Paul Graham On Startups, Ambition, and Great Founders
The "Jumped the Shark" Critique Is a Permanent, Self-Refuting Feature of YC's History
Paul Graham points out that claims of YC's decline have been constant since almost the beginning, and structurally, critics are forced to admit YC was once great in order to say it's declined. This is a durable pattern rather than a real signal, and he notes it happened even when the batches were far smaller than today.
"People are always saying that YC has jumped the shark, you know? ... They can't say it was always terrible. They have to admit it was once great. So if they want to attack it, they have to say it's fallen" 00:01:15
He extends this: "People used to say that back when we had 40 startups in the batch... The old days are gone. There's 40 startups in the batch, right? People are always saying that." 00:19:29
Ambition Is Innate, Not Trained — and Rarely Develops Mid-Batch
Graham states that the quality of ambition in founders is almost always present before YC, not cultivated by the program. Cases that look like a founder "becoming" ambitious during the batch are actually founders who were ambitious all along but suppressed it due to upbringing or environment.
"Those stories are rare because most of this quality is inborn... Sam Altman famously was extremely formidable the first time I met him before we even accepted him. And that's more the rule than the exception." 00:07:12
"The one case where you see founders, it's not that they lack ambition. It's more like they've been trained not to show it... Maybe you have pushy parents who tell you to do what they want... But even then, I would argue that the person was ambitious. They'd just been trained not to show it." 00:07:12
Wealth Isn't the Day-to-Day Driver — Fear of Disaster Is
A genuinely non-obvious insight: founders aren't motivated moment-to-moment by the prospect of becoming rich; they're driven by fear of immediate failure, and often are blindsided by their own financial outcomes.
"You know what actually motivates founders day to day? The fear of failure... You're thinking, oh no, the server's crashing. It will be a disaster... You just don't want your model train set to break." 00:05:54
"Sometimes I'm the first to tell them. Like I do the math. I'm like, wait, according to your last round, you're a billionaire. And like, oh yeah, I am." 00:06:55
AI Development Inverted Everyone's Expectations of How Intelligence Would Be Built
Graham describes a widely-held mental model in AI research (start with simple/perfect competence like a fly, work up to human) that turned out to be exactly backwards — LLMs arrived at human-like fluency first, with reliability/perfection still being chased.
"The way we thought it was all going to play out was you start out with like a perfect fly... And instead, what we got was basically a full-on human but full of shit, right? Like the first AIs... were like an undergrad trying to bullshit his way through a paper." 00:13:01
"So instead of starting with perfect and then working your way up to human, you start with human and then work your way towards perfect." 00:13:58
AGI Isn't a Line, It's a "Smear" — a Reframed Definition Worth Adopting
Rather than a discrete threshold, Graham argues AGI is better understood as an uneven frontier where some capabilities are far advanced and others lag badly — which explains the "jagged frontier" phenomenon directly.
"We thought back in the day that AI would be sort of like this finish line and AGI would be this finish line. We'd cross it... But here you realize, OK, some bits of AI are way across the finish line. Some bits are like maybe in the middle of it... It turns out to be a smear and we're on it." 00:15:02
Shipping Speed Remains the Best Predictor of Startup Success, Even in the AI Era
Despite dramatically more powerful tools, Graham says the fundamental predictive variable for startup success hasn't changed — and many YC startups still under-ship despite access to AI.
"We have all these powerful AI tools and there's still a lot of startups in this batch that are not shipping fast enough... So far, almost everything is exactly the same. The only weird new thing is that companies have these giant AI bills." 00:16:18
Startups-as-Credential Is a Category Error With No Historical Precedent for Success
Graham directly attacks the growing trend of doing a startup for resume value, arguing it's structurally different from other prestige credentials because failure (the likely outcome) carries no value.
"Only if it fails. If it works, it's like your life's work... These people trying to get YC, trying to do YC as a credential, they had no idea what they were asking for... This is a really, really unsuitable credential." 00:08:10
"If you want to seem cool, like starting a startup is just about the least efficient way to do it... It's brutally hard. And you won't seem cool for years, years of brutal hardness before anyone will think you're cool." 00:09:17
Formidability as the Core Investment Thesis
Graham explains the origin and precise definition of "formidable" — someone who reliably gets what they want — and ties it directly to why this is the trait investors should optimize for, since aligned incentives mean investor upside rides on founder formidability.
"I think that it's someone who gets what they want. So that's the test, right?... If you invest in them and you have stock in their company... If they get what they want, you get what you want. So that's why investors want people who are formidable." 00:09:38
Contrarian Perspectives
Lean Startup Methodology Isn't Dead — and PG Never Engaged With the Underlying Book
Pushing back on Patrick Collison's Startup School comment that "lean startups may be dead" due to AI and easy capital, Graham argues capital efficiency remains achievable and that inference costs are structurally falling regardless of current GPU shortages.
"I haven't actually read it. I mean, why would I? Tolkien didn't read all those other fantasy books people wrote... Token prices will certainly... go down dramatically, like 30x a year or something like that... I think you can still start a startup on not much money." 00:10:20
Even Hard-Tech/Capital-Intensive Startups (Rockets) Can Start Nearly Broke
Graham argues that capital intensity doesn't require capital upfront — you sequence your ambition to your resources, using simulations, designs, and expert credibility to unlock funding stage by stage.
"It's harder to start a rocket startup on no money, but even a rocket startup you can start on not much money because what you do is you adjust for how much money you've got. You can't build an actual rocket. What you can build is a design for a rocket..." 00:11:56
Startup Ideas Are Nearly Irrelevant Compared to Founder Quality
Graham makes a strong claim that the specific idea behind the next trillion-dollar company is almost secondary — it is a function of who is building it, not what's being built, effectively downweighting idea-selection as a variable in venture outcomes.
"It's not some particular idea... It comes from the next trillion dollar founders is where it comes from. And they probably have good ideas, right? Whatever they're working on is probably promising." 00:20:20
Founders Won't Look Different in the Future — Twenty Years of Data Say So
Against the assumption that AI-native or younger generations of founders will look meaningfully different, Graham states flatly that the founder archetype has been constant for two decades and gives no reason to expect change.
"They still look exactly like they used to 20 years ago. Why should they change 20 years from now?" 00:21:00
Companies Identified
OpenAI — AI research company, incubated as a side project connected to YC's ecosystem. Cited as the realization of Graham's 2012 "frighteningly ambitious" prediction of "a new Google" — succeeding not by attacking Google directly but by making its underlying model obsolete. "That's what OpenAI is, right? Like I realized very early on, like after using OpenAI, like, whoa, I don't use Google anymore." 00:04:15 / "Yeah, yeah, yeah, yeah. One of Sam's many side projects." 00:04:54
StarCloud — Startup founded by Philip, in space/satellite computing. Mentioned as an example of raising capital on almost nothing but a white paper and a booked launch, aided by the founder's pre-existing domain credibility. "They basically just wrote a white paper and booked a launch and they raised the money for StarCloud." 00:12:31 / "Nothing as convincing as booking a launch, right? It's got to happen if you've booked a launch. But he's a huge famous expert in the stuff he works on... It's easier for him to do that than some kid who just graduated from college." 00:12:39
GitLab — Referenced through co-founder Sid, whose personal battle with cancer inspired a related startup. Mentioned as a case where "founder mode" thinking was applied to a health crisis, not a company. "He approached it like a startup, right? And these were essentially his co-founders in that de facto startup." 00:03:15
Unnamed on-demand cancer research startup (current YC batch) — Startup providing individualized, on-demand research for cancer patients, founded by the same people who helped GitLab's Sid fight his cancer. Highlighted as a "death of a thousand cuts" approach to a notoriously hard problem, and notable because it's literally the team Graham once wished existed. "There's a startup in this batch that is basically doing on-demand research for people with cancer... they're sort of inflicting on cancer the death of a thousand cuts, which may be the way to defeat it." 00:02:18 / "Not only is this startup happening, it's the people who actually did it for Sid. So it couldn't be better." 00:03:44
Unnamed "intercontinental ballistic cargo" startup — Described as a cargo-delivery rocket concept ("like an ICBM, except instead of exploding, it lands"), cited by Graham as a hallmark example of the kind of frighteningly ambitious idea current YC batches now attract. "It's like an ICBM, except instead of exploding, it lands. And drops stuff off. Yeah, that's serious, right?" 00:01:43
People Identified
Sam Altman — Co-founder of OpenAI, cited as the archetype of an inborn "formidable" founder who displayed extreme ambition even before being accepted into YC, and credited with incubating OpenAI as a side project. "Sam Altman famously was extremely formidable the first time I met him before we even accepted him." 00:07:12
Sid (GitLab co-founder) — Fought cancer using startup-style methodology ("founder mode"), inspiring both Graham's belief someone should build a company to replicate his approach for others, and directly inspiring the founders of the current YC cancer-research startup. "It's funny because I remember thinking after that... somebody should start a startup to do for everyone what Sid did for himself. And now, not only is this startup happening, it's the people who actually did it for Sid." 00:03:15
Philip (StarCloud founder) — Domain expert in space/satellite technology who raised funding based on a white paper and booked launch, leveraging pre-existing credibility in his field. "He's a huge famous expert in the stuff he works on, right? So he comes with automatic cred." 00:12:39
Patrick Collison — Referenced for a claim made at Startup School that "lean startups may be dead" due to AI enabling parallel work and easier fundraising; Graham uses this as a launching point to disagree. "Patrick Collison recently said that lean startups may be dead." 00:10:22 (Paul Graham, paraphrasing Collison)
Maya and Jessica — Co-founders alongside Graham of Y Combinator (Jessica Livingston referenced implicitly via "Maya and Jessica's private language"), credited as originators of YC's internal vocabulary, including the term "formidable." "A lot of the language of Y Combinator is Maya and Jessica's private language, right? And so we already had this word before we started YC, like for a certain kind of person." 00:09:38
Alan Turing — Referenced for the Turing Test, which Graham uses as still the best working definition/benchmark for AGI discussions. "Could you do better than the Turing test? The Turing, I think Turing was pretty good, right?" 00:15:02
Operating Insights
Sequence Fundraising to Milestones, Not Ambition Level
Graham's tactical framework for any capital-intensive startup: don't wait until you have "enough" money to build the real thing — build the smallest convincing proof (a design, a simulation, a booked commitment) that unlocks the next funding round, then repeat.
"You just do what you can on the money you got. As long as you can get to some kind of milestone, then you can convince investors to give you more." 00:12:24
Use Batchmates as a Captive, High-Quality Early Customer Base ("YC GDP")
A specific, repeatable tactic for early-stage founders inside any accelerator-like structure: your cohort itself is a built-in customer base of early adopters who decide fast and are somewhat obligated to hear you out.
"You have some sort of product you can sell it to the startups in your batch. Almost no matter what it is, you can find some. And they're the exact kind of users you want. Early adopters who decide quickly." 00:17:39
Use Peer Cohorts to Solve Technical Problems Faster
Beyond selling to peers, Graham highlights that batchmates function as a distributed technical support network — if you hit a technical wall, someone else in the batch has likely already solved it.
"You have some technical problem. Probably somebody in the batch had that problem. You can just ask them how they solved it." 00:17:11
Overlooked Insights
The Standardization Insight Behind YC's Founding Was About Paperwork, Not Just Capital
Buried in the origin story is a detail that's easy to skip past: YC's founding innovation wasn't merely "fund startups early," it was the creation of standardized investment paperwork at a moment when no "angel firm" category existed — individual angels and late-stage VCs existed, but no institutional, repeatable, standardized small-check vehicle did. This is arguably the actual structural innovation that made YC scalable (independent of the batch model, which came later "by accident"), and it's a governance/process innovation, not a sourcing or brand innovation.
"There were VC firms who would do these giant late stage round. And there were angel investors who were individual people investing their own money. But there were no angel firms. That was the original idea... invest small amounts of money earlier and use standardized paperwork." 00:18:02
Token Deflation Is an Underappreciated Structural Tailwind, Not Just a Cost Line-Item
Graham tosses off a specific, numerical prediction — 30x annual price decline in inference at a given quality level — that is stated almost as an aside but has major implications for capital efficiency calculus in every AI startup pitch deck, arguing against the current narrative that "AI is expensive" is a durable moat or barrier.
"Tokens are very expensive at the moment, but that's just because there's a shortage of GPUs... inference prices at any given level of inference go down dramatically, like 30x a year or something like that. So even token prices are somewhat misleading because you're getting higher quality tokens over time." 00:11:21