How YC fosters AI Innovation with Garry Tan

0:05So today I know Pryor is we're very lucky to have Gary Tan. Gary is currently the CEO, is that the right title? Yeah, president and CEO. The president and CEO of Y Combinator. Prior to that, he ran, initialized, early stage venture fund. He started Posterous, which was acquired by Twitter, and has had a storied career working at a variety of Silicon Valley companies, including Palantir and others. Thank you so much for joining us today, Gary. Thanks for having me. Well, this is going to be our first Collabo episode. So part one on this podcast, part two over at the Gary Tan YouTube channel. So thanks for having me.

0:37Thanks for doing the collab. Super excited about it. Love to work with like a real creator. Um, okay. Let's start with like, you obviously need no introduction, but people want to know the story. Like, how'd you go from being a founder to an investor at all? I think, uh, Paul Graham basically took a chance on me as a founder. And then later I got burnt out. And, uh, YC itself, I like to think of as, you know, today people sort of look at it as this sort of institution. It's sort of the defining place where people want to start their billion dollar companies. But I feel like in 2008, when I first found out, found out about it, it was a little bit more like, uh, you know, a punk club in the nineties in Seattle and, you know, sort of grunge was just getting started or something like, like these things are basically subcultures and, uh, they sort of start off of the internet.

1:27And so one of the reasons why Y Combinator was so dominant and, you know, start, even had its chance to start was that Paul Graham's essays for, you know, sort of time immemorial, you know, from his experience creating Yahoo stores, that he would write these essays that basically spoke to the experience of trying to start a company at that moment. So when you go around sort of the legends of, you know, YC, whether it's, you know, Patrick Collison at Stripe or my buddy Harge, who I work with, who co-founded a company with Patrick. You know, lots of people, even me. I mean, I basically found out about starting a startup, you know, from the news, like that suddenly, hey, there's this person who could sit in front of a computer, create software that could then touch a billion people.

2:19And I was like, oh, it's like perfect mimesis. You're like, oh, I want that. How do I do it? And then you go into Google and you type in start a startup and you find Paul Graham's essays. And then increasingly they're finding this pod and they're finding my YouTube channel and they're finding the Y Combinator YouTube channel. And, uh, you know, I think all of that is sort of the direction we're going in, but that's how it started. It was like, literally starting a company was a very strange thing. Like when all of us came into tech, that was, I think more true today. I think it's something that a lot more people aspire to.

2:53And that's a really, really good thing. And that's sort of what we want. Yeah, it's funny. In 2007, I started my first company. And I remember at the time, people, if you talk to the investor community about Y Combinator, they'd say things like, oh, there's probably adverse selection and who would take that deal and all this stuff. And that's right when Airbnb got funded by YC, when Dropbox got funded by YC. And so it's always striking how right when people start saying something, that's actually the moment you know it's working. I'm a little bit curious, in 2008, when you were going through YC, who were some of the other people who were in your batch or in that same year?

3:23Yeah, I remember, you know, actually, there were probably three or four companies that ended up, quote unquote, making it, I think. And how big was a batch at that time? Oh, it was only, I think it was 25 companies total. So, which is funny because that's the number of companies that a group partner now has. We have 14 group partners at YC now and each of them fund between 15 and 25 companies each. So I like to say that we basically have 14 of those sort of 2008 era batches happening all of the time. Let's see. There was a company that actually also sold to Twitter, Mike Montano and Chris Golda's startup.

4:03And I think it was in basically like comment analytics. And then Mike ended up being VP of Eng at Twitter for many years. Chris still does a ton of YC investing. That's right. Chris is doing incredible investing now. So who else? D. Scott Phoenix is now, I think, over at 50 years. The Vicarious. Yeah, he did Vicarious, which was one of the top AI startups for some time. And then that sold to Google. So, you know, it's kind of cool. Like, I mean, you get started in the biz. You, like, defriend people over pizza and beers. And then years later, you're just, like, sort of trying. You're just still at it, still doing things.

4:42And then now we're trying to help the next generation, which is fun. Yeah, people forget how much longevity there is sometimes in Silicon Valley. And they forget how you act early really impacts almost how people perceive you 10 years later, 15 years later, etc. And so if you're very non-zero summer early, it really matters in the long run. Because all these relationships end up unfolding over time. And so it's one of those things that feels very under-discussed. Yeah, that's definitely true. So still, posturous age Gary, when do you start making angel investments? Oh gosh, it was much later. I mean, let's see.

5:15Posturus was funny because we had a chance to be Instagram, but we didn't know it at the time. Obviously, nobody did. I mean, you were working in the social space too a lot. That was a really interesting moment in internet history. I think it kind of parallels what people are probably experiencing right now with AI, isn't it? We all knew that something very different was happening with the way humans were sort of communicating. There was six degrees. There were like, I mean, all of the same parallels. Like there were a whole generation of social networks that failed. And then those were the obvious ones to point to and say, oh, the whole trend is bunk.

5:56Like it's not going to work. You know, six degrees didn't work. That was a good enough team. What happened to that? Or like, oh, blogger, you know, it's funny because sometimes it's the same characters, right? Like Evan Williams worked on Blogger. It was mostly forgotten at Google. and then he just never really like lost the bug to work on this social space. Probably the craziest thing was, do you remember what Facebook, like meeting Facebook people was like back then? Like it was like meeting cult members. Yeah, they had a really strong culture. And to your point on the waves, you know, before they had Friendster that everybody thought was going to be the winner and then MySpace that everybody thought was going to be the winner.

6:33And then I remember when Facebook came out, everybody kind of poo-pooed it because they're like, well, it's just colleges and who's really going to care and MySpace is for everyone and all this stuff. And then of course, MySpace eventually just kind of died in part due to over monetizing the site with too many ads and things like that. But also the founders were not technical. Yeah. Which actually, I think that's even, you know, going back to the YC, the cult of YC, I'm starting to realize that is actually one of the core things that we've always believed that extremely technical people, those are the people who probably should be like the Zucks and the Larry and Sergeys of the world.

7:07Like those are the people who we should give extra support to and actually fund. And, you know, basically the hard part about playing chess is not knowing how the pieces move. It's being smart. So let's go find the smart people and then we can teach them like the community will teach them how to, you know, play chess, the game itself. Your analog to AI and mobile is really interesting because I remember during the early days of mobile, like 2009, 2010, I remember seeing half a dozen different companies all go from zero to a few hundred thousand users in like a week or two for mobile uploads, photo uploads.

7:41And then they die. And Instagram is the only one that sustained. Yeah. And its growth was actually a little bit more linear. How do you all think about that in the context of the batches that you fund at YC? Because you have, you know, over 200 companies per batch now. I think a large proportion are AI centric. Is it? About 70 % right now. 70%. Yeah. So quite a lot. And that's got to be really different from a year and a half ago. Yeah. I mean, it started back then, and then now I guess the secret is out. I think GPT-4 was actually a really big breakthrough. If you talk to Jake Heller at Case Text, he was one of the first people to actually get access to GPT-4.

8:17And then when he comes and speaks at the batch, you guys should have him on the show. He's actually, I mean, his story is really crazy, actually. they actually only two of the co-founders had access to it and then they went off into a room to play with it they had gotten access to three and 3.5 before but four was the first one where definitively they almost never got hallucinations and so they said this is actually the moment where they can apply large language models in a practical way and actually charge money and so it was actually fascinating to hear. Like, you should hear the story directly from him.

8:56But I'm just obsessed with it because they were just very pragmatic about here's this API. It literally is a form of intelligence on tap. And then because Jake was a lawyer, he's able to sort of go into the mechanics of what a lawyer would do. Like, you're basically just very, even like tailored time in motion level, like detail of, well, this is what a lawyer does when they look at a brief or when they're writing a chronology, specifically he's doing X and like they're opening, they're, they're reading paragraph by paragraph. They're converting that into a score. Like that's what highlighting is at the end.

9:34You, you know, take the outputs of that. And then, so he would, you know, map the specific things that a human would do in this very pure form of knowledge work that pays a lot of money. And then he would turn that into prompts and workflow and testing and scoring. And there's just like a billion different things that frankly, like all, like probably half of our portfolios right now are in the weeds trying to make those things work. Totally. And I think like this sort of like, if you have somebody who is really technical and understand what the models can do, right. And they might not be a researcher, but like sees it and has intuition for how to manipulate them.

10:11And you have, as you said, like the time in motion domain person. It's kind of an obvious pattern. Like we're both investors in Harvey. Like it's actually a very related story. But I bet, I mean, I want to hear about some of the things you're selecting for in like YC founders in this era. That hasn't changed though, by the way. I mean, we're just selecting the same type of people that we want, like all the time. We want highly technical people who are very clear communicators. And so, you know, the bad version of it, which, I mean, you guys get pitches like this all the time. Like you go to an event and someone's like, oh, I'm working on this thing.

10:41And you dig a little bit and it's like, they haven't talked to customers. They haven't talked to users. There's no, uh, it does everything. It's a general purpose thing. Um, it's, there's not like a thin edge of the wedge. There's not, you know, some set of users that will absolutely love it. And then I guess like I was speaking about this at, what was it? the startup grind conference recently that like, I'm realizing there are just really two totally different ways to think about the world. And then the founders that really do it, do it almost entirely from first principles. And the, for a startup founder, that what that means is actually going to talk to founders, not talk to other founders or even talk to investors.

11:25You know, that's all sort of secondhand, like to go to pick a very specific set of people. It's like, not just all lawyers, but like, I'm just going to work with, um, litigators, right? Because, and then I'm actually going to focus on corporate litigators who, you know, do this particular segment. I would almost view it as like, um, talking to customers as firsthand, talking to their founders as secondhand, talking to investors as thirdhand. And if you ever go and talk to an investor and an investor in LP or whatever, it's like fourth hand. And so I'd help you if you're like trying to, um, make, make your way by like reading the wall street journal or tech crunch or even Twitter.

11:59It's like, then that's fourth or fifth hand. The fourth or fifth hand retelling of things that happened six or nine months ago. They don't know. Yeah. So if you look at the YC batches, so you have 70 % of companies now doing AI. If you have, say, 200 to 250 companies a batch, that's somewhere between 150 and 200 startups. And to some extent, you could argue each startup is in some sense a vote on what's interesting in AI. If you kind of aggregate up all those startups, are there common themes that are emerged? Is there sort of things that are more topical or interesting from a startup perspective?

12:27I almost view it as a lens of a giant founder voting machine. Yeah, totally. So I'd say 70 % are somehow related to AI, and then two-thirds of them, you would sort of argue are SaaS wrappers, actually. Wrapper is sort of the pejorative, but I definitely believe that the chat GPT wrapper thing is actually just wrong. It's like saying that all of SaaS and cloud are basically MySQL wrappers. It doesn't make any sense. a pure technology that can be implemented. And then essentially it's a concentrated form of intelligence. Um, and then the funny thing about it is like, it's probably only 85 IQ actually.

13:10Like if you look at who would you actually hire in your real workplace, like. Sorry, 85 in terms of what you need or 85 in terms of what people are using today? Oh, 85 IQ in terms of like, when you ask it a question, like, I mean, it's able to give you an answer, but. So the model quality today. I think it's like a eager intern right now. Right. Yeah. It's definitely someone who doesn't know that much. It's not that the intern is often better. Uh, I mean, you would hope that your interns would be higher IQ than that. Right. So a hundred IQ is average IQ. Right. And then, um, the wild thing is, I think that we're right at that moment where hopefully, you know, looking at 4.0, looking at cloud three, like maybe we're just past average intelligence at this point, it's still very limited.

13:54Like you still have to be very constrained. You still have to do the tailored time and motion study to like break down the workflow to, you know, as literal as possible and interpretation of what a knowledge worker does. But I think that that's going to bear fruit. Like they're just companies that, you know, I think last batch, the average ARR of a YC company was$6 million at the beginning. And at the end, it was north of 30 mil and this was over three months. So to get that kind of growth in a very short amount of time, like I actually don't think that we've ever had a YC batch be able to get that type of revenue growth.

14:32I mean, in aggregate, and these are all very early startups. These are still like two or three people, the first demo, like it's literally the first version of the product, but to be able to, uh, you know, grow at that multiple in such a short time is sort of, um, a very good indicator. How do you explain that? Right. Like why are, why are the companies working so quickly? Oh, I mean, I think you can walk in and solve a problem very directly. That's basically a human being like would knock their head into head into that sand. Yeah. They would have to knock their head against that wall, you know, always, or you're already paying, uh, you know, an outsourcing firm.

15:08You're already, you already have a team, uh, offshore that's doing this thing and then here's this thing that allows you to do that plus more with even less. And it's just like the pure leverage of software. Yeah, it's really interesting because it feels like it's almost like two things are happening. You have a technology disruption that opens up new capabilities for software and then you have this really interesting top-down openness to try things because it's AI. I've noticed a lot of enterprises just want to do things because they're doing AI and there's a CEO mandate and they have no idea what to do about it.

15:36And so everybody's scrambling, what does a CISO do that's AI? What What does the customer support do that's AI? And so it kind of opens up a lot of things. Are you seeing any sort of commonalities in terms of people working at the application layer versus infrastructure? I don't know that you see as many people building models, for example. So I don't know how you think about that stack of stuff. Yeah, it's about, I mean, two thirds is still very practical implementation of ideally frontier models. And then maybe one third increasingly are some sort of infra. So tooling, testing, prompt collaboration.

16:08proliferation like things sort of in the lang chain lang fuse like there's just this huge proliferation of tools i'm sure you guys see it all you're in a bunch of the great ones too it's like uh it's super interesting to see that the really interesting thing to see also is that you'll have um this parallelism of the like closed source uh one that raises a lot of money and then you'll have the open source scrappy one that will come and commodify it and so all of that is sort of happening simultaneously right now. So one thing about the scrappiness is, you know, YC is like, I think, respected for like the scrappiness culture, right?

16:46Hacker culture. And, you know, one question is like, do you expect companies in YC to do pre-training and do their own foundation models in any way? Because, you know, there is a compute-driven narrative that you need a hundred million and then a billion in next generation closer to 10 in order to compete there. Totally. I think people are starting to do it. And then I guess what, you know, it's kind of like when Cruise came through YC, here's this giant like mega research project and you need a hundred million dollars to do it. And then you work backwards from that. Like what are the milestones that we can get to so that, you know, we can draw a line from, you know, we have nothing to we have something to have to really have something.

17:30You know, that's sort of the question always. So we absolutely have companies that are building foundational models. Diffuse Bio is doing over on the bio side. There's a company working on robotics foundational models. There's so many things that you could do. And then the half a million dollars during the three or four months, it's sort of what Kyle did for Cruise. He had to figure out, well, how do I make a$10 ,000 automated driving attachment to an Audi A4? And how do I get people to pay for it It's how do I, you know, all he did was actually make a demo that funny enough resembles full self-driving in any Tesla car today.

18:12But this was in 2013. You could drive up and down 101. It didn't do streets. It didn't do cities. It only did, you know, driving up and down the highway. And that was enough to sort of show that and to, you know, attract talent, to attract capital. And then it basically is a self-fulfilling prophecy. So that, you know. I think that's one of the things that YC founders tell me often is that they feel that YC is incredibly good at pushing near-term milestones. So every week, what have you done? Have you actually followed up on the things that you said you were going to do? Are you talking to customers?

18:46It's very tangible and that pushes progress. And I think Cruise is a really good example of here's a tangible thing I need to ship. And I think there's a very strong shipping culture. Where does that come from? Is it the partners that you have? Is it originally from Paul Graham? I think there's a really strong culture of a lot of these things that really help drive progress forward in the earliest stages of a company. Yeah. I mean, I think it was sort of like the blend of a whole lot of different things. Like, I actually don't, I think Paul Graham in particular was always about helping people see how big the idea could be.

19:18Like, you're doing this thin edge of the wedge, but if you do that really well and you do these other intermediate things, like, wow, you could be a$10 billion,$100 billion company. And so I think Paul Buchheit, like creator of Gmail, he was one of the first people to make group office hours, which is sort of using social pressure. It's almost like Alcoholics Anonymous as like, but for founders where it's like you really don't want to you. I mean, it's a mild form of coopetition. Like you're in competition with your friends, like you want them all to succeed. But at the same time, like if you come and you don't come correct, you're going to feel bad that week.

19:56And, you know, our realities are very socially - No, I mean, the YC founder, like founder friends will tell me, like, they're very stressed about group office hours because they have to show up. Yeah, they have to show up. And like, everybody's going to look around and say, what have you done? Yeah. Did you hit your goal this week, up or down? And one of the things I was joking about, like, I haven't done yet, but I might still do is force people to do pushups if you miss your goal for that week. It sounds ridiculous. But I mean, I think that, yeah, our realities are basically socially constructed.

20:25And then if you can learn how to run fast, then that translates into mega success. You know, basically, if you can run fast for a long time in the marathon, then compounding takes care of the rest. In terms of just distinguishing, like, what are ideas that you would actually be willing to back? You've used this analogy of, like, it's a bunch of startups that are mice running around the feet of the elephant. Oh, yeah. How do you decide what's happening? What isn't going to get stomped? We just ask them. And then if it's plausible and makes sense, why not? I mean, that's the amazing thing about early stage.

21:03You can basically, if we learn something from people, if they are in direct contact with what's really happening, then that's the most powerful thing. What has been the most surprising version of that where you thought this company, this YC company is definitely going to get stomped by an incumbent and they just ended up winning or doing really well? That's a good question. I mean, there's just so many examples, right? Like scale might be a good example. Like, you know, there's really, you would have assumed that like, wouldn't Google just do this? Or there are like so many other people who have more or less like infinite access to information.

21:41Like Amazon could have done that at some level. Like they already have Turks. Like there's so many other people who could have done it. It's true. I mean, Google had all these organizations where they're paying massive numbers of labelers internally and like had the operational expertise and the need. Yeah, exactly. So, you know, I think that that might be the good news for startups because, uh, actually the, the mega tech companies have such unassailable moats and they make so much money that they're actually somewhat incapable of like stomping anymore. It's a, it's a weak elephant. It didn't do its pushups every week.

22:16It just doesn't have to, right? Like the, you know, necessity is the mother of invention, You know, I used to believe, for instance, that, you know, when I was first coming up, like writing my first checks in 2012, 2013 as an investor, I didn't really understand what people were saying when they said, oh, like, don't raise too much money. I thought it was only a tactic. Like a negotiation. Yeah, I thought it was a negotiation thing. I was like, surely that's not what happens. And of course, like, you know, many years later, I feel like we have seen what that means. Like lots of, you know, some of our best companies with the best technical founders, they might have raised too much money.

22:57And then when it's time to cut, like, you know, I mean, that's a tall order to say it's time to let go 50 to 80 % of your staff. Like, are people really willing to do it? Like, these are people who you've hired and you've managed and, you know, you've like fought alongside for this whole time for sometimes for years. And, you know, it's actually too hard to do it. And so I can see that now, like, you know, too much money equals potentially just not doing the right things. And then the worst thing is like not even being able to cut to the point where you can get back on track. Yeah. And companies are so path dependent that now you still have a series of companies that raised a bunch of money maybe later in their life cycle over the last five years.

23:45And the war chest was so large that they're still sitting on it. They're sitting on a valuation that's very painful and trying to think about the down round and all of that. And the question is like, you know, some of these companies, there's fundamental goodness in there. Can they make the hard decision to get through it? Yeah, that's right. How have things shifted in YC over time? So I know in the earlier days, if I recall correctly, a subset of the companies would actually merge. Somebody would have an idea that wasn't quite working and they'd join one of the teams that was working or they'd shift ideas midstream.

24:17Also, it feels to me without any data, so I could be wrong, that the demographics used to be younger. In other words, if I look at the early batches of YC, Sam Altman was in there when he was a Stanford dropout. There's other people, Patrick and John from Stripe. People joined really young. Yeah, totally. Have there been big shifts in those sorts of trends in terms of how many companies can actually raise money out of the program, the demographics of the people? Yeah, great question. Wait, and how old were you when you did YC? I was 27. Okay. A grown up, almost. Yeah, yeah, yeah, totally. I think that that's actually the average age of YC at this point.