The Exchanges

Q What sort of policies would you think about, like, how would you think about setting that up to attract capital, to attract the right mindset of people, to attract and retain these people?

A I think what I want for San Francisco, for instance, is I think the rent should be lower. And so rather than subsidizing demand, we actually need to increase supply, like fairly radically, actually. And that just hasn't happened. I was, I think I was looking at it for the entire last calendar year. Uh, I think, you know, maybe Scott Wiener had just posted this on X that Literally, there were no new housing starts in all of, you know, San Francisco proper for the last year. So how are we supposed to actually bring down the rents and make this place, you know, actually livable? You know, if San Francisco is the microcosm where, you know, people build the future, and it is sort of the siren song for, you know, You know, techno-optimistic ideology, um, and it's also where they are most likely to succeed, uh, society, and certainly, you know, America is not serving society the right way if we're getting in the way of these smart people trying to solve these problems, trying to build the future.

AI assessment note: “rather than subsidizing demand, we actually need to increase supply, like fairly radically”

Q Can you give us an example of an eval? Like, make it tangible for people?

A Oh yeah, it's really straightforward. It's just a test case, right? So, given this prompt and this data, you know, evaluate the prompt to see if, you know, and it usually maps directly to, like, Something that is, you know, true, false, yes, no, like something that is pretty clear, like, you know, let's say there's a deposition and, you know, someone makes a certain statement, right? You might have a prompt that is like, you know, is this, uh, you know, is what this person said, um, in conflict with, uh, you know, any of the other witnesses, or I don't know, I'm, I'm totally making this example up, but like, this is the kind of thing that you can do, um, You know, at a very granular level, you might have thousands of these, and then that's how, uh, you know, Jake Heller figured out he could create something that would, you know, basically do the work of hundreds of, you know, lawyers and paralegals, and it would take, you know, a day or an afternoon instead of, you know, three months of discovery.

AI assessment note: “It's just a test case, right? So, given this prompt and this data”

Q For the first time in my life, and they're super ambitious. To what extent do you think that that environment just creates, um, better success or better outcomes?

A Oh, that was definitely true for me. I mean, um, without that, I feel like what my, I mean, I had a good, a really great community at the end of the day. Like it was, um, You know, my fellow Stanford grads, but I guess the weird thing to say is that, like, being around people who are really earnestly trying to build, ah, helps, ah, you know, 10 x more. Um, the, the default startup scenario out there is not about signal, it's about the noise. Like, you're playing for these other things, like, how much money can I raise, and what, you know, high status investor, like, You know, some people sort of float off and they become scenesters. They're like, oh, let me try to get a lot of followers on Twitter. That's the most important thing. And then what we really try to do at YC during the batch and then afterwards and, you know, in our office hours working with companies is, like, when we spot that kind of stuff, it's like, oh, no, no, like, maybe don't do that. Like, you know, let's go back to product, market, Uh, actually building, and then iterating on that, getting customers, uh, you know, long-term retention. All of those things are the fundamentals, and everything else is like the trappings of success, or And those will always feel, I, what's funny is, like, in other communities, uh, all of those things will always feel more present to hand, and they're easier. Like, you can just…

AI assessment note: “being around people who are really earnestly trying to build, ah, helps, ah, you know, 10 x more.”

Q I like that. What do you wish founders knew about sales coming in?

A Oh, how hard it is, and I mean, you know, like it or not, you know, the ideal founder is someone who has lived, like, 20 lifetimes and has the skills of 20 people. And, uh, the thing is, you know, you can't get that. And so, um, probably the first conference that we had, the first mini conference we have when we welcome the batch in is the sales mini conference. And, um, essentially it is don't run away from the know. Spencer Skates of Amplitude has this great analogy that he, uh, told, you know, uh, some companies when he came by to speak recently that I've been thinking a lot about, which is sales is about, um, you know, having a hundred boxes in front of you, and maybe five or six of those boxes has, uh, a gold nugget in them. And if you haven't done sales before, you think, I really, I'm gonna gingerly In a very gingerly way, open that first box and hope, hope, hope that, you know, I have a gold nugget, and then, you know, I don't, I almost don't want to know that there isn't a gold nugget in there. Like, I'm so afraid of rejection. It's sort of remarkable how often, um, high school and family and, uh, you know, the 10,000 hours of human training people get from their childhoods comes up in Paul Graham's essays. I always think about that because I think that Most people's backgrounds just don't prepare them for, ah, sale. It's a very unnatural thing to do sales, but then th…

AI assessment note: “Oh, how hard it is, and I mean, you know, like it or not”

Q And then at the end, you sort of like, I guess the last filter is like this 10 minute interview. Yeah. So what do you ask in 10 minutes to determine if somebody's going to be part of Black Combinator?

A I guess the, the surprising thing that, um, has worked over and over again, ultimately, is, um, in those 10 minutes, either you learn a lot about both the founders and the market, or you don't. So we're looking for incredibly crisp communication. So I want to know, um, You know, what is it? Um, and, you know, often the first thing I ask is not just what is it, but why are you working on it? Like, I want to sort of understand where did this come from? Did you just read about it on the internet? Or a much better answer is, you know, well, I, I spent a year working on this, and I got all the way to the edge of, you know, what people know about this thing. And, you know, what's cool about You know, the biographical is that then it invites more questions, right? It's, um, the best interviews in 10 minutes, like, you learn about an entire market, you learn about, um, a set of people that, you know, Normally, you might not ever hear of. Um, it's like you're traveling. It's like you're traveling the idea maze with, uh, the people you're talking to. This is all over Zoom, and, um, you know, at the end of those 10 minutes, like, sometimes the 10 minutes becomes 15, like, you want to talk to people longer, because that's what a great interview feels like to me. It feels like, uh, I'm a cat, and I see a little yarn, And I'm just pulling on the yarn. I'm just pulling on the thread because i…

AI assessment note: “often the first thing I ask is not just what is it, but why are you working on it?”

Q Oh, I love that. Thank you. Good advertising. Uh, I want to ask just a couple random questions before we wrap up here. What, what are some of the lessons that you learned from Paul Graham that you sort of apply or try to keep in mind all the time?

A I think the number one thing that is very hard, um, but is so, I mean, you can see it and read it in his essays. It's, um, to be plain spoken and to sort of, um, Be hyper aware of, uh, artifice of, um, kind of like bullshit, basically. Like, don't let bullshit, you know, I think, um, like, it creeps in here and there. I'm like, oh yeah, you know, I, um, you know, I sometimes am in danger of, like, caring too much about, like, the number of followers I have and things like that. You know, whereas, like, actually I shouldn't be worried about that. Like, what I should be worried about is, um, and, you know, I spend a lot of time with, uh, our YouTube team and our media team at YC talking about this. It's like, if we get too focused on just view count, um, we're liable to just, yeah, like, optimize for the wrong audience. Um, if we're not being authentic to ourselves, or, you know, if we're just trying to, like, follow trends, or, you know, do things that get clicks, it's like, that's not helpful to them either. Like, then we're just on this treadmill, right? Um, yeah, basically, like, trying to be very, very high signal-to-noise ratio, You know, the thing that I probably struggle with most, and, you know, I don't know, maybe some of the listeners here might feel this. It's like, sometimes I think out loud, and then, you know, really, really great ideas are not, like, thinking out …

AI assessment note: “to be plain spoken and to sort of, um, Be hyper aware of”

Q If I understand it correctly. Why Silicon Valley? Why San Francisco? And why haven't other countries been able to replicate that success inside?

A Well, um, at YC what we hope is that people actually come to San Francisco and, uh, you know, we do strongly advocate that they stay, but it's no requirement. Um, and then what we hope is that if they do leave, they, they end up bringing the networks And, uh, know-how, and culture, and, you know, frankly, vibes, and they bring it back to all the other, um, your startup hubs in the world, and I think that that's some, some of the stuff that has actually come about. I mean, um, Monzo was started by now my partner Tom Blomfield, uh, he's a partner at YC Now, but he started, you know, multiple startups, and a few of them, you know, multiple unicorns, actually, and both of them are Some of the biggest companies in London, for instance. So what we hope is that, uh, San Francisco becomes sort of really Athens or Rome in antiquity. You know, send us your best and the brightest. You know, ideally you stay here. One, one thing we spotted is that, uh, 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.

AI assessment note: “they end up bringing the networks And, uh, know-how, and culture, and, you know, frankly, vibes”

Q What would you guess is the strategy behind that?

A It's kind of funny because, uh, my critique on Meta would be You know, um, they very openly make everyone, they put in everyone's faces, right? Like you can't use Facebook or Instagram without, or even WhatsApp without seeing like, hey, Meta has AI now. But, um, the funniest thing is like, I'm very surprised that they don't think about sort of like the basic product part of it. Like I went to Facebook Blue app recently, and I was going to Vietnam, and I just wanted to say, okay, Meta AI, you're so smart. Tell me my friends in Vietnam. And it didn't know anything about me. I'm like, this is some basic rag stuff. Like, I get it. Like, you're already spending billions of dollars on training these things. How about, like, you know, spend a little bit of money on, like, the most basic type of, you know, retrieval augmented generation for me and my, you know, like, it's, they're just sort of sprinkling it in, and it's a little bit of a checkbox. Um, so I, you know, I'm a little bit mystified, right? Like, if they were very unified about it, I would really get it. Right? Like, clearly, the way that we're going to interface with computers is totally going to change. What Anthropic is doing with computer use is, you know, I think that, um, 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 Anthropi…

AI assessment note: “they're just sort of sprinkling it in, and it's a little bit of a checkbox”

Q Um, so just to explain to people, synthetic data is when the model creates data that it trains itself on?

A That's right, and so, um, I guess the other really big shift is actually test time compute. Like, literally, a one pro is this thing that you can pay 200 dollars a month for, and, um, It actually just spends more time at the sort of query level. It might come back, you know, five minutes, 10 minutes later, um, but it will be much more correct than sort of the, you know, predict next token version that you might get out of, you know, standard chat GPT. Um, yeah, from, from what I can tell, that's where a lot of the wilder things might come out. Um, you know, level four AGI, um, As defined by open AI is, uh, innovators. So we have, you know, lots of startups, both YC and not YC, that are trying to test that out right now. They're trying to apply, um, the latest reasoning models from open AI that are about to come out, you know, like O-three and O-three mini, and, uh, they're trying to apply them to actually, you know, scientific and engineering use cases. So, you know, There's a, um, a cancer, uh, vaccine biotech company called Helix that did YC a great many years ago, but, um, what they've figured out is they can actually hook up some of these models to, um, actual wet lab tests, and, you know, That's something that I'd be keeping track of, like over the next couple years. Like, if only by applying, you know, dollars to energy that then goes into these models, will there be, ah,…

AI assessment note: “That's right, and so, um, I guess the other really big shift is actually test time compute.”

Q Kind of like a plane, right? Like we want a physical pilot, even though the plane is probably better off by itself.

A Yeah, that's right. And that, that might be what, what ends up happening. Like even if 90% of the time you're using the autopilot, like you always need a human in the loop. And, you know, I'd be curious if that turns out to be one of the things that society learns. Uh, one of the crazier ideas, uh, I've been talking to people about that, like, I feel like would be a fun sci-fi book would be just speculation playing out on, um, you know, sort of how this interacts with nation states. Um, like, you know, China obviously is run by, uh, a central committee and arguably Xi Jinping. Um, you know, seemingly if you had ASI, you would only want, you know, sort of the central committee to have it. And so, that might turn into, like, a very specific form of, you know, it's, you know, China might end up having one ASI that is totally centrally controlled, and then everything else about it, you know, sort of comes out of that, and then you might end up with, I mean, controversially, like, I think Often they're trying to be benevolent, right? Like, if you spend time in China, it's incredibly clean. It's, you know, I'm sure there's all sorts of crazy stuff that happens that is quite unjust. You know, I have no idea. It's not really even my place to, like, uh, argue one way or another, um, what, what it's like to be in China, but, um, that's an interesting idea. It's like, You know, that socie…

AI assessment note: “Yeah, that's right. And that, that might be what, what ends up happening.”

Q Amazing. I want to just end with the same question we always ask, which is, what is success for you?

A I think looking back, I mean, growing up, um, I always just looked, looked up to the people who made the things that I loved. And, you know, Steve Jobs, Bill Gates, like the people who really created something from nothing. And, um, I just think of Steve saying, uh, you know, we want to put a dent in the universe. And, um, ultimately that's what I want. Like that's, you know, success to me is how do we bring forward? You know, actually, this is actually when Paul Graham, uh, came to recruit me to come back to YC, I had actually left and started my own VC firm, uh, you know, got to three billion dollars under management, like in the end.

AI assessment note: “we want to put a dent in the universe. And, um, ultimately that's what I want.”

Q Do you think that'll be different under Trump? I don't tend to get into politics here, but so many people in the administration are already incredibly wealthy.

A Oh, yeah. That's the hope. I mean, uh, we're friends with a great many people who are in the administration. We're very hopeful, and we're, you know, wishing them, we're hoping that really great things come back, and You know, uh, in full transparency, like, I think I was too naive and didn't understand how anything worked in 2016. That's not what I was saying in 2016. I was fully, you know, an NPC in the system. Um, but, you know, also that being said, I'm a San Francisco Democrat, so I really have a very, very little, uh, you know, I have very little special knowledge about how the new administration is going to run, um, except that I You know, really am rooting for them. I'm hoping that they are able to be successful and to, you know, make America truly great. Like, I am a hundred percent, you know, even though I didn't vote for Trump, I am a 110%, you know, down for making America truly awesome.

AI assessment note: “Oh, yeah. That's the hope. I mean, uh, we're friends with a great many”

Q better. I was on the hold for, like, three hours the other day for, like, a 15 minute question that I had to answer, and it's like, well, this could be, A, it could be done so much quicker by somebody who's not a human, and probably more securely and reliably, and more consistent, regardless of who's on the other end or how they're talking. How would you define AGI?

A Um, I guess the funniest thing is, uh, Microsoft, I think, is defining it when it gets its hundred billion dollars back, but, uh, I, you know, am sort of skeptical of that because, uh, you know, I think basically only Elon Musk then would, you know, qualify as a, uh, human general intelligence, I think. Um, Like AGI, the thing is, like, in limp, in a lot of domains, it feels like it's here, actually. I mean, you know, can it have a conversation with someone and take, you know, give incredibly good wine pairing recommendations, and have a perfectly fine, indistinguishable from a real human, you know, sort of, or even better than human Uh, sort of interaction, and also, like, take orders for very expensive wine, and have that just work.

AI assessment note: “in a lot of domains, it feels like it's here, actually.”

Q Let's go deeper on the regulation, then come back to sort of competition. How would you regulate AI, or how do you think it should be regulated, or do you think it should be regulated?

A It's a great question. Um, I guess there are a bunch of different models that I could see happening. Um, you know, I think what, what's emerging for me is that, um, the two things that I think, I think the first wave of people who are really worried about AI safety Not to be flippant, but like my concern is that they basically watch Terminator two, you know, and I'm like, I like that movie too, but, um, you're right now, you know, there's sort of that moment in the movie where they say suddenly the, uh, the AI becomes, um, self-aware and it becomes, you know, it, it, it takes agency, right? And, um, I think the funny thing, at least as of today, you know, these systems are, it's just matrix math, and there is no agency yet. Like, there's basically, they're, they're equivalent to incredibly smart toasters, and some people are actually kind of disappointed in that, and personally, I'm very relieved, and I hope it stays that way, um, because That means that there's still going to be a, you know, clear role for humans, um, in the coming decades. And, uh, you know, I think it takes the form of two very important things. One is, uh, agency. I mean, people often ask, like, what should we be teaching our kids? And, you know, the ironic thing is we send them to a school system that is not designed for agency. It is literally designed to take agency away from our children. Um, and maybe …

AI assessment note: “these systems are, it's just matrix math, and there is no agency yet.”