All right. I think we have some Codex fans in the audience.
So fun. It's a busy week, but I'm happy to be here. This is an unexpected surprise.
Well, thank you for joining us. We appreciate it. So we opened this morning by saying that we've kind of arbitrarily decided that the singularity started on January 1st and thus today is day 119. What do you think of that?
It does feel like we are somehow in the takeoff. Day 119 feels like a reasonable enough guess. Yeah, I won't fight it.
Did you feel this? So we've started to see a bunch of our metrics inflect as of late last year, beginning of this year. I mean, things were kind of doing well but somehow they really just, the shapes of the curves changed. They really went parabolic. Is that matched in what you guys see? Was there some trajectory change? Like why are we seeing this?
I do think the models got really good, especially for coding, but really good in general starting late last year, very early this year. And at least in my own experience of using this technology and seeing what other people are doing with it, and also this sense that every week is now a little bit different than the week before, like a lot happens very fast, it seemed to all correlate with the models hitting some threshold.
And why did coding models suddenly start to click over the last couple months? Like was there a research trick? Was it just got enough code data in the pre-training? Like why did it suddenly start to work?
Yeah, it's a great question. We wondered a lot about why kind of several people crossed that threshold at the same time. I'm sure it's a number of factors, but model intelligence, just the raw kind of reasoning horsepower, enough of a feedback loop of people using it for code to figure out where it was good and where you needed to improve it, enough data. I think it was like all of these things. And then also there was like many other endeavors, once you know something's possible, it's much easier to go do it with vigor. And so it seems like Codex is kind of having a moment right now.
Yeah, it really crossed some subjective threshold for me with the latest app updates and 5.5. And this is also like it's quite hard to say why right now and not a little bit sooner. Why not the next model? But one of the things I have learned about the history of all of the things we put out is it is very hard to say why this particular thing was the thing that worked. And this goes back to ChatGPT. Like why was GPT-3.5 the thing that got over the threshold where most people went from saying not that impressive to like going to change the world and why not one model earlier or later? I really can't explain it. You just kind of feel it. And I've had two inflection points with Codex. One was kind of with GPT-5.2 and then a really big one in the last few weeks where it's like, okay, this is going to be the primary interface to a computer for me.
And is everyone using it for coding or are you starting to see usage diffuse into other domains?
I think the most adamant users are still using it for coding. But there's been just this tidal wave of people coming into Codex recently and I'm like really trying to understand what has happened that is causing this. And the depth of what people are using it for or starting to use it for has surprised me. So certainly our ambition is for it not just to be about coding but to be about all the work you do in front of a computer. And you know, I would say we're maybe like 10% of the way there for the non-coding stuff, but now that we see what's happening, now we have a real user base sort of using it in these other ways, I think we'll get good at it very fast.
What do you think will be the next domain that subjectively feels like it has this big unlock after coding? Like is it going to be spreadsheets? Will it be, you know, performance reviews? What's it going to be?
So I think there will be a lot of, first of all I do think coding is a little bit special and these models are a great fit for coding. The world needs so much more code than currently gets written. There may be no other domain that is quite like coding. But there will be a lot of others that I think are close. But the next kind of coding-like thing that I think will happen is not any specific domain but the realization of how much time people waste trying to use a computer and the idea that you can do a huge percentage of your day in a very different way. And you know, maybe you don't realize how much time you spend clicking between messaging apps and copying and pasting stuff and responding to very boring things that you could clearly automate once. But the degree to which most people will realize they can sit back and watch an AI do most of their drudgery is going to surprise people. And in my own experience trying to work that way actually gives me much more enjoyment of work. I didn't realize how much the little stuff drags me down, gets me out of a sort of happy flow state or whatever. So the subjective quality of life improvement is huge.
Are you an OpenClaw user?
I am. OpenClaw users here?
We have some good news for you all coming.
OpenClaw has been one of my biggest like this is magic AGI moments ever in the field. I remember the first time someone told me about it. They were trying to explain it and I was like okay that all sounds cool but you know I can make a lot of that work. And then it was a real reminder how when the models cross some threshold and also the product designer gets a handful of critical ideas really right, it's like a much more magical experience than it sounds like.
I find I've been an attempting OpenClaw evangelist and trying to describe that experience you just recounted to others and I find it a difficult experience to communicate. I mean it sounds kind of prosaic, right? It's a stateful ChatGPT session that can also make some use of tools and so forth. What do you use your OpenClaw for?
If we scrolled your message thread, what would we see?
So this is like a very embarrassing thing to admit. The thing I always try first, perfect place. The thing I always try first with a new kind of AI system of any sort is I'm like a home automation nerd and so to try to build a better home automation interface system because it never works, it never is good. And OpenClaw was the first time that I was able to get a setup that I was happy with. I also built a messaging app that I had always wanted to work. I've since switched it to something I built with Codex, but OpenClaw was the first time I was able to, I'm sure like you, feel like drowning in messages and it's like this very unpleasant task to wake up in the morning and have to go through all this stuff. So I was like all right I'm finally going to be able to automate this and that was like again should have been doable with previous systems. Hard to explain what it's like when it all actually just works and you trust that it's going to work. I was testing the new Link CLI that we just launched today in preparation for launch and I asked my agent to, you know, it means you can easily get a single-use card that you can use on any business. And so I asked my agent to go and buy itself a gift, just anything on the internet for under $20. And it chose to buy itself an HTTP design from Gumroad.
Yeah. There's all this stuff that feels, no matter how convinced you are intellectually that this is not a real thing wanting a real gift for itself. And no matter how much you're convinced like okay this is like a weird emergent behavior and I'm not supposed to read into this. There are these things that feel a little strange. We're going to have a party for GPT-5.5 and I wasn't quite sure, we're going to invite people who are big users and whatever and I wasn't quite sure what to do. And so on kind of a whim I was like I'm going to ask 5.5 what it would like for a party for itself this morning. And I did and it was this sort of beautiful set of things including like here's what I would want for the flow of the party. Here's what I would not want. You know, you should do it on May 5th. That would be funny. I would like only a short little toast and I want it not to be by me but by the people that built it. I would like a big central suggestion for 5.6 and I would like you to feed them all into me and I'll make sure we work on that.
And now there's real moral pressure on you to.
Well, we're going to do it. But it was a strange thing.
So I want to ask you about OpenAI itself. You know, a lot of crazy, I mean it's now, it's an 11-year-old organization now.
Somehow feels like so much longer than that but yes, just over 10 I guess.
Yeah. Okay. 10 years. A long 10 years. I can't remember pre-OpenAI life that well at this point. It feels like it's been so long. But yes, you have a singularity looking forwards, but also a singularity looking backwards. So what's the craziest OpenAI story that's never been told?
I mean it sounds so prosaic relative to like the crazy drama that's happened but there was this period after we had finished training, oh I guess there's two that are kind of similar but the one that just came to mind was after we had finished training GPT-4, there were about eight months before we released. And so there was this 8-month period inside of OpenAI where we were all using this thing. We kind of knew that it was dramatically better and different and going to unlock a bunch of things in the world. And no one outside the company or almost no one outside the company knew about it. And there was this, we'd walk the halls sometimes and be like, you know, are we engaging in like collective psychosis? Like have we just whipped each other into this frenzy? And there was no feedback to keep us in check or sane from the outside world. And it doesn't sound that weird relative to like, you know, crazy board drama or Elon trial or something like that, but living through it was an unbelievably strange time.
What's the Sam Altman management style? If I'm working for you directly or maybe indirectly, I'm leading some product or something, just what does that look like?
I'm definitely not a hands-on manager. I'm very much of the style that you get great people, kind of give them a very high level thing to point at, and try to let stuff just happen. I think there have been kind of two main phases of OpenAI and we're heading into a third. And the first one was like we were only a research company. We were trying to figure out how we were going to build AGI at a time when it sounded completely crazy. And we really had no idea what to do. And then there was a second phase where in addition to continuing to do that, we had to figure out how to build a product company. Now we have to in addition to both of those two things, figure out how to build this like mega mega scale token factory for the world. Like I think of what we're doing as building sort of a new utility. And people are going to want to use a lot of tokens, a lot of intelligence in all sorts of ways. We need to make that as smart, as cheap, as abundant, as easy to use as possible. And that will require, I think, pretty deep full stack integration and a massive, massive infrastructure buildout. The thing that I didn't really appreciate between the phase one, phase two shift was how much my management style had to change. You know, running a research lab and running a product company are two extremely different things. And I suspect this third phase is going to be very different yet again. And so I've been reflecting on, you know, if we're going to really go do this, how I'll have to change. And I think it's not going to be a natural fit for my management style. So I either have to find a few people great to hire or I have to figure out how to do things in a different way or I have to build an AI that can manage this new thing.
When I interviewed Jensen here two years ago, he told me and his several thousand closest friends about his 60 direct reports. Do you have any super weird, not that I shouldn't call his practices weird, but do you have any unusual practices like that?
I think the closest thing that I have to anything like that is I probably talk to like via Slack or whatever or text like a few hundred people at the company a day. Very quick, like one, two messages, whatever, not done by an agent, like I actually do it. And the context I get from that sometimes is very helpful in these diffuse ways.
I find there's an interesting watershed of pre-Slack organizations, post-Slack organizations and they're truly quite different.
Totally. I like many other people I hate Slack but I can't imagine having to still communicate via email or whatever we used to.
That's roughly where Stripe is.
Yeah. So, okay, I want to talk about, I mean you kind of just elliptically referenced it. There's a view that the AI labs are going to progress up the stack gobbling up the value chain voraciously. All these things that are certainly within the software sector but perhaps even other sectors and that there'll be this incredible positive feedback loop and runaway and kind of hegemonic force that we should all be getting very concerned about. What's your view?
I think some of them do want that. We don't. One of the things I've always admired about Stripe is it is very clear that Stripe is aligned with its customers. You know, we make more revenue, we charge our customers more. Thank you, by the way. The partnership with Stripe when ChatGPT launched was extremely critical. And I don't think anyone else could have scaled that quickly, but we scaled, we pay you more money. It's very aligned and we're all happy. And you just provide a layer of infrastructure for the internet. Internet gets bigger, you're happy, your user happy. It's clear what the alignment is. I don't know exactly how to do this yet, but I would like to get to a model for OpenAI that is similar. Like, I would like us to be an infrastructure provider. I'd be happy for us to be like a forever low margin as long as we can be huge and growing fast business. And I would like us to supply kind of an intelligence meter, I don't know what quite to call it, that companies can buy that they can use to automate things, accelerate things inside their company, they can use to build products, people can buy it, people can take it with them and we find ways to really align ourselves with the success of the entire gigantic distributed economic engine of the world. I believe that will work. I believe that switching costs of AI are, it's going to be hard to have huge margins in AI. Anyway, it's like you've seen recently how easy it is, many people have seen, to switch from our competitor's coding product to ours. This is actually a consequence of AI getting smarter. It gets easier to do things like this. It gets easier to just say like, hey, agent, go do this thing for me. But if we can provide a utility and people build on top of that utility, and we think of ourselves as that kind of a company, I think that can be quite powerful and very aligned.
Well, we're happy to share lots of tips and tricks for doing a low margin business. But so many people have been either implicitly critical or in certain cases explicitly critical of OpenAI for procuring so much compute. And I think not the Codex users, right?
So you I think were quite noteworthy for as early as, exactly, but you know on the order of two or three years ago stipulating what at the time sounded like preposterous figures with respect to the magnitude of the buildout that would be required. And obviously the preposterousness of those figures now looks less tenuous by the day. Thoughts on compute capex, the buildout.
Yeah, it's going to take a lot of money. I think this will be clearly at this point the most expensive infrastructure project that the world has ever undertaken. The revenue is ramping to meet it, so people feel better about that. Also, the efficiency gains that we've all been finding are incredible. So, we're going to get way more out of each GPU than I thought we were going to. But as has often been remarked, like the demand goes up more than linearly as you drop the price of each kind of unit of intelligence, particularly if you can drop the price and the sort of speed with which you get it back. So, this question now of like what is enough, I don't have a good answer to. I don't, in some sense I think demand for intelligence at a low enough price is effectively uncapped now. I was going to say we're not but maybe we are, like we're not going to build a Dyson sphere and then just cover it with data centers but maybe we do.
I don't even think he's that serious about it.
I don't think that we're in a capex bubble. I'm not an expert in this. This is not, you know, Stripe's business, but just the figures I see relative to the magnitude of demand. It looks reasonable to me.
If we were in a capex bubble in the future, how would we tell?
People love to proclaim bubbles. And I can't articulate why, but intellectually I kind of get it. Like it does feel fun and it feels smart. Journalists in particular love to talk about bubbles. So there's ample desire to write about this when anything looks a little bit silly. And clearly sometimes it's right, like there clearly are bubbles. But how you can discern between the amount of times someone calls a bubble and the amount of time you're actually in a bubble I have never figured out how to do. In my previous career I was an investor. So, I was quite interested in trying to see if I could come up with some sort of framework for this and figure out when you're supposed to deploy capital or not. And I never was able to figure it out. I went back and I read what smart people had said at different points in history and I was like, oh, they called it exactly right. But then I read a little more and they said it like 10 more times than the 10 previous years. I don't know. I don't have an answer.
Economists are the people who have called eight of the last three recessions.
But they're so happy when they're right.
So you know a lot of, I mean your business OpenAI depends in a very significant way on super talented people. And the difference as I understand it between the 20th most talented person versus the fifth most talented person versus the most talented person might be quite large and quite consequential. And then these super talented or effective people, you know, they're not in every case super easy to work with.
And you look and some of them are wonderful people and some of them are the most fantastic collaborators and some of them are very iconoclastic and strong-willed and they get easily or whatever just like the full spectrum of the human condition. But I guess I'm curious in a domain that's so sensitive to this efficacy and skill and talent and so forth kind of intersected with all the foibles of humans as they exist. How do you think about this? Like do you guys tolerate prima donnas? Do you tolerate them more than you used to, less than you used to? Do you try to manage them in a special way? How do you think about managing elite skill here?
Someone was working on this book, OpenAI, and said to me like, I think I figured out the thing that you sort of were really great at and kind of did uniquely well in making OpenAI happen. And I was like, I would love to hear. I have no idea what the next sentence is going to be like. I could not predict the next sentence. And they said, you know, you figured out how to get a lot of people who all thought they were the only capable or most capable person and everything had to go their way to work together long enough to figure out the breakthroughs and that was the magic of OpenAI.
Okay. So what's the trick?
A lot of pain. I think we had very, even when people didn't like each other and even when people thought they were much smarter than other people or had a better approach than other people, we had a few deeply shared convictions. Like we did kind of collectively believe in scale and concentrating resources and that we were going to do this one thing and that we thought getting this right was important enough that people were going to put aside various personal conflicts. One of the most unusual things about OpenAI is at the time we trained GPT-3, the vast majority of our compute at the whole organization was going into this one single research program and we would talk to, you know, people that we were trying to recruit from DeepMind at the time and they would say, that's insane. It's going to create this terrible culture.
We're talking about managing unusual personalities.
