All right. All right. Let's do it. Super excited about this. We obviously are a YC company. It kind of feels like our DNA. Let's start off with YC. Maybe we can talk a bit about the current batches. Can you give a rough idea like of the categories, the breakdown? What sort of companies are you seeing come into the batch?
Yeah, definitely. You know, it's about 75% software, 25% hard tech, which is very noteworthy for YC because I think even going back five years, maybe that number was only three to five percent hard tech. And so that's a really big shift. You know, if I had to guess, it's really this anticipation for, you know, more or less AGI loops in things like Navier-Stokes, like literally pushing forward the ability to do brand new R&D that's never really been done before. It's the kind of research that at one point was done in universities. The people working on hard tech by and large have PhDs. About a third of the founders working on hard tech have full PhDs, you know, not even ABD, like they did their dissertations, you know. And so it's really, it's a different moment right now, which is actually very exciting, right? Like software is great, but hard tech is like the thing that creates abundance.
And actually, like, does it, is it because your confidence in software is lower or your confidence in hardware is higher?
So, I mean, my confidence in YC, in software has never been higher. I mean, this is the golden age of being able to build. Like, I got in trouble with you guys for saying like one person can do the work of 400 people. I retract the 400, but like I think you can do like eight times, 10 times, maybe 50 times. I mean, I might go as far as like there are probably people in this room who can do the work of 100 people, right? You know who you are. So, it's everyone here.
So then, what do the models bring that gives you so much more confidence in hard tech that you're willing to increase the number of companies in the batch?
I guess the wild thing to us is, I mean, the most surprising trend that we're seeing is that not only are the hard tech companies applying and the quality of the people working on them are like so, you know, this PhD level, you know, all of the right, you know, it's not, you know, necessarily like the software companies are all like a wide range, like 18-year-olds all the way up to 50-year-old, I mean 60s, I mean just like, you know, classically much younger on the software side, but you know, on the hard side, like people coming from national laboratories, people coming from, you know, the Bell Labses of the world. And what they're realizing is like, oh well, I can't get the compute. I can't get the support. I can't use AI the way I want to use it in like literally the universities and the R&D centers. And I think that's probably the biggest shift is just like you actually need compute in that agentic loop, like, you know, GPT-6, Astro with Navier-Stokes, like that kind of thing is going to be happening all around in society and I'm really excited about that. You know, it costs tens of millions of dollars to achieve that result. But the really cool thing about the scaling law is that, you know, that same tens of millions of dollars today, this year, might only be a million dollars next year. It might only be $100,000 the year after that. And you know, sometime in the next 24, 36 months, like does that mean that almost every serious problem facing humans could be addressed and, you know, someone in this room might pair with like a PhD researcher and like go and make that happen? Like I'm really excited about that.
Are there some categories of like hard tech that you think are like nearer term, some that are longer term? Like let's say the Navier-Stokes example, because math you can wrap some formal methods around it, so suddenly it's very good. But when you get some unconstrained maybe physics or something like that, they might be a bit harder. Like how are you trying to judge the companies that, oh, these guys can achieve it within this time frame rather than in the next 10 years?
I mean, the reality is with a startup, like these are not research organizations necessarily. You know, we're trying to be pragmatic. So, you know, the best version of these things have some sort of real pragmatic use case in like today, right? I was meeting with a, you know, I did office hours just now with a photonics company and they told me this crazy story about how data centers today run on fiber optics, but the very largest data centers with, you know, hundreds, 100,000 GPUs, like the Colossuses of the world, sometimes they only run 30 minutes because too many lasers burn out during the operation that, you know, the actual number of GPUs turned on and that can communicate to the main network is like, you know, 20 or 30% of them are like totally offline at any given time. And it's literally like because there are tens of thousands of lasers. And so, you know, how cool is that that you could sit down in front of a computer, come up with a new type of laser technology, and then that's very pragmatic. Like if you could go into data centers and solve that problem for that set of customers, like that's insane, right? And like these are things that teams of two or three people with frontier models, maybe developing your own like child models around like your own version of AlphaFold for photonics or whatever it is. Like I think that there's just going to be like this AGI loop around, you know, these frontier models that have literally, you know, superintelligence available on tap. Like just you have, you can have just a team of intelligence working on, well, here's the next step. Like you're supplying it what you need or what you want. It's like, no, no, I actually need, you know, 500-nanometer lasers. I need, you know, one-watt lasers. You know, it's not physically possible today, but if you wanted it, how would you get it? And then you would have, you know, you would have in silico simulation and then you have experimentation. And you know, I mean, Periodic Labs, for instance, a fantastic company doing that in material science. Like that's, you know, the low, they were working on a what, post-training 01, I think, and then they said, you know what, like this is the lowest hanging fruit, we're going to go do that. It's like I think that model is going to happen across all of science and engineering, and that's really exciting.
Do you think you have the most interesting job in the world where you just can sit down with people and basically see what the future is going to look like?
I mean, you hope that's what the future is. I mean, we give them money and we try to give them a network and, you know, often they'll do it, which is awesome. Like what that can happen at all is like the remarkable part.
Your office hours now, I imagine, are so much more interesting. You're like, we were talking to Michael backstage and he was saying that, you know, in the past a lot of the same problems and now he's learning a lot from office hours. I can imagine it's the same with you. There's a lot of new stuff coming in that's just super fast.
I feel like I've always learned a lot every single time. It's like, you know, how the funniest thing about society is how many conversations you have. Like I think you can tell how much fun, you know, I get to have by like how many dumb questions I get to ask of people who are, you know, way more knowledgeable about something than me. And so in which case I have the best job in the entire world because I can ask infinitely dumb questions and learn everything from people who really, really know that stuff. It's like, um, what's the word? There's like a word for like not being sad that you can't live like a thousand lifetimes or something like that. I forget. It's like saud or it's like one of those type of... Yeah. It's probably a German word. And it's like this is the perfect job to like quell that feeling because it's like I, you know, I can only live one lifetime, but when you meet like the best technologists in the world who are obsessed with a thing that other people need, like I mean that's just the beginning. Like seeing them, giving them money and a network and trying to help them, like that's very rewarding. It's super fun.
So now, like in these office hours, what advice are you giving to the founders, and especially contrasted to two years ago? So what is different from two years ago in terms of the advice that you're giving to the founders?
Remarkably, it's like all the same. I mean, you know what's funny is Paul Graham comes and speaks to every batch, and then the number one question everyone in the batch always asks is like, what advice did you write about that, like, you know, changed because AI is here? And he's always like, nothing. Like it's all the same. There's nothing changed. Like you still have to understand your user. You have to actually talk to them. You have to create, you know, an initial version of it. You have to put it in front of them. And it's like, hey, did it work? Did it not work? Like the fundamentals are all the same. Like my favorite part increasingly, I think, you know, you guys do this really, really well. It's like the meme marketing part, right? Like being able to have brain, you know, get brainworms into people's heads so that like when they have that problem, they remember, oh, like I need to use Supabase, right? Like that's actually really important. Like that's one of the things that brought everyone here. It's like, well, there are lots of ways to store data, but I just know, you know, this is the shelling point. Like everyone on the internet knows like this is the way to do it actually.
And then what about during the batch? I guess mostly on the software side. Are you finding people, a big part of our batch is like finding product market fit, or every YC batch? Are people able to find product market fit faster now because of the models?
Yeah, that's happening way faster. I mean, that's the thing. You used to have to sit there and code the thing, whereas now you can open, I use Cappy.ai now for all of my agent coding. It's like, you know, you can open 10 windows and just like any speculative thing. It's like, I want to do this, like let me do another project over here. You, I mean, you can basically have in parallel like 10 branches of like all...
Oh, I mean, let's see. So in January I built garyslist.org, which is me attempting to recreate Posterous, but like as sort of a news site and like a community site. And then I spent a lot of time like learning about retrieval and like it ingested like my entire X feed and all, I was kind of focused on politics. And so, you know, all of the white papers, all the articles ever written about like, you know, pretty gnarly things like the, you know, the drug crisis down in Tenderloin, like, you know, who are the players, what's the non, you know, which nonprofits did what, what are the policies, like what are the, what's the research behind it. And so it was sort of this agentic RAG project. And then I built that and then what's funny is then I discovered OpenClaw. And so I had written what I call a Foxconn factory for my agents, right? So I wrote, you know, half a million lines of Rails code using Opus 4.5 and 4.6. And I treated LLM calls as the very, I mean, I don't know, it was very expensive like at the time because I was like, oh, like I have to spend this much, like, you know, you're, you, I tried to like chop up the use of intelligence and like control it in a really intense way, kind of in the same way that a Foxconn factory worker probably wakes up, does calisthenics, like their lives are like planned out to like the second. And then I met Peter Steinberger, he came by the YC office. He's friends with our partner, our newest general partner, Rafael Shad. And I tried OpenClaw for the first time. I'm like, oh, I was making Foxconn factories where like deterministic Rails code tries to control what the agent does. And like here's the input, here's the output, here's like a bunch of evals and like everything's all like locked into straitjackets. It's like I have total control over what the LLM does. And then when I saw OpenClaw, I'm like, oh, this is like...
Yeah. No, it's like an ethos. It was like a bright open field. It was like free-range intelligence is like the, you know, you tell it what you want and like it's able to code what it needs. And then I heard about Karpathy's LLM wiki and how he, you know, has a memory system. And so then I created, I had created GStack as an attempt to understand like how to get the most out of, you know, Claude Code and agent coding. And it just was like, it actually started as two markdown files that I would just find myself pasting into Claude Code all the time. And it was like CEO review and review, and it was like make sure that their tests were running and that it was like properly tested. And then GBrain was like taking my Gary's List agentic retrieval stuff and then making it work inside OpenClaw. And then suddenly like that took a life of its own because what I realized was, you know, we already have a form of superintelligence in the working memory, like our whole society. Like if you think about every standup or every status meeting or every exec meeting that you ever have, it's built around like squishy humans who can only keep seven plus or minus two things in your head. That's why phone numbers are like the number of digits they are, right? So, you know, like we're kind of limited as human beings. Like I accept that now, but, you know, at the same time, an LLM can have a million tokens, which is like I like to say three Harry Potter books that it's like totally usable. Like, you know, you can generally like with 95% accuracy pull out like any part of like three books of Harry Potter. But it's not a library. It's not like if you could imagine a room of this size as a database and it's full of books, right? So what's the bridge between what LLM can do today and the amount of data that is actually something that represents your life, actually represents your company? And then to me, like I actually think there's something very elegant and interesting about markdown files and the LLM wiki because like you don't, you know, you don't, the agents just like speak it inherently, right? And by default it uses GPT, but it turns out if you add, you know, graphs, if you add additional, I mean, basically a full RAG, you know, graph RAG sort of framework on top of it and you have dream cycles and all of these other like syntactic sugar things that make it better. That's what I want GBrain to be. It's like, and so I use an OpenClaw with GBrain. It has half a million markdown files. That is literally like my entire life. And it makes my life really simple. It's like I, you know, I don't, I generally don't even have to open Slack and like click through anymore. Like the second someone, like if there's a crisis at one of my companies, it's like, oh, hey, most important thing you need to do right now, it just like, you know, it telegrams me right away and I can respond to it within like 30 seconds. Like I, you know, I'm going to walk off stage here and if anything happened, like I'll be able to respond right away instead of like, oh, next time I look at Slack, like, you know, I'll eventually get to that DM, right?
