker reports a 400x increase in personal output since 2013 due to AI agents, even after applying a "pathological verbosity penalty," with YC data showing a quarter of the Winter 25 batch having codebases that are 95% AI-generated.
- Speed of Adoption: Fastest-growing YC companies are not using AI for autocomplete but treating it as a workforce, leveraging the same underlying models but differentiating via high-quality context and workflow integration.
- Spinoza's Definition of Joy: Joy is technically defined as "the feeling of your power of acting increasing," while sadness is the feeling of that power decreasing; the speaker claims an agent completing a week's work in an afternoon triggers this technical definition of joy.
- Personal AGI Definition: Defined strictly as an agent running on user-owned infrastructure, reading from a user-owned memory bank, executing user-written procedures, and compounding value daily through usage, unlike corporate AGI which resets or pivots.
- Working Memory Constraints: Humans have a working memory limit of roughly seven items, whereas AI agents can hold a million tokens (approx. three Harry Potter books) simultaneously, allowing for synthesis across vast amounts of information that a human cannot hold at once.
- Gbrain Architecture: The speaker's personal system includes a 220,000-page "library" of 25 years of diaries, emails, and notes, processed by agents to provide context-aware briefings, pre-meeting prep, and research before the user wakes up.
- Skill Files: The fundamental building block of this system is the "skill file"—a markdown document written in plain English that defines a specific task (e.g., transcribe meeting, extract commitments), which the speaker argues makes anyone a programmer if they can write clear instructions.
- Latent vs. Deterministic Computation: A key architectural principle is that judgment and vague requests are handled in "latent space" (the AI model), while arithmetic, data queries, and logic must be handled in "deterministic space" (SQL, scripts) to avoid model hallucination errors.
- Compounding via "Skillifying": A critical discipline for success is "skillifying" tasks: converting one-off completed work into reusable skill files so that no question is ever asked twice, ensuring the system grows smarter every day rather than suffering from "amnesia."
- Economic Shift: New YC companies like Emergent ($15M revenue with 15 people) and Retail ($60M revenue with ~40 people) are leveraging this "one founder + agents" model to achieve revenue-per-person figures previously impossible in software or heavy industry.
- Ownership of Cognition: The speaker distinguishes between two futures: workers who own their "skill files" (externalized judgment) and retain them when changing jobs, versus workers whose judgment is captured in company repos, effectively resulting in the extraction of their cognitive labor.
- Spinoza's 1673 Rejection: Spinoza declined a full professorship in Heidelberg with a salary and freedom of speech because the contract required him not to disturb established religion, rejecting the offer to maintain "under your own power."
- Security Argument: Consolidating personal data into a private, user-controlled repo is argued to be more secure than the current default of scattering life data across ten cloud companies with misaligned incentives, provided the user holds the encryption keys.
- Open Source Strategy: The speaker open-sources the entire stack (Gbrain, OpenClaw, Hermes) to democratize the "technology of leverage," arguing that giving away powerful tools prevents the creation of a "priesthood" and fosters a renaissance rather than restricting power.
- Real-World Impact Example: A father built a Personal AGI for his son with rare epilepsy by creating an 80,000-file markdown library of every medical paper and interaction, enabling immediate access to a curated knowledge base that no lab or grant would provide.
- Future Outlook: The speaker predicts the collapse of the need for intermediaries (funding, headcount, permissions) for 7,000 "strivings," asserting that the technology now allows individuals to execute the "most excellent" work without a team.
- Final Directive: The talk concludes with the assertion that "All things excellent are as difficult as they are rare," but argues that the difficulty has collapsed for founders who own their context and tools, leaving only the rarity of execution.