Sam Altman: How OpenAI Wins, AI Buildout Logic, IPO in 2026? (YouTube Video Transcript)

(00:00:00) Your YouTube transcript will appear here (00:00:00) You know that 1.4 trillion you (00:00:01) mentioned, we'll spend it over a very (00:00:02) long period of time. I wish we could do (00:00:04) it faster. I think it would be great to (00:00:05) just lay it out for everyone once and (00:00:08) for all how those numbers are going to (00:00:09) work. Exponential growth is usually very (00:00:11) hard for people. OpenAI CEO Sam Alman (00:00:14) joins us to talk about OpenAI's plan to (00:00:16) win as the AI race tightens, how the (00:00:19) infrastructure math makes sense, and (00:00:21) when an OpenAI IPO might be coming. And (00:00:24) Sam is with us here in studio today. (00:00:26) Sam, welcome to the show. (00:00:28) >> Thanks for having me. So, OpenAI is 10 (00:00:30) years old and crazy to me. (00:00:32) >> Chachi PT is three, but the competition (00:00:35) is intensifying. Um, this place we're at (00:00:38) OpenAI headquarters was in a code red is (00:00:41) in a code red. Um, after Gemini 3 came (00:00:44) out and everywhere you look, there are (00:00:47) companies that are trying to take a (00:00:49) little bit of OpenAI's advantage. And (00:00:52) for the first time I can remember, it (00:00:53) doesn't seem like this company has a (00:00:56) clear lead. So I'm curious to hear your (00:00:58) perspective on how open AI will emerge (00:01:01) from this moment and when first of all (00:01:04) on the code red point we view those as (00:01:06) like relatively low stakes somewhat (00:01:08) frequent things to do. Uh I think that (00:01:10) it's good to be paranoid and act quickly (00:01:13) when a potential competitive threat (00:01:14) emerges. This has happened to us in the (00:01:16) past that happened earlier this year (00:01:17) with Deepseek. Um and (00:01:20) >> there was a code red back then too. (00:01:22) >> Yeah. There there's there's a saying (00:01:25) about pandemics which is something like (00:01:26) when when a pandemic starts (00:01:30) every bit of action you take at the (00:01:32) beginning is worth much more than action (00:01:33) you take later and most people don't do (00:01:35) enough early on and then panic later and (00:01:37) certainly saw that during the covid (00:01:38) pandemic. Um (00:01:41) but I sort of think of that philosophy (00:01:42) as how we respond to competitive (00:01:44) threats. Uh and you know it's I think (00:01:48) it's good to be a little paranoid. (00:01:49) Gemini 3 has not or at least has not so (00:01:51) far had the impact we were worried it (00:01:53) might but it did in the same way the (00:01:55) Deepse seek did identify some weaknesses (00:01:57) in our product offering strategy and (00:01:59) we're addressing those very quickly. I (00:02:01) don't think we'll be in this code red (00:02:03) that much longer. Uh you know like these (00:02:06) are not these are historically these (00:02:09) have been kind of like six or eight week (00:02:10) things for us. Um (00:02:13) but I'm glad we're doing it. Uh just (00:02:15) today we launched uh a new image model (00:02:17) which is a great thing and that's (00:02:18) something consumers really wanted. Um, (00:02:20) last week we launched 5.2 which uh is (00:02:22) going over extremely well and growing (00:02:24) very quickly. Uh, we'll have a few other (00:02:26) things to uh launch and then we'll also (00:02:29) have some continuous improvements like (00:02:30) speeding up the service. But, you know, (00:02:32) I think this is like my guess is we'll (00:02:35) be doing these once maybe twice a year (00:02:37) for a long time and that's uh part of (00:02:40) really just making sure that we win in (00:02:42) our space. Um, a lot of other companies (00:02:45) will do great too and I'm happy for (00:02:47) them. But, you know, CatchBT is still uh (00:02:50) by far by far the dominant uh (00:02:54) chatbot in the market and I expect that (00:02:56) lead to increase not decrease over time. (00:02:58) Um, (00:03:00) the the models will get good everywhere, (00:03:03) but a lot of the reasons that people use (00:03:04) a product, consumer or enterprise, uh, (00:03:07) have much more to do than just with the (00:03:08) model. And we've, you know, been (00:03:11) expecting this for a while. So we try to (00:03:13) build the whole cohesive set of things (00:03:16) that it takes to make sure that we are (00:03:18) you know the product that people most (00:03:20) want to use. Um I think competition is (00:03:22) good. It pushes us to be better. Uh but (00:03:25) I think we'll do great in chat. I think (00:03:28) we'll do great in enterprise and in the (00:03:29) future years. Other new categories I (00:03:32) expect we'll do great there too. I I (00:03:34) think people really want to use one AI (00:03:36) platform. People use their phone at (00:03:39) their personal life and they want to use (00:03:40) the same kind of phone at work most of (00:03:42) the time. We're seeing the same thing (00:03:43) with AI. Uh the strength of chatgbt (00:03:45) consumer is really helping us win the (00:03:46) enterprise. Uh of course enterprises (00:03:48) need different offerings but people (00:03:50) think about okay I know this company (00:03:52) open and I know how to use this chat GPT (00:03:54) interface. Um so the strategy is make (00:03:58) the best models build the best product (00:04:00) around it and have enough infrastructure (00:04:01) to serve it at scale. (00:04:03) >> Yeah there is an incumbent advantage. uh (00:04:05) chat I think earlier this year was (00:04:07) around 400 million weekly active users. (00:04:09) Now it's at 800 million reports say (00:04:12) approaching 900 million. Um but then on (00:04:14) the other side you have distribution (00:04:16) advantages at places like Google. And so (00:04:20) I'm curious to hear your perspective if (00:04:21) the models do you think the models are (00:04:24) going to commoditize? And if they do (00:04:26) what matters most? Is it distribution? (00:04:28) Is it how well you build your (00:04:30) applications? Is it something else that (00:04:31) I'm not thinking of? I don't think (00:04:35) commoditization is quite the right (00:04:36) framework to think about the models. (00:04:39) There will be areas where different (00:04:42) models excel at different things. For (00:04:44) the kind of normal use cases of chatting (00:04:46) with a model, maybe there will be a lot (00:04:48) of great options. For scientific (00:04:50) discovery, you will want the thing (00:04:51) that's right at the edge that is (00:04:52) optimized for science perhaps. Um so (00:04:55) models will have different strengths and (00:04:58) the most economic value I think will be (00:05:02) created by models at the frontier and we (00:05:04) plan to be ahead there. Um and we're (00:05:06) like very proud that 52 is the best (00:05:09) reasoning model in the world and the one (00:05:10) that scientists are having the most (00:05:12) progress with but also um we're very (00:05:14) proud that it's what enterprises are (00:05:15) saying is the best at all the tasks that (00:05:18) a business needs to to you know do its (00:05:20) work. Um (00:05:23) so there will be you know times that (00:05:24) we're ahead in some areas and behind in (00:05:26) others but the overall most intelligent (00:05:28) model I expect to have uh significant (00:05:31) value even in a world where free models (00:05:33) can do a lot of the stuff that people (00:05:34) that people need. The the products will (00:05:37) really matter. Distribution and brand as (00:05:39) you said will really matter. Um in (00:05:41) chatbt for example personalization is (00:05:43) extremely sticky. People love the fact (00:05:46) that the model, get to know them over (00:05:47) time, and you'll see us push on that uh (00:05:49) much much more. Um, (00:05:52) people have experiences with these (00:05:54) models that they then really kind of (00:05:57) associate with it. Uh, and you I (00:06:01) remember someone telling me once like (00:06:03) you kind of pick a toothpaste once in (00:06:05) your life and buy it forever or most (00:06:07) people do that apparently. Um and people (00:06:10) talk about it. They have one magical (00:06:12) experience with ChachiPT. (00:06:14) Healthcare is like a famous example (00:06:16) where people put their um you know they (00:06:18) put a blood test into Chachi or put the (00:06:20) symptoms in and they figure out they (00:06:21) have something and they go to a doctor (00:06:23) and they get cured of something they (00:06:25) couldn't figure out before. Like those (00:06:26) users are very sticky. Uh to say nothing (00:06:28) of the personalization on on top of it. (00:06:31) Um (00:06:33) there will be all the product stuff. uh (00:06:37) we just launched our browser uh recently (00:06:41) and I think that's pointing at a new uh (00:06:45) you know pretty good potential mode for (00:06:48) us. Uh the devices are further off but (00:06:50) I'm very excited to to do that. So I (00:06:52) think there'll be all these pieces and (00:06:53) on the enterprise uh what creates the (00:06:55) the mode or the competitive advantage um (00:06:58) I expect that to be a little bit (00:06:59) different but in the same way that (00:07:01) personalization to a user is very (00:07:02) important in consumer there will be a (00:07:04) similar concept of personalization to an (00:07:06) enterprise where a company will have a (00:07:09) relationship with a company like ours uh (00:07:12) and they will connect their data to that (00:07:14) and you'll be able to use a bunch of (00:07:17) agents from different companies running (00:07:19) that and it'll kind of like make sure (00:07:21) that information is handled the right (00:07:22) way and I expect that'll be pretty (00:07:24) sticky too. Um we already have more than (00:07:27) uh a million people think of us largely (00:07:28) as a consumer company but we have (00:07:30) >> we're going to definitely get into (00:07:31) enterprise. (00:07:31) >> Yeah. You know like (00:07:33) >> share the stat. (00:07:35) >> Well actually (00:07:36) >> a million (00:07:36) >> we have more than a million enterprise (00:07:37) users but we have like just absolutely (00:07:41) rapid adoption of the API. Um and like (00:07:44) the API business grew faster for us this (00:07:47) year than even Chad GPT (00:07:49) >> really. Um so the enterprise stuff is (00:07:51) also (00:07:54) you know it's really happening starting (00:07:55) this year. Can I just go back to this (00:07:58) maybe if commoditization is not the (00:08:00) right word model some maybe parody for (00:08:03) everyday users (00:08:04) >> uh because you you started off your (00:08:05) answer saying okay maybe um everyday use (00:08:08) it will feel the same but at the (00:08:10) frontier it's going to feel really (00:08:11) different. Um when it comes to chat (00:08:14) GPT's ability to grow um if I'll just (00:08:18) use Google as an example. If Chat GPT uh (00:08:21) and Gemini are built on a model that (00:08:23) feels similar for everyday uses, how big (00:08:26) of a threat is the fact that you know (00:08:28) Google has all these surfaces through (00:08:29) which it can push out Gemini whereas (00:08:32) Chat GPT is is fighting for every new (00:08:34) user. (00:08:35) >> I I think Google is still a huge threat (00:08:38) uh you know extremely powerful company. (00:08:41) If Google had really decided to take us (00:08:44) seriously (00:08:46) in 200 (00:08:49) 23, let's say, we would have been in a (00:08:51) really bad place. I think they would (00:08:53) have just been able to smash us. Um, but (00:08:55) their AI effort at the time was kind of (00:08:57) going in not quite the right direction (00:08:58) productwise. They didn't, you know, they (00:09:00) had their own code red at one point, but (00:09:01) they didn't take it that seriously. (00:09:03) Everyone's doing code reds out here. (00:09:04) >> Um, and then (00:09:06) >> and also Google has probably the (00:09:08) greatest business model in the whole (00:09:10) tech industry. (00:09:11) Um, and I think they will be slow to (00:09:14) give that up. Um, but bolting AI into (00:09:18) web search, I don't I may be wrong. (00:09:21) Maybe like drinking the Kool-Aid here. I (00:09:23) don't think that'll work as well as (00:09:25) reimagining the whole, this is actually (00:09:27) a broader trend I think is interesting. (00:09:29) Bolting AI onto the existing way of (00:09:31) doing things, I don't think is going to (00:09:32) work well as redesigning stuff in the (00:09:34) sort of like AI first world. was part of (00:09:36) why we wanted to do the consumer devices (00:09:38) in the first place, but it applies at (00:09:39) many other levels. Um, if you stick AI (00:09:43) into a messaging app that's doing a nice (00:09:45) job summarizing your messages and (00:09:47) drafting responses for you, that is (00:09:49) definitely a little better. But I don't (00:09:51) think that's the end state. That is not (00:09:52) the idea of you have this like really (00:09:54) smart AI that is like acting as your (00:09:56) agent, talking to everybody else's agent (00:09:57) and figuring out when to bother you, (00:09:59) when not to bother you, and how to, you (00:10:01) know, what decisions it can handle and (00:10:03) when it needs to ask you. So (00:10:06) similar things for search, similar (00:10:07) things for like productivity suites. I (00:10:09) suspect (00:10:11) it always takes longer than you think, (00:10:12) but I suspect we will see new (00:10:15) products in in the major categories that (00:10:17) are just totally built around AI rather (00:10:20) than bolting AI in. And I think this is (00:10:22) a weakness of Google's even though they (00:10:23) have this huge distribution advantage. (00:10:25) >> Yeah, I' I've spoken with so many people (00:10:27) about this question. uh when Chetchup PT (00:10:29) came out initially, I think it was (00:10:30) Bendic Devon that suggested you might (00:10:32) not want to put AI in Excel. You might (00:10:35) want to just reimagine how you use (00:10:37) Excel. And to me, in my mind, that was (00:10:39) like you upload your numbers and then (00:10:41) you talk to your numbers. Well, one of (00:10:43) the things people have found as they've (00:10:44) developed this stuff is there needs to (00:10:46) be some sort of backend there. (00:10:49) >> So, is it that you sort of build the (00:10:51) backend and then you interact with it (00:10:54) with AI as if it's a new software (00:10:56) program? (00:10:57) Yeah, that's kind of what's happening. (00:10:58) >> Why wouldn't you then be able to just (00:11:00) bolt it on on top? (00:11:01) >> Yeah, I mean, you can bolt it on on top, (00:11:03) but the (00:11:06) >> I spent a lot of my day in various (00:11:08) messaging apps, (00:11:10) including email, including text, Slack, (00:11:12) whatever. I think that's just the wrong (00:11:14) interface. So, you can bolt AI on top of (00:11:17) those, and again, it's like a little bit (00:11:18) better, but what I would rather do is (00:11:21) just sort of like have the ability to (00:11:23) say in the morning, here are the things (00:11:25) I want to get done today. Here's what (00:11:27) I'm worried about. Here's what I'm (00:11:28) thinking about. Here's what I'd like to (00:11:29) happen. I do not want to be I do not (00:11:32) want to spend all day messaging people. (00:11:33) I do not want you to summarize them. I (00:11:34) do not want you to show me a bunch of (00:11:35) drafts. Deal with everything you can. (00:11:37) You know me. You know these people. You (00:11:38) know what I want to get done. Um and (00:11:41) then you know like batch every couple of (00:11:45) hours updates to me if you need (00:11:47) something. But that's a very different (00:11:50) flow than the way these apps work right (00:11:52) now. (00:11:53) >> Yeah. And I was going to ask you what (00:11:55) ChachiBT is going to look like in the (00:11:56) next year and then the next two years. (00:11:59) Is that kind of where it's going? (00:12:02) >> To be perfectly honest, I expected by (00:12:04) this point Chachi BT would have looked (00:12:06) more different than it did at launch. (00:12:07) >> What did you anticipate? I didn't know. (00:12:09) I just thought like that chat interface (00:12:11) was not going to go as far as it turned (00:12:12) out to go. H like we I mean it was put (00:12:16) up (00:12:17) it looks better now, but it is broadly (00:12:20) similar to when it was put up as like a (00:12:22) research preview. was not even meant to (00:12:24) be a product. We knew that the text (00:12:26) interface was very good, you know, like (00:12:28) the everyone's used to texting their (00:12:30) friends and they like it. Um, the chat (00:12:33) interface was very good, but (00:12:35) I would have thought to be as big and as (00:12:39) significantly used for real work of a (00:12:43) product as what we have now, the (00:12:45) interface would have had to go (00:12:48) much further than it has now. I still (00:12:51) think it should do that but there is (00:12:53) something about the generality of the (00:12:55) current interface that I underestimated (00:12:57) the power of. Um (00:13:02) what I (00:13:04) think should happen of course is that um (00:13:07) AI should be able to generate different (00:13:09) kinds of interfaces for different kinds (00:13:10) of tasks. So if you are talking about (00:13:11) your numbers it should be able to show (00:13:12) you that in different ways and you (00:13:14) should be able to interact with it in (00:13:15) different ways. Um (00:13:17) it and we have a little bit of this with (00:13:19) features like canvas. It should be way (00:13:21) more interactive. It's like right now, (00:13:22) you know, it's kind of a back and forth (00:13:24) conversation. It'd be nice if you could (00:13:26) just be talking about an object and it (00:13:28) could be continuously updating. You have (00:13:30) more questions, more thoughts, more (00:13:31) information comes in. Um, it'd be nice (00:13:34) to be more proactive over time where it (00:13:37) maybe does understand what you want to (00:13:38) get done that day and it's continuously (00:13:41) working for you in the background and (00:13:42) send you updates. And you see part of (00:13:43) this the way people are using codecs (00:13:45) which I think is one of the most (00:13:46) exciting (00:13:49) things that happened this year is codecs (00:13:51) got really good. Uh and that points to (00:13:57) a lot of what I hope the shape of the (00:13:58) future looks like. Um (00:14:01) but (00:14:04) it is surprising to me. I was going to (00:14:06) say embarrassing but it's not. I mean (00:14:08) clearly it's been super successful. Uh (00:14:10) it is surprising me how little CHBT has (00:14:12) changed over the last three years. (00:14:14) >> Yep. It the interface works. (00:14:16) >> Yeah. (00:14:18) >> But I guess what the guts have changed (00:14:20) and you talked a little bit about how (00:14:22) personalization is big uh to me and I (00:14:25) think this has been one of your (00:14:26) preferred features too. Memory has been (00:14:28) a real difference maker. Um, I've been (00:14:31) having a conversation with ChachiPT (00:14:33) about a forthcoming trip that has lots (00:14:35) of planning elements for weeks now and I (00:14:38) can just come in in a new window and be (00:14:40) like, "All right, let's pick up on this (00:14:42) trip." And it it has the context and it (00:14:44) knows knows the guide I'm going with. It (00:14:46) knows what I'm doing. Uh, the fact that (00:14:47) I've been like planning fitness for it (00:14:49) and can really synthesize all of those (00:14:52) things. How good can memory get? I think (00:14:56) we have no conception because the human (00:14:59) limit like even if you have the world's (00:15:01) best (00:15:02) personal assistant (00:15:05) they don't they can't remember every (00:15:07) word you've ever said in your life. They (00:15:09) can't have read every email. They can't (00:15:10) have read every document you've ever (00:15:12) written. They can't be you know looking (00:15:15) at all your work every day and (00:15:16) remembering every little detail. They (00:15:19) can't be a participant in your life to (00:15:22) that degree. And no human has like (00:15:23) infinite perfect memory. (00:15:25) Um, (00:15:27) and AI is definitely going to be able to (00:15:28) do that. And we actually talk a lot (00:15:30) about this, like right now, memory is (00:15:31) still very crude, very early. We're in (00:15:33) like the, you know, the GBT2 era of (00:15:34) memory. (00:15:36) But what it's going to be like when (00:15:41) it really does remember every detail of (00:15:43) your entire life and personalized across (00:15:45) all of that and not just the facts, but (00:15:47) like the little small preferences that (00:15:50) you had that you maybe like didn't even (00:15:51) think to indicate, but the AI can pick (00:15:53) up on. Uh, (00:15:56) I think that's going to be super (00:15:57) powerful. That's one of the features (00:15:58) that still maybe not 2026 thing, but (00:16:01) that's one of the parts of this I'm most (00:16:02) excited for. (00:16:03) >> Yeah. I was speaking with a (00:16:05) neuroscientist on the show and he (00:16:07) mentioned that you don't you can't find (00:16:10) thoughts in the brain. Like the brain (00:16:11) doesn't have a place to store thoughts, (00:16:13) but computing there's a place to store (00:16:15) them. So, you can keep all of them. And (00:16:17) as these bots do keep our thoughts, um, (00:16:21) of course there's a privacy concern. And (00:16:24) but the other thing is something that's (00:16:25) going to be interesting is we'll really (00:16:27) build relationships with them. I think (00:16:29) it's been one of the more underrated (00:16:31) things about this entire moment is that (00:16:33) people have felt that these bots are (00:16:35) their companions, are looking out for (00:16:37) them. Um, and I'm curious to hear your (00:16:40) perspective. Um, when you think about (00:16:43) the level of I don't know if intimacy is (00:16:46) the right word, but companionship people (00:16:48) have with these bots, um, is there a (00:16:50) dial that you can turn to be like, oh, (00:16:53) let's make sure people become really (00:16:55) close with these things, or, you know, (00:16:57) we turn the dial a little bit further (00:16:59) and there's an arms distance uh, between (00:17:02) them and and if there is that dial, (00:17:04) >> how do you modulate that the right way? (00:17:06) There are definitely more people (00:17:09) than I realize that want to have, let's (00:17:12) call it close companionship. You I don't (00:17:14) know what the right word is like. (00:17:15) Relationship doesn't feel quite right. (00:17:16) Companionship doesn't feel quite right. (00:17:18) I I don't know what to call it, but they (00:17:20) want to have whatever this deep (00:17:21) connection with an AI. There there are (00:17:23) more people that want that at the (00:17:25) current level of model capability than I (00:17:29) thought. And there's like a whole bunch (00:17:31) of reasons why I think we underestimated (00:17:33) this, but at the beginning of this year, (00:17:35) it was considered a very strange thing (00:17:36) to say you wanted that. Maybe some a lot (00:17:39) of people still don't revealed (00:17:40) preference. (00:17:42) You know, people like their AI chatbot (00:17:46) to get to know them and be warm to them (00:17:47) and be supportive and there's value (00:17:50) there even for people who in some cases (00:17:53) even for people who say they they don't (00:17:54) care about that uh still have a (00:17:56) preference for it. I (00:18:00) I think there's some version of this (00:18:01) which can be super healthy and I think (00:18:03) you know adult users should get a lot of (00:18:05) choice in where on the spectrum they (00:18:07) want to be. There are definitely (00:18:09) versions of it that seem to me unhealthy (00:18:11) although I'm sure a lot of people will (00:18:12) choose to do that. Um and then there's (00:18:15) some people who definitely want the (00:18:17) driest most effect efficient tool (00:18:20) uh possible. So I suspect like lots of (00:18:25) other technologies, (00:18:27) we will run the experiment. We will find (00:18:29) that there's unknown unknowns, good and (00:18:32) bad about it. And society will over time (00:18:36) figure out (00:18:39) how to how to think about where people (00:18:42) should set that dial and then people (00:18:43) have huge choice and set it in very (00:18:45) different places. (00:18:46) >> So your your thought is allow people (00:18:47) basically to determine this. (00:18:49) >> Yes, definitely. But I I don't think we (00:18:51) know like how far it's supposed to go, (00:18:54) like how far we should allow it to go. (00:18:56) We're we're going to give people quite a (00:18:58) bit of personal freedom here. Um there (00:19:03) are examples of things that uh we've (00:19:05) talked about that, (00:19:07) you know, other services will offer, but (00:19:09) we we won't. Um like we're not going to (00:19:12) let we're not going to have RAI, you (00:19:15) know, try to convince people that should (00:19:16) be like in an exclusive romantic (00:19:18) relationship with them, for example. (00:19:19) got to keep it open. (00:19:20) >> But I'm sure that will No, I'm sure that (00:19:22) that will happen with other services, I (00:19:24) guess. Yeah, because the stickier it is, (00:19:26) the more money that service makes. The (00:19:27) whole all these possibilities kind of (00:19:30) they're a little bit scary when you (00:19:31) think about them a little bit deeply. (00:19:34) >> Totally. This is one that really does (00:19:36) that I personally, you know, you can see (00:19:39) the ways that this goes really wrong. (00:19:40) >> Yeah. Uh, you mentioned Enterprise. (00:19:42) Let's talk about Enterprise. you were at (00:19:44) a lunch with some editors and CEOs of (00:19:47) some news companies in New York last (00:19:49) week and told them that enterprise is (00:19:51) going to be a major priority uh for (00:19:54) OpenAI next year. (00:19:55) >> U I'd love to hear a little bit more (00:19:58) about um why that's a priority, how you (00:20:01) think you stack up against anthropic. I (00:20:03) know people will say this is a pivot for (00:20:06) OpenAI that has been consumer focused. (00:20:08) So just give us an overview about the (00:20:10) enterprise plan. Our strategy was always (00:20:12) consumer first. Uh there were a few (00:20:14) reasons for that. One, the models were (00:20:16) not robust and skilled enough uh for (00:20:20) most enterprise uses and now now they're (00:20:22) they're getting there. The second was we (00:20:24) had this like clear opportunity to win (00:20:26) in consumer and those are rare and hard (00:20:29) to come by and I think if you win in (00:20:30) consumer it makes it massively easier to (00:20:32) win in enterprise and we are we are (00:20:35) seeing that now. Um but as I mentioned (00:20:37) earlier this was a year where we (00:20:39) enterprise growth outpaced consumer (00:20:41) growth. Uh and given where the models (00:20:44) are today where they will get to next (00:20:46) year we think this is the time where we (00:20:48) can (00:20:51) build a really significant enterprise (00:20:53) business quite rapidly. I mean I think (00:20:56) and we already have one but it can it (00:20:58) can grow much more. Um (00:21:00) companies seem ready for it. The (00:21:02) technology seems ready for it. the, you (00:21:05) know, coding is the biggest example so (00:21:08) far, but there are others that are now (00:21:11) growing, other verticals that are now (00:21:12) growing very quickly. And we're starting (00:21:14) to hear enterprises say, you know, I (00:21:16) really just want an AI platform. (00:21:18) >> Which vertical company? (00:21:19) >> Um, finance science is the one I'm most (00:21:23) excited about of everything happening (00:21:25) right now. Personally, um, customer (00:21:27) support is doing great. Uh (00:21:32) but but yeah the the (00:21:36) we have this thing called GDP though. (00:21:38) >> I was going to ask you about that. Can I (00:21:39) actually throw my question out about (00:21:40) that? All right. Cuz I wrote to Aaron (00:21:42) Levy the CEO of Box and I said I'm going (00:21:44) to meet with Sam. What should I ask him? (00:21:46) He goes throw a question out about GDP (00:21:48) val. Right. So this is the measure of (00:21:49) how AI performs in knowledge work tasks. (00:21:52) And I said okay. I went back to the (00:21:53) release of GPT 5.2 to the model that uh (00:21:56) you recently released and looked at the (00:21:59) GDP valid chart. Now this of course is (00:22:00) an open AI evaluation. Um that being (00:22:03) said the uh GPT5 thinking model so this (00:22:07) is the model released in the in the (00:22:09) summer. It ti it tied uh knowledge (00:22:12) workers at 38% of test or tied beat or (00:22:15) tied (00:22:16) >> um GP so 38.8% GPT 5.2 2 thinking beat (00:22:22) or tied at 70.9% (00:22:25) of knowledge work tasks and GPT 5.2 pro (00:22:30) 74.1% (00:22:31) of knowledge work tasks and it passed (00:22:33) the threshold of um being expert level (00:22:37) it it handled it looks like something (00:22:38) like 60% of expert tasks uh of tasks (00:22:42) that would make it you know on par with (00:22:43) an expert in the knowledge work. What (00:22:45) are the implications of the fact that (00:22:47) these models can do that much knowledge (00:22:49) work? So, you know, you were asking (00:22:51) about verticals, and I think that's a (00:22:52) great question, but the thing that was (00:22:53) going through my mind and why I kind of (00:22:54) was stumbling a little bit is that Eval, (00:22:57) I think it's like 40 something different (00:22:59) verticals that a business has to do. (00:23:02) >> There's make a PowerPoint, do this legal (00:23:04) analysis, you know, write up this little (00:23:06) web app, all this stuff. (00:23:08) >> And and the eval is do experts prefer (00:23:12) the output of the model relative to (00:23:14) other experts (00:23:17) for a lot of the things that a business (00:23:18) has to do. Now, these are small, well (00:23:20) scopeed tasks. These don't get the kind (00:23:22) of complicated, open-ended, creative (00:23:24) work that, you know, figuring out a new (00:23:26) product. These don't get a lot of (00:23:28) collaborative team things. But (00:23:31) a co-orker that you can assign an hour's (00:23:34) worth of tasks to and get something you (00:23:36) like better back 74 or 70% of time if (00:23:38) you want to pay less is still pretty (00:23:41) extraordinary. If you went back to the (00:23:43) launch of Chat TBT 3 years ago and said (00:23:46) we were going to have that in 3 years, (00:23:47) most people would say absolutely not. (00:23:49) Um, and so as we think about how (00:23:52) enterprises are going to integrate this, (00:23:54) it's no longer like just that it can do (00:23:56) code. It's all of these knowledge work (00:23:58) tasks you can kind of farm out to the (00:24:01) AI. uh and (00:24:06) that's going to take a while to really (00:24:09) kind of figure out how enterprises (00:24:11) integrate with it but should be quite (00:24:13) substantial. I know you're not an (00:24:14) economist, so I'm not going to ask you (00:24:16) like what is the macro impact on jobs, (00:24:17) but let me just read you one uh line (00:24:19) that I heard uh you know in in terms of (00:24:22) how this impacts jobs uh from Blood in (00:24:24) the Machine on Substack. Um this is from (00:24:27) a technical copywriter. They said, (00:24:29) "Chatbots came in and made it so my job (00:24:31) was managing the bots instead of a team (00:24:33) of reps." Okay, that that to me seems (00:24:36) like it's going to happen often. But (00:24:37) then this person continued and said once (00:24:39) the bots were sufficiently trained up to (00:24:41) offer good enough support then I was (00:24:43) out. Um is that is that the is that (00:24:47) going to become more common? Is that (00:24:48) what bad companies are going to do? (00:24:50) Because if you have a human who's going (00:24:51) to be able to sort of orchestrate a (00:24:54) bunch of different bots then you might (00:24:56) want to keep them. I don't know. How do (00:24:57) you think about this? So I I agree with (00:25:00) you that it's clear to see how (00:25:01) everyone's going to be managing like a (00:25:03) lot of AI uh doing different stuff. Um (00:25:09) eventually like any good manager (00:25:10) hopefully your team gets better and (00:25:11) better but you just take on more scope (00:25:13) and more responsibility. I am not I am (00:25:16) not a jobs dumer. (00:25:18) Um short term I have some worry. I think (00:25:20) the transition is likely to be rough uh (00:25:24) in some cases but (00:25:28) we are so deeply wired to care about (00:25:32) other people what other people do. We (00:25:34) are so we seem to be so focused on (00:25:38) relative status and always wanting more (00:25:40) and to be of use and service to express (00:25:44) creative spirit whatever whatever (00:25:45) whatever has driven us this long. I (00:25:47) don't think that's going away. Now I do (00:25:50) think the jobs of the future or I don't (00:25:52) even know if jobs is the right word. (00:25:53) Whatever we're all going to do all day (00:25:55) in 2050 probably looks very different (00:25:57) than it does today. Um (00:26:00) but I but I I don't have any of this (00:26:02) like oh life is going to be without (00:26:04) meaning and the economy is going to (00:26:05) totally break. Like we will find I hope (00:26:08) much more meaning and the economy I (00:26:10) think will significantly change but I I (00:26:14) think you just don't bet against (00:26:15) evolutionary biology. Um (00:26:18) you know I think a lot about how we can (00:26:20) automate all the functions at OpenAI and (00:26:22) then even more than that I think about (00:26:23) like what it means to have an AI CEO of (00:26:25) Open AI. Doesn't bother me. I'm like (00:26:27) thrilled for it. I won't fight it. Uh (00:26:30) like I don't want to be I don't want to (00:26:31) be the person hanging on being like I (00:26:33) can do this better the the handmade way. (00:26:34) >> AI CEO just make a bunch of decisions to (00:26:36) sort of like direct all of our resources (00:26:39) to giving AI more energy and power. It's (00:26:41) like (00:26:42) >> um I mean no you would really put a (00:26:44) guard rail on (00:26:45) >> Yeah. Like obviously you don't want an (00:26:48) AI CEO that is not governed by humans, (00:26:51) but if you think about (00:26:54) if if you think about maybe like (00:26:58) a this is a crazy analogy, but I'll give (00:27:02) it anyway. If you think about a version (00:27:04) where like every person in the world was (00:27:07) effectively on the board of directors of (00:27:09) an AI company and got to, you know, tell (00:27:13) the AI CEO what to do and fire them if (00:27:16) they weren't doing a good job at that (00:27:17) and, you know, got governance on the (00:27:19) decisions, but the AI CEO got to try to (00:27:21) like execute the wishes of the board. (00:27:24) Um, (00:27:26) I think to people of the future that (00:27:28) might seem like quite a reasonable (00:27:29) thing. Okay, so we're going to uh move (00:27:31) to infrastructure in a minute, but (00:27:32) before we leave this section on models (00:27:34) and capabilities, when's GP GPT6 coming? (00:27:39) Um, I expect I don't know when we'll (00:27:42) call a model GPT (00:27:45) 6. Uh, (00:27:47) but I would expect new models that are (00:27:49) significant gains from 5.2 in the first (00:27:52) quarter of next year. (00:27:53) >> What does significant gains mean? (00:27:56) I don't have like an eval score in mind (00:27:58) for you yet but uh more enterprise side (00:28:01) of things or definitely both the there (00:28:04) will be a lot of improvements to the (00:28:06) model for consumers uh the main thing (00:28:09) consumers want right now is not more IQ (00:28:11) enterprises still do want more IQ so uh (00:28:14) we'll improve the model in different (00:28:15) ways for the kind of for different uses (00:28:18) but uh I our goal is a model that (00:28:21) everybody likes much better (00:28:22) >> so infrastructure you have 1.4 trillion (00:28:26) thereabouts and commitments uh to build (00:28:28) infrastructure. I've listened to a lot (00:28:31) of what you've said about (00:28:31) infrastructure. Um here are some of the (00:28:34) things you said. If people knew what we (00:28:36) could do with compute, they would want (00:28:38) way way more. You said the gap between (00:28:40) what we could offer today versus 10x (00:28:43) compute and 100x compute is substantial. (00:28:46) Uh what what can you help flesh that out (00:28:49) a little bit? What are you going to do (00:28:51) with uh so much more compute? (00:28:53) Well, I mentioned this earlier a little (00:28:54) bit. The thing I'm personally more (00:28:56) excited, most excited about is to use AI (00:28:59) and lots of compute to discover new (00:29:00) science. I am a believer that scientific (00:29:03) discovery is the high order bit of how (00:29:05) the world gets better for everybody. And (00:29:07) if we can throw huge amounts of compute (00:29:09) at scientific problems and discover new (00:29:12) knowledge, which the tiniest bit is (00:29:14) starting to happen now, it's very early. (00:29:15) These are very small things but you know (00:29:17) my learning in history of this field is (00:29:19) once the squiggles start and it lifts (00:29:21) off the x-axis a little bit we know how (00:29:22) to make that better and better. Um but (00:29:24) that takes huge amounts of compute to (00:29:26) do. So that's one area we're like (00:29:28) throwing lots of AI at discovering new (00:29:30) science curing disease lots of other (00:29:32) things. Um, (00:29:35) a kind of recent cool example here is we (00:29:38) built the Sora Android app using codecs (00:29:43) and (00:29:45) they did it in like less than a month. (00:29:47) They used a huge amount. One of the nice (00:29:49) things about working at OpenAI is you (00:29:50) don't get any limits on codecs. They (00:29:51) used a huge amount of tokens, but they (00:29:54) were able to do what would normally have (00:29:55) taken a lot of people much longer and (00:29:58) Codex kind of mostly did it for us. And (00:30:02) you can imagine that going much further (00:30:04) where entire companies can build their (00:30:06) products using lots of compute. (00:30:10) Um (00:30:12) people have talked a lot about how video (00:30:14) models are going to point towards these (00:30:16) generated real-time generated user user (00:30:19) interfaces. That will take a lot of (00:30:21) compute. Um (00:30:24) enterprises that want to transform their (00:30:26) business will use a lot of compute. uh (00:30:28) doctors that want to offer good (00:30:31) personalized health care that are like (00:30:32) constantly (00:30:34) measuring every sign they can get from (00:30:36) each individual patient. You can imagine (00:30:38) that using a lot of compute. Uh it it's (00:30:41) hard to frame how much (00:30:44) compute we're already (00:30:46) using to generate AI output in the (00:30:49) world. Um but these are horribly rough (00:30:52) numbers. So, uh, and I think it's like (00:30:54) undisiplined to talk this way, but I I (00:30:56) always find these like mental thought (00:30:58) experiments a little bit useful, so (00:30:59) forgive me for the sloppiness. Um, let's (00:31:03) say (00:31:06) that an AI company today might be (00:31:08) generating something on the order of 10 (00:31:11) trillion tokens a day out of Frontier (00:31:13) models. Um, (00:31:16) you know, more, but not it's not like a (00:31:19) a quadrillion tokens for anybody, I (00:31:21) don't think. Um (00:31:25) let's say there's 8 billion people in (00:31:26) the world and let's say on average (00:31:28) someone's these are I think totally (00:31:30) wrong but let's say someone the average (00:31:32) number of tokens outputed by a person (00:31:33) per day is like (00:31:36) uh 20,000. (00:31:40) You can then start to and the token you (00:31:42) can to be fair then we have to compare (00:31:44) the output tokens of a model provider (00:31:46) today not not all the tokens consumed (00:31:47) but you can start to look at this and (00:31:49) you can say hm we're going to have these (00:31:54) models at a company be outputting more (00:31:56) tokens per day than all of humanity put (00:31:59) together and then 10 times that and then (00:32:01) 100 times that. And you know, in some (00:32:05) sense it's like a really silly (00:32:06) comparison, (00:32:08) but in some sense it gives a magnitude (00:32:10) for like how much of the intellectual (00:32:13) crunching on the planet is like human (00:32:14) brains versus AI brains. And that's kind (00:32:18) of the relative growth rates there are (00:32:21) are interesting. And so I'm wondering (00:32:24) are do you know that there is this (00:32:26) demand to use this compute like (00:32:28) potentially like so for instance would (00:32:30) we have surefires like scientific (00:32:32) breakthroughs if you know open AAI were (00:32:35) to put double the compute towards (00:32:37) science or or with medicine like are (00:32:40) would we have you know that clear (00:32:42) ability to assist doctors like (00:32:44) >> how much of this is sort of uh (00:32:46) supposition of what's to happen versus (00:32:49) clear understanding based off of what (00:32:51) you see today IC (00:32:52) >> everything everything based off what we (00:32:54) see today is that it will happen. It (00:32:56) does not mean some crazy thing can't (00:32:58) happen in the future. Someone could (00:33:00) discover some completely new (00:33:01) architecture and there could be a 10,000 (00:33:03) times you know efficiency gain and then (00:33:05) we would have really probably overbuilt (00:33:07) for a while. But everything we see right (00:33:10) now about how quickly the models are (00:33:12) getting better at each new level, how (00:33:13) much more people want to use them, each (00:33:15) time we can bring the cost down, how (00:33:16) much more people really want to use (00:33:18) them. Um, (00:33:22) everything about that indicates (00:33:26) to me that there will be increasing (00:33:29) demand and people using these for (00:33:32) wonderful things, for silly things. Um, (00:33:35) but (00:33:37) it it just so seems like (00:33:41) this is the shape of the future. Um (00:33:45) it's not just like it's not just you (00:33:48) know how many tokens we can do per day. (00:33:49) It's how fast we can do them as these (00:33:51) coding models have gotten better. They (00:33:52) can think for a really long time but you (00:33:53) don't want to wait for a really long (00:33:54) time. So there will be other dimensions. (00:33:56) It will not just be the number of tokens (00:33:57) that that we can do. Um but the demand (00:34:00) for intelligence across a small number (00:34:03) of axes (00:34:05) and what we can do with those you know (00:34:08) if you're like if you have like a really (00:34:10) difficult healthcare problem do you want (00:34:12) to use 5.2 or do you want to use 5.2 pro (00:34:15) even if it takes dramatically more (00:34:17) tokens I'll go with the better model. I (00:34:18) think you will um can let's just try to (00:34:21) go one level deeper. Um (00:34:24) going to the scientific discovery, can (00:34:26) you give an example of like a scientist (00:34:28) it doesn't have to well maybe it's one (00:34:30) that you know today that's like I have (00:34:32) problem X and if I put you know compute (00:34:35) Y towards it I will solve it but I'm not (00:34:37) able to today. There was a thing this (00:34:39) morning on Twitter where a bunch of (00:34:41) mathematicians were saying they were all (00:34:43) like replying to each other's tweets. Uh (00:34:45) they're like I was really skeptical that (00:34:47) LM's were ever going to be good. 5.2 is (00:34:49) the one that crossed the boundary for (00:34:51) me. it did it you know figured out this (00:34:54) it with some help it did this small (00:34:56) proof it it discovered this small thing (00:34:59) but it's this is actually changing my (00:35:00) workflow and then people were piling on (00:35:02) saying yeah me too I mean some people (00:35:03) were saying 5.1 was already there not (00:35:05) many (00:35:06) >> um but (00:35:08) that that was like that's a very recent (00:35:10) example this model's only been out for 5 (00:35:12) days or something where people are like (00:35:14) all right you know the mathematic (00:35:16) >> the mathematics research community seems (00:35:17) to say like okay something important (00:35:19) just happened (00:35:19) >> I've seen Greg Brockman has been (00:35:21) highlighting getting all these different (00:35:22) mathematical scientific uses in his feed (00:35:24) and something has clicked I think with (00:35:27) 5.2 um among these communities. So it'll (00:35:31) be interesting to see what happens as as (00:35:32) things progress. (00:35:34) >> We don't (00:35:36) like one of the hard parts about compute (00:35:38) >> at this scale is you have to do it so (00:35:40) far in advance. So you know that 1.4 (00:35:43) trillion you mentioned we'll spend it (00:35:44) over a very long period of time. I wish (00:35:45) we could do it faster. I think there (00:35:47) would be demand if we could do it (00:35:48) faster. Um, but (00:35:52) it just takes an enormously long time to (00:35:55) build these projects and the energy to (00:35:58) run the data centers and the chips and (00:36:00) the systems and the networking and (00:36:01) everything else. Um, so that will be (00:36:03) over a while, but you know, we (00:36:06) from a year ago to now, we probably (00:36:07) about tripled our compute. We'll triple (00:36:09) our compute again next year, hopefully (00:36:10) again after that. um revenue grows even (00:36:14) a little bit faster than that but it (00:36:15) does roughly track our compute (00:36:19) fleet. Uh so we (00:36:23) we have never yet found a situation (00:36:25) where we can't really well monetize all (00:36:27) the compute we have. Um if we had I (00:36:29) think if we had you know double the (00:36:30) compute we'd be at double the revenue (00:36:32) right now. (00:36:32) >> Okay let's let's talk about numbers (00:36:34) since you brought it up. Um revenue is (00:36:36) growing. uh compute spend is growing but (00:36:39) compute spend still outpaces revenue (00:36:42) growth. Uh I think the numbers that have (00:36:44) been reported are OpenAI is supposed to (00:36:46) lose something like 120 billion between (00:36:50) now and 120 and 2028 29 where you're (00:36:54) going to become profitable. Um so talk a (00:36:57) little bit about like how does that (00:36:59) change? Where does the turn happen? I (00:37:01) mean, as revenue grows and as inference (00:37:05) becomes a larger and larger part of the (00:37:07) fleet, it eventually uh subsumes the (00:37:10) training expense. So, that's the plan. (00:37:12) Spend a lot of money training but make (00:37:14) more and more. Uh if we if we weren't (00:37:17) continuing to grow our training (00:37:19) costs by so much, uh we would be (00:37:22) profitable way way earlier. Um but the (00:37:26) bet we're making is to invest very (00:37:28) aggressively in training these big (00:37:29) models. The whole world is wondering um (00:37:32) how your revenue will line up with the (00:37:35) spend. Uh the question's been asked if (00:37:38) the trajectory is to hit 20 billion in (00:37:41) revenue this year and the the spend (00:37:43) commitment is 1.4 trillion. Um so I (00:37:47) think it would be great just over a very (00:37:49) long period. (00:37:49) >> Yeah. Over and that's why I wanted to (00:37:51) bring it up to you. I think it would be (00:37:52) great to just lay it out for everyone (00:37:54) once and for all how those numbers are (00:37:56) going to work. It's it's very hard to (00:37:59) like really I I I find that one thing I (00:38:03) certainly can't do it and very few (00:38:04) people I've ever met can do it. You (00:38:06) know, you can like you have good (00:38:08) intuition for a lot of mathematical (00:38:09) things in your head, but exponential (00:38:11) growth is usually very hard for people (00:38:13) to do a good quick mental framework on (00:38:16) like for whatever reason there were a (00:38:18) lot of things that evolution needed us (00:38:20) to be able to do well with math in our (00:38:21) heads. Modeling exponential growth (00:38:24) doesn't seem to be one of them. Um so (00:38:28) the thing we believe is that we can stay (00:38:30) on (00:38:32) a very steep (00:38:36) growth curve of revenue for quite a (00:38:38) while and everything we see right now (00:38:39) continues to indicate that we cannot do (00:38:41) it if we don't have the compute. uh (00:38:43) again we're so compute constrained uh (00:38:45) and it hits the revenue line so hard (00:38:48) that I think if we get to a point where (00:38:51) we have like a lot of compute sitting (00:38:52) around that we can't monetize on a you (00:38:55) know profitable per unit of compute (00:38:56) basis be very reasonable to say okay (00:38:59) this is like a little how's this all (00:39:01) going to work but (00:39:03) we've penciled this out a bunch of ways (00:39:06) uh we will of course also get more (00:39:08) efficient uh on like a flops per dollar (00:39:11) basis as you know all of the work we've (00:39:13) been doing to make comput cheaper comes (00:39:14) to pass. Um, but (00:39:19) we see this consumer growth, we see this (00:39:20) enterprise growth. There's a whole bunch (00:39:22) of new kinds of businesses that (00:39:24) have we haven't even launched yet but (00:39:26) will. Um, but compute is really the (00:39:28) lifeblood that enables all of this. So (00:39:31) we, you know, there's like checkpoints (00:39:33) along the way and if we're a little bit (00:39:34) wrong about our timing or math, we can (00:39:37) we have some flexibility, but (00:39:40) we have always been in a comput deficit. (00:39:43) It has always constrained what we're (00:39:44) able to do. Uh I unfortunately think (00:39:47) that will always be the case, but I wish (00:39:48) it were less the case and I'd like to (00:39:50) get it to be less of the case over time. (00:39:52) Uh because I think there's so many great (00:39:53) products and services that we can (00:39:55) deliver and it'll be a great business. (00:39:57) Okay. So, it's effectively training (00:39:59) costs go down (00:40:00) >> as a percentage go up overall. But yeah, (00:40:03) >> and then your expectation is through (00:40:05) things like this this enterprise push (00:40:07) through things like people being willing (00:40:09) uh to pay for chat GPT through the API, (00:40:12) OpenAI will be able to grow revenue (00:40:14) enough to pay for it with revenue. (00:40:16) >> Yeah, that is the plan. (00:40:18) >> Now, I think the thing so the market's (00:40:20) been kind of losing its mind over this (00:40:23) um recently. I think the thing that has (00:40:25) spooked the market has been the debt has (00:40:27) entered uh into this equation. And the (00:40:31) idea around debt is you take debt out (00:40:33) when there's something that's (00:40:34) predictable. Um and then companies will (00:40:37) take the debt out, they'll build and (00:40:38) they'll have predictable revenue. (00:40:40) >> But it's it's the this is a new (00:40:43) category. It's it is unpredictable. Um (00:40:46) is is that how do you think about the (00:40:47) fact that that debt has entered the (00:40:49) picture here? So, first of all, I think (00:40:51) the market more lost its mind when (00:40:55) earlier this year, you know, we would (00:40:56) like meet with some company and that (00:40:58) company's stock would go up 20% or 15% (00:41:00) the next day. That was crazy. (00:41:01) >> That felt really unhealthy. Um, I'm (00:41:04) actually happy that there's like a (00:41:06) little bit more skepticism and (00:41:08) rationality in the market now cuz uh it (00:41:10) felt to me like we were just totally (00:41:12) heading towards a very unstable bubble (00:41:14) and now I think people are some degree (00:41:17) of discipline. So I actually think (00:41:18) things are I think people went crazy (00:41:20) earlier and now people are being more (00:41:21) rational on the debt front. I I think we (00:41:27) do kind of we know that if we build (00:41:31) infrastructure we the industry someone's (00:41:33) going to get value out of it. And it's (00:41:37) still it's still totally early. I agree (00:41:40) with you. But I don't think anyone's (00:41:42) still questioning there's not going to (00:41:43) be value from like AI infrastructure. (00:41:46) And so I think it is reasonable for debt (00:41:49) to (00:41:52) enter this market. I think there will (00:41:53) also be other kinds of financial (00:41:54) instruments. I suspect we'll see some (00:41:56) unreasonable ones as people really you (00:41:59) know innovate about how to finance this (00:42:02) sort of stuff. But you know like lending (00:42:04) companies money to build data centers (00:42:06) that that seems fine to me. I think the (00:42:07) the fear is that um if things don't (00:42:10) continue at pace like here's one (00:42:11) scenario um and you'll probably disagree (00:42:15) with this but like the model progress (00:42:16) saturates uh then the the infrastructure (00:42:20) becomes worth less than the anticipated (00:42:22) value was and then yes those data (00:42:25) centers will be worth something to (00:42:26) someone but it could be that they get (00:42:28) liquidated and someone buys them at a (00:42:30) discount. Yeah. And and I do suspect by (00:42:32) the way there will be some like booms (00:42:33) and busts along the way. These things (00:42:35) are never a perfectly smooth line. Um, (00:42:39) first of all, it seems very clear to me, (00:42:41) and this is like a thing I happily would (00:42:43) bet the company on, that the models are (00:42:45) going to get much much better. We have (00:42:47) like a pretty good window into this. (00:42:48) We're very confident about that. Even if (00:42:50) they did not, I think the (00:42:54) there's like a lot of inertia in the (00:42:55) world. It takes a while to figure out (00:42:57) how to adapt to things. The overhang of (00:43:00) the economic value that I believe 5.2 2 (00:43:03) represents relative to what the world (00:43:05) has figured out how to get out of it so (00:43:06) far is so huge that even if you froze (00:43:09) the model at 5.2 to how much more like (00:43:12) value can you create and thus revenue (00:43:13) can you drive? I bet a huge amount. In (00:43:16) fact, you didn't ask this, but if I can (00:43:18) go on a rant for a second. Um, (00:43:22) we used to talk a lot about this 2x2 (00:43:24) matrix of short timelines, (00:43:27) long timelines, slow takeoff, fast (00:43:28) takeoff, and where we felt at different (00:43:31) times the kind of probability was (00:43:32) shifting, and that that was going to be (00:43:34) you could kind of understand a lot of (00:43:36) the decisions and strategy that the (00:43:39) world should optimize for based off of (00:43:41) where you were going to be on that 2x (00:43:42) two matrix. Um, (00:43:48) there's like a Z-axis in my head in my (00:43:50) picture of this that's emerged, which is (00:43:53) small overhang, big overhang. And (00:43:57) I I kind of thought that (00:44:00) I guess I didn't think about that hard, (00:44:02) but uh like my retro on this is I must (00:44:05) have assumed that the overhang was not (00:44:06) going to be that massive that if the (00:44:08) models had a lot of value in them, the (00:44:11) world was pretty quickly going to figure (00:44:12) out how to deploy that. But it looks to (00:44:15) me now like the overhang is going to be (00:44:16) massive in most of the world. You'll (00:44:18) have these like areas like you know some (00:44:20) some set of coders that'll get massively (00:44:22) more productive by adopting these tools. (00:44:25) But on the whole (00:44:27) you have this crazy smart model that to (00:44:30) be perfectly honest most people are (00:44:32) still asking this similar questions they (00:44:33) did in the GPD4 realm. Scientists (00:44:36) different coders different maybe (00:44:37) knowledge work is going to get different (00:44:39) but but there is a huge overhang and (00:44:43) that has a bunch of very strange (00:44:44) consequences for the world. I we have (00:44:46) not wrapped our head around all the ways (00:44:48) that's going to play out yet, but is (00:44:50) very much not what I would have expected (00:44:51) a few years ago. I have a question for (00:44:53) you about this uh capability overhang. (00:44:55) Basically, the models can do a lot more (00:44:57) than they've been doing. Um I I'm trying (00:45:00) to figure out how um the models can be (00:45:03) that much better than they're being used (00:45:05) for, but a lot of businesses when they (00:45:07) try to implement them, they're not (00:45:09) getting a return on their investment. (00:45:11) >> Um or at least that's what they tell (00:45:13) MIT. I'm not sure quite how to think (00:45:15) about that because we hear all these (00:45:16) businesses saying, you know, if you 10x (00:45:19) the price of GPT 5.2, we would still pay (00:45:21) for it. Like you're hugely underpricing (00:45:23) this, we're getting all this value out (00:45:24) of it. (00:45:24) >> Um, (00:45:26) so I don't that doesn't seem right to (00:45:29) me. Certainly, if you talk about like (00:45:31) what coders say, they're like, "This is, (00:45:33) you know, I'd pay 100 times the price or (00:45:36) whatever." Um, (00:45:36) >> just be bureaucracy that's messing (00:45:38) things up. Let's say you believe the GDP (00:45:40) valve numbers and maybe you don't for (00:45:42) good reason maybe they're wrong but let (00:45:43) let's say it were true and for kind of (00:45:46) these wellsp specified not super long (00:45:49) duration knowledge work tasks seven out (00:45:51) of 10 times you would be as happy or (00:45:54) happier with the 5.2 output. (00:45:57) You should then be using that a lot. And (00:46:00) yet it takes people so long to change (00:46:01) their workflow. are so used to asking (00:46:03) the junior analyst to make a deck or (00:46:06) whatever that they're going to like it (00:46:10) just that's stickier than I thought it (00:46:12) was. You know, I still kind of run my (00:46:15) workflow in very much the same way. (00:46:18) Although I know that I could be using AI (00:46:19) much more than I am. Yep. All right, we (00:46:22) got 10 minutes left. I got Wow, that was (00:46:23) quick. I got four questions. Uh let's (00:46:25) see if we can lightning round uh through (00:46:27) them. So, uh, the device that you're (00:46:30) working on. We'll be back with OpenAI (00:46:33) CEO Sam Alman right after this. Um, what (00:46:36) I've heard, phone size, no screen. Um, (00:46:41) why couldn't it be an app if it's the (00:46:43) phone if it's the phone without a (00:46:46) screen? First, we're going to do a f a (00:46:47) small family of devices. It will not be (00:46:49) a single device. uh there will be over (00:46:51) time a (00:46:55) this is this is not speculation so I'm (00:46:56) may try not to be totally wrong but I (00:46:58) think there will be a shift over time to (00:47:00) the way people use computers where they (00:47:03) go from a sort of (00:47:06) dumb reactive thing to a very smart (00:47:10) proactive thing that is understanding (00:47:11) your whole life your context everything (00:47:12) going on around you very aware of (00:47:16) the people around you physically or (00:47:19) close to you via a computer that you're (00:47:23) working with. And I don't think current (00:47:26) devices are well suited (00:47:30) to that kind of world. And I am a big (00:47:32) believer that we like we work at the (00:47:34) limit of our devices. you know, you have (00:47:38) that computer and it has a bunch of (00:47:41) design choices. Like it could be open or (00:47:43) closed, but it can't be, you know, (00:47:45) there's not like a okay, pay attention (00:47:47) to this interview, but be closed and (00:47:50) like whisper in my ear if I forget to (00:47:51) ask Sam a question or whatever. Um, (00:47:54) >> maybe that would be helpful. And there's (00:47:56) like, you know, there's like (00:47:58) a screen and that like limits you to the (00:48:01) kind of same way we've had graphical (00:48:04) user interfaces working for many (00:48:05) decades. And there's, (00:48:07) you know, a keyboard that was built to (00:48:09) like slow down how fast you could get (00:48:11) information into it. And these have just (00:48:14) been unquestioned assumptions for a long (00:48:15) time, but they worked. And then this (00:48:17) totally new thing came along and it (00:48:21) opens up a possibility space. But (00:48:24) I don't think the current form factor of (00:48:28) devices is the optimal fit. It'd be very (00:48:31) odd if it were for this like incredible (00:48:33) new affordance we have. Oh man, we could (00:48:35) talk for an hour about this, but um (00:48:37) let's move on to the next one. Cloud. (00:48:39) You've talked about building a cloud. Um (00:48:42) here's a an email we got from a (00:48:44) listener. At my company, we're moving (00:48:47) off Azure and directly integrating with (00:48:49) OpenAI to power our AI experiences in (00:48:53) the product. The focus is to insert a (00:48:55) stream of trillions of tokens powering (00:48:58) AI experiences through the stack. Is is (00:49:01) that the plan to build a big big cloud (00:49:03) business in that in that way? (00:49:05) >> First of all, trillions of tokens, a lot (00:49:06) of tokens. And if you know you asked (00:49:08) about the need for compute and our (00:49:09) enterprise strategy like (00:49:12) >> enterprises have been clear with us (00:49:14) about how many tokens they'd like to buy (00:49:15) from us and we are going to again fail (00:49:18) in 2026 to meet demand but the strategy (00:49:21) is companies (00:49:23) most companies seem to want to come to a (00:49:26) company like us and say I'd like the (00:49:28) name of my company with AI. I need an (00:49:31) API customized for my company. I need (00:49:33) Chach Enterprise customized for my (00:49:35) company. I need a platform that can like (00:49:37) run all these agents that I can trust my (00:49:38) data on. I need the ability to get (00:49:40) trillions of tokens into my product. I (00:49:42) need the ability to have all my internal (00:49:45) processes be more efficient and (00:49:51) we don't currently have like a great (00:49:52) all-in-one offering for them and we'd (00:49:54) like to make that. (00:49:55) >> Is your ambition to put it up there with (00:49:57) the AWS and Ashers of the world? (00:49:59) >> Uh I think it's I think it's a different (00:50:01) kind of thing than those. like I don't I (00:50:04) don't really have an ambition to go (00:50:07) offer whatever all the services you have (00:50:09) to offer to host a website or I don't (00:50:11) even know but uh but I I I think the (00:50:13) concept (00:50:16) yeah my my guess is that people will (00:50:19) continue to have their (00:50:22) call it web cloud and then I think there (00:50:25) will be this other thing where like a (00:50:27) company will be like I need an AI (00:50:28) platform for everything that I want to (00:50:30) do internally the service I want to (00:50:31) offer whatever (00:50:33) and you know like it does kind of live (00:50:35) on the physical hardware in some sense (00:50:37) but (00:50:38) I think it'll be a fairly different (00:50:40) product offering. Uh let's talk about (00:50:41) discovery quickly. Um you've said (00:50:44) something that's been really interesting (00:50:45) to me uh you that you think that the (00:50:48) models or maybe it's people working with (00:50:50) models or the models will make small (00:50:51) discoveries next year and big ones (00:50:53) within five. Is that the models? Is it (00:50:56) people working alongside them? And what (00:50:58) makes you confident that that's going to (00:50:59) happen? Yeah, people using the models (00:51:01) like the the models that can like figure (00:51:04) out their own questions to ask that does (00:51:05) feel further off. But if the world is (00:51:08) benefiting from new knowledge like we (00:51:11) should be very thrilled and you know (00:51:13) like I think the the whole course of (00:51:18) human progress has been that we build (00:51:19) these better tools and then people use (00:51:21) them to do more things and then out of (00:51:23) that process they build more tools and (00:51:24) it's this like scaffolding that we climb (00:51:26) like layer by layer, generation by (00:51:28) generation, discovery by discovery and (00:51:30) the fact that a human's asking the (00:51:33) question I think in no way diminishes (00:51:35) the value of the tool. All right. So, I (00:51:36) I think it's great. I'm all happy. Um (00:51:40) I at the beginning of this year, I (00:51:42) thought the small discoveries were going (00:51:43) to start in 2026. They started in 2025 (00:51:45) in late 2025. Again, these are very (00:51:47) small. I really don't want to overstate (00:51:48) them, but (00:51:51) anything (00:51:53) is feels qualitatively to me very (00:51:55) different than nothing. And certainly in (00:51:58) the when we launched three years ago, (00:52:00) that model was not going to make any new (00:52:02) contribution to the total of human (00:52:03) knowledge. um (00:52:09) what it looks like from here to five (00:52:10) years from now. This journey to big (00:52:12) discoveries, I suspect it's just like (00:52:14) like the normal hill climb of AI. It (00:52:16) just gets like a little bit better every (00:52:18) quarter and then all of a sudden we're (00:52:20) like, whoa, humans augmented by these (00:52:23) models are doing things that humans 5 (00:52:25) years ago just absolutely couldn't do. (00:52:28) And (00:52:30) you know, whether we mostly attribute (00:52:32) that to smarter humans or smarter (00:52:34) models, as long as we get the scientific (00:52:35) discoveries, I'm very happy either way. (00:52:38) IPO next year. I don't know. Do you want (00:52:41) to be a public company? (00:52:44) >> Um, you seem like you can operate (00:52:46) private for a long time. Would you go (00:52:48) before you needed to (00:52:52) terms of funding? (00:52:52) >> There's like a whole bunch of things at (00:52:53) play here. I do think it's cool that (00:52:58) public markets get to participate in (00:53:00) value creation and you know in some (00:53:03) sense we will be very late to go public (00:53:06) if you look at any previous company. Um (00:53:10) it's wonderful to be a private company. (00:53:12) Uh we need lots of capital. Uh (00:53:16) we're going to you know cross all of the (00:53:18) sort of shareholder limits and stuff at (00:53:20) some point. So, (00:53:22) am I excited (00:53:24) to be a public company CEO? 0%. Um, am I (00:53:30) excited for Open Eye to be a public (00:53:31) company? In some ways, I am. And in some (00:53:34) ways, I think it'll be really annoying. (00:53:37) I listened to your Theo van interview (00:53:39) very closely. Uh, great interview. (00:53:42) >> He was really cool. (00:53:43) >> Theo really knows what he's talking. (00:53:46) He's (00:53:46) >> He did his homework. You told him, this (00:53:49) was right before GPT5 came out, that (00:53:51) GPT5 is smarter than us in almost every (00:53:54) way. Uh, I I thought that that was the (00:53:58) definition of AGI. Does is that isn't (00:54:00) that AGI? And and if not, has the term (00:54:03) become somewhat meaningless? These (00:54:05) models are clearly extremely smart on a (00:54:08) sort of raw horsepower basis. You know, (00:54:10) there's all this stuff on the last (00:54:11) couple of days about GPT 5.2 who has an (00:54:13) IQ of 147 or 144 or 151 or whatever it (00:54:18) is. It's like, you know, depending on (00:54:20) whose test it's like it's some high (00:54:22) number and you have like a lot of (00:54:24) experts in their field saying (00:54:28) it can do these amazing things and it's (00:54:29) like contributing it's making it more (00:54:31) effective. You have the GDP things we (00:54:32) talked about. One thing you don't have (00:54:36) is (00:54:38) the ability for the model to not be able (00:54:40) to do something today, realize it can't (00:54:43) go off and figure out how to learn to (00:54:45) get good at that thing, learn to (00:54:46) understand it, and when you come back (00:54:47) the next day, it gets it right. And that (00:54:50) kind of continuous learning like (00:54:55) toddlers can do it. It does seem to me (00:54:58) like an important part of what we need (00:55:01) to build. Now, can you have something (00:55:03) that most people would consider an AGI (00:55:04) without that? I would say clear. I mean, (00:55:06) there's a lot of people that would say (00:55:08) we're at AGI with our current models. (00:55:10) Um, (00:55:13) I think almost everyone would agree that (00:55:14) if we were at the current level of (00:55:16) intelligence and had that other thing, (00:55:17) it would clearly be very AGI like. Um, (00:55:21) but maybe most of the world will say, (00:55:26) "Okay, fine. Even without that, like (00:55:27) it's doing most knowledge tasks that (00:55:29) matter. um smarter than us in mo most of (00:55:32) us in most ways. We're at AGI. You know, (00:55:34) it's discovering small piece of new (00:55:35) science. We're at AGI. What I think this (00:55:38) means is that the term although it's (00:55:40) been very hard for all of us to stop (00:55:41) using is very underdefined, right? (00:55:45) I I have a I have a a can like one thing (00:55:49) I would love (00:55:51) since we got wrong with AGI. We never (00:55:53) define that that you know the new term (00:55:54) everyone's focused about is when we get (00:55:55) to super intelligence. Um so my proposal (00:55:59) is that we agree that you know AGI kind (00:56:02) of went whooshing by. It was didn't (00:56:05) change the world that much or it will in (00:56:08) the long term but okay fine we've built (00:56:10) AGIs at some point you know we're in (00:56:12) this like fuzzy period where some people (00:56:14) think we have and some people think we (00:56:16) have and more people will think we have (00:56:17) and and then we'll say okay what's next? (00:56:19) Um, a candidate definition for super (00:56:23) intelligence is when a system can do a (00:56:26) better job being president of United (00:56:29) States, CEO of a major company, you (00:56:32) know, running a very large scientific (00:56:33) lab than any person can even with the (00:56:37) assistance of AI. (00:56:39) >> Okay, (00:56:40) >> I think this was an interesting thing (00:56:41) about what happened with chess where (00:56:44) chess got it could be humans. I remember (00:56:47) this very vividly. uh that deep blue (00:56:49) thing and then there was a period of (00:56:51) time where (00:56:53) a human and the AI together were better (00:56:56) than an AI by itself and then the person (00:57:00) was just making it worse and the (00:57:02) smartest thing was the unaded AI that (00:57:04) didn't have the human like (00:57:07) not understanding its its great (00:57:09) intelligence. Um (00:57:12) I think something like that is like an (00:57:13) interesting framework for super (00:57:15) intelligence. I think it's like a long (00:57:16) way off, but I would love to have like a (00:57:18) cleaner definition this time around. (00:57:20) >> Well, Sam, look, I I have uh been in (00:57:22) your products uh using them daily for 3 (00:57:25) years. Um (00:57:27) >> definitely gotten a lot better. Can't (00:57:29) even imagine where they go from here. (00:57:30) >> We'll we'll try to keep getting them (00:57:32) better fast. (00:57:32) >> Okay. And uh this is our second time (00:57:35) speaking and I appreciate how open (00:57:37) you've been uh both times. So, thank you (00:57:38) for your time. (00:57:40) >> Thank you everybody for listening and (00:57:41) watching. If you're here for the first (00:57:43) time, please hit follow or subscribe. We (00:57:46) have lots of great interviews on the (00:57:47) feed and more on the way. This past (00:57:49) year, we've had Google DeepMind CEO (00:57:51) Demisabus on twice, including one with (00:57:54) Google founder Sergey Brin. We've also (00:57:56) had Dario Ammoday, the CEO of Anthropic. (00:58:00) And we have plenty of big interviews (00:58:02) coming up in 2026. Thanks again, and (00:58:04) we'll see you next time on Big (00:58:06) Technology Podcast.