Topics Pioneers Insight Method Author
Back

Garry Tan Decides Who Gets to Build the Future

·The Builders

Rendering the long image…

Topics Early-Stage VC Investing Dynamics Battle for Frontier AI Moats Personal AI Assistants Adoption and Monetization AI Workforce Displacement and Labor Disruption
Pioneers Garry Tan

Summary

  • YC pairs a rapid selection process with $500,000 in funding intended to broaden founder access. Tan says 80,000–90,000 applications compete for about 800 spots, with a 10-minute final interview testing the user, problem and solution: “We just want to know what it is and why it’s awesome.” Isaacson argues that lacking a financial safety net can keep people out of startups; Tan calls meritocracy an aspiration YC will never fully achieve.
  • Tan sees antitrust and interoperability as conditions for startup growth. He credits the DOJ’s action on Microsoft with helping Google emerge, then argues Google now closes markets through self-preferencing. Isaacson points to successful new AI entrants and Google’s Gemini; Tan agrees and worries AI companies are becoming “the new big tech” and might gatekeep frontier models.
  • Tan estimates spending about $100,000–$200,000 annually on tokens for his personal AI experience. His agent prepared for Isaacson in roughly 15 minutes by downloading, transcribing and diarizing Isaacson’s top 10 podcasts. Tan says it “seems inevitable” that everyone uses such assistants in two to three years—a “Jiminy Cricket on your shoulder”—while acknowledging that current costs make his experience a privilege.
  • AI-driven job creation remains Tan’s hope, with room for near-term displacement. His open-source project, gstack, aims to teach founders the tools he uses for “400 times” the output; people whose roles become redundant could combine domain expertise with software skills to start companies. He warns that CEOs choosing only cost cuts could deepen wealth concentration.
  • Control of personal agents could determine who captures the next wave of commercial value. Tan expects another “browser war” next year and cites AI-lab insiders’ “five kings” scenario, in which potentially all commercial activity collapses into five kings. His preferred alternative is a Linux for personal AI, inexpensive and simple enough for everyone to choose.
  • Tan says investment returns must coexist with responsibility for the businesses YC helps scale. He recalls an applicant focused on ADHD medication for children whose commercial prospects did not establish a societal need. Isaacson’s closing reflection is that AI should address cancer and climate change, with YC accountable for the future its selections help produce.

Deep dive

1. YC tests for outliers while trying to widen access

  • Tan says YC receives 80,000–90,000 applications annually for about 800 spots; 8–10%, sometimes fewer, get interviews. The final interview lasts 10 minutes because Paul Graham thought five could be enough, but five minutes felt rude. The test is to explain “a real person who has a problem” and how the technology solves it. Tan says 10–15% of self-made billionaires age 50 and under are YC alumni.

  • Tan describes YC’s institutional test as whether it could say yes in 10 minutes to founders such as Musk, Gates, Jobs, Drew Houston, Brian Chesky, Brian Armstrong or Tony Hsu; if not, “I’d have to rewrite everything.” Walter’s Musk hypothetical specifies seeing him before his success was known.

  • Isaacson argues that people who need to pay for their next meal may not be able to gamble on startups. Tan answers with YC’s $500,000 funding, versus $18,000 when he did YC, intended to let founders participate without rich parents. He contrasts that with seed funding of $30 million at a $300 million post-money valuation for what may be just an idea and one person.

  • Tan calls meritocracy an aspiration: “it would be a mistake to say that we were actually a meritocracy,” but “it is a virtue to attempt to achieve it.” On immigration, he says about 50% of startups that “really make it” have an immigrant co-founder. His mother’s family came from Burma as refugees; his father came from Singapore.

  • Tan says his experience founding Posterous—“dead-simple blogs by email”—is important to his work as an investor. Its demo made immediate sense to YC, and the company got funded. At Palantir, Alex Karp taught him how buyers infer product quality. Young engineers wanted Motel 6 rather than Washington, D.C.’s Mandarin Oriental; Karp explained that a general or Fortune 500 executive might ask where they were staying. “If you say Motel 6, they’re going to think that they’re buying Motel 6 software.” Palantir, Karp insisted, was “Mandarin Oriental software.”

2. Startup opportunity depends on dominant platforms opening markets

  • Tan argues that the DOJ’s consent decree on Microsoft helped Google emerge by checking Microsoft’s use of Windows to dominate browsers. He now sees Google self-preferencing, pointing to the shrinking share of Google search results occupied by non-Google properties. Europe’s DMA is a model for opening platforms through data transparency and privacy, portability and reverse engineering. His railroad analogy is that a monopoly over transportation can let its owner set prices across the commodities carried on it.

  • Isaacson points to OpenAI and Anthropic’s success, Apple’s inability to embed its own system in its phones, and Google’s effort to push Gemini in Search. Tan agrees, then calls AI companies “the new big tech.” He shares critics’ concern that, as the AI competition progresses, frontier companies might withhold access to their best models. Openness, he argues, would support startups and consumer choice.

  • Tan says YC’s policy work responds to groups funded by astroturf money and hired actors who claim to represent small software companies while advocating positions that mainly help big tech. He says YC spends about one-hundredth of what even a single big-tech firm spends on policy, with one full-time hire in D.C., Luther Lowe. Its leverage includes more than 2 million YouTube subscribers: finding founders comes first, and ensuring they can access markets comes right under that.

3. Personal agents offer broader agency at an exceptional price

  • Tan rejects treating AI as merely chat. YC’s internal agent, Straylight, has a “company brain”: context on every founder YC has backed, people who have emailed YC and all its applications. Tan contrasts a human’s ability to track “7, plus or minus 3” things with an AI agent’s capacity to work across a stack of 1,000 pages—roughly three Harry Potter books—and says that capacity could grow to five, 10 or 20 books in another year. Isaacson challenges the example: retrieving facts is useful, but “that’s just data retrieval.”

  • For a more concrete example, Tan describes an agent downloading and transcribing Isaacson’s top 10 podcasts, diarizing them, then using an LLM to extract themes line by line and timestamp by timestamp, about 15 minutes before their interview. Citing Peter Steinberger of OpenClaw, he says Claude Code and Codex currently trade off some of the appealing, humanlike parts of agents in favor of repeatability and safety.

  • Tan says his agent has read his email, Dropbox, Apple Notes and therapy notes going back 20 years. The ambition is anticipatory help: a “Jiminy Cricket on your shoulder” that knows its user and can help them see around corners. Tan says widespread use in two to three years “seems inevitable,” but estimates spending $100,000–$200,000 annually on tokens himself: “that’s a crazy privilege.”

4. Employment gains require retraining and new businesses

  • Tan says he does not know for certain whether AI will create more jobs, though that is his hope; in the short term, it may not, and some positions will become redundant. He argues technology is over-vilified and that the response should target abuses of technology, not stop it. He also warns CEOs against treating cost cuts as the endpoint, which could produce a “dystopian moment” of wealth accumulating without being shared.

  • Tan’s open-source project, gstack, works backward from teaching YC founders to use Claude Code for “400 times” their previous output. His retraining example pairs someone with experience in a field such as medical billing with a 22-year-old who can build software—or has the experienced person learn to do so. The result might reproduce their former work or be “two clicks away”; either way, Tan says, it requires more agency and applying knowledge from lived experience.

  • Tan recalls an applicant focused on ADHD medication for children as an example of an ethical concern: the founders were impressive and the business was likely to make money, but the need for greater distribution of the medication was unclear. He says the company was later sanctioned by the FDA and shut down, and that there may have been criminal charges. Walter says he has not checked recently. Tan says YC must not only make money, but fund “the things and the futures that we want to live in.”

5. Personal AI needs an open alternative to concentrated control

  • Isaacson proposes social media that strengthens society. Tan imagines Facebook and Instagram accessed as APIs through a user-controlled agent, with a personal “cognitive filter.” Isaacson objects that personalization already exists while business models reward outrage. Tan’s answer is that the algorithm’s owner has the control today; personal AI could put control over what enters someone’s mind back in the user’s hands. He gives his own example: wanting to focus on himself and see local, but not national, politics.

  • Tan predicts next year will bring another “browser war,” centered on whose agent people use, with probably a Mac-like and a PC-like option. Isaacson asks whether consolidation to two or three providers would be troubling. Tan says the market is already concentrated, citing AI-lab insiders’ “five kings” scenario, in which potentially all commercial activity collapses into five kings. His counterweight is “Linux for personal AI,” easy and cheap enough for people to choose.

  • Tan argues regulation should open access, citing the Open Markets Act and the possibility of apps operating on iMessages. He rejects Apple’s claim that only Apple can be safe enough, calling it a smokescreen for market power that can prevent the next Apple from emerging. He also says there are “zero open-source AI companies in the United States” and argues regulation could allow distillation, an advantage he attributes to Chinese companies over American and European companies.

  • Isaacson asks whether left-wing populism or concentrated technology worries Tan more; Tan says both outcomes are terrible and argues for a middle way. He rejects Bernie Sanders’s proposal to seize assets, and argues for more Googles, choice, prosperity and jobs rather than either big-tech oligarchy or mass asset seizure. Isaacson’s closing reflection is that AI should build “what humanity actually needs,” including curing cancer and fighting climate change. Choosing who builds the future carries responsibility for the future they build.