Key points
- Garry Tan says personal AGI runs on owned context, not a $20-a-month rented chatbot.
- In YC's Winter 2025 batch, a quarter of companies had codebases 95% AI-generated, and the batch is on track to be among YC's fastest-growing and most profitable.
- Tan says rented frontier models plus owned context and a harness can give one person 400x output.
Garry Tan's case for owning your own AGI
Garry Tan argues that AGI will not arrive as a single event in a data center. It spreads into individuals, each running a rented frontier model over their own context with a harness wired between them. He frames this through Spinoza, excommunicated in 1656 for teaching that God is diffused through everything rather than seated on a throne. The evidence is his own output, 400 times what he shipped at YC in 2013, and portfolio companies reaching nine figures with a few dozen people.
Takeaways
- Rented models plus owned context make personal AGI. Tan's equation for the next decade is a commodity frontier model getting cheaper each quarter, plus your own context, plus a harness such as Claude Code or Codex. Corporate AGI is the 20-dollar-a-month chatbot that resets when you close the tab.
- Tan's output rose 400 times in twelve years. In 2013 he was a Y Combinator partner building Bookface at night and shipped about 14 useful lines of code a day. Running YC full time now, he puts his output at 400 times that, with eight times as a pessimistic floor.
- Small teams now reach nine-figure revenue. Emergent, from the summer 2024 batch, reached nine figures of revenue in eight months and had 15 people at 15m dollars annualized. Retail, winter 2024, hit 60m dollars annualized with about 40 people.
- Quarter of YC's W25 batch had AI-written code. About 25 percent of companies in the winter 2025 batch had code bases that were 95 percent AI generated. That batch is tracking as one of the fastest growing and most profitable in YC's history, though he concedes he cannot prove causation.
- An agent holds a thousand pages at once. A person holds seven plus or minus two items, which is why teams, funding, permission and credentials existed as workarounds. An agent holds a million tokens, roughly a thousand pages. Tan's G Brain holds 220,000 markdown pages and G Stack has 123,000 GitHub stars.
- Markdown is code and English is programming. Anyone who can write clear instructions in English is a programmer, and the compiler is a language model. YC's finance staff compiled about 100 Excel workbooks into one app she built with an internal agent, and media and events staff build skill files.
- Skill files must live in repos you control. A skill file is a piece of your cognition written down and made executable. Tan's hypothetical support engineer Maya teaches her agents 40 skills over two years. If she controls the repo she keeps them. If not, her job becomes a skill file.
Tan's five steps to personal AGI
- Run an agent on your own machine rather than through a vendor's interface.