What It Actually Takes to Build an AI Company [Ft. Francois Chaubard, Y Combinator]

LMs will be better than us on specific things, and they already are. But in terms of fluid intelligence, we still are far superior. Given this advent of Claude code and vibe coding, it brings the activation energy to start a company way down. And so back to the something of somewhere, there's nothing of nowhere. You know, if you were going to start a company, what would you focus on as a mentor? What would you advise? The first thing is do a company. The skills that you need to be an entrepreneur, one of them is not to be Einstein. Like, think Einstein would actually be a very bad entrepreneur. So why did Steve and Bill Gates and Jensen, why are they so good is because they've failed so many times, and there is no shortcut. There just isn't. What advice do you have for folks that feel like YC is maybe not accessible to them?

There's two things I'd probably say. Number one This is AI twenty thirty. Today I'm with Francois Chobard, a visiting partner at YC, and I'm Chad Lorelei, founder of Quadra AI and SDX. Where are you today? Woodside, California. It's like a tree house. This is a Francois tree house. Yeah. This is the, the, the offshoot of my house to get away from the crying babies. Yeah. Where I can finally get some solace and work done. I love it. It's a, it's such a peaceful place. First time over here and we were driving up from kind of like through Stanford and, that's where you're doing your PhD, obviously.

But excited to chat today. Thanks for having me. Yeah. We'd love to just get in a bit of your backgrounds. Obviously, we've we've met a few times, had you down in San Diego For a YCSDX event. But let's learn a little bit more about Francois today on, your story, how you came from New York, studied at Delaware. Yep. And then made your way to Stanford, started a company. Now back at Stanford. Tell me the trajectory. Sure. So yeah, I grew up in, upstate New York, a town called Highland Falls, New York, right outside of West Point for the military folks out there. And then went to University of Delaware, undergrad, was the first person in my immediate family to go to college.

It was kind of like, was I going to go to college at all? Was kind of the question. And then I was like, Yeah, don't know, maybe I would go to college. It didn't seem like a thing to do. Then, so I went to Delaware, did math and mechanical engineering, and then graduated. Got what I thought was the coolest job of all time. At Lockheed Martin, I got to work on Blisset Missile Defense, like watching missiles hit other missiles two miles above the Earth's crust, going thousands of miles an hour. That seemed pretty cool. And then quickly realized working for the government as a defense contractor is probably the most boring job. Very unionized. People work forty hours a week and then yell at you if you work more than forty hours a week. And I kind of wanted to work one hundred hours a week, genuinely.

And so then after about two, three years in the Engineering Leadership Development Program at Lockheed, they actually forced you to go apply to grad school. So I applied to grad school, got into Stanford. To my shock, I had no idea that that was gonna happen. And then, Is that where you wanted to go? Was that like Yeah. I honestly, I yeah. I thought I've I thought it would be, you know, worth a shot. You know, it's a free call option. Why not throw your application into it? Like, I did well on, you know, the standardized tests and all that stuff. And then, I got in, but, like, just to show how little my, you know, immediate circle knew about Stanford, I told my mom I got into Stanford. She's like, is that in Connecticut?

And I was like, no, that's Stanford. But, you know, that's not the community that we grew up in. It's not a very where education is a big piece of my upbringing. And then so I got to Stanford and then, you know, was thrown it was twenty twelve, fall of twenty twelve, was the best time to you know, in, Fei Fei Li's computer vision lab. And then the AlexNet paper got published, and it was kind of like off to the races and been doing back, deep learning back prop stuff ever since. Yeah, that's amazing. I've got a very similar, but different background. So, spent time at JPL and came to the same realization. Working in government is too boring, or I should say there's too many limiters.

And, instead of applying to Stanford, I started doing stuff in crypto, which is completely different and then made my way back to ML. But the only experience I have at Stanford was CS231n. Yeah. And, that was probably one of the most pivotal, like online courses that I took, but like one of my entry points into machine learning. Yeah. So my view of that at the time, it was actually, I had done at the same time Andre was doing CS231n, I was doing CS231m with Silvio, Fei Fei's husband. And it was like, you know, I gotta respect Stanford for this because every other major university has a long lag from when things happen in the real world or academic world, and then when it actually makes its way into curriculum.

And to Stanford's credit, Fei Fei in particular allowed Justin and Andre, which Justin doesn't get a lot of credit for that class, he was co taught it and co created Andre went to Fei Fei and said like, this needs to be a class. This is like just we're still teaching bag of visual words. Like we're still teaching hog and sift descriptors. We need to like not be doing that. Like it's just completely different now. And Fei Fei saw it quickly, but I still remember very serious professors, even at Stanford, saying, Deep Learning maybe will work for computer vision, it's not gonna work for anything else. And then Richard Socher, who was a PhD at the time while I was creating that two thirty one ms class, wanted to do a class, the same type of class, but for deep learning for NLP.

And even Chris Manning was like, I mean, I'm not sure this is really worth a full dedicated class, but I agree we should teach it, but I'm just not the one wants to teach it. And so then Richard got the nod to go teach it. And then Richard asked me to co create it with a few others, and we created CS224D, and then two years later I found out they gutted all of the bag of words model stuff, like all the CBOW, all that stuff, and then replaced it all with our content. And so now the main MLP class at Stanford is still has my name on the courses like that. The cool thing that I observed was how quickly Stanford professors were willing to completely gut the curricula and then apply it with the latest greatest.

And like, You go to, I don't know how San Diego is, but University of Delaware, for example, I'd be shocked if they're teaching transformers yet. Totally. Yeah. That's the difference in education. You're just getting it so much quicker to the bleeding edge. Yeah. I think it's such a problem. I remember, when I was, so I went to university of, California, San Diego and my girlfriend at the time was hesitant to go because their engineering program was not accredited. And, you know, the, you know, for some engineers, you have to get a professional engineering degree. She was doing like mechanical engineering or something. I didn't care. Was gonna do computer science. I was like, just give me the most bleeding edge stuff. Right? But that contributes to some of the lag time. Right? It takes time to, like, roll these courses out.

By the time you you graduate, you're just, like, far behind. And I think, right now with, obviously, with AI and and all the things going on, like, if you're not at the cutting edge or even able to use, like, LMS to do your work, you're already, you know, kind of kneecapping yourself. Yeah. Yeah. A hundred percent. Do you believe like Stanford is continuing like to just be on the cutting edge and developing courses at a faster rate or at the same rate or, how are they adapting to change? I think that they are maintaining that by accident. I'm not sure that there's an overt effort to like, well, what's the bleeding edge? Let's make sure that we have the, correct curriculum there. I think the by accident is twofold. Number one, a lot of the industry leaders want to be teaching courses at Stanford.

And so there was that like Coachella lineup of this frontier systems or whatever that everyone saw on Twitter. Like, you're getting like Satya and Sam and like all these other people to come and teach at Stanford. And so like, granted, there's that. And then the other side of this is that every single Stanford CS professor has a company that has raised over three hundred million dollars. Every single one. Like, I don't think that there's an exception at this point. It was like an extreme anomaly, when in twenty twelve where you had Daphne and Andrew, and Sebastian were the only ones that had companies. I might be forgetting some, but Fei Fei didn't, Silvio didn't, Yuri Leskovic just got there. He didn't have a company yet.

It wasn't a thing. And then now it is fully a thing. And the and there's a there's a pro and con to this. One, the the con is that you get this reverse classroom issue, which is like they you show up to class and just and just ask questions. And like, I would say that that's a really bad thing that I went to your Leskovits class to CS229G, I think it's what it's called like for graph neural networks. So yeah, don't know anything about graph neural networks. Let me go there and I'll learn. And then he showed up to the first lecture, didn't show up, sent his grad students for the rest of them. It's like, that's not so good, But the content and like, he is great because it's completely from state of the art and he's actually using it and deploying it and so much so that it's worth, it's worth having a company and that company can raise.

James Zhu just raised like a billion dollar evaluation doing autonomous research. Is he a professor? Chris Wray, obviously, is very entrepreneurial, has done a bunch of companies. You get from the bleeding edge, Chelsea Finn, obviously, physical intelligence. The con is that you get them less involved and they're focused on their startups because running a startup is not a part time job. And being a professor is, it could be a part time job if you want it to be, but it probably should be a full time job. And so that's the pro and the con. I would think on the margin, if I had my choice, would take, give me the bleeding edge. I want my PI to be like, you know, on the bleeding edge. To publish in Europe, you have to have novelty.

You can't just copy someone else. And in startups, copying is like recommended. In fact, pioneers get errors in the back. It's much more likely, you know, what did Instacart do novel to the world? The web van came around way before them and invented a lot of things. And then like DoorDash after Instacart and then Wrappy after DoorDash and Instacart. I mean, we were in the Pet Class of Rappi, Winter sixteen, at NYC, and Simon said that got on stage and said, we're the Instacart of LATAM. And I was like, why isn't Instacart the Instacart of LATAM? Right. Right. You know? Turns out Instacart can't be the Instacart of LatAm because they don't understand lat LatAm. And so it requires someone from Colombia to be the Instacart of LatAm. Yeah.

We we have a we have quite quite a few, Colombian team members, and, like, that's the first time I learned about RAPI. Went to visit them, and they're like, I I use RAPI. I was like, where's Instacart? Right? That's so funny. Yeah. Yeah. It's it's fascinating. Think, yeah, I think PG said this, but it's, you know, get as many touch points with reality as you can, you know, and that seems to be kind of like the differentiator and some of, like, the the YC ethos. Let's talk about a bit of what you're doing at YC, your visiting partner, how long have you been there? What are some of the trends that you're seeing today, especially as we've gone from ChatGPT to rappers, to agents, to what's next? After nine years in the cockpit at Focal, which was about seventy years in startup life, I wanted to get to PG's Default Alive.

Then we signed this sixty million dollars deal. After that, was like, I'm done. We're default alive. I hit my goal. And I took a step back. And I was kind of thinking what I wanted to do next. And so my options were I could go start another company, which I actually had a lot of pressure to. A lot of the initial investors in Focal wanted me to let me know when you have the next company. I'll write your term sheet now. Presuming that was going do it, I obviously had a bunch of ideas. I had some employees that were like, hey, tell me, I know you're leaving to start another company. What do you want to do? I'm down to leave. So there was that going on. And then my wife said, You have to take at least a year.

And then so, What am I going to do for a year? I just wanted to get really back deep into research. I forgot how a screen session worked, how do I get in and out of a screen session, how does TMUX work, how do I Docker compose. I forgot all these things. I haven't done it in so long. And so I just wanted to get back technical. And I left the deep learning field, at least from a hands on standpoint, when training on one ten eighty or V100 was a big deal. And then now you have multiple GPUs, you have multiple nodes, you have multiple data centers. And it was like, you know, I don't know how to do that. And so I wanted to learn torch distributed and stuff like that. So I went back and started doing research. And then I talked to my professor, my professors were like, you should apply back in.

And so I was kind of like, Okay, that's kind of an interesting idea. So I applied back in, got into the CS PhD program, and then have been doing research there. And then I caught up with Jared Freeman, who is my partner at YC, just catching up and showing him my Arc AGI results. I was like, look, loss curves going down. This is kinda cool. Then on Fridays, I would do investments. I invested in a couple of YC startups, I did some demo days, something like that. And he's like, know, Francois, we're trying to deeper into AI. You seem to be wanting to help startups. Why don't you just do that at YC? Would you be interested in that? I'd be like, I didn't even know that was an option. I mean, that's super fun. That would be like a dream job.

And so, put together a little slide deck for Gary, presented to him of some ideas on how to nudge YC deeper into the AI community, and he was like, accept all changes. So I was like, okay. So you're using G Stack, right? Exactly, exactly. So then that was six, seven, or probably eight months ago now, and then now I've been at YC, I think, since October last year. We did my first batch with Jared. That was super fun, and it's just been, we focus, obviously, a lot more on the AI stack and, like, what's coming. And, you know, we have our points of view. And it's been, like, you know, a dream come true. It's been super, super fun to help. You know, I love YC.

I love what YC stands for. We give opportunities to folks that just no one else would have funded. No one would have funded Airbnb. If PG didn't fund Brian Chesky, they literally said, We're just going to go back to Rhode Rhode Island and start designing some PowerPoints or something. You know? Yeah. We wouldn't have Airbnb. Yeah. And so there's so many companies and examples like that. So it's just super fun to be helping, you know, these companies, these great founders. And generally social mobility. Yeah, for sure. Yeah, you talk a lot about how YC has just increased social mobility, right? Yeah, I think compared to what other company has done more for social mobility, I would have said that before I joined back in, but just watching it firsthand.

Take Simon Adrapi, for example. He created a multi billion dollar company from an idea that he had living in Colombia. I mean, if he could have funded that elsewhere, he would have. And so, enables these companies to pop up and contribute meaningfully to job growth, meaningfully to GDP, and that's what's required. I'm curious with Simon's story, how did he hear about YC? Like, during his interview, like, know, could have gotten funding in presumably Colombia or, you know, there's My co founder was Bolivian, and so they were close. And we would always grab dinners and stuff like that. You know, I just recant what I heard, but he had very little investment.

This is not his first company. Think he did like two or three before this. And I think he applied to YC like three to four times or something like that. And he just had little bits of funding to keep him going, like twenty ks, fifty ks here and there. And then we got onto YC. He thought, I finally made it. Got on stage. He did not fundraise well. It was not like, Oh, here's thirteen million dollars from Andreessen. It was of crickets for the same reason, because we were all on the Peter Thielism, The Something of Somewhere is the Nothing of Nowhere. So that was like, again, everyone was thinking that same mindset, which is like the Instacart of Columbia should be Instacart. Turns out, no. And then Rappi just exploded. Simon is just a great entrepreneur, and then he started getting and then he finally got Sequoia to invest meaningfully, and I think Andrew did that investment, and then they just took off from there.

Yeah, but yeah, just an incredible entrepreneur that but again, don't think Sequoia it would have been on Sequoia's radar if he was based in Cartagena, and he was just like building it from Medellin or something, you know. I'm not sure There's no way it would have been on Square's radar. I guess just because we have so many team members and a presence in Columbia, there's so many entrepreneurial folks down there, think they've got an AI tinkerers down there and a lot of great ML talent. Think Andrew Ng taught down there, something down there. He he did work with Landing AI. What advice do you have for folks that feel like YC is maybe not accessible to them because they're in South America or some other part of the world?

We are actively trying to get more of the world involved with YC. And so, explicitly, just in the last two weeks, Jared, Ankit, and Harshita, I think, just went to India and did a two thousand person India startup school over there, and trying to get more and more entrepreneurship participation all around the world. And so, we do get a lot of applications. Think two thirds of YC is international at this point, you know. We do fund, people from all over. I think that with there's two things I'd probably say: Number one, I don't think people even know like my mom not knowing about where Stanford was, you know, my parents didn't know who Warren Buffett was, Right?

So, even knowing about YC is step number one. Even knowing that's a path to go. And that's what we definitely want to promote. And you really don't get there unless you show up. And that's why we came to San Diego. That's why we're doing a lot of these college tours. It's amazing how few people think that, Oh, I could start a company. And it's like, Or I should just go work at Google and be another cognitive machine. Or even, worse, Lockheed Martin. And the second thing I'll say is, given this advent of Claude Code and vibe coding and things like that, it brings the activation energy to start a company way down. And so back to the something of somewhere, there's the nothing of nowhere. The reason why Instacart couldn't have been the Instacart again for LATAM is because there's so much bespoke knowledge on how the world works, how drivers would think, how the customers think, how to price, what are really things that are important to to do mopeds.

That's what they do. They have mopeds with these, like, you know, backpack things. Is that like massive backpacks? Massive backpacks that are like orange and like, you know, very, very well branded with the little must the wrappy mustache. To even think to do that, like, unless you grew up down there, you wouldn't know that that was the way to do it. And so now there can be so that's geographic segmentation of Uber, which there's like there was a Middle Eastern Uber, there's a bunch of other Ubers, there's a New York Uber, there's like you know, there's so many different like that had to be geographically distributed. Same thing with with delivery, and even in New York, there's a seamless we've had seamless in New York forever.

That geographic distribution, it allows even more fragmentation because anyone else can start with local knowledge and just do Epsilon better locally for that local community, and your startup capital is much less. So understanding local, dynamics allows much better product market fit, is what I mean. And then now you can it would be implausible to do that to make a better you know, this is a dumb example, but like, it might make sense to build a better search engine or a better Yelp for every single country. I don't know. And that was impossible because the amount of software engineering you had to do to build Yelp and branding would be very difficult.

And so the activation energy doesn't make But now, because it's so much lower, now you can make a Yelp in Bulgaria, and it's bespoke to the Bulgarian community, which I know nothing about. You know, so anyway, so that's my rant on startups internationally. Go for it. Now is the best time, and it's going to be one of those things where the first mover advantage is going to matter a lot. So it sounds like, take an idea that works in some other country, and because the activation energy to go and build it using Cloud Code or Codex is so low, just go and build it, go to market, use your local knowledge or cultural knowledge to just see if that's a viable option, right? And novelty is not important in startups.

In fact, it's actually a con, not a pro. Like, it's because it's unproven. And and even if you prove it, you already got so war torn and try all these other things and how to learn so much that you're like, you already raised so much money. Your cost of capital is gonna be higher now, whatever. And so, like, you just copying what works in America and doing the rocket what what is that? The the the rocket and ventures guys in in Germany that just copy all the American companies that do it in Germany. Yeah. Like, just do that. Yeah. Like, honestly, I I'm not necessarily thinking it's a good thing, but it's a great business. And this is back to the idea that we started on is academics suck at this because they wouldn't do it. Well, They'll never do it. That's what was in my mind.

It's like, they wanna be novel, right? Like, you do research, you you wanna figure out something new about the world, you know? And so taking an existing idea and and and trying to kinda, like, retrofit it for a new culture feels maybe not novel. Yeah. And so that's that's the copying. And then even for the novelty ideas, like we we have I I do this table prompt for most dinners that you'll have with me, with and my wife hates it, but it's actually been quite fun. We just did a, like, surf trip in Sayulita, and we forced all of our friends, and before we get our appetizers, whip out Lovable and come up with a startup idea, and it has to be built by the end of dinner. And then we have to, like, vote on it, and whoever whoever wins, we're like, try to all try it out or whatever it is.

And so one of the girls is super deep into astrology and way deep, more deep than you'd expect. I'm sure you've lived in LA, so you have some people who are super I'm a Pisces, theorizing, something. Exactly right. Deep into this stuff. Then she has all this knowledge and used the right terms in the Lovable thing and then had this serious math equation that Lovable came up with to see your compatibility with your partner and gave it a score. Then, every word is going to lead to three divorces or something like It was kind of cool. It had these reasons, like here's the friction points that you're gonna have in your marriage.

It's actually really true. Weirdly, it's really true. But would people pay a dollar a month or a ten dollars one time thing for that? There might be a real business there and she built it in one hour with Lovable. And so, like, is there a bunch of these, like, non venture scale businesses that will pop up that are certainly lifestyle businesses, but they might make three million a year? That's pretty good. I'll take with three million a year. Yeah. That's fine. Right? What's your cogs? Zero. Right. Right. You know? So What's changed in terms of YC's, like, investment thesis now that the bar to build is so low? Is is the bar higher on what you build, or is it more important to get distribution quickly, to go to market? Is it still just betting on founders?

Like, has anything fundamentally changed? Gary's definitely not AI pilled at all. No. Mean, we're definitely probably the most AI pilled, of the venture community. I think we're, like, every single partner is every single partner is very technical. Every single partner has is an ex founder. Every single partner is an x y c founder. And so that's a very unique group of people. And so everyone has their claw like, open claws set up. After you guys all wore all wore your claws. Just thought that was so funny. That's why I love YC, because it's like, I wouldn't say go ahead and do that, I don't think so. Maybe it'll be a little bit too strict, but it's just very funny like that.

So anyway, I think that the aperture of what's fundable is wider. I don't think YC does not fund ideas, generally, because even I had this experience. It was actually crazy, but out of the thirty one companies we funded, maybe three or four of them pivoted before they even got to batch. So January sixth, they start, we fund them December, in a month they've already pivoted, right? And some of them already had revenue on the new idea, which was just like Odd Voice did. That was an insane story we can get into. What we have always funded and now it's even more true is what Gary calls agency and taste.

And so are you a person who's going to say, Well, we tried Or is it like, I tried six hundred Maybe Ah! Voice is a great example of this because I love those guys. There's two Thai brothers. We funded them. They were a voice agent company. And we funded them in December fifteenth. They show up on January sixth and have office hours with them and they say, So we pitted ideas to do Harvey for architecture. And I was like, Okay, that's an interesting idea. Do you know any architects? So I was like, we already have like six architects paying us like a thousand a month. Like, What? That happened like three weeks ago. Yeah. And so that's an example of an insane agency. They not only pivoted their idea, then they built it, then they got customers, went and talked to them, got them the pay, enough to actually pay.

I mean, that's the type of agency that YC loves. And so, there's so many people, especially the academics, will sit in their corner, they'll code, code, code, keep working on some benchmark and never actually release anything. And that is the antithesis of YC's ethos. And so the cursor just talked to this podcast that I just loved. It was like we shipped our initial version four months after launch. We had a big jump. Then we had the trough of disillusionment, the valley of despair, etc. Everyone turned off of cursor. But they were building it themselves so that they could dog food it and try it themselves, which is such an important thing. Most times, can actually dog food even if you're building software for grocery stores, we could dog food.

We actually built a grocery store in our office so that we could dog food. Argument that you can't ever dog food because it's like you're building something for a new vertical, it's like you almost always can dog food. And then because they were dogfooding, they really built it out. They were wondering if they should pivot to bug fixing only and doing a sentry bot kind of thing. And then they just kept working on the thing that they wanted until it got better and better. Until it got good. That's the thing that most people think that they're too smart for that. I can just skip it. I know what's good. I'll just build it myself because I'm Steve Jobs. No, you're not. Steve Jobs is Steve Jobs because he has failed three thousand times by launching it out.

He's built a mental model of what's going to stick and what's not. You haven't. He wasn't that good when he was twenty one. He shipped a lot of products, the Newton and all this other stuff that completely flopped. And the last thing I'll say on the cursor side that is certainly within the YC ethos is he said, We knew that every ten things that we were going to launch, one of them was going to work. And back to the PM stuff we were talking about for the pod, PMs that work at big, slow organizations, you get fired if you launch nine products and they all failed, but that's exactly what you need to do. I would fire you if you didn't launch ten things in a month, and if nine of them failed, if I looked at them, I'm like, Those are pretty good ideas, but I would have thought they would have worked, But that's how bad of a predictor people are and how good products are going to do when they actually get out to market.

Yeah, that's such a strong intuition. One out of ten are going work, you just keep chasing that. But I think with cursor, and maybe you felt this too, you could tell that it was being used by the team. Yeah. You know? And you could just read, like, I think, Armand's tweets, you know, and he's he's super technical. And every so often, he would post, like, this long thread on, like, you know, how they pivoted from, like, semantic indexing of the code base to GLOB and GREP and how that worked better, or whatever the case is, right? And you could just feel the passion of the founders and that they were using it and getting it and that was fueling the growth. Yeah, back to the academic versus the founder mentality. I'm just obsessed with making coding easier.

I'm just going to keep working on that. You're going to able to work on that for one hundred years. It's never going to be solved. Versus we have this new patent on indexing code, and then we're going to commercialize it, whatever, and it's like, no one cares about all that. The biggest innovation in Cursor that is now worth sixty billion dollars was the user interface. It was like actually making the UX the way it was. And that was the main innovation that I can arguably say I didn't see before Cursor. Every single other we had Sourcegraph, we knew how to index code bases, That company has been around for ten years. There's all these other pieces that were there. Auto complete by hitting tab, we've had in IDEs for a long time.

And so Cursor Tab wasn't that innovative. I think showing the user experience, but all the things that were right, it had to be fast. Had not just to be good, but it had to be fast. It had to adapt. They had to push the next thing, the next thing, the next thing. And that's what makes great companies in that mentality, not the look at my black box. I remember people freak out when the UX would change frequently on Cursor. They'd be like, my god, I'm so used to this, now it's moved. And they would listen. And that takes courage too. For you to change a design pattern that is already working, that takes product courage. One of the things I love about Boris on Claude Code is he's constantly thinking about cutting products. He's constantly cutting slash commands.

And he's like, Yeah, I think we're going to cut that one. But it has usage. And it's like, Yeah, but it's not enough. Did you get a rare buddy or I didn't get a buddy. No, I actually tried. It didn't work. Yeah. Yeah. What do you use, day to day? I use Cloud Code pretty much, exclusively with Gemini Pro as my last step before they come back to me. I'm gonna say before you come back, like, I have this big fork of Karpathy's Auto Research. And then I call it Auto Research with Good Ideas. Since, like, the way that his basically comes up with the next idea when you do a new branch is like, just come up with another idea. That is not the way that research kind of works.

It's like every single time you finish one experiment, you have zero to N new ideas. And you should always be appending that to some queue, rank sorted by EV, expected value, and the return, right? And so like, you should main so I have this separate thing called an ideas. Md that I am constantly running a priority queue over. And then at the end of that, you run this whole reflection period where you're kind of thinking about what did I just learn? And and so like that. And so and so the Gemini part helps with that reflection. So, like, doing chain of thought on, like, your your your prioritization. And look through the results. And, like and then and then that will lead to a bunch of one off scripts that you might run and be like, okay.

Well, like, here's a here's a thing that happened, and then here's another thing that happened. Here's another thing that happened. What's the rule? Oh, ****. Woah. I actually found the rule. That's what research is. And it's like, oh, cool. Well, like, I just figured something out that, like, wasn't before known. And so that's what I'm trying to get the models to do. And I do think, to Gary's credit with G Stack, there's a lot of juice to squeeze with a harness. And I think imbuing the human processes of how we do research into it. James Zhu is talking a lot about this. He's kind of been leading that charge at Stanford. Chris Wray is doing a lot with this. And I've been toying around with it a lot as well.

But I think we're going to get a lot of scientific research done just by imbuing these scientific research processes, around with a harness, essentially. Yeah. Something something that, I've been kind of thinking about, is, you know, given, you know, a harness or, you know, FMLLM and a very, you know, procedural, like, SOP or, you know, just steps to get some goal accomplished and, perfect context or access to, like, tools and APIs. Like, what is insoluble in the software realm with a harness and an agent? There's definitely classes for intelligence that, like, in complexity, like automata theory, that we have these nicely formed classes of PNP, NP hard, very nice clean classes.

We don't really have that in intelligence. I think we're certainly bounded by the bounds of, let's say, the interpolation of human knowledge. We're like, let's say, a lot of innovations came from going cross discipline by merging a neuroscientist and a computer scientist, like, you know, whatever, there's a lot of, new knowledge to be created by calling too big spikes, right? And so I think that like, that's the way I think about knowledge is like spiky ball. And like, and there's an innovation to be had by sampling from both and then finding some linear interpolation between the two to, you know, form this like beach ball, let's call it. And getting outside that beach ball, I don't think you can do. And so I think that that's where unless you can figure out how to get them get these models to think in a much deeper way, where like, Francois Chaule talks a lot about the fluid intelligence of these models.

Basically, there's almost none. There's no fluid intelligence in these models. But when you say outside of the ball, you mean outside of like the training distribution, or the samples? Yeah, thinking about it in terms of train and test distributions or out of distribution is usually unhelpful because it's like, well, it's the full world. It's like all of common crawlers, what's outside of the distribution? And so, it's kind of hard to think about, but it's evident when you give it a new task, how quickly it can adapt to that new task. And I've been trying to harness my way to win RKGI three, and you actually can't do it. It's like, it's very, very difficult to do because the pattern if you look at the way that these things learn, and you put all the way that these things learn in one column, and look all the ways that we learn in another column, you just realize we're like way far.

And so, here's an example. I give you a new game. It's called Francois Go. It's a new Go version, but like I changed all the rules. It's like, how quickly will you learn it? How quickly will the agent learn it? How could it learn it? You would learn intra game, like one, you would move, and then I would take your piece like, oh, I get it. Like, you'd have that epiphany. So I call that batch size one learning. If you try to update weights with batch size one, the gradients will be massively unstable, and you'll get NaNs very quickly. It just will not work. So batch size one cannot work. And so then you, oh, well, we'd have to like increase the batch size, critical batch size is like twenty, forty eight, or whatever.

So I need you to think about two thousand, you have to load up in your head to be equivalent to that. Two thousand and forty eight other experiences in your head all at the same time while you're learning. There's no chance we're doing that. It's just not even close. We learn stably and meaningfully with batch size one. And that is very human. But I think that that's the key. There's something very different that we're doing. And so the counter example that people give is, well, just throw it into context, in context learn it. And the thing that no one knows, if you grab a random person at NeurIPS and ask them, what do you think is going to happen if I give one experience into context, the second experience into context, the third experience into context, the fourth, fifth, whatever, hundredth experience into context.

It doesn't actually monotonically improve. It actually improves, and then it starts falling, falling, falling, and then it gets no performance because you've exceeded the context window. And so, does horribly. And also, it wasn't trained to consistently improve as I increase the context, as I increase the number of examples, as I ICL. So from one, two, or three examples, it wasn't learned to do that. It hasn't learned to do that at train time. And so there's recursive self improvement people that are trying to get it to learn to do it, but it wasn't learned to do that at pre trained. And so, like, it's inherently in the model to not be able to do this extremely well. It's amazing that it works as well as it does. But for many real examples, it doesn't do that.

So we do. How? Yeah. And the more I play chess, the better I do. And like and it's monotonically improving, which isn't monotonic, maybe is a fair argument. So what what's like again, this is kinda hitting my limits of of, you know, what I know in this space, but I have been listening to the same podcast, like, every time because it's down it's the only thing downloaded on my phone for some reason. It's on the flights. Yeah. There's this Torcash episode with Ilya, and he talks about the, like, the value function. Right? Is this somewhere to kinda, like, how we get to, like, batch size one? Or I think it's a training procedure issue. Yeah. Like, yeah, I agree that there's a value function where we have that's a target. That isn't a update rule. And so the actual update rule, if you're using Robins Monroe setup, which is like SGD is an implementation of Robins Monroe.

There's many different implementations of of Robins Monroe, but this is like your stochastic gradient descent method is one way to achieve that. For those that don't know, Robins Monroe, is that hill climbing generalized or? It's generally just a recursive update rule where I'm stepping on some estimate of the gradient times the learning rate to update the weight and hill climb into some local optima. And so that's historically nineteen fifty nine Robins Monroe framework, and then we've been working off of different versions of that for the last seventy years to find the right procedure. Like, I work on a lot of zero order stuff and a lot of perturbations off of pure vanilla SGD to try to find an update rule.

But one of the things is clear to me. Number one, we are certainly not doing back prop. It is very difficult to see how I'm doing the transpose of the weight matrix in my head. And it's very difficult to do back prop through time because I would mean I would need like C versions of my brain. And it's it's just where do the bits go? Right. That information cannot exist in in the brain. See see copies of the activations and then doing back prop through it. We would see it in, some type of, like, MRI plots and things like that. And so it's not there. The since nineteen fifty nine, Hebbian David Hebbes, you know, we've had Hebbian learning rule, and we know it looks much more like that, which, the the more proper version that is bounded, Hebbes is unbounded, it's the, learning rate times x times y, it's you have to subtract some normalization term, which actually turns it into online PCA, and that's OHA's rule.

And OHA's rule is actually quite stable and it can update. The problem with OHA's rule from my experimentation and why we can't replace it replace gradient back prop with OHA's rule. It's a very weak learner. And so, like, the amount of generalization you get from one example is quite small. An online PCA with principal component one is actually just very weak. And so, like, you just the sample efficiency you get from it is just not very strong if you just take exactly problems in a row and apply it. But there are other learning procedures that are that look and smell a lot like Oha's rule that are gonna get us closer and closer to closing that gap. And that's the stuff that I generally work on. Yeah. Maybe maybe to talk a little bit more about what you're working on and and some of, if you can talk about your current research and what you're working on for NeurIPS.

Yeah. I think that that'd be interesting to dive into. Yeah. So I think, there's two, research, angles I work on. One is the the batch size one learning stuff. And so instead of doing ICL, there's some other cool ideas that we could actually take inspiration from. And so there's a paper out of my lab called cartridges. And so it's basically this, if you have like a big code base that doesn't change that much, and let's say, or maybe you have a bunch of experiences, and you want to compress all that in kV space in the time space into a much smaller, compressed version of memory, let's say, a cartridge, you can do so with, like, some procedure of self study reflection.

And if I think about what I'm doing, there is that kind of spirit of like, let me reflect on what's been going on. And this is what Sabri calls self study. Is this similar to the Google's recent paper on like KV compression and Yeah. It's in the spirit, but it's very unique. This cartridges paper is very unique in that it updates it learns it starts from a random context. It's literally around numbers, and then I I freeze the model, and then I learn my my context in continuous space. Learning a context is very strange. It's like, whatever. But I'm using SGD to to learn a context and it actually works. And Sonata says k v here. Yeah. Yeah. And so I'm learning what k v would maximize.

So I what over all possible k v's, what KV, if I made it this small, would perform well on this self study objective freezing my model? And that's what it basically learns. Well, that's pretty cool. If now if you can do that in an online way, that would be really cool. It's an offline algorithm. And maybe this is like I'm sleeping and I'm compressing my KV cache into this little cartridge. So there's stuff like that that are pretty interesting. And my idea is to, in the same spirit, get rid of that idea because it's very expensive. And what you can learn is a little adapter that, and there's something called recursively squares, where you are mapping from where you're outputting now to your current output distribution, you freeze the model, and you just learn this little W adapter that's steering more towards some new local thing.

And this is like test time learning is the spirit of things that Like similar to LoRa or Yeah. So like, Laura would be an example of actually touching the weights. Instead of doing that, I spawn a little adapter, a little small weight matrix w, and then my output head would output these logits, and I'll output, this w, and I'll and my actual output will be like my original output plus alpha w, that original output. And I'll I'll steer the distribution of my output logits just a little bit towards some direction. And as I get more and more training examples, can quickly learn a w matrix with recursively squares with literally order one every single time I get one example.

And with order one training, I can, like, get meaningfully better performance than just ICL. And that is for saves on compute and all these great things. So that's like research direction one. The other one that I think is like, you know, going, is my Hail Mary that I've been working on for two years that everyone deep learning says, why are you working on this? Go work on, you know, back prop. And is that there has to exist a learning procedure that is not the transpose of the weight matrix. Back prop is extremely expensive. I have to store the activations. These are all things that the brain is not capable of doing. And so, there's a better learning procedure. And then some people will actually argue, Well, I think that if evolution could have discovered backprop, then it would have preferred to find backprop.

And I say, Hell no. And so there's a lot of loss functions where backprop, which is the local infinitesimally small local gradient, gives me the direction to step from this infinitesimally small perturbation. There's a lot of loss functions like, for example, a zero one. If it's zero and one, there's no local information. The only information that you can observe is by massive perturbation, and then I know which direction to go, and I have to get lucky. I have to find a sample from the zero, I have to sample from the one. And that is almost all of RL. Whether I won the game or not, whether I got eaten by a lion or not, whether I met my wife and reproduced or not, and the the loss function of life is extremely zero one.

And whether or not you reproduce and you pass on your DNA before getting eaten by a lion, that is not a differentiable loss function. It just isn't. And so, I just take huge pause from that. And that's where I think other learning procedures will be better. They don't require activation retention. They don't require an absolute gradient estimation, and they actually work better in a lot of high dimensional lost landscapes. And so that's what I'm trying on the hunt to find is one that actually scales to multi billion, a hundred billion, eight hundred billion parameters like the size of your brain. And I think we're getting there. And so Yeah. I can give you some of that.

Do do you, you know, some some will say, and and I'll play a bit of devil's advocate here that, you know, the the way to get to to to better performing models or just closer to whatever we call AGI SI isn't will not stem from, like, how humans do it. Right? Or, like, we think a lot of a lot of, like, kind of your statements are anthropomorphic in nature. Like, you know, the brain does not do back prop. Right. Do do you believe that, you know, the solution you're looking for will stem from, you know, biology or at least be influenced by by biology? I think the the the yon lacoonism on this is is to map the conversation to flight. And like, if you knew a map reducibility to flight, where like, did we need flapping wings to fly?

No. Did we need two wings? Yeah. The Wright brothers needed two wings. Yeah. And so like, taking looking at all the things that fly in the world, we're trying to, you know, figure out how do we achieve flight, and ignoring all of it and going off in another direction is probably a bad idea. You should take some inspiration. So to say it specifically, to be academic in my statement is the set of all things that can achieve intelligence is certainly bigger than us. And the learning procedures, the way it looks and smells, the way it learns, the way it operates is going to be very different in the way that a rocket and a helicopter and a jet and a turbine and a turboprop, like the way that they achieve flight are all very different and a drone, right?

Those are all different ways of achieving flight. But I wouldn't take a drone to Cartagena, I wouldn't take a helicopter to Cartagena, and I probably wouldn't take a rocket to San Francisco. So like, there's certain things that will be better for certain, situations. And so LLMs are are going to be better in the limit. LLMs will be better than us on specific things, and they already are. But in terms of fluid intelligence, we still are far superior. And if you took a LLM and, you know, tried to play a brand new game against it, as the Arc AGI three benchmark shows, humans just destroy it.

Like, the the Ray bench benchmark, R H A E You want a Ray test? It's the the the metric for Arc AGI three. I mean, the best I've been able to achieve is point o two, which means that humans are five hundred no. Fifty times more efficient in terms of number of actions to get to a specific level. So 50x. And most of the time, it can't even solve it. Like, it actually never solves it. So like, in terms of fluid intel, and these are these are games where like a random, you know, Craigslist responded for in a fifteen dollars ad and then beat the LLM. Yeah. That's this is like still there's still glaring holes. And the other problem that I've worked on a lot is funny.

Try to make an LLM funny. I actually realized it's it's by definition impossible. Yeah. And then people like scoff, like, can you know that? The way that if you take a Mitch Hedberg joke, I think the joke is, my fake plant died because I forgot to pretend to water it. The token pretend has one in two in one thousandth probability. And if I'm doing your top p sampling, it will never be sampled. It will never be sampled. And so then it will never be funny. And the only way things are funny is high predictability, complete, like, con like like shock, surprise.

Yeah. And then it recovers. And that's what every single one of Mitch Hedberg's jokes look like. It's massively high probable, very low probable, very high probable. And neural sample nucleus sampling will never sample that trajectory, so it will never be funny. Even with the right system problems and the right You can't get there. You like, it's how will you sample Yeah. The funny one that will lead to it? So I've worked on look doing some type of beam search to try to get that exact shape. It just, like, doesn't work. Were you the one who tweeted, there's only, like, two to three joke like, funny jokes that an LM can No. I did tweet this whole phenomenon that like, you know, this little viral thing that went out there.

But you can't it's you need a new sampling method. That's at minimum required. But I think it's because we think in a higher level than a token space, that this idea space that we can't train, and then it would make more sense. And what you kinda wanna find is this contrastive thing of like, what's the minimum change in the sentence that will meaningfully change the semantic? And we those things can't do it, but we can do it. That's what joke writing is. And I actually think that that's it's a deeper meaning, because it's actually tells us about why LMs can't write good movie scripts, good poems, you know, that are really deep and profound and meaningful and hit us hard, and maybe do great research.

Yeah. I don't know. That is fascinating. We were just talking about, you know, you can spin up all these grad students, right? Agents. And you're really just limited by compute. And you, you've been working on this project of, you know, mapping out the, number of GPUs per student. Right? And, you know, that seems to me like it's becoming more and more important, right? If you wanna go and do research or go learn things, you need to have compute, right? And why would you go to university that doesn't give you compute? Why would you go work at a job that doesn't give you compute? One hundred percent. I mean, I think, take Stanford for example, I there's like four thousand CS students across undergrad, grad, PhD, and they just launched Marlowe, which is two fifty H100s.

If you start at OpenAI, you're given that as an intern. That's for the entire campus. It's embarrassing. We have a forty billion dollars endowment. We can't give a billion dollars to NVIDIA to get a real cluster. And so, like, I started, like, wrap and then we had this dinner with a bunch of professors at NeurIPS. And I was just asking, like and I experienced this firsthand. Luckily, I had some an exit at Focal that I can actually, like, spend on GPUs. If I didn't, I wouldn't be able to even do any of my ideas. And so like a lot of great minds in my lab, are unable to get compute, and I realized that like that's the limiting step. And so then I started like down this rabbit hole, and I asking them like, hey, where are you guys at on compute?

I realized this isn't a Stanford problem. This is, an academic problem. And it's true in all of academia. So I just, like, you know, grabbed I I went home that night at, like, eleven o'clock, grabbed, like, six universities, Princeton, Harvard, whatever, and just took like what I could find online for the h h one hundreds or equivalent that they had access to and divided by the number of students, and just posted it online. And then I heard someone yelled at you. Yeah. I got Yan Lakun yelled at me and said, this is total BS, which I inferred meant total batch size. And then because NYU wasn't there, and so then I added NYU to the plot, and it was actually, like, in the bottom. Yeah. Because they have yeah. They do have a thousand GPUs, but they also have six thousand, seven thousand student c c s students across Tandon and Kuran. Yeah.

How do we bring more awareness of this? Because I think it's important. Right? And Yes. As a student, like, if I were to go back to to grad school, like, that's probably one of the first things I would look for is do I have enough compute to run experiments? Because I can't afford, you know, thirty thousand dollars of, anthropic bills, and I probably want to play with some open source models anyways, and host them and run them and so forth. Yeah, I think that, like, in society, you know, we pay a lot in tax in America, and spending a lot of money on a lot of stuff. And if you think about like, what's going to cause the greatest amount of good, what's, you know, going to I think AI is largely a cause for good and can cure a lot of disease. If you read the EVO stuff, I think EVO has the capabilities of like, solving all disease.

They literally can't get funding to like go beyond an eighty bill model. And if you went to an eight hundred bill model, like maybe all disease would be solved. Is that worth it for society? Can you, what is EVO for? EVO is basically taking your LLM stack of like predicting the next token, and then replacing it with DNA, and then predicting the next base pair. Okay. And so your ACTG, predict the next one, predict the next one. And so then from that, you get these rich embeddings, like, you know, pre training on DNA. Why is that useful? Well, it's not. But why it's useful is that the embeddings it produces, you're literally in this rich embedding space that, you can train a small little soup con contrastive learning thing to if I have like ten people with a disease, I can literally detect that disease now forever with like ninety nine percent accuracy, with just the DNA.

And like, then you identify, well, we actually can solve this with a CRISPR, one CRISPR edit, a bunch of these major diseases, especially rare diseases that have like no research. And so, it's insane, the results out of this. But again, we're limited, we're not limited by ideas, we're largely limited by compute. So how do we solve that problem in academia? Endowments exist. Yes. We tend to you know, we have funding. And it's you know, there might be an issue in mapping from what ideas we should do back to the research good auto research with good ideas to compute, but largely the issue is we just need more compute. And like the the all name a problem in the world, like the probability of that problem getting solved if we add more compute towards the problem is like ninety nine percent.

Yeah. And and and, like, when you say compute here, you're you're largely talking about, like, GPUs. Right? Yeah. Like NVIDIA, for example, not flake energy. I tried to train my models on CPUs, but NVIDIA Jensen does a good job. Yeah. It's hard to compete. I won't get into the the Dwarkash pod and, you know, some of those things. I didn't wake up a loser. Yeah. I did not wake up a loser. Exactly. I say that every day in the mirror. That's funny. I wanna end the conversation on, just focusing on, on young founders, or young potential entrepreneurs and what you believe they should be focused on what they should be.

Right. And what AI has has opened up, you know, for them to go and experiment and try things and build the next generation companies. What do those companies look like? Right. I think, you know, the term product is kind of ephemeral nowadays because we can build so quickly and we can build just so many features for so many people. Right? So like what is product, as well as like services is sort of, overloaded term these days. And we're sort of seeing the mesh of these two, but, you know, if you were, going to start a company, like what would you focus on? As a mentor, what would you advise? Yeah, I would say the first thing is do a company. Like, I think that's the larger step that I wish more people would consider as an occupation, the people that are successful entrepreneurs, like, you know, I wouldn't call them in the top ten, you know, top ten percent of the intelligence of people that I've met, and they're wildly successful.

And so like, you know, the, the, the, the skills that you need to be an entrepreneur, one of them is not to be Einstein. I think Einstein would actually be a very bad entrepreneur. You know? And so, people think that you have to or you had to come from the right background or something like that, or you had to go to Stanford or whatever. It's all hogwash. The best entrepreneurs don't have, like, traditional backgrounds at all. So I'd say that's my biggest advice is, like, go do it. Like, have the cojones to go and do it, and don't take no for an answer. And then once you take that plunge, you know, I think you should probably take advice. There's no shortage of advice in the world out there. And if you want to get to the moon, like the only real people that you should talk to are people that have been to the moon and like listen to astronauts, like, and every single other part, your mom, your dad, your parents, unless they've been to the moon, weight their advice to zero.

Literally not like zero point one, zero. And it's hard for people to do that. My mom, know, once I had my job at Lockheed Martin, told me explicitly, she was like, Oh, you have this pension now. In forty years, you're going to have this amazing pension. You're going have a job for life. That sounds awful. Can you just kill me now? You're going to prefer it. That understanding of where to allocate your like how much you listen to who. And so like, there's a lot of advice out on the web. Find entrepreneurs that you really admire and go just like listen to every single thing they say. And then the last piece of advice is most of the time you learn by doing.

And so why did Steve and Bill Gates and Jensen, why are they so good? Is because they've failed so many times and there is no shortcut. There just isn't. And so like the faster you get up there, launch something and fail, just count do that a hundred times, and, like, you know, that's the only way you're going to get good. So that's my my advice, is throw a lot of crap out there. Don't be embarrassed by it. Why do like, there's a whole thing of, like, people that have, like, you know, maybe, like, less social awareness are gonna be better entrepreneurs. It's because they don't care about throwing it out and looking stupid. And it's because who cares? And so, like, just do it. Throw it out there.

Not everything has to be beautiful, amazing, and just, you know, keep learning. And that's the best way to become a great entrepreneur. I love that. I think that's a perfect way then. I'll send so well. Thank you for the time. Alright. Cool.