The Exchanges

Q on, right? You can imagine a world in which ChatGPT does all this or Google's Gemini and its consumer grade products that people are using to figure out what to do with their, with their weekend plans that it's natural for them to then ask the same question for. It's not an unfamiliar tool. How do you, when you do this, how do you compete with, with these bigger players?

A Yeah, absolutely. I mean, I guess I have two answers. One is obviously the one that, um, we're, you know, at Y Combinator, we, ah, you know, we believe in that we see it happen. Like, we just see small teams of people go out into the world and create, you know, they, they're obviously using these incredible frontier models, but they're adapting them to the very specific things that real people in the economy actually need, and so they're not necessarily glamorous Scenarios. Uh, there are customer support scenarios for, um, HVAC consultants, for instance. This fragmented industry, but a very big industry, um, And they're taking, they're building software, you know, there's a company called Avoca that we work with at YC. They're, they're doing exactly this customer support for HVAC, but V-one of it was basically Service Titan. So Service Titan is a incredible public company, but, um, they're basically software. And, uh, HVAC consultants and firms spend about one percent of their dollar wallet, you know, for every dollar For every hundred dollars they bring in in revenue, they spend about a dollar on software like Service Titan, but they spend five or six dollars on actual people picking up the phone and doing scheduling and doing all that stuff. So the wild thing that we're seeing is that if you, like, scope what you're doing and make the thing that is perfect for that set of peo…

AI assessment note: “they're adapting them to the very specific things that real people in the economy actually need”

Q Would you edit this? Or is this like a bad first draft that you then Get to go and put your spin on, or are we talking about 80%?

A Oh, it's usable. Like, I mean, basically, yeah, well, I mean, the, initially it was bad, and then, um, what I would do is do this process, get the script to where I felt really good about it, and then at the last point, I would say, given, uh, what we did in this session to improve the prompt, output the next version of the prompt. And so now I've done this about 20 or 30 times, and so now I have a thing that has all of the different tricks, like I even, you know, it started off as just like, here's a format, here's, ah, sort of specifically how a good video might work, and then now what's crazy is because of the, once the reasoning models came in, ah, now I can actually give it a grab bag of tricks, some things that, um, I mean, what's funny is, like, it's not entirely the AI coming up with it. It's not entirely me coming up with it. Like, in the course of co-writing something like 10 or 15 scripts, like, it's figured out all of these grab bag of tricks. Like, a pop culture cold open. A 15 to 45 second film TV news clip that mirrors the thesis before the hook. Like, uh, an authority pillar. Like, I love quoting Paul Graham, or Alan Watts, or Naval Ravikant.

AI assessment note: “Oh, it's usable. Like, I mean, basically, yeah, well, I mean”

Q I also wonder if that's great for you, because what you're looking for at YC is not to have a bunch of small companies where the founder can live a good life, maybe buy a second home, but where they're building the, uh, the Mark Benioff size successes, right?

A Yeah, that's right. I mean, I guess YC is funny because even if someone doesn't end up making the Salesforce size thing, like, they often sell their, I mean, that was true for, um, Posturus. My, my YC startup ended up selling to Twitter. My, our YC batchmate, um, back type, uh, it was Chris Golda and Mike Montano's company. They sold to Twitter and then the, that team ended up creating Twitter ads. I think like our, our old teammates at Posturus ended up making the, you know, uh, making the first Twitter, uh, Twitter mobile apps and, or working on that team. So I don't know, there is like a creative destruction aspect and then on a sort of Day to day career basis. Like, it's better for people to, ah, become founders, learn how to create things for other people, and then either, you know, you manage to get product market fit, and you figure out a moat so that you can be, you know, as big a company as possible. Or even if you don't, like, everything about your life and career moves ahead by, you know, five or 10 years faster than, It would have been, or, you know, we have lots of friends who, instead of starting companies, they stayed at Microsoft, and, um, It's better to be directly in the face of real users and shipping real code and product, and then learning how to support that, um, because that's just actually valuable. And I think there are lots of other jobs out there that…

AI assessment note: “Yeah, that's right. I mean, I guess YC is funny because”

Q What was one of the hard things to cut back on?

A Obviously, what's great is, like, all the people sort of involved in continuity are working on their own funds, and they're doing great, and, uh, we think the world of them. Um, but yeah, that was probably the hardest thing. You have a great team that's executing on a strategy, and then at some point, we realized, actually, like, YC should be about the, the initial batch, and, like, rather than Treat like group partners as kind of like camp counselors. It's like, oh no, no, those people are actually the partnership. Like we're an equal investing partnership, similar to benchmark, but we have 15 people. Like that's actually the core of what YC is. And then we also have incredible staff. We have the world's best media team. We have the world's best software team. And, um, those are sort of like the pillars of YC. And then it's just, So much simpler. It's just like, let's do what we uniquely do the absolute best. Let's stop competing with all the other VCs in, you know, let's be their partners actually.

AI assessment note: “all the people sort of involved in continuity... that was probably the hardest thing.”

Q I think there was also like a Q&A component of this, right? So they could go and talk to other lawyers. We're going to get into how things get better. Why wasn't that enough?

A Some things can become huge and drive billions or tens of billions of dollars in revenue every year. And some things really could only get to They only provide value that, you know, and then you multiply it out by all the people who need it, and that might only total up to 10 or 20 or fifty million. Like, that's, you know, weirdly quite common. I think a lot of founders are worried about that early, but my sense is maybe it's premature worry because embedded in that is also the case text pivot that, you know, they got users and an understanding and a useful corpus of data All of which turned into a tremendous moat for them literally right at the correct moment as technology itself shifted, ah, something that could only make, you know, tens of millions a year could suddenly become something that could make hundreds to billions of dollars per year. And that was the dawn of the large language model in 20, 23.

AI assessment note: “and that might only total up to 10 or 20 or fifty million.”

Q So then he starts adding this on, once it's ready, once it's ready for lawyers, what was the original use, and then how did it take off?

A I believe he basically started using it for, um, being able to answer specific questions about legal cases, um, and once he got access to GPT-IV, he realized that if you cut down the size of the question to small enough, um, and today we call that context engineering, but at that moment, he realized if you asked a very long-ranging question, um, Like, is the defendant guilty or something? You know, it's like such a big question that, uh, even GPT-IV, I mean, you could argue that, uh, some of the reasoning models today are actually much more capable of doing it, but back then you didn't have multi-stage, like, test time compute reasoning. Um, at that moment, uh, if you chopped it down to a bite-sized chunk, like you gave it some amount of, uh, context, That a human being, given the same context and the same prompt, would answer in a certain way. He found that he could, ah, you know, given inputs and outputs, have output that was usable, useful, and reliable, and not a hallucination. But it required you to chop that down into, um, a particular small enough step. I think of Jake a little bit like the first man on the moon. You're like, oh, you can chop it down, and then you should actually have tests For a bunch of different inputs and outputs, and you should have evals that actually, um, give you a sense and certainty about specific tasks. So you would sort of do tailored time an…

AI assessment note: “he basically started using it for, um, being able to answer specific questions”

Q But Gary, even if it's the same, like I, I'm wondering that because of all the vibe coding apps, I keep seeing vibe coded apps from people. Will they turn into something significant? Or is it going to be like most of the YouTube videos where there's no business from it? It's just fun to create. Or does that even matter?

A My argument would be, I mean, especially vibe coding, um, The Claude Code team apparently writes, 95% of their code is written by Claude, which means very directly that each engineer working on Claude Code themselves is doing the work of 20 people. That's sort of a direct quote from a recent, like, Lenny podcast with one of the co-founders. And so, I think that that's actually the good news. You know, I think if you look at tech across, like, 10, 2030 years, um, It's actually that, like, the access to good software is incredibly inaccessible, and one argument I often make is, if you use an iPhone, you probably have hit, ah, bugs in Apple Calendar, and it's, like, very frustrating, because, come on, guys, like, this is the built-in thing to Apple, the, the iPhone, like, the iPhone is, the Apple is, like, one of the most dominant tech companies in the world, and yet, they cannot find good enough software engineers to fix the basic bugs that Still exist in Apple Calendar. And so that's been true for time immemorial. Like, if that's true for Apple, how could you possibly imagine an HVAC person ever getting access to good software? And, you know, that's the difference today. It's like, hey, you can have it now, and it can be customized to you. And if anything, like, the funniest thing is, if AI and CodeGen gets even better than it is today, um, people can, like, that might be one of…

AI assessment note: “that might be one of the vectors by which, uh, HVAC people like compete”

Q don't even know if you know that I'm publishing the video. Cause in the past I didn't publish our videos this I'm I'm publishing, but it didn't matter. You set it up beautifully for me. What am I not seeing? This is the, obviously the exterior stuff. This is an indication of like you modernizing the way the Y Combinator communicates, but what am I not seeing underneath the surface?

A Yeah. I mean, I guess, uh, I felt really, really inspired and just like sort of filled with fire actually by, uh, one of our board members at YC is Brian Chesky of Airbnb. And so He was part of the selection process when they were looking at candidates for this job, and, ah, the second I got in the role, like, I mean, my board is, ah, Brian, and obviously, Paul Graham, Jessica Livingston, the original founders, ah, Carolyn Levy, and, ah, you know, Harsh Tagger has actually just joined as a observer recently, and so, yeah, these are sort of like the stalwarts of YC, and then they basically just really Enabled us to, I mean, think about it from first principles. Like what does, I feel like YC, one point O was the creation of Paul and Jessica and Trevor Blackwell and Robert Morris. I mean, the original founders of YC really set Like, the, the vibe, and the course, and like, what YC is about, and they built it up, and then the second decade, you know, was really with Sam and Jeff, and Sam created, he took Google and turned it into Alphabet, and it was, you know, a lot of different competing things that all, like, sort of raised the, um, ambition level of what YC was.

AI assessment note: “they basically just really Enabled us to, I mean, think about it from first principles.”

Q is there is a focus. It is companies like, uh, where is it? Um, market silver bullet for trade compliance to give you an example of what I see. I see another one Bluma automation, automating short form video ads at scale. So you really are still saying, I'm going to be focused narrowly on a vertical. Am I right? Or am I just looking at a handful and drawing?

A This is also about like making individual founders successful, right? I guess famously, I think at some point, Sam Altman came out and was, while working on OpenAI, he was, you know, sort of rethinking whether like the classic, ah, YC advice was correct. I hadn't, I mean, obviously we're friends and we like hadn't, like we had some exchanges about it. You know, I think that he's sort of changed his tune a little bit in that he's seen now that like AI, like all the startups out there using his APIs are sort of his commercialization arm, and that's not a bad thing, right? Um, there was a time when I think he said he just wasn't sure if, um, All of the advice around make something people want and like being lean was quite the right thing. And then to me, I think looped had to be lean. You know, a lot of people who start really huge companies had to start companies that were much more specific. Elon Musk had to start zip too. I think the reframe for us at YC is that we actually want people to be Uh, directly in control of their own destiny to the extent they can.

AI assessment note: “A lot of people who start really huge companies had to start companies that were much more specific.”

Q What is like, how, how do you stay in touch with people to keep guiding them after the time that they're in the program?

A Oh, well, I mean, these days, uh, everyone who gets into YC, they have one particular primary partner. And, um, obviously when you apply, it goes into sort of this giant pool, but then, uh, we have 15 equal partners on the investing side who like, uh, are in, we basically are in there trying to fish. Like we're, you know, in there with our nets and like, let's, uh, take a look, like, let's watch the video. Let's try the demo. Let's read everything about the founders, what they've done and What do they know about their users, about, uh, the product? And then we try to figure out, well, who do we meet? And then anyone you choose, you meet. And then when you meet, it's up to you, uh, whether or not you accept them. And then when they're in, like, you always have one, at least one person. Who is sort of like your investor, uh, at YC and that.

AI assessment note: “everyone who gets into YC, they have one particular primary partner.”

Q it. I can create a business coach for you. He did it overnight. It is good. He's making it better by adjusting it to me. Do you think one of these tools can, that's built on using, let's say, 11 Labs agent feature, do you think they could become a business that eventually ends up on Y Combinator? Are we looking at stuff that's always going to be too small?

A I mean, honestly, uh, we, we would, I, I think that even founders who, uh, have access to a lot more capital or better resources, um, we've been working with a lot more alums in the batch now. Daniel Kahn, for instance, who created, co, co-founder of Cruise Automation, he's in the batch, and, uh, I'm pretty excited about it because whether it's like your first time or your fifth time, Like, being next to a bunch of people who are moving extremely fast, like, that, that's probably the biggest thing that second-time or multi-time founders maybe struggle with, is that, like, the next time you have plenty of resources, but the one thing that's actually the most important is time. And so there, there's almost the only thing that matters up front is, like, can you speed up? And then I, I do think it takes a village to actually properly speed up.

AI assessment note: “we've been working with a lot more alums in the batch now”