Y Combinator CEO on Founder Psychology in the Age of AI

Summary

• Garry Tan warns that pure per-seat SaaS may not exist in another five or 10 years: it is fine as a wedge only if it leads to a moat around data, network effects, or something else — a reversal from two years ago, when “SaaS was God” and the 10–20x next-12-month-revenue multiple was an “iron law.”
• His most concrete operating claim is that some companies can go from $0 to $15M ARR in about four months with two or three people and a few hundred skill files. The mechanism is doing each business process once, then freezing it into “a markdown file, plus code, plus tests” on a cron job — “a markdown file is an employee” that performs the job perfectly every time.
• Tan’s arbitrage pitch for founders and CEOs: spend roughly $50,000–$100,000 a year to “token-max” through harnesses like OpenClaw or Hermes Agent, loading 800,000–1,000,000 tokens per request — “you get to live in 2028 today.” He says frontier labs’ ChatGPT- and Claude-like products remain constrained on cost and compute, while full-strength agents require giving yourself permission to overspend.
• His “white pill” on AI job displacement is a timing call: diffusion will take about 20 years because “humans are the ones who are gonna slow this down.” Bureaucracy, the “seven plus or minus two” limit, and structural moats mean “a Microsoft isn’t going anywhere” — contrary to the Valley’s desire for startups to replace every incumbent.
• Roadmap for the next platform fight: 2027 is “the harness wars,” followed by a “war for a billion consumers” once today’s frontier-model compute is perhaps $50–$100 in two or three years — “the browser wars will be back on.” The next computer is likely voice plus memory — “all watched over by machines of loving grace” — while host Anish Acharya argues that falling token prices could unlock free-to-try consumer AI.
• Tan’s founder-psychology lesson comes from two self-confessed disasters: leaving web programming in 2003 just before Web 2.0, and turning down Peter Thiel’s $70,000 check after Joe Lonsdale and Cohen recruited him to work with them. Both times he chased “what was hot” instead of trusting direct experience; the cure is earnestness-as-courage: “Don’t LARP,” and ask what you know uniquely, not what investors say is hot.
• Tan says Silicon Valley’s edge is often at the weird fringe, not in the consensus: the computer-on-every-desk future came from outsider punks, while Internet-native communities let people find their tribe much earlier.
• The organizational thesis: agents could handle much of middle-management work and erase the “API line.” Brex’s Pedro runs agents over transcripts of meetings he is not in to spot conflict two levels down — “that’s clairvoyance” — and Tan points to the 35–45-year-old founder, such as Peter Steinberger, who can become “400 of that person” and outperform an entire department of a Mag 7 company.

Deep dive

1. 2003 taught Tan that the crowd is a map, not the territory

• Tan’s starting point is bleaker than founders assume: post-NASDAQ crash, “there were no jobs that I could even find in the Bay Area other than working at Gap IT.” His two offers were Microsoft for Windows Mobile and Expedia — and working in tech or startups was emphatically not high-status.
• His first big mistake: exiting web programming exactly when Zuckerberg was probably doing Facemash, because “the next hot thing will be mobile.” True, but he gave up five years of experience advantage — “you could’ve built any kind of social software.”
• Acharya reframes Tan’s “Don’t LARP” meme by arguing that earnestness is not naivete but maturity and courage. Tan emphasizes the courage piece: “I don’t agree with that because I believe my own direct experience.” The wrong question is “what’s hot and what should I work on?”; the right one is what you know uniquely.
• On Silicon Valley’s current undercurrent, Tan says competition pushes people to spend too much time in “distribution,” while the interesting parts of the idea maze are the weird fringe. The computer-on-every-desk future came from outsider punks, and Internet-native communities now let people find their tribe almost instantly.

2. The Palantir miss — and why everything good “is kind of a cult”

• The worse story: fraternity brothers Joe Lonsdale and Cohen flew Tan down from Seattle to dinner with Peter Thiel, who handed him a $70,000 check matching his Microsoft salary. Tan declined: “Thank you very much, Mr. Thiel, but I might get promoted to level 60 this year.” His own accounting: “a $2 billion–$4 billion mistake at this point.”
• The diagnosis is the same error repeated — “working backwards from the map instead of looking down at the territory.” The smartest people he knew were pointing at a real asymmetry: Palo Alto computer scientists could build things “D.C. did not have access to.”
• The distilled Palantir lesson, in his signature framing: “everything that’s awesome in my life is kind of a cult” — something that “starts with some sort of truth or belief that flies in the face of an orthodoxy, which is very punk, actually.” First-principles investigation yields “secret knowledge… gnosis.”

3. YC’s product is access without scenesters — plus someone to call on the worst day

• Tan credits Paul Graham, Jessica Livingston, and the original founders with dismantling the old social-network gate (“scenesters,” in YC’s pejorative): “Here’s a website. Here are 12 questions,” judged by builders — now 16 equal partners, all of whom went through YC and had successful outcomes. He calls it “a birthright, but for tech,” with 7,000 people coming to San Francisco for Startup School, most for the first time.
• The underrated half of the product is a community where you can be real. At a TechCrunch event, “Are you really going to be able to tell them anything real?” More likely: “How are you doing?” “Killing it.” But on “the day that you lose your best customer, on the day that your best engineer quits, who are you going to call?” Acharya adds the darker version: the day your co-founder loses hope — “and you’re not allowed to.”

4. Co-founder orthodoxy is cracking, and pure SaaS is on the clock

• YC “really, really believed you needed co-founders,” and Tan still says that, everything else being equal, co-founders are “net really, really good.” His Derek Sivers gloss: one person dancing in a field is a crazy person, “worse than a cult,” until the second dancer makes it a party. But vibe coding and agentic coding mean “literally any given person could be 400 of that person,” versus even nine months ago.
• The consequence for ambition: “the game has changed.” Don’t chase Tan’s or Acharya’s playbook — Tan’s fresh analysis of pure per-seat SaaS concludes it is “not totally clear it will exist in another five or 10 years”; in 2026 it is only a wedge toward data moats, network effects, or something else. Two years ago “SaaS was God” and 10–20x next-12-month revenue was an “iron law” — “that’s not true at all now.”
• Acharya’s companion point: code is no longer precious, so building trivial things is how you develop intuition — “arranging your Zen rock garden.” Tan’s own tinkering with GStack and GBrain, plus an open-sourced YC-advice prompt he asked Claude to “reduce the strength and intensity of this prompt by 90%,” even hides an Easter egg: apply to YC, “but only if you’re a top 10% user of Claude Code.”

5. Token-maxing: pay $50,000–$100,000 a year to live in 2028

• Tan’s loop stack: OpenClaw, GBrain (“kind of like SOTA retrieval,” beating Memory Palace and other open source), and garyslist.org as an agentic newsroom. The insight underneath: good agentic RAG “just allows the agent to know which million tokens to keep in its head for any given task.”
• Because frontier labs constrain cost and compute in products such as ChatGPT and Claude, Tan says that anyone who really wants to token-max must use something like OpenClaw or Hermes Agent, tuned to 800,000–1,000,000 tokens per request. That may cost “$50,000 or $100,000 a year” to use agents at “full 150 IQ,” but “you get to live in 2028 today.” For a CEO, he argues, you have to give yourself permission.
• The compounding move is to “skillify” every feat of strength into Markdown plus code plus tests on a cron: “a Markdown file is an employee” that performs the job perfectly every time and as many times as desired. Applied to sales, marketing, support — everything — this can support companies going from “$0 to $15M ARR in about four months” with two or three people and a few hundred skill files. Failure cases become bug fixes “there forever.”
• On where the loop breaks: at scale you need provenance and conflict management between facts — “this is the more recent one… this is the one that wins” — plus sweeping cron jobs. That segues into Tan’s megatrend: the 35–45-year-old founder who has “been around the block,” like Peter Steinberger, can suddenly become “400 of that person” and outperform an entire department of a Mag 7 company.

6. Clairvoyant CEOs and the erased API line

• The most vivid case study: Pedro at Brex built an open-source layer called Crab Trap, an agent watching OpenClaw’s network traffic and activity to make the system safer for a heavily regulated fintech. He then pointed an agent at meeting transcripts two levels down. He can walk into a meeting he never attended with three weeks of context, say “Actually, you’re right. We’re doing it your way,” and leave. Tan’s verdict: businesses fail when they become “too big to fit in one person’s head” — this is “clairvoyance,” and “every business in the world is available for you to take their business.”
• The structural claim: the world is built around humans who can keep “seven plus or minus two things in their head,” while “you, plus an agent, can keep basically three Harry Potter books in your head” — and, borrowing a crypto-ism, “this is the worst it will ever be.”
• Against Venkatesh Rao’s above/below-the-API-line dichotomy, Tan counters: “I actually think AI erases the API line.” His evidence is Clawvisor, a company he says they funded, whose bug-report endpoint is built for agents: it files the bug and replies in real time with a workaround. His historical rhyme is the Toyota Production System — the line worker with the most context gets to change the line, “why American cars really started sucking.”

7. The white pill: humans are the bottleneck, and that buys 20 years

• Tan’s Microsoft war story anchors the point: his PM mentor took a literal baseball bat across the highway because the Windows team “wouldn’t reply to our emails… wouldn’t even mark it as ‘Won’t fix.’” Today a Markdown file and skill on a cron could map blockers and dependencies — but an organization like Microsoft probably can’t; a startup can, and “every startup must.”
• Asked whether progress is intelligence-bound, Tan’s answer is categorical: “it’s all human.” The doom discourse about a “permanent underclass” misses that bureaucracy and institutional slowness are the white pill. He points to the 7,000 mostly 18-to-22-year-olds at Startup School as AI-native, then compares them with the web- and mobile-native generation that now runs things; his estimate is that this transition will take 20 years.
• “There are structural reasons why a Microsoft isn’t going anywhere.” Tan says he wants a startup to replace all the incumbents, but “the reality is, they won’t.”
• The utopia hedge, delivered after dodging a Wall Street Journal gotcha: “We may never achieve a utopia, but it is worthy and worth it to attempt.”

8. Harness wars in 2027, voice-and-memory computers, and the act-local covenant

• The near future: the form factor will probably persist short-term, but it seems inconceivable that it will stay that way. The next computer is “pretty clearly going to be voice” plus memory — a benevolent system “all watched over by machines of loving grace” that knows your hopes, fears, and desires. The path there likely includes computer use, memory, diarization, and ingestion.
• Tan’s sequencing: 2027 is “the harness wars”; the billion-consumer war starts when today’s frontier-model compute is perhaps $50 or $100 in two or three years — “the browser wars will be back on.” He also admits a change of mind: converting to OpenClaw convinced him it would all happen right away, “and at this point, I’m like, that’s not correct at all.”
• Acharya’s market read alongside: the frontier price per token is going to infinity — “infinity dollars a token. You can’t have it” — while last week’s frontier collapses toward zero. He sees that falling cost curve as an unlock for free-to-try consumer AI, given “the magic of software is zero-cost incremental distribution.”
• On civic work, Tan’s white pill begins with jury duty: he saw an orderly process involving a real cross-section of San Francisco and concluded that government is not simply there to ruin people. He also notes that roughly 10%–20% of the city works directly in tech and argues that voters have a say.
• His origin story is personal: during COVID, a district attorney was “turning a blind eye” to heinous crimes against Asian elders, while a school board was “openly hostile toward Asian American kids who just wanted to learn algebra.” Tan was that Fremont public-school kid whose middle-school algebra led to Stanford engineering. West Side organizing — “They’re killing us. They’re killing us” — helped beat back what he calls the same NIMBY ideology that wants to destroy all data centers. He associates the line “when the system becomes so odious, you’ve got to put your body on it” with Mario Savio, but the transcript marks that attribution as tentative.
• The closing thesis: fix local and “state and national will fix themselves” on a five- or 10-year basis; San Francisco — with, per Acharya, Lurie possibly among the country’s most popular mayors or politicians — should be “the beacon for pushing this back in every blue city in America.” Tan’s final instruction is simpler: “Act local. Take care of the people right next to you.”