Greg Brockman on Navier-Stokes, Astra's breakthrough in computer use, ChatGPT health at 300M weekly users, and OpenAI's startup-mode refocus

Speaker 2: very, very fun to see. So, I'm sure that they will continue cooking, and we will have to check-in with them soon. But we have Greg Brockman, the cofounder and president of OpenAI with us. Welcome to the show, Greg. How are you doing? Doing great. Thank you for having me. Thanks for hopping on. Huge day. Can we start with the with the math advances? What is what what is happening? Why is this important? There's a lot of back and forth in the timeline, but I'd love for you to just set the table for us on what actually happened with Navier Stokes today.

Speaker 6: Well, it's always a huge day in AI and modern scientific progress, I would say. Yeah. Today, we announced that our model had solved the Navier Stokes problem that we found a counter example or a sort of proof that you can actually that these these theoretical equations do have a singularity or kind of breakdown under certain circumstances. Interesting. And this is a problem that has been open for a very long time. It's one of the seven millennium problems Yeah. Which are kind of some of the deepest, most important problems in mathematics. And I I think that the problem itself is important. Right? This is new knowledge for humanity, that the proof itself is actually very elegant and beautiful and I think that there's a lot to learn from it. Yeah. The the equations have lots of application and fluid dynamics in in other areas. But to me, what's even more important is about what this represents about where we are in terms of model capabilities. Yeah. And the fact that we can actually generate new knowledge, that we can learn from these models to have them help us solve problems that are otherwise outside of reach or would take us very long to solve. Yeah. I mean, where should I actually go with this? Is this going to help me book a flight? Is this going to help me cure cancer? Is this just going to help you recruit researchers who are fascinated by this stuff? Because I think that there's this has taken over the technology world,

Speaker 2: but I imagine that this will not be something that gets talked about at backyard barbecues with friends and family that are three clicks removed from Outside of SF. Yeah. Yeah. Well, look, I think that there's, first of all, the applications

Speaker 6: of this specific result or the equations themselves, right? Which are things that let us better understand phenomena from ocean currents to airflow around aircraft Sure. Around turbulence and things like that. But it's really about the broader insights and methods Mhmm. That can help scientists and mathematicians further accelerate their research. And I think that again, of if we have models that can help solve this kind of problem, then what happens from here? Like what other problems that are immediately applicable? And I think that talking about curing diseases and new medicines we're going to be able to develop, all of that starts to become much more real when you have models that are at this level of assistance and capability and I do think that there are going to be real changes to think about in terms of we can have so much more ambition with the kinds of of challenges that we can hope to to tackle now. Yeah. I I mean, I think even even if this doesn't break through to the broader

Speaker 2: the broader world, the weekend definitely will because it seemed like everyone was talking about Blender, talking about Astra, building stuff. Take us through the launch of Astra, what the feedback has been, what you've learned. It seemed like there were a couple resets, the model scaled very well. How was this launch different than previous launches?

Speaker 6: Well, first of all, I've just been blown away by the community reaction to Astra. It's been really amazing and very humbling honestly to see all the creativity and the different ways that people have been applying the model. Mhmm. And I think that it's very clear that we've reached a new threshold of computer use, so this model is able to really work with different kinds of applications in a way that was not previously possible. People are taking full advantage of that fact and really thinking about how to create. Lots of people showing off three d creations and mapping out physical locations and turning them into into these three d models and thinking about can you use this for design of physical parts. Someone talked about how they were designing some mechanic for catching hair in a shower drain and that they were able to

Speaker 2: super cool, right? That they're now able to actually manufacture that. Shower hair super intelligent. Yeah. No, that stuff's so mundane but it's so important. I I feel like, a lot of this stuff gets lost. Right? Exactly.

Speaker 6: And I think there's a core there that's very important which is that we are talking about these grand challenges sometimes or very esoteric applications, but really the everyday, the number of problems that you have in your life Yeah. That you would love to solve. Yeah. It's now possible, right? Yeah. That that we're really trying to empower the individual to make it so that you you can be you can have superpowers, that you can accomplish more. Yeah. And I think that that, you know, really trying to benefit people, empower people, build tools that can really help you and help you in your your daily life. Yeah. That's all part of what we're we're we're working on. So yeah. I mean, it seems like a huge

Speaker 2: number of people in in tech effectively rebuilt their entire house in Blender and plan to remodel this weekend. Thanks to Astra. And but I am wondering about the merge and how you bring together Codex, Chatuchi PT work, Chatuchi PT on desktop. I have a gaming PC now with a NVIDIA card in it. I have a Mac mini. Like, I have all these things, and I can imagine that I'm just I'm doing that unnecessary or, like, it's fun for me, but that early adopter work of going and unhobbling it a little bit here and there. But in the future, this will all just be tucked in one, prompt box, and it might build a three d blender model for to answer my question of should I remodel my house or not. But how do you see the capabilities that we saw on display from sort of light power users over the weekend actually making their way into consumers who might not even know what Blender is?

Speaker 6: Well, I think you're exactly right that we really want to shift these tools from requiring kind of a low level sort of Yeah. Access or guidance to really having the human be able to fly, right, to really be empowered, for you to be able to set the goals and the objectives and that you still should feel like you can get into those details and you can understand them, you can provide that oversight because ultimately, you should feel accountable for the outcomes, but you have this this absolute amplifier, right? Like a trampoline or like a, you know, rocket ship for the mind, like however you want to to to analogize it. And I think what that means at a practical level so first of all, this year we've been really seeing this shift from just pure chat use cases Mhmm. Through Jentic use cases. But I would also keep in mind that chat is alive and well. I mean, we're we're now well over a billion users every week that you can see that the market share of ChatGPT is starting to climb once again Yeah. Because we've been investing so hard in in so many use cases that are important for people in education Yeah. And health, and a variety of other areas. Then at the same time, these productivity deep knowledge work use cases, those are really taking off. We've had this like almost vertical wall of agentic adoption since we launched ChatGPT work. Yeah. And I think that the fact that these are two distinct modes, that that is actually a point in time. Is something that we're continuing to unify and merge and that we're starting to see that there's a new emerging form factor for how people want to consume AI. And I think that it's almost like that the promise of AI has always been that you have something that you can talk to and really delegate work to that's proactive and persistent. And that I what we were promised, if you were to rewind five, ten years ago, is never a low level language model that you have to think about context windows, and you have to select thinking strength, you have to select different models. Like, none of that. None of that is is the future. And so I think that we're moving towards real amplification, real giving you time back, real having computers that are able to operate according to your goals, to your desires. And I think that that is a core of it. We're developing this safely. That's that's one of the core commitments that that we make and how we think about this. But but we really see the power of these tools starting to really really start to increase in terms of what people are capable of, and that under the hood, utilizing tools like Blender so that that is almost a detail that fades into the background is absolutely the direction of travel. It feels like in AI particularly,

Speaker 1: there's been a almost like a first mover disadvantage in that billions of people have tried ChadGPT, and some percentage of them tried it for the first time and have a certain impression of the product and what it can do. And then, you know, even in the last few weeks, there's been new agents and products that come online, and people try it, and their mind is just completely blown. And I think it's funny because I'm like, well, as somebody who's trying to get the absolute max out of ChadGPT, I'm like, well, I've been running like I've had an agent running that, for example, will tell me every time a SpaceX launch is gonna happen and if it gets delayed. Right? These sort of persistent agents that are running the background. Strategically, I feel like it's a new kind of challenge because you have this As capability has been scaling, first movers need to be almost like constantly reminding the the market of all these just like new ways The ability to overhang.

Speaker 6: Yeah. We we think about this a lot. And I think that there's this discovery problem that we as a field need to really encounter in a first class way. And we haven't done it fully yet, but I think we have a real shot at solving it better than any products before because the the thing that that right now we kind of rely on, you think about chat, gbt, chat, gbt work, these are both text boxes. Mhmm. Right? And it's like, well, this new text box is way more powerful than the old text box and but, oh, there are some reasons that you still want to use the old text box. It's not you know, it's like far too confusing. Yeah. People just want something that can help them solve their problem. The whole point is to get your time back, not for you to have to go and become an expert in all these these internal details. But at the same time, we also have a model that understands what you're trying to accomplish. Right? That you're explaining to it. Here's what I want. It has a lot context on you, and so it should also be able to proactively say to you, hey, actually, if you ask me this other way or if you added added this connector or if you authorize me to do this or if you, you know, hook up your credentials in this in this way, I can go and do this other thing for you. So we're thinking lot about that self knowledge, that onboarding process. I think that is a huge, huge opportunity. I think that there is both the disadvantage that you cite of people tried it, they form an impression, and it's changed that it's something new. But there's also an advantage. I mean, ChatGPT, like over a billion users every week, like that is unique. One has that kind of use on these models. And I think that the number of people who have tried ChatGPT before, I think it's another billion, billion and a half, something like that. That's a huge opportunity as well for us to go back to those users and say, Hey, we can now solve the problem for you. We can now help you in ways that you didn't see before. And I think that it's it's true. It's real. Like, you look at how many people use chat for health, 300,000,000 people every single week with health queries. Right? And that that's really making a difference in people's lives and that of their loved ones. And so we have such opportunity Yeah. Can you stay there with health and and and explain where these goes? One note before that. Something that I think is really interesting and I think something that OpenAI can can do a lot better is, like, when people talk about like, everyone in AI wants to be like the apple of AI from a marketing standpoint. And when they when you think, like, Apple marketing, you're thinking like Mac versus PC or you're thinking 1984,

Speaker 1: these big branding campaigns. But the actual thing that Apple does really, really well with marketing is they just hammer really, really specific details about their products. Right? They're like, they're advertising the new camera. They're advertising Memoji. Right? They're advertising like certain features in Safari. Right? And so it's like, with AI, the surface area of like things that you need to communicate is actually like an order of magnitude greater because it can do so many different things. So I think that it's such an opportunity for the company to focus advertising.

Speaker 6: Yeah. The brand campaigns are awesome and like the the launch video for ASTRO was amazing, but it's like there should be billboards running of like very specific things that ChachyBT can do to give you back your time. Yeah. Yes. This this has actually been a real sort of realization or just like something that I have really come to over the course of of of this year. And if you look at even for example, the you know, we just announced ChatGPT images 2.5. Mhmm. And if you look at the launch video there, the thing that I love about it is it shows here's someone creating an image and here's like a bunch of different variations of it, and then here's them taking their favorite one and having it in the world. Like someone said, here's a cool like little sketch of a candle holder, you see an awesome visualization of it, and then you see the physical candle holder and you're just like, that's what you want. Right? It's like you it speaks to you immediately. And I think that that really showing people here's a use case, and the thing that was also a little surprising to me is that we've sometimes highlighted esoteric use cases, something that appeals to someone in particular and it's amazing for that person, but people don't then say, oh, because it's this powerful, I can also do this other powerful thing I've been waiting on. Like that connection is is is is something that is less sort of easy to make than I than I'd realized and it makes sense. What you want is for there to be use cases where people say, I actually want that particular thing. Now let me go try it myself. Then from there you start exploring and you start to find, I actually do have this powerful use case that I didn't even realize was tip at the tongue. Yeah. The other

Speaker 1: is probably like the best example of that working really really well. The other thing is is reminding people to ask the AI what it's capable of. Oh, yeah. Like, I was I was having lunch with a buddy who's a real estate developer, and he is using chat all day long for different like deal memos and to understand like projects that he's working on, all this stuff. And he'll ask me he'll ask me all the time, can can Chad GPT either do this or that? I'm like, I'm happy to answer you.

Speaker 2: But I'm like, have the thing that will just explain exactly how to do the thing that you wanna do or or maybe not, but it probably can't. Yeah. I I I did the same thing. I was kicking off this blender thing and I was like, should I run this as a local codex thread or in the cloud? Let me know which one's better based on my system, and it gave me a good answer, and I was able to go for it. On on Images,

Speaker 1: when when Images two came out, I had some moments where I thought, okay, Images is solved. Yeah. Like, where do you where do you think Images actually go as a category? Because it felt like this has been something that has maybe one shot me more than anything else. Specifically with like when it when it released, I was I was spending hours like does like on a Saturday trying to design furniture, right? And just going through, like, hundreds and hundreds of prompts. But how far can image models go and where are they going and what are the ways in which you think they can be applied? We were talking earlier too about the downstream impact of image models. If you can take, you know, a a physical space somewhere and take a picture of it and imagine it as all these other variations, there's so much, like, real world activity that will be driven from that because people can see this thing Yeah. Visually and say like, now I want to go make that reality, which I think is really cool.

Speaker 6: Well, think that's exactly the right way to think about it is as you hit new thresholds of capability, my experience has always been that fundamentally new applications become unlocked in ways that you almost wouldn't have thought about ahead of time. And so I think that within, for example, knowledge work, professional work, marketing, all those areas, you just need to be above the quality threshold. If you're below it, it's a cool concept, but you can't actually use the final material. Right? That that then means that you you haven't really solved the problem. And that having precise edit control Mhmm. Being fast, and really being creative, and having diversity of different results, and also being able to have this this good interplay back and forth with the person, I think that that really unlocks whole new use cases. And I think there's a huge market there, and even for example, the kinds of things you may not think of naively, but actually start to be really important applications we're seeing happening is slide creation or making awesome websites. Right? Being able to have that image generation capability in the middle is something that's very unique to OpenAI relative to some of our competitors, I think that you're able to then produce much better artifacts downstream. And so we really view images, we view voice, we view coding, all these capabilities as one package that are going to come together to create an AI that empowers you. That means you can create anything that you imagine, and I think it's going to be something that's just unlike anything out there. Let's go back to health.

Speaker 2: I think most people already are aware that you can synthesize some lab data with some sleep scores. But your vision that you laid out recently for where that product goes is much more complex, much deeper. So tell me where ChatGPT Health is going in the future.

Speaker 6: Well, I would think of it as there are three sides to what we do on health. There's the consumer side, again, 100,000,000 Yeah. People every week with health queries. There's the clinician side, which is bottoms up, and that's really about think about a chat GBT that's really tuned for clinicians that gives them direct citations to medical literature, things like that. There's a third pillar which is the enterprise side of selling directly to hospitals and them enabling it, and you can see things like we have an integration within Epic and really trying to bring each of these three pillars the best possible service independently. But you think about as those really build momentum that you actually are able to get synergies across them, right? That there's something that actually makes the health experience and the ability to really transform healthcare in America and the world on the table because there's so many the thing about how much work you as a patient have to do if you're talking to different specialists, you have to carry your medical record from one to the other, you have to explain again, here's the issue, and that ultimately you're on the hook. You're the doctor who has to make the decision whether you like it or not. And actually being able to have just good sharing of that information across different providers, that becomes possible if everyone's on one platform. Or think about clinical trial enrollment. That's a huge bottleneck to drug development and finding people who are eligible and will benefit from being enrolled in a particular trial. And if you have that kind of data, if people are willing to sort of entrust you with that information, that that's something that can actually really benefit them and benefit the world at the same time. And so what I view us as building is really trying to build the world's best healthcare platform to really be able to bring healthcare into the AI age. And I think that it's something that is going to be absolutely transformative to many people's quality of life, really uplift so many people, and we're seeing it already in such concrete ways. Some of my favorite stories about chat GPT are people who say, hey, information I got from chat helped me save my own life, helped me save that of a loved one, that there was this medical issue that someone had and that I if, the doctor told me one thing, was able to double check that and understand what they were saying and be able to push back on it and got to a good outcome. And that happens every single day. So I think that health with these AIs is something that we're still just scratching the surface of what's possible and I think it's one of the most positive applications of AI that you can think of. Yeah, I'm very interested to see how

Speaker 1: the advancements in memory intersect with health because I expect that ChatGPT with where memory has gone, where I'll be in a new thread and it will bring up just the right information or tie back to a thread that maybe happened three weeks ago, when you actually apply that to health related queries, it may be able to pick up patterns that sometimes would take a human years to figure out, like a certain ailment or something like that, where it's like, hey, you're asking about all these different things and maybe you thought they were not connected. Turns out they they actually are and you should go down this sort of like rabbit hole. Mhmm. Last question. I'm sorry. Please. Alright. I was I was gonna say, I think that's absolutely right. We're seeing that very concretely and we've seen here's in my own personal life.

Speaker 6: My wife, you know, we talked about some of her medical conditions publicly, but it was really this five year journey of talking to many specialists, each one who was kind of touching one part of the elephant and would try to address that one part and it was only finally her allergist who said, hey, I think all these symptoms you're seeing are connected and you have this genetic condition that affects all of your subsystems. And that's the kind of thing where it's really hard to say how many people have similar kinds of conditions and just never find out. How many people have these areas where it's like if you just sort of are functionally specialized that you're never going to bring together the whole diagnosis. And I think that is one of the powers and potentials of an AI that really deeply is able to help you across all parts of your life and also is a deep domain expert in all areas of medicine. Yeah.

Speaker 2: I have one last question. What how do you tell the story of Operator? It feels like it was a failure or a side quest, but it feels incredibly important now given the advances in computer use. Is there a clear lineage there? What was operator? Does that still exist somewhere within ChatGPT? How does how did how did computer use get solved?

Speaker 6: Yeah. Well, I would look at all these things as timing and all about iterative deployment. Yeah. Right? That there's a moment where you need to where the capabilities aren't quite there, but actually learning from real world deployment is very helpful. Sure. Right? I think operator was just kind of below threshold in terms of the model capabilities. It was a cloud based system that operated with computer use, but it was slow, it wasn't fully accurate, it was like pretty painful to to use. Some people got value, but it really wasn't above threshold. If you look at what's happened, one thing OpenAI does very well is we make long term investments on things that really matter and we do the grind. The team this year, I think really started to build momentum that we put in a lot of effort to go and sort of burn down a long list of issues. We're able to really focus on solving computer use and I think that they delivered in a significant way. Yeah. And there's more to do, never done all those things, but it's a true milestone. I think people are really appreciating what's possible because we've been in a world with these agents using computers through connectors, right? Through these very painstakingly coded systems that are so different from how humans use computers whereas humans can already use everything on a computer, right? Everything is designed for people. So if you have an AI that can operate that way, and even from the very beginning of OpenAI, we had a dream that one day we could create such an AI. It becomes able to help you across everything that you would be able to do with a computer yourself. And so I think we're there with Astra. I think that there's just so much more that people are going to uncover in terms of application where this can go, but it's an example of long term focus, doing the work, and not giving up, even when the going gets tough. Yeah. And it feels like Astra,

Speaker 1: my view is it really felt like all these different bets coming together Mhmm. You know, at the right at the right time. Right? The the advancements in Yeah. In the model itself, the the computer use, voice, all these things. And it it's all it's all making sense. So Yeah.

Speaker 6: Together, it's focus. It's something that I think this company does extremely well when we really put a challenge in front of us and think about how to accomplish it safely, well, and to really deliver the value.

Speaker 1: The thing. Put differently, it felt like when you look at last year, it felt like OpenAI was operating like a big company, and was a big company, but the way in which product experimentation was happening was, you know, when you think of, like, a hyperscaler, they'll launch a new product thinking, okay, if there's a 20% hit rate or even a 10% or a 5% chance, that's okay. And then this year, it feels like, actually, the entire company switched back into actual startup mode, which is like, no, focus, focus, focus. All these things need to come together. The whole team needs to be rowing in the same direction, and then the difference in momentum and growth and all these things coming from that, and actually taking the company from operating and shipping more like a big company to shipping again and focusing like a startup has been feels like an impossible task and it's been incredible to watch.

Speaker 6: Yeah. Has. Thank you. No. It's been a real real effort from many many people at OpenAI to really bring together and something that that I really value that I think we really value as a company. And I think that we're just so laser focused on our mission and really thinking about every piece of what we do should add up to helping us accomplish it. Well, thank you so much for taking the time to come chat with us. Great to see you. We'll talk to you soon, Greg. Thank you. Talk to you Appreciate it. Rest of your day. Cheers. Goodbye.