Greg Brockman (Part 2)

[music] Tectrogrammerton. [music] The first website I ever built that got users was this amazing experience. I had this idea to build what I call the reverse-turing test. So in the Turing test, to determine is a machine intelligent, you have a human who talks to another human and talks to an AI, and the goal is to figure out which of these are the human, which is the AI. So I built a website that turned this into a competitive game, where you have both humans are talking to each other, and they're each talking to an AI. They don't know which terminals which, and the objective is to figure out which of your terminals is the other human, before the other human does. And so the optimal strategy is to ask questions that kind of discern, "Am I talking to a human or a bot, but while still acting kind of bot-like, if you're acting to a human like, then you'll lose, because the other person will figure out who you are." Wow. This was 2008, and I just taught myself how to code. I'd gone online to W3 schools, tutorials, did HTML, Geodescript, PHP, CSS, and I remember that I built this game, and it's a two-player game, so I was sitting in the lobby. In case anyone showed up, they'd have a good experience and have someone to play against. Lobby where? I made it so that there was a game lobby just on my website. So I just sit there, just waiting, waiting with this open screen, just sadly waiting for someone to show up, and for two weeks, no one showed up. But then one day it was the most glorious day. I got 1,500 hits from Stumble upon. Wow. Yeah, if you remember Stumble upon, it was like an early, you know, and send people to random websites. And it was an amazing moment where that day, there were like always three or four games going constantly. I'd sit in the lobby and someone joined within a couple of minutes. And I remember this feeling that this thing was in my head, and now it's in reality. And now these people are all enjoying what I built, and I want to keep chasing it. Yeah. The more you played, would you get better at that game? Yes. Yes. I got quite good at it. It was interesting actually. I focused a lot on improving the bot. My bot was like very rudimentary, and the way that it would work is I kept a database of all the previous games. And then I tried to, in any particular conversation, match that conversation to the most similar one, and then reply with what the human had said then. And it actually kind of worked. For any sort of chit-chatty thing that kind of it's done already, then you have a pretty good database of replies. But anything more sophisticated, and of course it just would follow. Can you remember it all where the idea came from for that game? Well, the way that I got excited about AI was by reading Alan Turing's paper on the Turing Test. So this is his 1950 paper called Computing Machinery and Intelligence. I was reading it shortly before building this game, and it had the most inspirational ideas in it. Because first he asked the question, can a machine ever be intelligent? And he says, look, I don't know what intelligence means. Everyone's going to have their own definition. So let's define a test. You define the Turing Test. But then he says, how will you ever program an answer to this test? You will never program it. It's just too hard to write down all the rules of this person says this. And that instead, you will need to build a machine that can learn its own answer to this. You have to build a learning machine. So if you can build a machine that is a child machine that learns like a human child, and you can then have a human who gives it rewards and punishments as it does good things and bad things. Then that's how you will solve this test. And here's the wild thing. That is exactly what we've been doing. This is like Alan Turing in 1950 projecting how we will first build these unsupervised models that learn, sort of observe the world, have all this knowledge in them. And then we do this reinforcement learning process. We give the machine rewards and punishments in order to achieve the objective that you have in front of it. And I remember talking to my co-founder, Ilya, about this. And I was like, how did Turing know? And he said there's a reason that Turing is Turing, right, that he is just so smart. And that's why he really took so long from 1950 to now. Very simple answer. It's compute. It just was not enough compute. It's like no matter how smart Turing was, he did not have a computer that he could implement his test on. Was there a feeling in the early days for you that AI was sort of a failed thing? The ideas have been around since the 50s and it hasn't really worked. So this for me was always a great mystery. It was so clear the Turing test vision was so clear. It was like, that's the thing you need to do. Because as a programmer, I have to understand the solution to the problem. But the point that Turing makes is that you can have a machine that comes up with its own solution to the problem. And I was like, think of how many problems I have no idea how to solve. And maybe the machine could do it. That's what I want to do. I want to help that thing come into existence. And I remember after building this game, I showed up at college at Harvard in New Still 2008. And I was very excited to do research with a natural language processing professor. We studied math at Harvard. I was going to study math. And I got computer science nerds night. But my number one thing was going to be math. I actually was originally intending to do math, chemistry, and philosophy, triple major. But I ended up realizing that I just loved the practicality of building with computers. And so I went to this professor. And I asked him if I could do research with him, he said, yes, no problem. And he gave me this parse trees problem. And I remember looking at the parse trees, I was like, this is never going to scale. What is that? I don't know what that is. parse trees, they were like a old school NLP, like natural language processing approach. So the idea of like, imagine you take sentences and you figure out like where the object is and where the noun is. Kind of like the things that people do in elementary school. Like linguistics. Yeah, like linguistics as an approach for how you're going to do natural language processing. And you can get some kind of simple looking stuff out of that because it'll make reasonable looking sentences. But that's never going to scale to having a conversation like this. It was just so clear to me. I was like, this is not what Torin was talking about. I'll go do things that are useful. And instead I actually got into programming languages. So I was very excited about that. I took a class. I wanted to do more research there. And I think about programming languages. And you have all this power of like a compiler is a computer program that takes a different program and makes it better in some way. So it usually takes it from a high level form and puts it into a form of the machine can understand. Usually optimize it so it's faster and really take the intent of a human and translate into a form that the machine can then. We take care of some of the details. Because for me that was the spirit is I want a machine that can solve problems that I can't that will empower me in ways that I am unable to to reach the bring these new heights. And not just for me, but for everyone. And so I would do that. I got very into the Harvard computer society where we would build services for the Harvard community. So we would host email and web hosting and other technical services. We built different web applications for for the community. And so this was very much the ethos that I had. And it really wasn't until I was already at stripe where I was doing a startup and building things, but paying attention to the community. And I just kept seeing people talking about deep learning. It felt like every day if you went on hacker news, which was this website that many engineers would would post content on. You'd see something about deep learning for acts. And I remember wondering what is deep learning. At the time it was basically impossible to figure out because you go to deep learning dot com or dot org or whatever it was. And just a deep learning is a new approach to AI. But you can say what it was. It was made no sense. But I remember that I had a friend in the field. And so I went and talked to that person. And he started introducing me to other people in the field. And I just kept getting introduced to all my smartest friends from college because they were all on the field now. How many people were in the field at that time in the world? It was a small community, very small community. I don't know the overall number of thousand. And it was rapidly growing because the thing that I learned is that there was this moment in 2012 that really unleashed the current deep learning revolution. And in many ways everything had been building up to that moment. But this was the moment that really cemented the there's something real here for many people. And this was the creation of Alex net, which was a image recognition paper that competed on the benchmark. Explained what that is. So the idea is that in I think 2006, 2008, a lab at Stanford created a competition where they gathered millions of high resolution images from across the web. At this point, we would consider a small data set at the time as massive and unprecedented. And that categorized this image into a thousand different categories that humans that would label them and say this is a specific type of cat. And this is a specific type of bird. And this is a specific type of airplanes with thousand different categories of images. And the goal was create a machine can create a program that can categorize a new image into one of these thousand buckets. So can you recognize whether the cat or dog in an image. And that people would compete in this and it would compete hard and they would take all of these 40 years worth of computer vision research ideas very similar to like the feeling of those parse trees right you would have these different techniques that would do like edge detection and things like that that are like very specific. And if you think about what are the rules for recognizing a cat it's like well, maybe you look for a eye and another eye but how you recognize that there's an eye and you look if there's a nose and then but okay, the orientation could make it very complicated to actually program relationships. Exactly. You have to talk about these relationships in the very higher article. If you think about the process of you have to see how all these different pieces fit together and then how those pieces really other pieces. Very complicated and so people were not getting very good results here felt like a total loss. We really need to study the real world. It wasn't really technical. It was observational. Yes, because the thing that people would do is that they would say how do I think humans do this or what's the process I have in my head or let's write down some symbolic way to tell the machine how to pursue a process. And this was very successful in some domains, right, for example, chess is one that in the 90s that we built great machines to solve it but very unsuccessful in other domain like computer vision. There's only one page of rules exactly the game of chess simple rules and a small search space. So you can have a computer that would basically just say you know I'm going to look through all the possibilities and that is how I will win. Which by the way was not enough for go right go simple rules but massive search space and so you needed something more like human intuition on top of the search. And for computer vision you needed something that was entirely like intuition and so that's where the neural nets came into play. And so a team of researchers who were Jeffrey Hinden, Ilya Sutskiver, Alex Orsevsky created a neural net that won this competition and it didn't just slightly win it. It just blew everything else out of the water like a massive jump. Did they have a different vision of it than everyone else? Their approach was neural nets. No one else believed in it. I see. And it's really it's actually very funny the inside story on how that result came to be because Alex Orsevsky was a grad student in Jeff and in the lab and he was working on very fast. And he was a very fast convolutional kernels for GPUs so he basically was programming graphics processing units GPUs which now are what people use for for deep learning. And everyone felt bad for him. It was like that's just an engineering project. He's just writing these very fast kernels who cares. He's just an engineer that's not valuable and he had some cool results on like the small image recognition data set and people didn't really care. But Ilya saw that and he instantly knew what to do with these kernels right that he realized this is a breakthrough in the making because when the image net had come out this big data set. And he had felt like this was this grand challenge that was just so impossible if you could solve it. It's so great. But you just need to be able to put enough compute into it. And he sees these kernels that we're going to use the computer GPU very efficiently. He said we need to put these two things together. Don't apply it to this other data set that you're using image that is the thing. Jeff Hinton's contribution was a management trick because Alex of CF ski really hated writing papers. He had a review paper coming up and Jeff told him each week that you get a 1% improvement on the data set. I will push back the deadline on your review paper by one week. And he did this like a dozen times two dozen times in a row. And so it was just one of these things where I just like Alex just kept grinding at the problem and the numbers got better and better. And so you see the way the progress in this field happens is you need the right theory. You need the right objective, the right sort of underlying approach. You need the right engineering right you need to really implement it you need to push hard on the problem you need to not give up even when it feels impossible. And you need the right spirit right you need to know that like it's worthwhile and you need to have that desire to keep going even in the face of everything else. And so I think those three things together were what unlocked this particular result in this particular moment. And they submitted the competition ever in the computer vision field is like what just happened right this impossible problem has basically now been solved. What was the reaction to that it was one of these things where within the computer vision community it was seismic right people very quickly went from saying. Neural nets are to those dead end like you're kind of a fraud if you're doing neural nets to only neural nets. And when you were a fraud if you were doing neural nets. Absolutely. Yes. Until the breakthrough. That's right. And actually the history here is also very fascinating. So there's this paper you can find on Wikipedia somewhere from 1995 that talks about the history of the deep learning. Looms and busts so it's really before all the current waves and if you read it that the things they're saying in there at the exact same things people would say to us throughout the history of open AI these neural net people. 1965 they would say these neural net people have no new ideas they just want to build bigger computers. And so it makes you realize that history is written by the victors that there was a very concerted campaign waged against this whole direction saying symbolic systems yes neural nets know and that the neural net people knew what they wanted to do they want a bigger computers deeper neural nets. And that the symbolic systems people got in very cozy with funding agencies and really just poison the well and said this whole thing is kind of BS. And so that's what killed it for the 70s was that they claim that it over hyped there are all these groups working on it that there's no results and that to put the nail in the coffin that there is this result. And so that a single layer on the net couldn't solve a particular problem so therefore the whole thing is dead and of course the neural net people rate but just let us go to not single layer like we know what to do. But it was all like one of these things where the establishment killed it these sort of centralized and he said no. And then the crazy thing is in the 80s what this paper says is the reason neural nets came back was because of the democratization of compute. It went from being that you have these professors who guard all the compute to suddenly all these PhD students have their own computer. And so professor can't tell them they're not allowed to neural nets and so suddenly people are doing it again. And so to me that is so interesting that this theme is universal it's like always been true it's not a new thing over the past 10 years it's something that's been there for 60 70 years. So what happened after the adoption well so after the Alexander result that people outside the field of computer vision. Would still poo it they would say well it works for computer vision but. No, that's I'm not going to do with machine translation right because you have these fixed size images and stuff like that and for machine translation you have these variable windows and things like that. 2014 you have sequenced the sequence didn't get the same massive step function you did without the net but you could just see yeah you're going to push that and this is going to be the only thing you need. And what happened is you effectively had the walls between departments being torn down and this beautiful unification of you thought that you had all these different domains of computer vision machine translation speech recognition and as note you just have AI you just have deep learning. It sounds like any time that the fringe groups yes can come together something much more interesting can happen yes I remember with I think it's our third C car first law no it's if an elderly but distinguished scientist tells you that something is possible. They're almost certainly right but if an elderly but a single scientist tells you that something is impossible they're almost certainly wrong that's great I love that. Open AI started in your living room in San Francisco describes living room to make it was a big open space and we had a black wood table that was this big oval shape at some couches I did big screen TV day one there was no whiteboard and I remember two researchers were debating something they turned to write something on the whiteboard there wasn't one and I was like I can get a whiteboard and so I felt like I was adding value from day one. Yeah so who was in the room that first day in the room that first day would have been Sam Altman I'll just give her voice X Rambo was there I think Vicky Chang Pam Fagata John Schillman Andre Carpathy probably would have been there at the time some of them were finishing up their work elsewhere the PhDs. I apologize if I'm forgetting anyone else but that was you know the founding team the founding by was we had this great objective of we really wanted to help build the GI and have it be something that is a positive force for humanity and we did not have a thesis on how we would do it and so that was where we started. And at what stage of the I revolution were we when you were in that room what was known was not known so this was the very beginning of 2016 at this time so long ago not so long ago 10 years ago 10 years time flies yes that's why time it compresses and it's crazy we had gone through at that point four years of this deep learning revolution and so one thing that was clear was that. It was like there's this early phase where the fruit was just hanging on the ground because you could just take a GPU take a neural net you point at a new problem and it's going to work and it's going to get you awesome results and so new architectures was kind of the the heyday of basic research on a lot of ways so individual researchers being able to come with a novel idea. A few months prove it out get an awesome paper it would be an unprecedented thing to a kind of like define field so that was the moment that we were in. It wasn't yet the moment of grand engineering wasn't at the moment of large scale compute because you want to as much computers you can get but you couldn't really get more out of more GPUs right it really was that you had one GPU and orchestrating many of them together we didn't have good techniques for how to actually get good returns from that. I remember in the very early days working on the first engineering projects to support the researchers and I observed two researchers building with two engineers and the way that it would go is that researchers would say here's the system I want my requirements and then the engineers would go off and build something and come back a few days later and they would project it up on it and then they would go line by line spend a whole afternoon just baiting every single line. I remember looking at that and thinking this is never going to end it's so slow. It takes too long. Too long. Not going to work. So instead I ended up working on the project and I would work in a very tight loop with the researcher and I would say here are five ideas to say these four are bad and I say great that's exactly what I wanted. And so just really this not trying to push my own ideas but really trying to learn the other person's perspective the other person's view on the world and try to then say okay can translate in all these different ways and to try to just tease out what truth is what reality is. Typically are the researchers engineers as well or no in this field they are much closer to engineers and there are some people who are really at that intersection for example. Yacca Patroki who is our chief scientist one of the things that really distinguished him is that he really has his foot in both worlds that he has deep theoretical understanding his a PhD in optimization but he also really knows how to build systems and has done it many times. And so that it's a unique skill set it's very valuable and so the thing that I found was that for engineers to add value in this field you have a pretty high bar because these researchers they all know how to code they can build their own things so you have to do better than they would on their own. And that's a contrast if you're just building for doctor say right most doctors probably don't know how to code so the bar to do better building something for them than they could build on their own is relaxed and so that a lot of what we focused on is how do you make sure you're empowering and moving forward the researchers. How well did the handful of people in the room that they know each other so there's a subset of people who had. They all gone through PhD programs together some had intern together so there is a set of people who knew each other and there is a set of people who were newer to each other but at this point we'd actually already gone through some formative events so we had. Really throughout the back half of 2015 I've been doing all the recruiting to find just two of the best people in the field was the first meeting the meeting of open AI or was it a get together that turned into open AI I'd say that the very first moment that was really to get together that set things in motion. Was a dinner in July of 2015 and that that was with Sam that was with the Ilya that was with Elon that was with a handful of others and the question there is it is it too late to start a lab that can actually really get to a GI right that it's not too late well it felt like deep mine kind of had it right the deep mind had all the talent that they had as part of Google all the compute that it felt like maybe a GI was very close and can you actually get together a group of great people and really go for this. Did you start it in competition with Google would you say I don't think of it as competition but I do think of it as complimentary right that I think that my view on how AI should go yeah this is very foundational is that I think that AI is something that everyone deserves to participate in and to me it felt like we're going to build these incredibly powerful systems and that how they play out for humanity to uplift everyone is something that is the single most important thing. That can happen contributing to that and helping steer that in direction that make sure it actually is beneficial to everyone like that the thing that I want to do and so to me it felt like it's not about the back and forth on who has the best benchmark yeah it's really on how do we build systems and overall society that integrates with those systems right these things are going to co evolve together. It's a better world than the one that we have today and that that isn't something that any one group can do on their own. How big of a revolution did you know it was then could you see where we are now then or no then it felt like if it was going to work at all I kind of have to feel like this and I think we're not done. I think that I want to describe the personalities and strengthen weaknesses of every person in the room. Well I'd say that Sam I think is a visionary and I think that Sam is someone who gives permission to dream big ideas and I think he also cares a lot I think he cares a lot about people and I think that he is someone who is like always very optimistic about how we are going to do it. He's very optimistic about how we can find a way to configure any solution to a problem and so I think that he is someone where you feel like hey this is never going to work he will find a solution. But he is always someone who opens your mind that it can be solved. That's right and that maybe it's not the best solution yet. That's right but once you know it can be solved then you can work on a better one. It's not detached from reality right it's like kind of connected to like I think he's been he's like an excellent sort of facilitator for researchers and for this overall sort of shepherding of this technology into the world. So again I think is a visionary I think he is someone who I remember in that very first couple days he said I have this idea of been thinking about for how we can solve what was called unsupervised learning how we can observe the world. What is that so if you think about how a human baby learns just by observing the world right there's no one saying this is the right thing no one's doing input exactly it just happens. And this was always to me this always felt like a crazy concept of how can the machine ever learn without someone telling it whether it's doing a good job or not. Yeah. But we figured it out. I remember that he had a lot of ideas on how to really push into machine. Even in that room would there be people who think that's too far. So in that room we immediately started trying to write down ideas and that the energy was just palpable. One of the steps along the way to everyone coming to that room was this off site in November of 2015 so I mapped out the whole field or all the best people and kind of been asking people for who do you know who's great and they would introduce me to people. So I just kind of kept track of kept getting introduced to this guy called void chat. I was like all right, void checks probably this one I should go after. And we had narrowed down to a set of people but they were all kind of like okay well who else is joining I'm interested but who else is in you know like how do I collapse this. And I asked Sam what to do and he suggested to bring everyone to an off site. And at this point actually I remember it was very grateful to John Schollman who had said that he would be in so I was at least not the only one who had committed and I think maybe there was one or two others who were kind of there. But it really was a group of people who had not yet coalesced into a team. And we brought everyone out. We're in my apartment we got into the boss we drive up to Napa and it was just this day where everyone clicked right at the energy was just so smooth. Or is that flow state in human form and I remember that we wrote up on this flip chart and I have a picture of the flip chart. The plan there's a three step plan step one was solve RL which is reinforcement learning that is learning from these rewards and punishments to is solve you well that is unsupervised learning that is observe the world and just absorb the information and then three is gradually learning more complicated and quotes things. And this actually is what we've been doing for a decade actually crazy right is it like we really set out the division and so it's very tight to the original vision it's just the growth of the original vision it is it really is when I look at everything we have done it has been in service of the same goal with the same almost technical approach and the same ethos underlying it any of the other people in the room who had particular. And I think that we're going to have a lot of people who are going to need to expertise or views or that were different than the others that are worth describing. Yeah I'd say voice check and I still work very closely with him is also unique character he is extremely good at idea generation and he will come up with very creative ideas to any problem and then he is also someone who is not at all attached to his own ideas. And I think it's really hard to be that that generative what you want to be as good as it can be exactly I can't all be your ideas that's right what I usually do when working with and is really think about okay well what are the bounds on what we should think about right here the places that we don't want to go or here's kind of the place that we need to end up and then the idea generation process just ends up with great fly will. And then we end up with strike so while I was at MIT I was building more and I did more startups and from each one I felt like I learned another thing not to do and eventually I kind of felt like I knew enough that I could be successful at doing a startup but I was missing a key component which is having an idea. I was pattern matching off of my friends who had been in the computer club started their own startup the gun to master's program come with the cool idea there I was like well clearly that's what you need to do you need to go to grad school you saw the path yes but the path that I saw was actually too much of the beaten path right the beaten path is in your PhD program you invent something there and you turned that technology to start up. And I remember just feeling like okay I'm 21 I'm just too young to do real things in the world this path is the only thing that's possible but I the very least I can start meeting people doing startups because I had taken a startup class and it was just not useful I was like all right like this is not going to get me to where I want to go and so I decided I'd meet these people doing startups and literally the next day I got an email from some people working on a payment startup in Palo Alto I was like well my new thing is to meet these people and learn to pattern match every time. And I remember when I met Patrick we just clicked what was his background also he had been at MIT and so he and his brother you know I did not yet but we had mutual friends because he had gone to MIT John and gone to Harvard so is that community that they were asking around yeah and my name of course came up both circles. And I remember I flew out and it was like raining and kind of miserable and I remember opening the door and just like when we first started talking it was just an instant connection right I think we just like had this you know different backgrounds in many ways right but also we had a very similar technical perspective and like even for very nerdy think I mean we had the same like split keyboard that we use this can use this keyboard we both were devor act the actual layout that we use. Different and we were talking about how to like build the firewalls for the systems that they had and you're talking about you know different things in the kernel and so it's just like we had this like technical connection yeah and I think that what I also really appreciate about Patrick and and John is that they were my age and they're already out there doing things right that they're already doing startup and I was like I don't think that's possible. One of the two of us is wrong I really want to know who it is yeah and then what happened so we can was great they said you should join I said let me go think about it I went back to school John happened to be in town on Thursday and so I was like you know I probably should make my decision and I was like you know what I want to do this because again back to that algorithm of dream of the what if this works I'm like I don't know anything about payments like this is not the problem that I've grown up passionate about but these people. You get to solve a real problem solve a real problem exactly and these are the people that I feel like I can learn from or that you can work with and I remember John is like will you do it. I was like I'll do it and I just remember feeling like okay like that they want someone exactly want me it's a good feeling it was a really good feeling and so I decided that on Thursday Friday spent telling my teachers that I was out how did that go. It was a little tough was it emotional for you it was definitely it felt like the end of something yeah and you know Harvard when I told them I was leaving they said you're coming back you're coming back exactly and I think it's probably for their numbers and you know that kind of thing but it definitely felt like okay I see how this goes MIT was a little bit more like okay like talking to the the girl in person that they said that you know if you you have to check in every six months or so because if it too long goes by. Then you know maybe you'll have to reapply and we're not really sure and it just was a very different vibe and I remember talking to the professors the professors were supportive and I think this is the thing that they see and but they were also you know I think a little sad to to see me go like mid semester yeah I really appreciated the mentorship they provided like I was in the middle the operating systems course which is this famous course at MIT top by these like extremely good professors and it was just such a cool class. And I was very sad to not get to implement further projects and so it was definitely kind of giving up on on some sort of exposure to experience and it was also hard to say goodbye to the people right there are all these people that I'd gone there to learn from and work with and the funny thing is many of them ended up coming out to the valley got to work with them that strike got to spend a bunch of time with them in other ways so it was much less of the goodbye than I thought it was at the time of the day. Growing up where you grew up and moving as quickly as he did. Did you feel different. I definitely felt different yes is it lonely did you feel like an outsider like I definitely felt different I definitely didn't fit in right and there was like a lot of things that other kids were into that just didn't understand like I remember being on the school bus. And kindergarten and other kids were singing along to the radio I didn't know any of the words yeah and I just felt like I don't even I don't even know how to bridge this gap and so I had a number of moments like that where I just felt like there's just something about me that like doesn't quite match and some of it was about activities like a lot of the kids would hunt and that was not something that my family did at all and so there's just something very different I did play hockey for a little bit I was not very good. I was going so I started to try to you know do some activities that would do a match but I remember I was one thing I really cared about was carving an identity for myself because if you're different and you don't really feel like you have an identity then it's lonely but if you're different and you have an identity something that defined you yes then you're charting your own path and for me. I was being a smart kid like I remember in elementary school that we had a weekly spelling quiz and the way it would work is on Monday the teacher would give you the 10 or 20 words for the week and then at the end of the week you'd be tested on them and you'd kind of have this pre test on Monday where you'd have to write them down as the teacher says them and if you got it wrong you'd have to go ask one of the other kids who got it right how to spell it and normally I get them all right but I remember one week I got one of the words wrong. I remember another one of the smart kids of the class you got it right and it was so ashamed that I would have to go and ask him for the right answer because that would be just eroding this core identity that I had. Wow that sounds like it was really healthy though that you got to do that and it set you up to not have to have all the answers. Yes yes yes I think it was not necessary that was a good story that's right. So then when you got to meet someone like Patrick was it a feeling of oh he's like me. Yes. Must have been a great feeling. It really was it really was and he had done he done a startup before so he and John had previously. And he was 21 and he had already done a startup. Exactly. Yes. How did that happen? So they did a startup called oh man I haven't thought about this for a while. I'm not going to remember all the details but I think okay so Patrick and John grew up in Ireland and I think that Patrick had met. Patrick had met Paul Graham who runs Y Combinator this startup incubator through the list community. So through the program and I was that they were very into. And I think that that was his connection to then doing YC doing a startup that started got sold to a company called live current media. They worked for the acquire for some time but clearly that was not the thing that they wanted to do with their lives. How early were you in Stripe? So early days of Stripe there were four of us. There was John, there was Patrick, there was Deribuckley and there's me. And when I first was there there was some infrastructure and there was a payment processor. It also wasn't clear exactly how we're going to proceed because there was one idea was well what if we build some apps and then use this payment processing we're building to power those apps. And so that's how you actually build something and eventually the payment processor will probably be the thing but these apps could be two. And so there was a time tracking thing that Patrick had been working on that was one of the potential ideas. So it's still in this nascent form but still something that we could see like the direction of travel. And the funny thing by the way is that I mentioned that I did this startup class in MIT and that as part of that you're supposed to build like a mock startup. And the one that I ended up building was a payment processor. And so I spent all this time trying to figure out how do you do payment processing online and because it gets horrible it's painful. It's like trying to go to PayPal and trying to like read their documentation trying to figure out a sign up for something. It was just so opaque and you just realized how can it possibly be so bad. And that was the fundamental unlock for Stripe was this realization that you can do payments processing better. And so I was like, the initial website said payment processing doesn't need to suck right and it was just like like that is the ethos are just like it just doesn't need to but can be better. And I really like that spirit. I remember talking to the VC at some point this was pre launch but we had gotten a lot of buzz and he was saying, OK, what is your secret sauce? And I was like, well, we just do it better. He's like, no, come on. I know that's what you say to everyone but you're like, what's the actual secret sauce? And I'm like, I don't know what to tell you. Like that is what we do is just focus on every single detail and get it right. So it really wasn't creating something new. It was already PayPal. It was just a much better PayPal. Exactly. Exactly. Right. It was. And it was really focusing on the details of the whole experience and the end. Right. And really think about just because if you've gone through it yourself and you're like, this part, like, how do you sign up for an account? How do you actually been connected to the API, so the actual way the computers talk to each other? What are the different parameters you program against? How do you figure out which program in language to use or which rapper to use all these things? Each one of those can add a ton of friction. And so if you just focus on OK, I'm going to make this one good. This one good. This one good. This one good. Make it good for myself. Thank you. Good for my friends. Yes. Actually, this would be something that makes it good for everyone. And why did you leave Stripe? Well, I had been at Stripe for almost five years, four and a half years. And about four years in, I think I started to consider whether or not this was what I wanted to do for the long term. And the way that I kind of view things is like, OK, if you think if you're career in five year chunks and five years is about the right length because less than that is kind of hard to do something significant. And I remember feeling like, OK, I had gotten this company to a place where it was going to succeed with or without me. Yeah. And then the question was, do I continue or do I go and do something new? I was very excited about doing the idea of a startup. I remember talking to Patrick and I think he had some very convincing reasons to stay that one point he made is that it's very hard to assemble a group of people that can do significant things in the world. In best case scenario, you go start some startup and have some success with it. And then five years later, you'll have formed that group of people that can do stuff. But you already have it here. Right. We already had this tight group that was able to accomplish things. And so why walk away from that. So it was very tough. It was not easy. I remember I cried when I told Patrick that I was out. Like, yeah, it was very, very tough. He and John through a juice party for me. It was very, very nice send off. And the reason that I decided to remember feeling like as I thought that through I was like, well, if it really is so hard to build that group of people who can accomplish significant things, I got to get started now. Yeah. It sounds like you left because the mission of stripe wasn't the mission that you wanted to focus on. That is true. Yeah, it's a beautiful mission. And it's a lot of very much support. But it's different to say, is this one where I will just need to pursue it in any form. And did you leave knowing what was going to be next to do you leave thinking I'm going to figure it out. And with a ladder, I had a list of three different areas I might focus on. Number one was AI. Number two was VR slash AR. Number three was programming education. And for me, it was very clear if I can contribute to AI in some way, OK, I'm doing that. But it wasn't clear to me. Do I have the skills time, all those things. And so the other two were kind of backup options. And then there was a point in time that you wrote a chemistry textbook. That's true. When was that? That was also 2008. So after high school, I took a year off. And I had spent much of my high school doing academic. College courses, but also competitions. I got very into math competitions, got very into chemistry competitions. And I remember 10th grade. Because I'd taken chemistry in 9th grade, I took some chemistry competition that my mom had found online. It's kind of on the mark. And I got best in my region. And I think that's cool. I got my state didn't really have it. So I was in Minnesota to do it. I took the statewide one. And I got best in the state. And they invited me to the training camp for the international chemistry. Let me add, so top 20 kids in the nation. And for this, they send you some textbooks ahead of time. They say, please read chapters one through eight of this organic chemistry textbook. You know, the textbook weren't used for the camp. And so I read through chapters one through eight. Didn't take it that serious. I was like, look, I'm just destined to succeed. It's just going to work. It happens so far. And I remember I showed up at the chemistry competition, the chemistry camp. And the other kids had read not just those first eight chapters. They'd read all the books. They're not just the whole one book, but all these big fat like physical chemistry, all these other ones. And I was just like, wait, what? You can do that. These other people are doing that. And I remember just feeling so demoralized and crushed for the two weeks of the training because it was just like people knew all these things that I didn't know. And we were supposed to be studying for the final exam. I was playing cell phone games in my room. I just gave up. I was like, there's just no hope. And so they announced the top four to go to the international. They said these four people. They weren't me. Yeah, they announced two runners up. They announced those two. They wasn't me. I was like, I am obviously 20. Fortunately, they don't tell you the ranking of the rest. I was like, so bad. And I remember at the beginning of the next school year, looking back at how I'd spent that summer. And I felt like I had this amazing opportunity in front of me. And I'd squandered it. And it was a horrible feeling. It was feeling that I had just coasted that I just sort of believed that talent alone, which is get me to where I wanted to go. I didn't have to work hard. And I was like, I never want to feel that again. And so I took it seriously that year. I started taking physical chemistry. I took organic chemistry. I took all these things. I spent a lot of time on looking at old chemistry competitions and learning them and looking through a bunch of different material in order to become a real competitor. And that year I made the team at the top four, went to the International Chemistry Lampiade, got a silver medal. It was very, very awesome experience. I really enjoyed it. But for me, the big takeaway was this feeling of you always have to work hard. Yeah. And one of my favorite quotes is the cycling quote. It never gets easier. You just go faster. And I think that that is something I very much loved by. So how did you end up writing the book? Well, I felt like I'd come up with a very unique way of looking at chemistry, a very mathematical first principles approach to it. Because if you read most chemistry textbooks, it just says, well, basically memorize these reactions here in these chemical properties. Here's these compounds. You always ask, why? Why is it that way? Does it have to be that these atoms interacting this way? Does this compound have to be this color? How do you derive it from first principles? And so the approach that I'd taken in order to be good at the competitions was to really try to figure out the underlying rules, not the-- Which wasn't in the textbook that you read. It's not in the textbook. You have to still it down yourself. Maybe they try to communicate in some way, but it's like, it's just not the thing that they spend their time on. And so I structured the book. I felt like I had a different way of teaching chemistry. And I structured it in a way that was inspired by some of my friends who had done something similar in math in this math form called the art of problem solving, where it's a very stocratic method. So the book has just questions, but they're intentionally scope so that each one builds up. And so the first one starts from knowledge you should have. If you just think a little bit, you're like, oh, I can see how this works. And then the next one builds on the previous thing, the next one builds on the previous thing. In the case of chemistry, you need some experimental results. So you say, here's this double slit experiment. And then you're like, okay, well, what does it mean for particle versus wave? And then, okay, if it's both particle and a wave, then, you know, and so forth. And this was something that I really wanted to communicate to others. I really cared about not just this approach living in my head, but other people being able to benefit from it. Now, I never finished. I made it through about 100 pages. You can find it on my website if you're interested. But I felt like that ethos was something I really wanted to carry forward. You basically wrote the book that you wish you could have read. That's exactly right. That's great. That's. How is coding similar or different to other activities? So the way that I think about coding is that you deeply understand some process. You write it down in a very obscure way. We call program. And then anyone can get the benefit of that thinking. Right. People don't need to write the code. They don't need to understand the sort of mechanics of what went into it. And I think that there are other very cerebral domains that are like this like mathematics, right? Where you think hard about a problem, write it down in a obscure way. We call proof. But no one reads those right. Only like the five mathematicians who care about a particular domain will really deeply read it. And so I think that what makes coding stand apart to me is it's almost like magic. You sort of have this vision in your head. And you just by describing it somehow it comes to be. And so in some ways, it's like management, right? That you have a computer that is there to perform the function that you have in mind, the vision that you have. And that it carries it out in a very little fashion when you write a program. And I think that that to me is something I've never seen in really any other traditional domain like it feels like many other things that you might do. That you just don't get that same leverage, right? If you have this vision and somehow it comes into reality and that you don't have to physically move things in the world that just comes to be. Would you describe it more like a language or more like math? I would describe it more like math. But the thing about about math and I think there's a misconception for is that math is not one plus one right. It's not about these like mechanical calculations math is about the underlying structures of the universe. It's about understanding how different objects different ideas relate to each other in this deep conceptual way and the sort of symmetries and the relationships between object look very, very different. And I think the program is like that. It's really about understanding how should a website work or what is someone want and what are all the different ways that something in a corner case should behave. If there's an error, how should you handle it? All these things they feel some maintain, but if you really look at the underlying architecture, you're thinking about you have all these systems that are talking to each other, you have data stored in these different forms, you have ideas like encryption that are brought to play. And how do you orchestrate all of that in a way that delivers something useful. And so to me, I think it's it's about the beauty of mathematics that's reified into useful form is math. An overlay on nature or is nature an overlay on math. That's a great question. I think it math is the fabric of the universe in any ways. Like the thing I love about math is that is the set of rules that are true in any reality. Like I remember in middle school or high school starting to learn biology and you learn all of these details about how different cell types work and all these different processes about chlorophyll and photosynthesis. But I remember feeling like, okay, this is something that happens to be true right here for this specific organism or way of studying things, but is this universally true? Like does this have to be true always does this have to be how things work. And the thing that I love about math is that it is purely decoupled from observation. It is the set of things that must be true. And so I think for me, math is this deep, immutable understanding of what is even possible and the nature is an instantiation in a specific form. You think it was possible for nature to exist before the understanding of math. It was math needed for nature to exist. I don't have two minds at this one because to me math feels like it exists independent of everything else. It's really a first principles deep truth that doesn't matter what perspective you take. Like if we were to meet an alien from hundreds of light years away, they would have the same mathematics. We would probably have something in common in that way and maybe nothing else in common. But at the same time, I also feel like there's a question I struggle with sometimes of do we discover mathematics or do we invent it? Is it a deep truth or is it our perspective on it that makes it come into reality? And I think that a naive view is to say, well, it's already there without us. But in some ways, if there's no one there to appreciate it, it's a little bit like the tree that falls in the forest within the line to hear. It is the sound there. And so I think that there is something about nature and the fact that we are here and that we are observers and that we are thinking beings that causes math to have meaning rather than just be an abstraction that with no significance. If you were to explain what the coding process feels like to someone who doesn't code, how do you describe what it feels like you're doing? So the most beautiful part of the coding process is when you're in flow state. And there everything just kind of clicks. You have some objective, perhaps it's something simple, like you want to change the color of a button, perhaps something complex, like you want to build a database or a big distributed system with many computers talking to each other. But you have some objective in your mind that you want to see in reality. And you have a partial implementation, maybe you're starting from a bank slate, maybe you have written some code and it sort of works or it implements a subset of what you're looking for. And then you try it, you test it out and you see that there's a gap between what you have in your head and what you observed in the system. And sometimes it's a subtle gap, which usually you think of as a bug. You wrote the code. It was supposed to do something. It did something different. And then in that case, you start to form a mental model of, well, I know how the code is written. I see this observation that's different from what's expected, or perhaps was expected, but is not yet what I have in my mind. And you think about what's the, what's the gap? Maybe you know, immediately, oh, there's the specific line. I probably messed it up. This value is too high. I should go back and check. Maybe it's you don't quite know. So you think about, well, what would give me more information to figure out where it comes from. And so often adding what we call telemetry, so adding observability so that you can get numbers out from the middle of the code about how fast different parts are running. And then you can get little log lines that say that the code effectively saying, I did this. This happened. Here's some event. And then from looking at that trace, it's almost the history of what has happened within the program. And in other ways, it's almost like you're in a very detailed fashion, instructing someone else for how to perform some process. And you have some understanding of it. And you want to write out all the rules for exactly how it works in every single corner case. And then you can get it whenever you see some undesirable outcome, you go back and say, well, what did I get wrong in the rule book? And you try to really have enough understanding of the trail by which that outcome occurred so that you can go back and make the appropriate changes. And then what does that mean? And then you can see the results of the process. And then you can see that what is happening right now is software engineering is changing entirely. coded, this was for a live demo of one of our early coding models. We built a little web and interface and you could talk to the model to ask it to write some JavaScript. And so I, on a live stream, had it build a little game and asked the people watching the live stream to suggest a feature we implemented the feature just by talking to the computer. And to me, the deep thing that's happening is that computers have always been created in order to help humans, right? That's the whole point. And that we can toward ourselves to the machine when writing code, whether it's a low-level computer programming language or all this skill and machinery for how you actually tell whether computer did the thing that you wanted. And what vibe coding is, is it's moving the machine closer to the human. And so you instructed, you still need a good, depending on how good the model is, you need, didn't have difficult taskings. You need some mental model of how it's going to solve the task, right? If you just say, "Build me an awesome website." Well, what is an awesome website? So, okay, fine. I want there to be a button here or I want there to be this type of functionality. You're effectively acting as not a individual contributor, not a software engineer, but you're acting as a manager. But a manager who's still very accountable for the outcome. And I think what is happening is that the models have been incrementally getting better over the past year in a significant way. I think it was a real turning point in December of 2025, where it was the first time that for many of the expert engineers that I know, these models went from being kind of nice, kind of useful to, they can actually do incredibly hard pieces of work. And so it's really shifted from being a thing for demos. And just if you don't really know a programming language, you can get something done quickly, but it's not quite right to, this is driving serious work and really accelerating what people can do. Is there still a reason to code the old way? I think that coding by hand in some ways is like handwriting, like penmanship, like co-agraphy, that there is an art to it and that there is an understanding of how everything fits together by offering at that level. There's a way in which it's like mathematics where you still really want to load, even if you're not going to do a bunch of hand calculations, you probably should still know how to do long multiplication. And there's something about what coding is really about is understanding abstraction, right? It's really about understanding how systems fit together, how pieces will interoperate, and that is something that you want to become an expert at. And so if you're too far from the details and that you're just pushing a system and you don't really understand how it works on the inside, I think that that is a limiting factor. As the models get better, I think that the nature of what you as a human need to really take accountability for and responsibility for will change. The machine will be much better at a bunch of mechanical things, a bunch of the design things, how interfaces work. Each of those will come over time. But at the end of the day, it's your vision. And if you care about how your vision is implemented, then knowing the nuts and bolts of how the machine is going to do it, at least having the good mental model of it, that is something that will pay dividends. And you see a time when that won't be necessary. You tell the machine what you want, it codes it the way it wants, and you can still do iterations and improve it. But is there a time when the active coding will be maybe like lateness today? I think if it's almost like an ocean that is rising, right, where the ocean level is the capability of the models and you have these islands, and that every so often an island gets totally covered by the water, but there's other islands that are even higher. And I think that coding is like that, where the islands are difficulty of problems. And so I think what we're seeing is some islands have already been covered. That for example, I have a test prompt that I've used for every model for a number of years to build a particular website that was one of the first websites I ever built. When I first built it by hand, it took me months. Then when I used in early one of our codex models, it probably took me five hours, six hours, something like that. With our latest model, it just codes it up in a minute. And I just don't even have to touch the details. And it actually does a much better job than I ever did, even with the previous iterations. That's really interesting. You really feel the power of what is at your fingertips and the fact that you are now empowered to do even more. But there are still mountains that we have not covered. And the question is, are those mountains infinite or finite? And I do think that maybe the best analogy for where we're going is that people are going to become rather than individual contributors. They'll become managers of the agents. And then they'll become middle managers, right, moving up the food pyramid and eventually become CEOs of this organization of agents. And the one thing that I actually don't have an answer for is yes, all the mechanical skill, all of the deep debugging, all of the architecture, all of these things, those you can see how the machine will get very, very good at that. But owning the outcome, right, the accountability, the picture of is this doing what you want? I don't see a line of sight, not to say that you can never happen, but I don't know how that is something that you would ever transition out of the human. And I think that that is something that is deeply human and sort of something that is unique to being a person with an intention, with a life, with relationships. And so I think that there is an involvement in the process that you will always care about. And just like if you're someone who is having a house built, you may not care that much about where every nail is going. But you care a lot about the outcome and you care a lot about are the people building it, doing it in a way that you'd be happy with because you're the one who at the end of the day if it's built poorly is going to be on the hook. How different is a poorly built model versus a well built model in terms of how it functions? In other words, if you can describe something that you want to work a certain way, and if it works that way, and if the code is done elegantly, if you were to do it yourself, and maybe less elegantly, if the machine were to do it, but it still did the thing you wanted it to do, would it matter? So this is something we see very concretely within OpenAI. And we are taking a very agent first approach to software development. What does that mean? We have set a goal that by March 31st, though, very soon, that we want two things. One is that the default tool that all of our people reach for is an agent rather than a text editor or a terminal. So really that this becomes the tool that you find the most reliable and the first thing you apply to every problem. The second is that the default way that people use these tools is something that has been explicitly evaluated as safe and secure and that those two factors to me are what it means to be agent first. So it doesn't mean that you never go into the details, but it means that most of the time you don't have to and that you built both trust with the system, but also that we've architected in a way so that we can have trust in the whole organization as an outcome. Now if you drill into some of the individual points of how we direct people, one of the hard decisions was really about how do you make sure that the code base doesn't turn into AI slot. It's tough that works, but somehow it's just not very good. And we have many years of making sure that it doesn't turn into human slot, right? That we have individuals who write code and then you have people who review code and that you have a large system of incentives for if you're writing good code that over time is something that is maintainable and other people can build on, then you get promotions and that you have good performance reviews and all of these things. And we need to bring this to bear in the agent world. So that we have a mantra of say no to stop so that we tell all the human reviewers that there's still whoever is submitting the code needs to be accountable for it. You as a reviewer should hold an even higher bar than you would for a human on quality, make sure that the person actually understands what's there. It doesn't mean that every single line that they really know, but that they can really sign off and say, yes, this improves our code base rather than regresses it. And one of our best engineers, as he's been playing with these models, has found that the way he wants to balance is that he wants to control the interfaces. So he goes and still writes by hand, here's the components, here's how they fit together, maybe here's the file structure, but the details of how it gets implemented, which often are quite intricate and complicated, those he outsources to the machine. And so I think that the difference is really if you don't pay attention to a particular aspect of what the model is doing, will it be something when you look under the hood that you're proud of? And right now, we're evaluating that. We're seeing in what circumstances the answer is, yes versus no. But we're saying that our requirement is that if someone does go look, that the answer should always be yes, I am proud of this. Or alternatively, you can have some sections of code that you don't care if they're good or not, as long as they meet the correct and specification. And so if you have good ways of verifying that that section is correct, then you can have very highly optimized code that's extremely hard for anyone to build on. But you don't think of it as the way that we think about codes that's evolving. Artifact, you view it as a one off, I produce this, I'll produce a totally different version, never build on top of it. Is most of what's happening, like building blocks being made, and then maybe the way they link together is more casual? Yes. Would that be way to describe it? I think that's a pretty good way to describe it. And I think what's going to happen is that the size of the building box is going to increase over time, and the way in which the human overseas is also going to move up that level of attraction over time. Because right now you look at every individual agent, that's going to feel totally barbaric and slow even just this year, because you'll want an overseer agent that's looking at the work of all these different ones and flags to you. This particular detail doesn't look quite right over here. This agent seems to have gone off the rails. This doesn't look like what you wanted. And so I think that really figuring out how do you as a human manage a larger and larger fleet of agents. And in many ways, it's really about the fundamental measure is how much compute does an individual human marshal. That is going to be the most important metric for future productivity. What would you say the biggest technological revolutions you've witnessed over the course of your life, each one? Well, I remember growing up in North Dakota and reading a Katai magazine article about Silicon Valley and feeling like I was born too late. All the exciting things that was already happening. It was happening and I wasn't there. Yeah, it was too young. It just felt like there's only so many good ideas in the world, only so much innovation that's possible. And I can see it happening right now and I am not part of it. So that was a moment that I felt the FOMO. I remember later things like mobile phones and the shift there. It was never something I was very passionate about. There are a lot of people wanting to build apps. For me, it never really felt like the thing that I was deeply attracted to. And I think that there's so many different pieces of technology that add up to modern life. And I think that's what I was talking about. I think that's what I was talking about. In each of the pioneering situations that you've been involved in, how much do you know in advance of what it's going to be? I would say I always go in with a vision. I always go in with thinking about the just set aside all the reasons this might fail. Just dream of what it can be. What it can be? What it can be? What is even possible? Within the laws of physics, right? There's no point in dreaming of the impossible, but how it could go. I think that stripe was like that of thinking about, well, we could build something that will be this amazing payment behemoth, right, that actually makes payments more accessible. It makes payments more accessible. More things happen. That sounds great. And the people there, great people, you can learn and grow together and you could do more things in the world. That's something that felt like, okay, I see the, what this could be in that way. Many details to be figured out. Open AI, you think about AGI, it's like if you can actually build an AGI into existence in a way that lifts up everyone, there's no better thing to work on, right? That is the most amazing thing you could hope to do in a technological sense. And then for projects within Open AI, I think usually it has the same kind of flavor to it where I remember there are moments where I see a demo or I see a result. You see a little initial curve and you just realize this is going to work. I remember at some point actually, voice check, one of my co-founders saying that the thing about this field that is so remarkable is every idea works. As long as it's theoretically motivated the math works, it's going to happen. Well, actually succeed. You'll actually get good results. But the challenge is really figuring out which ones are going to the fastest path to your objective. And so it really feels like there's fruit lying on the ground. And because it's all lying on the ground, sometimes that's the challenge. There's such a massive opportunity space. But if I figure out the path through, you need these proof points. And I think we're sitting in the middle of one right now. Yeah. Very, very clear with these agents getting so good at not just software, but they're going to get very good at all knowledge work. I think this year we're going to have a very big transition at how work is done. I think it's one of the biggest issues with AI in general. Is that because it's so open-ended what it can do, it's hard for people to imagine what it's going to do. It's the tool that can do anything. Yes, you know? That's hard. That's right. Yeah, we have this problem with Chatchy BT. Right? You show up at Chatchy BT. There's a text box that can do anything. Yeah. But what do you want it to do? Many people just, you're stuck on the blank page. And I think that that actually, the dual of it, shows the opportunity with AI. And I think that ultimately, AI in my mind, AI is opportunity. That is why we build it. The opportunity is that if you do have a vision, if you do have something you want, if you do have a particular way you think things should be done, if you have agency that you wish to see enacted in the world, we have the tool for you. There's lots of questions as these tools get better. Where did the humans fit in? What does it mean to be human? All these things. I think that agency, that drive, that vision, those things, those are something that we as humans have to contribute. Yeah. What is AGI and is it very clearly delineated? I think of AGI as not just a system that can do any intellectual task that humans can do, but that can really be this force multiplier for an individual human to the extent that they can operate as the visionary, as the CEO, and so that that's that level of empowerment for the individual. And to me, it's really not just the technical abstraction of this system that theoretically can do something. It's how it's actually deployed into the world. That's a new definition. Yes. It's how I've grown to think about it. Because if you look at our mission, it's not just a technical problem. You can define the technical AGI, but I think for the very, very beginning of opening AGI, we were unsatisfied with the idea that we would just write papers. You could write all the papers you want that describe how an AGI would be built, but there's no impact. The impact is the instantiation into people's lives. And I think that what super intelligence will be is something beyond that signpost and capability. And I think that the outcomes we want from it are to help us solve problems that are totally out of our reach. Great. And I think that solving diseases, I think that space travel, I think that there are many, many problems that we see. Is chat GPT the primary product of open A.I.? No. What is? The thing that we ultimately are selling is intelligence on demand for your problem. Chat GPT is one instantiation of that and it's massively popular, almost a billion acts of users every single week. But that's not the end of it. We have an API that many business customers build on top of. And that is also growing absolutely phenomenally. Our newest product is called codex and it's transforming how software is built within open A.I. and really within the industry as a whole. And that again is something that we're just seeing this take off. And the thing about what codex is, it's really two things. It is a general purpose agent harness so that it's able to use tools for any kind of tools, any kind of application. And it's a system that knows how to write code. But if you think about that first thing, that is extremely valuable and repurposable for any knowledge work tasks you might want, anything you might want to do with your computer can be expressed in terms of an agent doing some things orchestrating some tools. And so we're starting to apply it to things like Excel, PowerPoint, being able to create these artifacts that people in business produce in various business functions. And we're spending a lot of time to actually make our model very, very capable at this. And so I think what we're going to see by end of year is a very different product surface where every knowledge worker will have this tool that really gives them super powers. What is agent to AI? I would think of a agent to AI as a model that you don't just talk to like in chat, but that it's hooked up to tools. And that at an implementation level, it's almost like the model it can chat with you, the human, but it can also chat with the system. And I can say, please run this command. Please create a spreadsheet. Look, this thing up in someone's email, so it needs access to the outside world. It's actually able to take action. It's not just like chat, you see its impact is more cerebral, right? It's you talk to it talks back to you for agent to AI that it's actually embedded in the real world and you take action. But one thing to note about agents is that we've been increasing the time that an agent can run quite significantly. So you can have agents now that do useful work over the course of the day and the space of how you apply agents is very large because you can apply lots of compute and parallel. You can have many agents that are working on one task and that are able to produce things that would take humans very, very long time to do. Tell me a bit about the mechanics of chat GPT. How does it work? At the core, chat GPT is powered by a language model and you should think of a language model as a system that takes in some text and then outputs some other text. Now, the text doesn't need to be literally English language. It can be images, it can be videos, it can be really any sort of modality. And the output similarly doesn't need to be literally not just language. It's not just language. How long has it not just been language? Well, to be fair, even from the very beginning, it wasn't just language because the real code. You could do code. Right. But it was always, it was always taxed. That was what we started with. So GPT for, we had a downstream project of that called GPT for V, which was the first time that we had vision and that you could have a model that actually recognized images and you put text in and images in and output text. And since then, we've trained models that take in sound, take in images, take in text, output sound, et cetera, et cetera. And so you actually can have a full voice conversation with models that we have produced. And that's part of chat GPT as well. And so I think of it as these models are general purpose intelligence processors. And when you create them, the way that you train them is that they look at data, but they don't really learn the data. They learn the underlying rules that created the data. And so that's what makes them smart is that they're sort of general purpose understanding machines. And you can point them towards whatever task you have representative data information training for. Do you have a name for the big brain? Well, we have names for our different models. So we call, but the models are the use part of it, but the thing that learns everything. We don't have a name for for the overall training system. So there is, there's names for various components. For example, you know, we have a training system that does a component of the system. We have different systems for how we actually take a trained model and serve it. But now also training and inference are sharing to come together because you do reinforcement learning where the model teaches itself. Similar to what Turing was talking about. And so there's the kind of an overall orchestration system. And each of these components, they have their own names, they have their own concepts. And there's really a whole industry that's built up around how to train models for helping people in particular ways. Is all of the knowledge in one big base? And then is it divided up to do these different things? So we thought that an AGI would be literally one giant model. And you come up with these architectures that there's this thing called mixture of experts, for example, where you think of it as it has rather than every time you're running information through the network that it runs through just smaller parts of the network. And so there's an opportunity during the training process to specialize. So the training process could choose to say, well, I want to specialize this for language and this for text and or this for programming in reality. And so it's a little bit more complicated than that. But I would say that the big unlock has been to realize that well, the most useful tools that humans are creating for language and for vision and things like that are models. And these big models, they're expensive to run, but they can use tools. So why not use smaller models too? And so I think that we're heading towards a world of this menagerie of different models. And you're seeing this almost came through an explosion right now in the field of all these open source models, people are training and that we train models of all sorts of different sizes. And so there's not really just one system anymore that there's these models that can talk to other models and that are specialized for different purposes. And that really introduces the diversity of approach and means that you can specialize in all sorts of different ways. What is pre training and what's post training? I would think of pre training as a phase where the model observes the world and learns from it. At a technical level, the way that we implement it is with what we call next step prediction. So you show a model a sequence and you ask what should come next and that sequence could be anything. For example, it could be the public post for some site on the internet. And so what is the word the human right next year? It could be Einstein's thoughts. It could be something very deep and significant. And if you think about it, if you can predict every word out of Einstein's mouth, you were at least a smart, a science time. And the important thing here is that what the model is incentivized to do is to learn not just where the nouns are, where the commas go, those kinds of surface statistics. It is incentivized to learn the underlying rules of this distribution, like where does this data come from? Why is it here? Because if you're just sort of parroting back what someone else said, it's not going to be helpful as soon as you're looking at something new, which is what these models are trained to do. It's really about the underlying rules and generation and deeply understanding any new situation that the model is placed in. That's what pre training is. And at a technical level, it is an extremely interesting problem that you scale these up to massive numbers of compute devices that the actual problem that you have to solve is you put data through the network. And then you pass it backwards. So you kind of see, how do I have to adjust all the connections in order to have gotten a slightly better answer from the forward. So the idea is as you pass data through, you've got to some some predictions, some output. And then you see, OK, it's almost like if you have all these wires that are connected to produce some, you know, some results. And how do you adjust the taughtness of all these wires in order to have gotten a slightly better result. And that's how the machine is ultimately programmed is that you do this forward and do this backward. And then you have a step called an optimizer step where you adjust all those parameters and you do it again and again and again, do this at very large scale. And so any individual data point, even any individual large source of data doesn't really matter, right, because you're talking humanity scale learning. This thing is observing the whole world. But like if there was some part of your childhood that you forgot, like you're not going to forget all this knowledge, like it's like really about the underlying background understanding of of reality. So that's pre training and at a technical level, you slice up this computation across many devices and yours always trying to pump more efficiency and trying to get. It's not just larger models. We also have all sorts of innovations on architecture so that we have models that are shaped in different ways and you map those. A lot of the engineering challenges is trying to understand how the hardware works, where its weaknesses are, how it fails, and then how do you design the architectures in a way that are most amenable to that. So the output of pre training and usually these runs could be a month, they could be multiple months, maybe our longest one was somewhere around nine months. So it's a big team of people in order to keep that thing running. GPD 4, for example, I was very involved in the pre training and built a lot of that training stack 2 am we wake up because the run was down, go and fix it. Like that's what you got to do. And it's like every hour that the job is down every minute that it's down, you just look at the number of GPUs that are sitting idle and you think about how many dollars are being wasted. But more importantly, just that your list opportunity, it's a missed opportunity. And you really feel that in a very, very, very great way. Post training. Post training, you take the output of free trainings, you take this model that knows a lot of things, it's seen the world. And you try to tell it how it should use that knowledge, what's the right behavior in different circumstances. So it's almost like when you have a child that's observed the world and learned a lot 20 years old now it's time to go to college. Now you specialize in teach for specific domain, but one thing it's different about these models is you're not putting new knowledge in necessarily you're really almost pruning down what it already knows because the not a compute in the pre training process versus in the post training process is typically very, very out of balance of the post training process usually a couple days, that kind of thing. And a lot of how we teach the model we've evolved these techniques, but classically we do it through feedback that we would have some way of saying that well let's train a reward model. So another AI that can judge what the say is doing and you train that one usually through saying these are good, these are bad or here's two possible generations from an AI which ones better than the other. And then this reward model judges the pre train model and then gives it feedback and from there it's able to shape its behavior. So we really have the ability to take this model that knows kind of everything I remember talking to Alec Bradford who is one of our researchers who we describe it as these pre trained models they're less like a human and more like a humanity like everything's in there. And then we do our best to steer it we don't always get it right, but I think that's a lot of what we've been working on in the pre training is like the library of Congress let's say I'd say pre training is like the library of Congress and then post training is almost taste giving the machine a sense of which of those books. It would like or what to do with that information once it's retrieved for pre training where's all the information coming from is it the internet. I would think of the classic approach has been publicly available data on the internet what's been changing is that as these as I've gotten much smarter that actually you really want to train them on their own data in some form for example reinforcement learning where the machine goes out and tries to solve a task. And you learn everything it learns in trying to solve the task is now part of that's the idea that's exactly and again it's not just the knowledge itself it's also the skills right and it's really that's what we're really getting at like our dream has always been to have a reasoning model is just pure reasoner. And it is able to in any new situation be able to figure out the right thing to do and to do a great job there and sometimes having background knowledge is helpful but it's really about those smarts and the ability to adapt very quickly that is like the real thing that delivers the value. How much of the human hand is involved in the post training. It's been changing as well it really used to be that we would have these large campaigns and sometimes we still do but that that most the data that we train on would be we'd have humans who would painstaking the label these different examples. And the thing is that as the tasks that the models are capable of has gone up there's way less to learn from most examples so you really need domain experts who are deep deep in their field and so some of the tasks that we produce are these like incredibly complicated like look up you know this finance report from this specific year and judge how this one compares that one wasn't mean for the strength of the business. You know paragraphs of prompt that the very specific answer and it requires some domain expertise to do and require a dozen hours from that domain expert to accomplish so we've really moved up the sophistication of the task and the way that we teach the machine and it makes sense right as the machines get smarter we really need to figure out where are they breaking down and it becomes much less about massive volumes of data much more about this very high taste very. Targeted like what are the most important problems that we want this AI to solve and trying to then teach the machine accordingly. My takeaway from the AlphaGo story was that the computer made a move that no human would have made and that's why it won yes and if you're teaching the AI how to act more responsibly as a human. Wouldn't that undermine its ability to make the move that the human wouldn't make and doesn't that undermine the whole AI premise so the objective of our research in many ways is to achieve that type of AlphaGo moment but in science in coding in these other domains because that is exactly what you want is you want new knowledge discovery and you're exactly right if all you're doing is just learning from the data. Learning from what's been done it feels like how are you going to go further but the thing that that perspective misses is that the way that we are now extending the training to reinforce not just take the public data take the humans providing us a no but actually have tools that let you test things out in the world like how humans discover new things sometimes we think deep thoughts but I think that usually it's through experiment. Right we tried something it didn't really work we realized out and I think that part of what's going on is the universe is almost as massive computer and that it has far more computing it than our brains do it is far more computing it than our GPUs do and so that's why there's something to be learned from experiment because there's this computational process that is just unimaginable in terms of how sophisticated it is and that you can kind of shape it to your will right you can kind of have a ball that rolls down a incline of this height and that height. And that height and then if the AI is able to propose that experiment and see the experimental results maybe has a robot that set that up for it. Maybe it's a human who performs it either way that is how you can actually discover new knowledge and I think that the idea of actually have experiment and then be able to compress all this into this model and for it to understand the underlying rules how all that data was generated. And that is something that we are just starting to see the fruits of if the AI proved something in physics that negates what's in the current textbooks is that a safety problem or is that breakthrough. It's already happened. Tell me the story so there's a physics professor who has been a vocal skeptic of AI yes and we convinced him to use our latest unreleased AI system and he gave it a particular hypothesis in quantum physics that he's been planning on working on this whole year with his collaborators it's a very hard problem that people are pretty sure that there's a particular answer to it. And our AI proved that actually the opposite was true and his reaction was that this is the first time that it's felt like the system is thinking right there's new knowledge in there there's something very, very novel and that paper. Submitting it to be published but I think it's a very significant moment and very representative of things to come. I think the most potential in AI is when it does things that humans don't know is the right answer yes that's what's exciting it is and if the corporate perspective is to prevent that from happening which I can see an argument for we can't rock the boat this is accepted the science is accepted. It's where I get most concerned about AI the biggest potential downside is that it can't do what it's able to do I think this is a really important point and I think that the bigger framework around it really is that this technology we have the ability to steer it and that's actually one of the deep motivations for why we created open AI is that. I think it will happen with or without us but we think that we can help be an influence on it playing out in a direction that we think is more positive for the world and that that is our ambition in our aspiration and there's a lot of value to be delivered by the actual creation of the technology the core really being that we think it's something it's be available to everyone and it should be something that just like humans can question the wisdom and it's always tough when humans question the wisdom right there's a lot of antibodies that they try to prevent that. But that is also how society moves forward and I think that we're going to have to make choices as a society it's not just for anyone company anyone individual to decide it's something that as a society we should decide what are the rules of the road and a lot of how we thought about it is there should be broad bounds that society decides in AI can never cross within that. You really need people to have the empowerment to pick the people need to have an AI to represent them whether it's about their values or whether it's about being able to question and there may be context where people don't want that that should be their right that should be their choice and there should be context where people are trying to discover new science or whatever it is and that should also be their choice and I think that we really have this philosophy is very different from from others in the field of self empowerment and this technology really is for everyone. Is there an AI bubble? I think that we will find that we were under ambitious on compute. Yeah. I think that where we are going and we're seeing the proof points of it is a world where knowledge work is amplified by compute power. I saw someone tweet saying if you're taking a job ask how many tokens are in your budget. It's kind of a joke right now it's not going to be a joke later this year and I think that the degree to which these tools are changing software engineering like the software engineers who have not tried this don't feel it yet but those who have they feel it they feel it in their bones and I think that we're going to see that across finance we'll see this in sales and people will be able to do so much more. And what we're already seeing is individuals within open AI who want 100 GPUs 1000 GPUs dedicated just to them. Yeah. And if you think about a thousand GPUs for an individual you have a thousand such individuals you're at a million GPUs already and there's not 10 million GPUs in existence. So you can only scale so far so I think that we are in a world where we are seeing what's coming what this technology is capable of when we look at how good the models are getting when we see our own revenue curves and we look at the things we cannot launch because we do not have the compute for all these things together. Mean that it is actually quite rational why all of the hyper scalars why we are all trying to build compute like I think where we're going is compute will be a basic human right. Like I think that people to be economically productive and even further own lives the more compute they have the higher quality of life they can have and so I think we're going to have to be in a world where everyone has access to compute. Will it always be GPUs or might something replace the GPU. I think the computer is fundamental and that the computer is always changing and we're already seeing lots of different approaches and even in video for example has acquired GROC which is a different approach to a compute device and there's lots of interesting approaches people are taking. And so I think that GPU is a good stand in for the different types of accelerators we are cells for example have a custom accelerator that we're working on called intelligence processor and I think that there is lots of room to improve the efficiency and scalability of this technology. What's the most expensive heart of the whole operation. Compute. Where is electricity fit any I'd say you can almost think of AI as a manufacturing process from electricity to intelligence and that we use electricity as one input to how we actually do the computation so electricity is almost like the water that like drives the whole the whole system is it get more and more efficient over time yes. But one note is that we also as we increase the efficiency and we increase it a lot like if you look year over year we tend to cut our prices for the same level of intelligence by 100 X sometimes like literally like you can just see it from our price drops. So like where we were for GPT 3 in 2020 that and that was a 175 billion parameter model you can get that level of intelligence I I've even looked but billion parameter model something like that like it's something that you could run on your phone for free no problem whatsoever. Open AI versus Anthropic versus Gemini versus GROC versus deep seek. Tell me about what's different about those companies. For me I really focus on us I think that what competitors are helpful for is almost as like a pace car just get a sense of how you're doing sometimes they can point out that oh here's a particular feature that we didn't think we can implement they did it okay we can probably do it so it's helpful for that but the way that we've always proceeded is that we invest the most in basic research in the actual paradigm shifts and you can see this where with language models with the reinforcement learning paradigm and here's upcoming paradigms to that we are. Embracing and capturing and really put long term investments in and I think that is one thing that really stands out is actually interesting I had a candidate today who. We're pitching him on a long term research project and using oh surprise I thought. From the outside looks like you guys are always doing things fast and I didn't realize that you make these long term investments and the reason that looks like that is because we have a pipeline of them and they come to fruition. One by one and I think we have the smartest models right if you look at the reinforcement learning stack I think we have something that is very unique and new. I think each company has its own niche like I think when it came to consumer chat to be teased by far use the widest I think Google did a very good job last year of building better models and they had previously and that they have a natural distribution like it's very. Always front and center to me the fact that Google is like they a lot of compute they have a lot of talent they have a lot of users they have a lot of these natural advantages and I think that and traffic is coming on the scene and they focused very hard on coding and one thing that I think that they did well was. We were focused on the bench parts in some ways we are focused on academic program and competitions we had amazing numbers there we didn't focus as much on how these models will be used in the world and so training on these like messy repose and how people are actually using it and that was a lesson that we. We're delayed on but we gather team focus very hard on that I think we are caught up and I think the people are really seeing the fact that we are definitely on takeoff so in some ways having the other companies does give you an idea of other things to be focusing on it's a big picture that's right and there's so much in this world like the whole space of knowledge work is so large that figuring out exactly what to focus on is sometimes the hardest problem since everybody's models are being optimized it seems for the same. It seems for the same benchmarks does that end up being a limitation on what's being done because everyone's focusing on this small group of. Tests it can be it and I think it has been in the past if you look at. Where things are going I think that we are in a bit of a post benchmark world. What really matters is ultimately the benchmark of our people actually using it yeah is your revenue growing like those kinds of benchmarks that are impossible to gain. The problem with the academic benchmarks is they're very easy to get 100% on them right you just train on the test set right you. At the numbers on their own don't necessarily mean much so sometimes you'll see a model from a particular company or your model comes out and it's got really good numbers they're too good to be true and it always is. Would there be an argument to not participate in the benchmark tests so actually there's two answers one is that the way that we develop these models is through very good e-vows so that we actually do want. Benchmarks that tell us they're never perfect they're always proxy metrics but it really tell us are we on track so we've made a system called GDP Val for example that's an e Val that shows how useful our models are on a number of knowledge work tasks. And that he'll climbing on that is actually great at some point you'll saturate and usually once you start getting to 80% on these benchmarks or something 90%. Usually it means that you're done and there's no point getting 100% because you know that means you're doing something very specific for this benchmark that doesn't make sense. And so I think it's as long as you're in this range where it's like giving you meaningful signal fantastic and you don't want to focus on too many of them because I think that there's usually you want well constructed benchmarks that give you some good signal in previous years there have been times when actually the right answer was you take an average across a bunch of benchmarks that are themselves not that good and then that gives you actually pretty good reliable signal too. So but I think that the important thing is that you don't wake up every day think about how do I move the number on this benchmark you use it as a proxy use it as a sideback sure if you focus on it too much then you can what we call good heart it. What do developers do and are they assigned jobs or do the developers pitch you ideas. So two big aspects to what we do there's research and deployment in research we are creating new models and that that requires both research and engineering to be joined up to hit in deployment that that is usually about taking the fruits of research and bringing it to the world. There there's a very high bias of engineering but you need to be high context you need to really understand that research because if you don't you're not going to do the job of bringing it to reality. What the engineers actually do is usually on the research side that there's some idea that we're pursuing and that we usually form teams sometimes these are very small teams to three people that are working on some novel idea that's very high risk may or may not work and there you just want to get some signs of life and you need to have enough patience because it never works the first time. Sometimes these are we know it's working we already have signs of life we scale it up so usually the life cycle of an idea is that once we have an idea we tried it out we got the signs of life we then put more people on it and a lot of the engineers are focused on the system is breaking too frequently let's go figure out why it's too slow that the data that it's importing. We're getting too small of a download rate trying to figure out what's going on there that they're often building distributed systems so that we're able to process things or be able to take outputs from the model and be able to observe them and be able to see how things are doing building dashboards so there's a lot of work to be done in the actual m l engineering is a very sophisticated. Art requires a lot of deep domain expertise and there aren't that many people who do it there's writing the actual kernels to the actual code that runs on the GPU that turns the global representations of all this data into objects that can be computed upon so it's almost like you know I think about how your your eyes turn visible light into some signal in your brain and that once you're there it's almost like we have these people who are doing the like the low level neuron engineering. So all of that is very sophisticated takes a long time to build up the expertise and there's a lot of techniques there at that intersection of the machine learning and the engineering for example as you scale the size of the model the way that people classically did this was wrong that it turned out that you would get. That results but you would always know right that should it be the case that as you scale up the model that you get very smooth curves no like deep reason a priority that you should think so but one thing we realize at some point is that actually we need to have different ratios between different parameters get those things right and we could develop this technique called new p for how we set the initialization and that actually gives you a much more straight line as you go and so there's a lot of these things where even knowing if something is wrong. Is not easy and deep expertise and partnership between people different expertise is what yields great results. Have you followed the club at open cost story at all of course tell me your thoughts I love it so I think that and I know I've spent time with Peter I think he's he's great he was the developer of open call and to me open call encompasses two things first of all it is a system that you can hook up tools and then it is always on and runs and is able to take action and I think that there's a hacker spirit that I really like about it of saying we have these tools we have these models there's a massive overhead there's so much more capable than what we're using it for yeah let's try let's see what it can do and there's a second thing to which is that there's a gap that we have to fill of what I call the security architecture right the trust architecture of well you have an AI that's hooked up to many things but how do you know it's going to do the right thing how do you know that it will send the right messages and people post things on Twitter that are kind of fun like someone who his wife was texting him and his cloud bought replied and you know it's 2 a.m. and the baby is crying and the cloud bots replying wife isn't having it like you know it's kind of funny to read the text but you think about where this goes we need better guard rails right we need systems that are engineered for trust and safety and that's a lot of what we focus on right that we really think about not just the let's build the technology make the capability but how do we actually bring this to the world in a scalable way that we're transforming our own enterprise on the basis of this technology we want to help transform every enterprise and so a lot of how we're thinking about this is that I think it's a great sign of things to come and I think that to really scale it to everyone requires making it incredibly easy for people to set up but also making sure that the default way that things are done is safe and that that is something that we are investing in pretty heavily and there are deep choices that we made for example there's one vision of how you could build a I we're winding 10 years which is that you keep it all secret and then you put together the finishing pieces and no one knows that you're doing it so you have no pressure to deploy so you really have time to like get it all right and then you push the button and you know question mark question mark but then you know benefit the world and I always looked at that plan and I was like I don't think I can sign up for this it just feels wrong you know first of all from a technical perspective if you never encounter a reality how will you be certain that you actually have put in place the right safety systems you think about cloudbot is nice that it's starting as kind of a fun project it's clear that for it to get hooked up to very sophisticated important systems being trusted with a lot of responsibility we need to develop new technology but we're going to learn and iterate and loop and so to me it always felt important to get this technology right we have to encounter reality each step along the way but the second is legitimacy that you really if you're going to build technology it's going to change everyone's lives I think people need to know about it people need to be included in that and so if you look at chat you be teased we made a very deliberate decision to say we think that this is technology for the world to be included in for our ethos for what we started this company for it's for everybody yes and it's a podcast tetragramatine is a website tetragramatine is a whole world of knowledge what may fall within the sphere of tetragramatine counterculture tetragramatine sacred geometry tetragramatine the avant-garde tetragramatine generative art tetragramatine the tarot tetragramatine out of print music tetragramatine biodynamic tetragramatine graphic design tetragramatine mythology and magic tetragramatine obscure film tetragramatine beach culture tetragramatine esoteric lectures tetragramatine off the grid living tetragramatine old spirituality tetragramatine the canon of fine objects tetragramatine muscle cars tetragramatine ancient wisdom for a new age upon entering experience the art work of the day take a breath and see where you are drawn