Apr 29, 2025 · 2h 20m · knowledge-project
Y Combinator CEO Garry Tan: Turning Ambitious Misfits into Founders
0:00 / 0:00
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this in-depth interview on The Knowledge Project, Y Combinator CEO Garry Tan sits down with Shane Parrish to discuss how YC discovers and trains transformative founders, the structural collapse of traditional blitzscaling in the AI era, and why authentic earnestness and open technology will fuel the next generation of builders.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Shane holds 16.6% of the talking time here. How this is scored →
Shane as informed peer 4.4 Guest teaching 5.6 Guest disagreement 1.5 Shane pushing back 1.4
speaking balance: gold is Shane, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 1:20:20 Garry slams AI safety think tanks as opportunistic rent-seekers
Garry forcefully dismisses NGO think tanks, criticizing employees who overnight rebranded themselves from bioterror experts to AI safety experts to capture billions in funding.
Hardest push from Shane ▶ 1:31:20 Shane challenges human superiority in sports play-callingShane pushes back against Garry's hesitation on AI decision-making by arguing sports coaches should already be replaced by AI models with massive correlation context windows.
Biggest teaching moment ▶ 1:45:40 Garry explains the Casetext stepwise prompt architectureGarry provides a detailed technical and historical explanation of how Jake Heller solved GPT-4 hallucinations by breaking legal research into microscopic human-like workflow steps and evals.
Shane holds their own ▶ 35:35 Shane synthesizes Berkshire Hathaway's structural simplicity with YCShane articulates how YC's enduring moat mirrors Berkshire Hathaway's disciplined simplicity, which Garry enthusiastically validates using Charlie Munger's weighing machine concept.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Shane as informed peer | Guest teaching | Guest disagreement | Shane pushing back | Why |
|---|---|---|---|---|---|---|
| The Origins and Core Philosophy of Y Combinator | 3 | 6 | 1 | 1 | Shane asks Garry to explain what makes YC uniquely successful and invites him to go deeper on YC's free content and demo day model. Garry gives an extensive historical breakdown of Paul Graham, ViaWeb, and the leverage of media and software. | |
| Filtering Tens of Thousands of Startup Applications | 3 | 7 | 1 | 2 | Shane probes into YC's filtering mechanics for 80k applications and what happens in a 10-minute interview. Garry educates Shane on why concise communication in 10 minutes reveals whether a founder actually grasps their market. | |
| YC's Batting Average and Founder Maturity Timelines | 5 | 6 | 1 | 2 | Shane highlights YC's unicorn batting average relative to other VC firms doing hundreds of hours of diligence. Garry details how YC focuses on early, low-status work and debunks venture myths about 9-month success horizons. | |
| Earnest Ambition Versus the Trappings of Startup Vanity | 4 | 5 | 1 | 1 | Shane explores why founders lose focus and asks how geographical concentration drives innovation. Garry explains how vanity metrics destroy startups and why moving to SF doubles unicorn odds. | |
| Ethical Boundaries and Mission-Driven Startup Selection | 4 | 5 | 1 | 1 | Shane asks about startup ideas YC rejects out of ethical concern, suggesting hypothetical cyber weapons. Garry explains partner autonomy and compares YC's stance to MIT's American institutional mission. | |
| The Weighing Machine and the Power of Serial Founders | 6 | 6 | 1 | 2 | Shane draws a parallel between YC and Berkshire Hathaway's simplicity. Garry builds on this with Buffett's 'weighing machine' analogy and reveals surprising data about serial founders and YC alumni. | |
| Earnestness in Founders: Brian Armstrong Versus SBF | 5 | 6 | 2 | 1 | Shane questions what personal traits correlate with long-term success. Garry introduces 'earnestness' as the key differentiator, contrasting Brian Armstrong's mission-driven persistence with Sam Bankman-Fried's fraudulent artifice. | |
| AI Disruption and the Demise of Traditional Blitzscaling | 4 | 6 | 2 | 2 | Garry argues traditional blitzscaling is obsolete in AI because lean 5-person teams reach tens of millions in ARR. Shane connects this shift to cloud computing removing capital infrastructure barriers. | |
| Founder Mode Versus Big Tech Organizational Bloat | 5 | 5 | 3 | 1 | Shane asks if big tech incumbents hold the advantage, offering an analogy about agile kids outrunning bloated companies. Garry passionately critiques Big Tech bloat, hoarded talent, and 'retired-in-place' engineers. | |
| AI Market Dynamics and Resisting Premature Regulation | 4 | 6 | 2 | 1 | Shane and Garry discuss AI regulation and educational applications like Synthesis Tutor. Garry dismisses early regulatory proposals as regulatory capture driven by Hollywood 'Terminator 2' fears. | |
| Sponsor Message: Accenture and Spotify | 3 | 6 | 1 | 1 | Following an ad break, Shane asks when models will begin self-improving. Garry explains synthetic data, test-time compute in o1, and domain-specific applications like Helix in biotech and Camfer in airfoil design. | |
| Intellectual Property, Paper Ingestion, and Legal Clarity | 6 | 4 | 1 | 2 | Shane pitches automated patent generation and research paper ingestion. Garry explores the legalities of copyright in generative AI and challenges redundant bioterror legislation. | |
| Open Source Ecosystems and Local Model Distillation | 4 | 6 | 2 | 2 | Shane asks about existential fears in AI and Meta's open-source strategy. Garry praises Zuckerberg's open-source push but critiques Meta's lack of basic user-facing RAG in apps. | |
| Pragmatic AI Governance and Regulatory Capture | 4 | 6 | 3 | 1 | Garry highlights DeepSeek's 1.5B parameter math model beating benchmarks on legacy chips, calling out rent-seeking think tanks pivoting from bioterror to AI safety. | |
| Reclaiming User Agency Through Open Algorithms | 5 | 5 | 2 | 1 | Shane asks if news articles will become hyper-personalized. Garry advocates for open algorithms where users control ranking parameters and calls out Apple's closed iMessage moat. | |
| AI in Decision Making and Institutional Politics | 5 | 4 | 2 | 2 | Garry expresses skepticism toward replacing human leaders with AI CEOs. Shane suggests sports play-calling in football would benefit immediately from AI analysis. | |
| Geopolitical Approaches to Superintelligence | 5 | 5 | 1 | 1 | Shane discusses human-machine hybrid intelligence in chess. Garry compares geopolitical ASI models, contrasting China's single centralized ASI with an American ideal where every citizen commands an ASI. | |
| Commercializing AI, Infrastructure Bottlenecks, and Labor | 4 | 6 | 2 | 2 | Shane asks what leading edge AI labs discuss. Garry highlights test-time compute, GPU shortages, and real-world labor displacement in overseas call centers. | |
| Prompt Architecture and the CaseText Blueprint | 4 | 7 | 1 | 1 | Shane asks about effective prompting techniques. Garry walks through Jake Heller's Casetext workflow, explaining how decomposing complex legal tasks into stepwise prompts and golden evals unlocked reliable outputs. | |
| Autonomous Reasoning Models and Enduring Moats | 4 | 6 | 1 | 2 | Shane asks how startups maintain moats when reasoning models automate prompt engineering. Garry explains that while chain-of-thought is absorbed into models, UX and customer workflows remain the true moat. | |
| Garry Tan's High-Conviction Tech Portfolio | 5 | 5 | 1 | 2 | Shane plays a hypothetical portfolio game forcing Garry to allocate 100% of his net worth across 3 stocks (Nvidia, Microsoft, Meta), followed by an exchange on YouTube algorithm tips from MrBeast. | |
| Paul Graham's Writing Philosophy and the Two-Sentence Pitch | 5 | 6 | 2 | 1 | Garry shares Paul Graham's philosophy on plain-spoken writing and formulating a two-sentence pitch. Shane links this clarity directly to sniffing out startup impostors like SBF. | |
| The Golden Age of Building and Accelerating History | 4 | 5 | 2 | 1 | Shane and Garry celebrate the current era as the golden age of building, acknowledging high startup failure rates while honoring Sam Altman and OpenAI for pulling forward AI history. | |
| Garry Tan on Success, Opportunity, and Returning to YC | 5 | 5 | 1 | 1 | Shane asks Garry for his definition of success. Garry reflects on returning to YC from his own fund to empower ambitious builders, culminating in a shared philosophy on equal opportunity versus equal outcome. |
Statements from this episode (73)
Garry Tan: Paul Graham created the first web browser-based application
“He actually basically created the first web browser based program. So he was one of the first people to hook up a web request to an actual program in Unix.”
Tan: Median YC startup raises $1M to $1.5M post-program
“At the end of it culminates in people raising, you know, sort of the median Raises about a million to a million and a half bucks for, you know, sometimes teams that are two or three people, just an idea starting at, ah, you know, at the beginning of that.”
Tan: YC companies raise ~$1B yearly with a 1% acceptance rate
“I think we have about a billion dollars a year in, ah, your funding that comes into YC companies, and that's because, ah, the acceptance rate to get into YC is only one percent.”
Tan: Y Combinator receives 70,000 to 80,000 applications per year
“Yeah, I think it's close to 70, 80,000 at this point.”
Tan: YC group partners read 1,000 to 1,500 applications per cycle
“Yeah, I mean, ultimately, the best thing that we can do is actually brute force read and on average, I think the group partner will read something like a thousand to 1500 applications for that cycle that they're working.”
Garry Tan: AI companies increasingly sell intelligence rather than traditional software
“Today, because people are not selling software, they're increasingly actually selling, ah, intelligence. They're like, you know, like it or not, like, these are things that you could not buy before.”
Garry Tan: Overseas call center tasks are the most vulnerable to AI
“Probably the most vulnerable things in the world today are things that you could, you know, farm out to an overseas call center. That's sort of like the low-hanging fruit today.”
Garry Tan: Founders unable to explain their startup in 10 minutes lack clarity
“One of the things I feel like I learned from Paul was that if in 10 minutes you cannot actually understand what's going on it means the person on the other end doesn't actually understand what's going on, and there isn't anything to understand.”
Garry Tan: About 2.5% of Y Combinator startups become $10B decacorns
“Yeah, about two and a half percent end up becoming Decacorns sooner or later, so.”
Garry Tan: Y Combinator's unicorn hit rate increased from 3.5% to 5.5%
“If anything, I think that the unicorn rate has gone up over time. You know 1015 years ago, I think it was closer to maybe three and a half to four percent, and now we're around five and a half percent. Some batches from You know, maybe 2017, 20 18, or, you kno…”
Tan: 50% of YC companies raise Series A, 25% after year 5
“About half of companies that go through YC will end up raising a series A. That's, you know, much higher than any other pre-seed or seed sort of situation that I know of. But about, A quarter of those who raise the Series A, they do it in year five or later.”
Tan: Startup founders often kill their companies by angel investing
“If you're a startup founder, and suddenly some, you know, people have heard of you, and people try to add you as a scout. Like, people kill their startups all the time by that, just by taking their eye off the ball.”
Garry Tan: YC teams staying in SF double their unicorn odds
“One thing we spotted is that the teams that come to San Francisco and then stay in San Francisco or the Bay Area, they actually double their chance of becoming a unicorn.”
Garry Tan: San Francisco had zero housing starts last year
“Literally, there were no new housing starts in all of, you know, San Francisco proper for the last year.”
Tan: YC general partners can pretty much fund whatever startups they want
“If you're a general partner at YC, you know, you pretty much can fund what you want”
Tan: YC passed on ADHD medication telehealth startup over ethical concerns
“There are lots of examples of, you know, maybe five or six years ago, there was a rash of telehealth companies that are focused on, for instance ADHD meds, and I distinctly remember one of our partners Gustav Alströmer, I, he met that team, and he said, you kn…”
Garry Tan: Y Combinator must be an explicitly American institution like MIT
“MIT is a an institution, and that institution is an American institution. And so being very clear about that, I thought was totally the right move for MIT. And you know, I think that YC needs to be a similar, you know, an institution of similar character.”
Garry Tan: Founders should aggressively seek early rejections in sales
“You should aggressively try to get to a no, and you know, you'd rather get a no so you can spend less time on that lead, and you can get on to the next one.”
Garry Tan: Competitors offering double cash cannot match Y Combinator's success
“There are clones of YC right now that give like twice as much money, for instance, but I don't think that they're going to see this level of success because they're not going to have As earnest people who become as formidable around you.”
Garry Tan: Serial founders built 40% of recent global unicorns
“About 40% of the unicorns from the last 10 years in the world were started by multi-time serial founders.”
Garry Tan: 60% of serial founders who built recent unicorns are YC alumni
“Did you know that 40, you know, of those 40%, 60% of those people, the people who created unicorns the last 10 years are YC alumni.”
Tan: Earnestness is the defining trait of durable, successful startup founders
“Like incredibly sincere, I think. Basically what you see is what you get. Like you're not trying to be something else. It's like authentic, but like, you know, even humble in that respect, right? Like I'm trying to do this thing. The opposite, I mean, and it's…”
Tan: Coinbase's Brian Armstrong is one of the best founders ever
“Brian Armstrong is, like, the best, one of the best founders I've ever met, and gotten, gotten the chance to work with and fund, and I think the world desperately needs more people like that, where, you know, in the background, just, like, consistent, doing th…”
Garry Tan: 5-person YC startups are hitting $12M revenue in 12 months
“We have companies basically
You know, going from zero to six million dollars in revenue in six months.
We have companies going from zero to twelve million dollars a year in revenue in 12 months, right?
And with under a dozen people.
Like, usually five or six p…”
Garry Tan: Sub-15-person startups reaching $100M revenue will be common by 2027
“I think we are seeing companies that in the next year or two will get to 50, a hundred million dollars a year in revenue.
really with under, you know, maybe 10 people, maybe 15 people tops.
and so that was relatively rare, and my prediction would be this bec…”
Tan: The majority of Y Combinator CEOs are now technical
“The CEO of the majority of YC companies, they are technical.”
Tan: Delegating without founder agency turns tech companies into political fiefdoms
“If
The founder and the CEO does not exercise agency, you know, then it's actually a political game, and then you have sort of fiefdoms that are fighting it out with one another, and the leader is not there, then you enter the situation where, ah, neither the l…”
Tan: Big tech talent hoarding decelerated overall technological progress
“I guess it felt like a little bit of a prisoner's dilemma, because I think the result is that you know, tech progress itself decelerated. You have, like, the smartest people of a generation basically retired in place, working at places that, you know, the worl…”
Garry Tan: Five or six competing AI labs prevent monopoly pricing power
“Sitting here a year after a lot of those attempts I feel pretty good, because it feels like there are five, maybe six labs, all of whom are competing in a fair market, trying to deliver models that, you know, honestly, any startup, anyone, you know, any of us …”
Tan: Current AI systems have zero agency and resemble smart toasters
“As of today, you know, these systems are, it's just matrix math, and there is no agency yet.
[3400] Garry Tan: Like, there's basically, they're equivalent to incredibly smart toasters”
Garry Tan: Modern school systems are designed to strip children of agency
“We send them to a school system that is not designed for agency. It is literally designed to take agency away from our children.”
Garry Tan: Screen time, Minecraft, and Roblox foster agency in children
“I'm actually personally pretty pro screens, and pro Minecraft and Roblox, and, you know, giving children, like, this sort of playground where they can exercise their own agency.”
Shane Parrish: El Salvador replaced its K-5 math curriculum with Synthesis Tutor
“And it's crazy, so it teaches the kids math, and my understanding just from reading it a little bit is El Salvador just replaced, like, the K through five math with synthesis tutor, and the results are, like, astounding.”
Tan: Social unrest will occur if AI benefits bypass majority of people
“In reaction to, you know, the nature of work changing that would be the outcome that, you know, the majority of people, if not all people, like, see the benefit in some sort of direct way, and if we don't do that, then there will be unrest.”
Garry Tan: AI reasoning models will drive biotech breakthroughs within a year
“Like, if only by applying, you know, dollars to energy that then goes into these models, will there be, ah, real breakthroughs in, you know, biological sciences, like being able to do new processes or come to a deeper understanding of you know, whether it's, a…”
Tan: YC startup Camfer used OpenAI o1 to improve airfoil design
“There's a YC company called Camfer that is trying to apply they actually were one of the winners of the recent YC O-one hackathon we hosted with OpenAI, and their winning entry was literally Hooking up a one to airfoil design. So being able to increase the sor…”
Tan: Humans will talk directly to superintelligent AI entities
“Like, we're going to have these, you know, basically super intelligent entities that we talk to.”
Garry Tan: Training AI on published data ultimately feels like fair use
“The funniest thing, my main response to all of that around, like, provenance of the data itself is, at some point, like, it feels like it actually is fair use, though.”
Garry Tan: My AI existential risk 'p(doom)' score is 1%
“My PDOOM score is one percent. Like, I'm not totally, you know, unworried, right?”
Garry Tan: Premature AI laws will stall clean energy and medicine
“It would also potentially be worse to prematurely optimize And basically make a bunch of worthless laws that slow down the rate of progress and prevent things like better cancer vaccines, or better airfoils, or, you know, frankly, like, you know, nuclear fusio…”
Garry Tan: Meta open-sourcing foundational AI models is 'God's work'
“What Zuck and Ahmad and the team over there are doing is frankly God's work. I think it's great that they're doing what they're doing and I hope they continue.”
Garry Tan: Every major AI lab will release computer-use agent capabilities
“What Anthropic is doing with computer use is, you know, I think that you know, what I've heard is basically every major lab is probably going to need to release something like that, whether it's an API the way Anthropic has, or literally built into this you kn…”
Garry Tan: $4 billion currently funds AI safety organizations and think tanks
“There's something like four billion dollars of money sloshing around think tanks and AI safety organizations”
Garry Tan: Sriram Krishnan as government AI czar is a positive development
“I think we're going into a very interesting moment right now with you know, the AI czar is Sri Ram Krishnan, who, you know, used to be a general partner at Andreessen Horowitz, and I think that that's a very, very good thing. Like, we want people who have the …”
Tan: Mainstream global users know ChatGPT but definitely not Anthropic or Claude
“If you go any, almost anywhere in the world I, you know, people maybe have heard of ChatGPT, they definitely haven't heard of Anthropic or Claude.”
Garry Tan: Elon Musk proved Twitter didn't need 80% of its staff
“Elon came in and quickly asked, like, 80 or 90% of the people, and it turns out you didn't need 80 or 90% of the people, so.”
Garry Tan: AI code generation makes engineers 5 to 10 times faster
“Today, with CodeGen, you know, today, engineers are basically, you know, writing code about five or 10 X faster than they would before.”
Garry Tan: Apple opened RCS messaging in reaction to DOJ pressure
“Apple even today still, you know, they're opening it up a little bit more with RCS, but, you know, it's those are actually in reaction to the work of Jonathan Cantor and the DOJ.”
Garry Tan: I didn't vote for Trump but I support the administration
“Even though I didn't vote for Trump, I am a 110%, you know, down for making America truly awesome.”
Tan: Automated systems will always require a human in the loop
“Like even if 90% of the time you're using the autopilot, like you always need a human in the loop.”
Tan: China will likely rely on a single, centrally controlled ASI
“That society probably, you know, unless there's other changes there, like that's, you can sort of count on a single, ah, artificial super intelligence, like sort of setting the, how everything works over there.”
Garry Tan: The US should issue every citizen an Artificial Superintelligence
“My argument would be, like, the most American version of it is that, like, you know, you and I have our own ASI. And, like, each, you know, each citizen should be, you know, issued an ASI and be taught how to get the most out of it.”
Garry Tan: Nvidia maintains a monopoly on the best price-performance AI compute
“I'm still pretty bull on NVIDIA, and that they more or less have the monopoly on, you know, sort of the best price performance.”
Garry Tan: Indistinguishable voice AI is currently actively replacing call center labor
“The speech to text and text to speech, those things are indistinguishable from human beings now, and you can train these things, the evals are good, the prompting is good you know, I, you know, going back to what we were saying earlier, like, what we're seeing…”
Tan: YC-backed Leaping AI automates 80% of German wine merchant orders
“I funded a company in this very current batch that you know it's called Leaping AI. They are they are working with Some of the biggest wine merchants in Germany, which is fascinating. So, I mean, that's another fascinating thing. Like these things speak all hu…”
Garry Tan: AI passing the Turing test will jump to 30% this year
“I think in a lot of domains, and this is sort of the year where, like, Maybe there's, like, five or 10% of things that, like, it's, you know, sort of hitting the Turing test and, you know, really satisfying that, but, you know, I think maybe this is a year whe…”
Tan: Casetext founder Jake Heller first commercialized GPT-4 in legal
“Garry Tan: He was one of the first people to get access to GPT-IV, and we think of him at YC as the first man on the moon, and that he was the first to successfully commercialize GPT-IV in the legal space.”
Garry Tan: Breaking knowledge work into micro-steps prevents LLM hallucination and failure
“Garry Tan: It's like, what is a specific thing that a human does? Break it down into the very specific steps that a real human would do, and then actually, basically, if it breaks, you're just asking in that step to do too many things, so like, break it down i…”
Garry Tan: Proprietary evaluations, not foundation models, are the true AI moat
“I mean, I think you know, ultimately the model itself is not the moat. Like I think that the evals themselves are the moat.”
Garry Tan: ChatGPT is the second-best software; custom UIs are first
“One of the funnier quips is that, ah, you know, the second best software in the world for everything is, ah, using ChatGPT, because you can basically copy and paste, you know, almost any workflow or any data, and it's like the general purpose thing that, you k…”
Garry Tan: Software moats remain unchanged, but startups only need six employees
“The moats are not different, actually, at the end of the day. It's still, you still have to build good software, you still have to be able to sell, you have to retain customers, you have to But you just don't need like a thousand people for it anymore. You mig…”
Garry Tan: AI infrastructure demand will scale to the Manhattan Project's level
“I mean, NVIDIA just, you know, has an out and out. Like, for now, they're just so far ahead of everyone else. I mean, it can't last forever, but I think that, you know, the demand for building the infrastructure for intelligence in society is going to be absol…”
Garry Tan: Energy availability is the ultimate bottleneck for advanced AI
“I think energy and access to energy is sort of the defining question at that point. Like everything else you could solve, like, and everything else you could sort of either, you know, if it's in the realm of science and engineering, like, you know, in theory, …”
Garry Tan: Microsoft's market cap heavily depends on OpenAI's success
“I mean, I think Microsoft has you know, just really, really deep access to OpenAI, and I think OpenAI is probably, you said public companies, right? ... So, you know, I think There's a non-zero, pretty large percentage of, like, the market cap of Microsoft tha…”
Tan: Meta Is an AI Dark Horse Accelerating Progress Cheaply vs AR
“I mean, I think Meta's sort of the dark horse, because, like, they are amassing talent, and then they have crazy distribution. And I think you know, I just would never count Zuck out. I think that he, you know, that it's so crazy that it's super smart that he …”
Garry Tan: MrBeast told me my YouTube titles and thumbnails sucked
“Look man, your titles suck, and your thumbnails are even worse.”
Tan: YouTube watch time matters most for feed distribution, subscriptions matter least
“Subs do very little, actually. There's no guarantee that you show up in People's feeds. If someone subs, it like helps a little bit. Liking helps more. Watch time helps the most.”
Tan: YouTube notification bell is almost as valuable as direct email
“Because if you hit the bell icon and they have notifications on, that's the only thing that is almost as good as having their email address and emailing them.”
Garry Tan: Sam Bankman-Fried's quirky public persona was a pure artificial affectation
“Like you might think of an SPF that's like, that's all artifice. Like, think about that guy. Like, that guy was, like, full of shakes and, like, the guy was, like, on meth, right? Like, the guy was, you know, everything about it was an affectation. Like, he wa…”
Garry Tan: In full transparency, the median Y Combinator startup still fails
“In full transparency, I mean, probably, you know, the median YC startup still fails, right?”
Garry Tan: Sam Altman pulled the future forward by at least a decade
“Sam Altman probably brought forward the future by, you know, five years, 10 years, at least 10 years.”