Sam, thanks so much for being here. Thanks for doing this. So we're at DevDay 2025. Can you give us a rundown of everything you’re announcing and what you're personally most excited for. I'm excited for all of it. I am Apps in ChatGPT is something I wanted to do for a long time, but I was just talking to people about what they're building with Agent Builder. There's like a lot of stuff, or Agent Kit, There's a lot of stuff I really want to use. But yeah, maybe if I had to personally pick one thing, I think Apps in ChatGPT would be great. What are you most excited about? I like the agent builder. I think Apps in ChatGPT will be cool though, we can start there. I think 800 million weekly active users. Congrats, by the way. Thank you. It's incredible. It's clear that ChatGPT is the next distribution platform, I think. How can builders or founders take advantage of the apps SDK to build on top of ChatGPT that I think I will take like a little bit of iteration to figure out how people are mostly going to use these, like, are people always going to mostly going to call apps by name? Will people mostly want ChatGPT to like, know what they use and keep suggesting those? So I think there will be a new distribution mechanic that developers figure out that works really well for them, but, you know, these things have a way of surprising you as you put them out into the world. And you guys are releasing docs, I believe, on how you can maximize getting recommended. We are, although it'll be with the standard caveat that brand new products change quickly and we'll learn together as we go. But yeah we're releasing this today. I'm excited to read that. Let's talk about the agent builder a little bit. So back to the first Dev day. Two years ago, you guys launched GPT builder, which was great. I think I was one of the first people to build a public GPT, actually. What breakthroughs have you made since then to now to build the agent builder? I mean, the biggest one is the models have just gotten so much better. So there's a bunch of other things we'll talk about. But it is I was reflecting back on that first day and, the difference in the model capability between then and now, it's like really for 22 months or whatever it's been come a super long way or 23 months. But then we've also learned a lot of other things about how people want to build these kinds of agents. Also that they want to build them on other services outside of ChatGPT. But I think, I think it's like quite remarkable how easily you can build something pretty sophisticated, like the fact that you can just sort of have a visual builder and upload a few files and give it access to, you know, data sources or tell it what you want and get it to follow that and deploy it, you know, in a few minutes, that the thing that I was most struck by, we were doing like a rehearsal yesterday and I got to sort of look at the whole watch, the whole thing for the first time, the the ability to build impressive software fast with stuff like Codex and agent kit and whatever else that it feels like it has gone through, like a tectonic shift. And I don't think I've gotten my head yet around what that's going to mean. if I understand correctly, you can kind of build agents with no code in the agent builder. You totally can. It's more helpful if you know how to code and you can build more specific stuff. But the average everyday knowledge worker can also build agents now. Totally. So it kind of seems like it's like almost like a no code revolution for agents. As you said, if you can write a little bit of code or a lot of code, you'll be able to do more impressive things. But the barrier to entry, the barrier to get start started is like clearly super low. What do you think this is going to do for like the next wave of businesses or builders. Yeah. This is what I'm trying to get my head around. I was you know watching Romain’s demo. It was like backstage while he was on. And I was like, this would have taken so long to do this stuff a year ago. And the fact that you can do it almost live right now. It's almost like, I don't think I can come up with ideas fast enough anymore. And again, I don't know I don't know how that's going to play out, but it is like the amount of software that's going to get written in the world is clearly going to drastically increase. The time that it takes to test and improve ideas is clearly going to just go down and down and down. But I don't I don't like I really do have a hard time trying to reason through exactly what this is going to mean. I think it will mean that stuff just happens faster and that you can that you can try a lot more stuff and figure out the better ideas quickly. But I don't know what else will change. I heard you have a betting pool amongst your friends on when the first billion dollar company run by agents will be, does the agent builder bring us to that level of autonomy yet? Probably not yet. Yeah. So it used to be we had this little betting pool for when the first one person, billion dollar company will get started, and now there is not, we've not formalized this, but there's, there's lots of speculation. But when it's the first, like, zero person company. Months, years, I would expect years. Okay. But the fact that, like, we can even credibly talk about the day, you can like type A prompt into a chat bot and like this is like a very it's very incredible. Right? Yeah I agree. But wondering if I was supposed to be the year of agents. We've seen some agent products that work okay, but they still need pretty significant human oversight. And any feedback When will we get the agents that can just work for a week without any feedback? I don't think Codex is that far away from a week of work. Which again, crazy thing to say. Like not not probably not a 2025 thing, but, I was talking to some people today and they're like, I can't believe that it's doing like day long tasks now. Like, how did that happen so quickly? And it does feel to me like, like not often does AI progress feel disappointingly fast to me, but that watching the rate of improvement in the length of a task Codex can do has felt like one of those times. And I would guess week long tasks are not super far away. What are the technical bottlenecks in order to get there? Smarter models, long context, better memory. Okay. So you have agents, all these model upgrades Codex use the API now. It seems like there's so much you can build. If you went back in time and brought 20 year old recently dropped out of Stanford to present day with all the knowledge you have now what would you build and what would you not build? I was thinking about that the other day. I, I'm envious of the current generation of 20 year old dropouts. Because the amount of stuff you can build them out, like the opportunity space is so incredibly wide. I have not had, like, a real chunk of free mental space in a couple of years to think hard about what I would build, but I know that there would be a lot of cool stuff to build. And it's amazing just talking to people here today about what people are building. Yeah. It's something I struggle with and I think maybe a lot of other builders here. It's like there's so much you can build now. Is there any advice you have on, like, a unique advantage that you can hold on to while you're building these products? Distribution data? Some kind of workflow mode? I always have, like, a hard time with this question in the abstract because I think the best unique advantages like our unique like you figure it out just for you. We worked super hard to figure out our OpenAI and, I think I think that, like the general case here, is difficult. And the best answer to this question is something that you figure out that only makes sense for like what you're doing, your product, your technology, your place in the market at the right time. And that's kind of like some substantial portion of the value you create by founding something new. The one general thing I will say is you figure it out as you go, like we there's there's a great business quote that I've always loved, which is let tactics become a strategy. You can start just doing things that work and surprisingly often in the process of that will emerge something that develops into a strategy. If you had asked me when we started ChatGPT what our enduring advantages were going to be. I would have said, I have no idea or I don't know, I would have had a bunch of guesses, but I wouldn't be confident. But the things that have proven maybe most exciting there, for example, memory is a really great competitive advantage for us and a reason people keep using ChatGPT. That was not on our minds at all at the time. So you start building features and then sometimes something emerges that like, oh, this is this can be a really a durable advantage for us. Are there any hints to what those advantages you should be building towards for GPT-6 are? Or like, how should I be thinking about building a product right now that lasts? I mean, that's kind of what you have to figure out. I, I, I'd be happy to, like, brainstorm about it sometime because it would be fun. But I the degree to which OpenAI is like taking over all of my mental space and I don't get to go think about how to build a new startup is a little bit sad. I think I changed a lot of things about the world, but the things that make. Make these advantages happen for companies. It doesn't shift that... I mean, there's network effects. There is like brand and marketing advantage. There is sort of a user's data. There's like the sort of can have like marketplace effects. I think if you made a list of what has worked for this over recent years, it would look much the same now. But there's probably new tactics to establish it. OpenAI recently launched the GDPval benchmark to measure how well AI models perform economically valuable real world tasks across major knowledge worker jobs. To me it was surprising that GPT-5 came second after Claude’s Opus model. And major props for you guys are still released that were you surprised at the results at all? I mean, first of all, I, I think it'd be really bad if we didn't if we weren't willing to release things where our model is second, like, we'll always be the best at some things and not the best at others. And the way you build a culture of getting better over time is to, like, cheerfully and directly, admit when on a, you know, given benchmark or eval or anything else someone else is better than you. I think they have done a remarkable job of really kind of understanding a lot of these enterprise use cases and really beautifully formatting outputs. So I wasn't surprised. I feel inspired to get better. Has that benchmark kind of shifted the way you're building GPT-6 at all? it'll shift the way that we do some post training, but I, I think our strategy for what will go into GPT six will not change. You recently said your definition for AGI is when it outperforms humans at most economically valuable work. What GDPval score would make you say we've achieved AGI? I've been thinking about that. So first of all. Like many other people have multiple definitions of AGI. It's a the closer we get to it, the fuzzier concept the concept becomes. But the one thing that I care the most about and to my like, great surprise, we're finally in the moment. Where it's starting to happen is when I can do novel discovery, when I can, like, expand the total human knowledge base and it's these things are very small. I really don't want to overstate it. But you see all these examples now on Twitter where scientists in these different fields are saying, like, it made a small discovery or it came up with a novel approach or it figured something out. And again, don't want to overstate, don't want to understand either. Like, this is this is like the thing and the fact that we're at the very beginning of that and that we are optimistic will be able to push hard on it in the coming months and years like that. That is a really big deal. That's probably the like AGI like thing I care about the most. Is there any particular, scientific breakthrough that you're particularly excited about solving or figuring out? I mean obviously cures for diseases, discovering novel physics would be awesome. Even the stuff but like, the little things that are happening now and, you know, mathematics, for example, it's just like. This feels momentous when when GPT-4 came out, I know there's like a lot of quibbles about exactly what the Turing test is, but the popular conception of the Turing test sort of like was this thing that felt impossibly far in the future, and then all of a sudden was like, okay, it passed and society updated bsically not at all. People had like, their two week freak freakout and then said, why is this f****** thing not faster? Or it does this or it does that, or, you know, make it better. And I think that says something like wonderful about humanity. But there was the thing that had been the AI test forever, just went whooshing by and we all adapted. And I sort of feel like the same thing is going to happen now. We're just going to get used to AI making scientific discovery very quickly. It's kind of similar to like Waymo. It's only weird once. It's only weird once, only weird for like three minutes, and then you get used to it. So I see the analogy. So the recent “workslop” study by Stanford, I'm going to explain to everyone really quickly. So it's basically a new term that describes low effort output that looks polished but creates more work through rework. So it basically surveyed over 1000 desk workers, 41% say they've encountered this workslop from their coworkers in the past month. Clean up times average an hour and 56 minutes per incident, and it cost about $186 per employee per month. So if AI can 10x output for some people, like many people in this room or at this event, what's missing from education onboarding that teaches others when to use it and when not to? First of all, a lot of humans do the equivalent of like work slop as well. So this is not this is not like an exclusive to AI phenomenon. There's a lot of I get, like a particular pet peeve is, like an email that only creates work, you know, like meetings that do the same thing. So this is not new, for a long time there are people who use tools to be more effective, and there are people who use tools to like create drag in an organization. And I think we should have no reason to expect AI to be any different. The economy is self-correcting, and people and companies that use tools to get more done will have more ability to influence the future than people who use it to slow organizations down and do less. And like with any new tool, I do expect there to be a learning curve. But I think it'll I think it'll be pretty fast. Is OpenAI doing anything for education or onboarding that helps people learn these like AI intuition skills better? Yes. We totally are. People use tools the way they're going to use tools. And one of the things that I have learned is you can make great educational content and do great training, and then people find out that they can, like, make it talk like a pirate or whatever, and they're going to go do that. But yes, we do try to create a lot of content to help people use it in their workflows. And and in some cases, like certainly in Codex, adoption goes super fast and integration into workflows in a very productive way happens over a period of like days or weeks for a whole company. On the topic of slop, you're officially the world's world's first CEO with viral deepfakes all over the internet. Yeah. Does seeing Sora memes of you like the target GPU heist freak you out at all? Actually, it was one of those things that I expected to be much stranger than it is. And there was a weird way in which, like scrolling through 100 of them was much less weird than looking at one. I, I was like, thinking I was so I was, I was like, on a big trip around the world before launch. And one of the guys on the team messaged me and was like, can we, like, make your cameo thing open access? And it was one of those moments where I was like, I'm going to respond to this in eight seconds without thinking, but probably it's something I should think more about. But like, you know, this is the new technology and I, you know, shame on me if I'm not willing to like, really sort of experiment with it. So I decided I would just do it. And, and then after that I had like some time on a plane or whatever, and I was like, I wonder if this is going to be a strange experience. And I was like, yeah, it's probably gonna be really strange. And then it launched, and I had already seen some of this because people inside of OpenAI like made a lot of memes of me and other people when we were testing. But I like I was, I was in Asia it was the middle of the night. I like woke up to like do my tweet and went back to bed. And then I woke up again like six hours later or something. And I pulled the app out and it was just like meme, meme, meme, meme. And like getting in that Waymo. It was weird for the first three minutes or less. And then I was just like, okay, this is like obviously like this is an app full of generated video and these things are funny and I can like, laugh at it. And, it, it did not have any lasting strangeness. The one thing that worries me is like the watermark remover stuff. So this morning, there's a couple companies that released it. Like a Sora watermark remover. Like, what is the incentive for me to allow people to cameo me if they're going to, like, possibly remove the watermark and post it on social media? Would that affect like that would affect my personal brand badly. What's that, what's the pull? So first of all, one of the reasons we like to release technology like this is we see something coming, and we know that in some number of months or years it's going to be widely available. There will be great open source models and anybody will be able to make a video of you. Doing whatever they want with video that's publicly available of you on the internet. Society will adapt to this, of course. But one of the ways we have found to sort of help society with these transitions is to release it early with guardrails, so that society and the technology have time to kind of co evolve. And that works with text. I think video is going to be harder because video like hits. But in a way the text doesn't always. But I think we'll learn to adapt and we'll learn very quickly that there will be a lot of fake video of you on the internet, generated with no watermark by some open source model and not not that easy to trace, and it's just gonna happen. So getting society inoculated to that probably has some value. Is indistinguishable AI video actually the goal? The the goal is like AGI. I think great video will be important to that for a bunch of reasons. Spatial reasoning the things we can learn from world models hopefully someday real progress with robotics will be important. But I think great video is a good thing. Like, I, I don't want the only interface of the future to be text like, I, I'm very excited for a future where you can kind of, like, interact in a real time video stream, and it's constantly generating a new kind of UX for you. So I think that will be great, but but most importantly, I think this is, a very valuable step on the path to to real AGI. On, on Friday, you released a blog post that said you're possibly exploring revenue sharing for people who allow the cameo of their face. Can you share any details of that? How might that work? Yeah. So like in many instances when you release a new product, you find that people use it in different ways than you thought. One of the things that we thought that would happen is there would be a smaller number of creators making these very, like, cool, delightful, complex videos and sharing them. And then a big audience would be looking at them. And there's definitely some of that happening. But there is a huge amount of use of people that are making videos just for them. And like three friends and, sharing them in a feed or whatever. Sorry, in a group chat or whatever, not in a feed. And I don't know if that's going to stick or that's just novelty. But if it does stick, it very much changes the model of how much compute we'll need relative to the engagement that we get with that. And maybe we'll do things like let people pay for video generation. And, you know, if you want to generate like 100 videos a day to send memes to your friends then that's fine. And if you want to, like, generate it with some famous person in there and they they want it to happen, maybe they get some of that pay per generation as well. So we'll we'll have to experiment here. And I don't like, it's very dangerous to make any pronouncements about a six day old product, because this may all still be the novelty. And maybe this won't be a use case, but at least so far it's been a big one. Have you explored the possibility of ads in the Sora app at all? Not yet, but I think there are interesting things to do there. I mean, there's like terrifying things to do there, to, but unlike ChatGPT, where we can run a nice business model where people pay us a subscription if if the version of Sora that people want to use is like looking at, you know, funny content or beautiful content in a feed, that probably is a more natural, ad driven model. If it's mostly DMs, that's a different kind of model. So I think, like optimistically, by the end of this year, I think or maybe that's too optimistic. Maybe by like the end of the first quarter, we could kind of understand what the product is going to look like, and then and then figure out a business model that goes along with it. I mean, there's like certainly I think charging per generation makes sense and is something we should go try, but all the other ideas would depend more on how the product evolves. That’s fascinating. On the topic of jobs. So, jobs are changing. I like to use the farmers analogy. So, like, if you tell a farmer 50 years ago that this magical thing called the internet, is going to create a billion new jobs, and you're sitting from a desk and there's a developer, a marketer, he probably wouldn't believe you. And similar to this era we're in now, I think you have many others have echoed that there's going to be many new jobs created. Yeah. So and it's really hard to see what those jobs will be. But the difference between this era and the internet era is that the internet era, there's a billion new jobs created out of kind of nothing. Whereas the intelligence era, a billion knowledge workers jobs are arguably going to be impacted first before new jobs are created. Does that worry you at all? Totally. I, I think you touched on something really important there that both makes me a little bit less worried and more worried in some other way. The thing about that farmer is, not only would they not believe you that this thing was going to happen. They very likely would look at what you do or I do and say like, that's not real work. Like that's you know, if you're like farming, you're doing something people really need. You're making them food, like you're keeping people alive. This is real work. And you, all you people of the future, like life just got too easy for you. There's, like, abundant food and abundant wealth and all this stuff. And, like, so you have both of you have access to all the food you need, and you're doing something that's like playing a game to fill your time and your need to feel important. And it's not a real job to us, though it feels real. Certainly to me, it feels real. And I'm grateful to get to work on something that is, like, satisfying and seems important. It's very possible that if we could see those jobs of the future, we would be like, okay, well, you know, maybe our jobs were not as real as a farmers job, but it's a lot more real than this game you're playing to entertain yourself. And so I can very much believe, like, I'm I think people will still have tons of meaning in their life, but I can very much believe that what we think of as work really changes. And from our reference perspective, it would maybe look more like less like work than what we do now. I can also believe that the sort of social contract changes a lot, especially if there's this transition where it happens pretty fast. But I so I have like some short term worries as we go through this transition. But man, I'm so willing to bet on human drives being what they are. And, I think we'll find plenty of things to do. Think we'll be exploring space? BCI? Is there anything specifically that you think we're going to like focus on after AGI? I mean, I kind of hope everything just goes in every direction. And we go. We go do all the things like space does seem very cool to me, but probably to you or you or you like there's other things that you think are really cool and I like. I hope this sort of like canvas of possibility is just super wide open. If you could set one global policy tomorrow, what would it be? It's really hard to narrow that down to a single one. But the thing that kept coming to mind was I think there's like a big debate about AI regulation and whether that's good or bad and whether it's going to advantage, whether it would advantage big companies too much. And I can see a lot of sides of that. But I do think as we get to super powerful models, there should be some global framework to reduce catastrophic risk. Just of like the very leading edge, how we're going to think about the biggest, like the really big safety issues. And if there could be a global policy for that, that would seem very good. And could you outline that like a high level? I would take just like a good testing framework to, to start. So another thing I want to touch on is WeChat. So in China they have kind of like an everything app you can shop on it, you scroll social meida on it, chat with your friends on it. It almost seems now like with ChatGPT you have shopping, web search, Sora, all these new features. Are you attempting to build like kind of an everything app for America? No. There's a bunch of reasons why I don't think that approach would work in this market in the same way, but no, we're trying to, like, build a really great AI super assistant. What's the goal of, like, launching all the features separately? Like, for example, Sora is a separate app. Why don't you launch it directly in ChatGPT? ChatGPT has become, for many people, like the most personal account they use, and I think combining like a social experience in there would have felt strange. There are things you can imagine, like messaging, because people do share these a lot and collaborate, but I think the way people think about their ChatGPT account and the way they think about like an entertainment app are extremely different and would have had some discordance. But a lot of things, of course, we do put in ChatGPT you. Let's talk about the agent builder again What do you think is, like, really important or useful agents that people can build that you're most excited for? I think looking at what's happened with Codex and thinking about that for a lot of other industries, like can there be a codex like experience for law or building financial models or any of the other things of which we've seen good existence proof? There's like amazing startups that have started doing these things? But as the technology has matured, if you can get those things to be as good in their industry as Codex is at coding, that's probably the area I'm excited most about. I mean, I can imagine a world where you can, like, start a startup, this one person startup, you can start a startup by just like talking to a bunch of agents. And I don't think Agent Builders is good enough. Agent Kit is good enough yet. But I can see a path from here to there. I know you briefly, you mentioned in the keynote that you think voice is kind of the final interface for AI or agents. Can you touch more on that? I don't think it's the final interface. Like there's a lot of times where voice is not the right interface. You know, if you're like on public transportation or something and you're that person, like talking to it, that's kind of annoying. But there are a lot of times where voice is the very natural interface language in general, but sometimes voice, sometimes typing it out is underexplored. I think we all got used to like, you know, smart speakers are sort of made fun of as a category, but we forget that a lot of people use them and really love them, and they're not very good yet. They just, we didn't have good enough AI. And we didn't have like a good enough infrastructure around them. So if you think about what it could be like if you just talk to your device and it does exactly what you want and then it kind of gets out of your way. That seems like that is a kind of computer I would like to use. Any hints at io? Like what are you guys building there, is it going to be you know, you interface with it with voice? It's going to take us a while like it is. This is a place where we just have to like, ask for patience, to do a new kind of device at a super high quality level. At massive scale. That's a totally new way to use a computer. We need the kind of creative space to figure that out. We have some ideas. We're really excited about. But nothing we're ready to talk about. And we're not going to be for a while, but we will try to make something that is very much worth the wait. Okay, well, that's it, that's all the questions I have for you. Thank you very much. Enjoyed getting to do this. Thank you. Thanks.