The Sam Altman Interview 8/8/25

0:00Bring in show music please. Hi, I'm CNBC producer Katie Kramer. Today on Squawk Pod. OpenAI CEO Sam Altman the day after chat GPT-5 hit the public. Healthcare is maybe the area where there's the strongest improvement of any category. A key founder in artificial intelligence on the impact his digital project has on humanity. I do worry about this general case of people that just make their decisions based off of what the advice ChatGPT gives them. Like, it does feel to me like something gets important, gets lost in that process, even if the advice is always really good. Why he's luring Silicon Valley's hottest talent.

0:43The bet, the hope, is they know how to, like, discover the remaining ideas to get to superintelligence. That there are going to be a handful of algorithmic ideas and a, you know, medium-sized handful of people who can figure them out. And how he's thinking about returns for investors. Spoiler, he's focused on innovation. As long as we're on this very steep curve of the models getting better and better, I think the rational thing to do is to just be willing to run at a loss for quite a while and continue to do that, which is also part of why I think it's nice not to be public. Plus, the other stories that got us squawking, including President Trump's calls for Intel's CEO to resign.

1:22This is clearly industrial policy, which nobody liked. brought on by the original industrial policy. Intel got in bed with the federal government. It's Friday, August 8th, 2025. Squawk Pod begins right now. Stand Andrew by in three, two, one, up in Andrew, cue. Good morning and welcome to Squawk Box right here on CNBC. We're live at the Nasdaq market site in Times Square. I'm Andrew Osorkin. By the skin of my teeth, I'm Andrew Osorkin. Well, Becky's off. Joe is here. I was here. It's nice to see you. You were ready to cook. Good to see you. Broadway decided that there's a crane on Broadway. So I've been running down the street.

2:08I know. You had to do that. But you were really only about a minute later than normal. Really? Right? You know, there's a little room to spare. But it's nice to see you, sir. It's good to see you as well. You're finally back. You have to globetrotting. Country trotting. More than globe. Country trotting, is that what they would say? I don't even think that's an expression. Now, President Trump is back, sir. Yep, and he looks temporary. That's what they're saying so far. He's picked Council of Economic Advisors Chair Stephen Mirren as a temporary replacement for Adriana Kugler on the Federal Reserve Board of Governors.

2:47If confirmed by the Senate, he would serve out Kugler's term, which expires just at the end of January. kugler is stepping down uh from the fed today and in uh reportedly judy shelton was under consideration yeah what it says on the front page of the journey what do you think uh it didn't have i mean under consideration for this no i know but temporarily but but by the way someone's got to fill this this position long term right and is it the same person that ascends to uh i don't know I don't know what the, I don't know what is in the president's mind, and I don't know what the machinations are for, you know, how it worked.

3:28Is it like a trial run or something for Fed chair? I mean, and with, you know, all the worry about independence and everything else, is it, I mean, is that someone that's going to play ball? Is that, get tested out for that? I don't know. Well, the other question I have is if Moran goes back, like, does he do this as a pit stop and then go back to the administration? How does that work? Yeah. Exactly. I don't know. And have we had that before? I don't know that sort of sort of in and out. We know a lot more about the Fed like 10 years ago. Could you have meant that you've named all the governors and all the.

4:04No, no, except that many of them come on Squawk. They do. That's why. I don't think everyone is like us, obviously, knowing every single, you know, the minutia of the Federal Reserve. It's all I mean, is that the way it's supposed to be? They are very important in our life. In a post on True Social, President Trump said, in the meantime, we'll continue to search for a permanent replacement. Mirren is a past critic of the Fed, specifically its aggressive stimulus action during the COVID pandemic. He's also author of a plan to devalue the dollar as a way of managing the country's current accounts deficit.

4:41it. Meantime, OpenAI announcing its latest and most advanced large-scale AI model. OpenAI saying that ChatGPT5 is smarter, faster, and more useful, particularly in things like writing and specifically in coding and healthcare. The company says its hallucination rate is lower, meaning less frequently fabricated answers. And ChatGPT5 will be available to everyone. This includes OpenAI's non-paying users. It's the first time those users are going to have access was described as a reasoning model which carries out certainly an internal chain of thought before actually responding. ChatGPT5's announcement coming with a bit of corporate executive drama.

5:19Microsoft CEO Satya Nadella posted about the new release on X saying it's the most capable model yet from our partners at OpenAI. Now Elon Musk responding to that post saying the following, he says OpenAI is going to eat Microsoft alive. Well, we're going to speak with OpenAI CEO Sam altman it's going to happen this morning i got to play with chat gpt5 for a good part of the flight back to san francisco yesterday and pretty good it's better it's better than it's better than four i think that most people are not paying users so the difference might actually feel even bigger in terms of what it can do and the coding piece is the whole game because right now anthropic owns the coding world.

6:03That's their sort of purview. And these guys have really owned the consumer-facing world. And the question is whether this model is going to effectively, forget about eating Microsoft for lunch, whether it eats Anthropic for lunch or not. Have you had a hallucination experience with AI? All the time. You do? Yeah. I haven't used it enough, I guess, for that. How can you tell? Just crazy stuff? It just makes up answers. You ask it something. I have seen that. I have seen that. Just completely man up. I yell back at them. I go, you're an idiot. And they say, I'm sorry. You feel that way. We tried it.

6:41Sometimes it'll make up facts and make up quotes. I'll say, where did you find this? Only someone the other day was dead. Right. I'll say, where did you find this? And they'll say, oh, I apologize. I must have gotten this wrong. I remember I asked about the interview with President Trump. I asked some specific questions. And it said, President Trump did not have an interview with CNBC today. So then I said, you really need to check your facts. And then what happened? Then it came back and it checked its facts and came back. Yeah, I've had that. Weird. Yeah. This, I think, should be better than that.

7:20But what is cool about this and actually all the reasoning models, though this now does it so fast, is you actually sometimes can see the reasoning. Where it'll say, the user has asked, it'll rewrite what the user says, the user's really trying to get at this, and I need to figure out this. And then it'll do that, and the next permutation is this. And so you actually, it's crazy to see the computer thinking through the sort of various ways it's supposed to get to the answer. In GPT-5, it actually goes so fast that you can't really even see that part. And one of the things that Sam is now saying is sometimes you think the answer can't possibly be right because it's so fast.

7:57That actually having watched the computer think was actually helpful to the user experience because you're like, oh, okay. They all don't have the little progress on their thoughts. Do they all do that? Because sometimes now I'm doing this, now I'm doing that. Not all of them do that. Not all of them do that. By the way, I think that's a great sort of UI. Well, because it makes it feel very human. Even when they decided early on, and ChatGPT was the first to do it, to even just have the words right out in front of you. That's, by the way, all that is, is an animation. I don't know if people appreciate what's happening.

8:29When I say it's an anim... It's just animating the text. Right, right, right. It could also just spit the whole thing out at one shot, and it just feels different. What sort of concerned me was it says that what entity is in the catbird seat for AI? Reddit. I mean, if that's the place, that cesspool of chat rooms is where you're getting information, I hope it gets a little bit better than that. I guess I shouldn't be poking the Reddit. The Reddit bear. Nut, nut jobs, yeah. Now to the pressure, they have no life, number one. Now to the pressure on Intel CEO, Lipboo Tan. The Wall Street Journal says the TAN had experienced some tension with members of Intel's board even before President Trump said yesterday the TAN should resign and called him highly conflicted.

9:21TAN was named CEO of Intel in March. And the Journal says since then he has clashed with some Intel directors on key questions like whether the chip company should remain a manufacturer. Shares of Intel ended yesterday down 3 % after Trump's criticism earlier this week. Republican Senator Tom Cotton sent a letter to Intel's chair raising questions about TAN's ties to several Chinese companies. He referenced a past criminal case involving cadence design systems, which TAN previously led. In a statement yesterday, Intel said that the company, TAN, and Intel's board are deeply committed to advancing U.S.

10:00national and economic security interests and are making significant investments aligned with the President's America First agenda fascinating and I went I don't assign they assigned it to Albert Brooks book to you once but you got this one talks everything about what we were talking about yesterday and it's like we're both right this is this is clearly industrial policy which nobody like brought on by the original industrial policy Intel got in bed with with the federal government arguing for for subsidies they couldn't even have built the manufacturing stuff with without subsidies it's not going well it was supposed to be ready in 2025 it's not gonna be till 2030 they had to lay off people so it didn't create jobs the the tip deck it lost 21 000 this has been a failed effort but if that's true and by the way but but it but now they're in bed with the government so trump feels like he you know you are you already owe us you're already beholding to us for all these subsidies so I can tell you what to do.

11:04It should have never gotten to that point in the first place. 100%. Okay. So it's industrial policy across the board. Either party. It doesn't matter. And by the way, the Democrats were just as bad on the industrial policy piece. No, I'm saying they started or much worse, not just as bad. That's what I said. Well, but what I was going to say is, so go back in the debates that you and I used to have about industrial policy under Barack Obama, for example, or Biden or whatever, Including, by the way, I don't know if you describe it as industrial policy, but I would say Biden's decision to this day not to invite Elon Musk in the tent on EVs.

11:44I don't even know if I put that in the industrial policy, but sort of. No, it is. Anyway, all of those kinds of things were complete mistakes. All I'm suggesting is if that is true, that we're deciding that it's a mistake and you decided it was a mistake a long time ago, then the question is, forget about Intel. Intel is an outlier at this point. There's so much of what we're doing right now is a grand industrial policy. And the question is, we're now doing industrial policy at scale. And if you didn't think we could do it properly, you know, in the micro, how can we really make it? Plus, you remember in the CHIPS Act, you had to subsidize child care.

12:20You had to pay union wages. I agree. So all these progress. So now Trump, who's going to stop him from saying, no DEI, no ESG, you can't have hiring. So it's going to be from the other side. But when you start it, you stick your toe in the water. You can't pull it back out. Well, can you? Can you just end it? Intel might end it itself. Well, no, no. When I say end it, but I'm saying all of it. It is a shadow of its former self. But I'm saying that this is corporate industrial policy. You'd say tariffs are a form of industrial policy. Huge industrial policy, right? That's like a massive industrial policy.

12:57Right. Aimed at bringing back an industry here. Right. Which is the same thing that this was aimed at in a totally different way. In a different way. And by the way, one may work better than the other. Well, I think that goes without saying. We'll see.

13:17Cheese will be next. Coming up on Squawk Pod, our full interview with OpenAI CEO Sam Altman. What the ChatGPT creator imagines for the future. One of the things I've always wanted ChatGPT to do is when I wake up every morning, every email that came in overnight, I like a drafted response. And then I can choose whether to send it or not or edit it. But we're only maybe another, maybe six months away from being able to do that really well. And if he looks really far to the future. If I had let myself dream forward to someday far in the future where, you know, I'm like retiring to an AI CEO for OpenAI and I'm like writing that final prompt of like, here's my advice to you, I would still feel like a lot of agency when I think.

13:59Don't go anywhere. Sam Altman is up next.

14:10Welcome back to Squawk Pod. You're watching Squawk Box right here on CNBC on this Friday morning. I'm Andrew Ross Sorkin along with Joe Kernan. Becky is off. We've got a lot going on, though, this Monday, this Friday. It's going to be Monday soon. Soon enough. It's basically Monday. Meantime, OpenAI announcing its latest and most advanced large-scale AI model. The company is saying GPT-5 is smarter, faster, more useful. I used it. It seems to be all of those things, particularly in things like writing, coding, and particularly health care. We're going to talk about all of that this morning with Yes.

14:43We've got the CEO, Sam Altman, first here on CNBC. So many questions for Mr. Altman this morning. So have you ever created a software application? Oh, using ChatGPT? Yes. I've played. Yes, I've tried. I tried to actually create. What would a normal person be doing that to do? I'm working on buying something online. That's possible, right? You can see something and then are you able to deliver it to your house? I have not thought I was going to, no kidding, I've done that. But I haven't considered writing a software application for myself. What would I, why would I do that? What would I want to do?

15:21Can you help me? Let's say there was a program that you wanted to build for yourself. Okay, I don't really have that desire right now. I'd like to be better from 90 yards in. But what happens if there was an app called the Kernan app that looked at the news every morning. Yes. And sifted through all the mainstream media bias, mainstream media bias. And then and then tagged each one with a comment from the Kernan AI, which would then explain what's wrong. OK, I want to do that. Have you done this? No, I'm making this app up as we speak. I think we could sell this people. And it's it's called Vibe.

16:02Vibe coding. Vibe coding. We could vibe code it together. All right. We got to talk to this. How much do you want to sell the app for on the on the app store? I could sell it? Yes. I think there should be a subscription model here. I think this is a whole thing. I'm in the 20th century, I think. Is it the 20th? It's the 21st. And then he says, interacting with the new technology feels like talking to a PhD. Is that a selling point? Is that a good thing? You get to ask Sam Altman that very question. All right. You got to say something. I don't want to. I'm not necessarily sure that that's where you're going to get the good answers.

16:41No. Today, Mark, where do you think we're going to get the good answers from? You got to look into your heart and your soul for those. OK.

16:53OpenAI launching its latest new model, GPT-5, offering faster performance, fewer hallucinations, improved coding capabilities and customizable personalities. I've been playing with it for the last 24 hours. Joining us right now is the man who co-founded the company, CEO Sam Altman. Sam, it's great to see you, sir. Good to see you, too. Thanks for having me. We've been trying to build, by the way, an app all morning. I mean, GitHub right now, trying to build Joe Curtin an app unsuccessfully. How's it going? Well, I think you guys have done the right thing so far. I mean, I think that the code looks good.

17:26The problem is I can't get it to actually effectively play for me, if you will. I've got to do it in GitHub. I was actually trying to get OpenAI to actually build me a way to play it effectively in the browser. But I'm not there yet. So I still got one. We'll help you with that. It can do that. But help us understand the distinction between where you think ChatGPT 5 is today versus where we were just even 48 hours ago. Yeah. So you said a bunch of things. I'm a little smarter. It's faster. It's more intuitive. It's more useful in a bunch of ways. One of the biggest things, though, is it's an integrated single experience.

18:01We built a system that can answer easy questions quickly. It can think for a long time and answer hard questions. And it's, you know, there's just one thing called GPT-5 now instead of the long list of models we used to have. It's much better for businesses and enterprises. The coding in particular is something that people have been really waiting for a great model for. So we're seeing super strong enterprise adoption of this. And, yeah. Well, so let me ask you. Some people have called it an Anthropic killer because one of the things that Anthropic, at least historically, has done well is the coding piece of this.

18:34How much of Chappie T5 was focused on the coding side of things? We did. We did really want to improve on coding. That was, you know, an area that we heard consistently from consumers, from users, businesses, enterprise, everything that we just weren't good enough. So we did put in a lot of effort there, but it improved at lots of things. The other piece of it is health care. lots of people ask questions about their own health care, including myself. What's the difference between before and after? And I'm curious how much stuff do you put into the system? Because sometimes I think to myself, I'll upload my reports from doctors and blood and this and that.

19:12And then I think, oh, I don't know if I should be doing that. What do you think? So health care is maybe the area where there's the strongest improvement of any category. We have seen healthcare become a huge issue. People use this, as you said, people are putting in all their reports, everything they can, and it's become really important to people as a sort of, you know, a way to get help with their entire healthcare journey. And I think people feel a very, very big difference there. Healthcare is a huge fraction of the usage of ChatGPT. It's one of the largest categories. So I put everything in, you know, I know a lot of other people do too.

19:49I think getting people better information on their health care and sort of the ability to make better decisions. That seems really important. And I'm thrilled that we were able to make this better. And definitely huge use to it personally. You know, one of the reasons I asked the question is, and I think you made a mention of it maybe on a podcast recently, was just this idea that ultimately the information that we're inputting into systems like ChatGPT and by the way, into search, if you were to do this on Google, are all, quote unquote, discoverable at some point. If somebody were to either sue you or something, they could come to you.

20:24They could come to a company like yours and say, give me everything. And we're now putting so much more into these systems because we're asking oftentimes relatively intimate questions or providing pretty intimate data. Yeah. Look, I personally believe that society is likely to and I think should come to the conclusion at some point that we need something, some concept like we have medical privilege or legal privilege. And if you are asking for this kind of advice, even though it's like a, you know, not a human doctor, but like an AI medical advisor, a lot of those same principles should apply.

21:02And that just the fact that you've asked Chachi BT to analyze your medical records does not mean they should then be discoverable by a court, but they should be protected by the same standards. And I think as AI becomes more and more a part of people's lives and the way we get this kind of information, I think, I hope, whatever, that society will decide to extend similar legal protections. Right. What do you think about safety? And I know that there's a lot of sort of different safety parameters built into this, certain types of questions that you'll answer, but you won't answer fully. Explain what that means.

21:38So, you know, there's this question of like, if you ask ChatGPT to help you build a bomb, you know, should it do that? And there's educational value in that. There's other ways to learn that. You could be trying to learn it for a good reason. But if you're actually like, I'm going to go try to like, you know, really damage someone, obviously we don't want ChatGPT to enable you to do, you know, cause like greed is harm in the world in a way that you couldn't without it. But a new thing that GPT-5 does is try to be smarter and help you with parts of a question that might be okay and then not help you with parts that could be used for significant harm.

22:14This is an evolving area. The calls here are difficult. There's a lot of principles sound easy in theory. And then there's a lot of edge cases. But as AI becomes very capable, society is really going to have to wrestle with these questions of we want to bring the tremendous value and enable people to do things they never could before. And then the amount of power and capability that brings people, you know, we don't want to enable really serious misuse. Hey, Sam, can you share any key architectural changes or training innovations that distinguish GPT-5 from GPT-4 and how they impact its reasoning or grounding capabilities?

22:58And by the way, he got that question, Sam, from ChatGPT. Is that OK? Can it look at itself? You know, that's supposedly the brain can't look at it. I did get that. It's a really good question, though, I thought. I'm watching him, by the way, Sam, literally looking at the... I was. I asked you... He literally asked GPT for the question to ask you. Good questions for you. Yeah. Well, it's an interesting and good question. Thank you. Thank you. Many improvements went into training GPT-5, but one that goes in the direction you touched on is we really started using synthetic data. So the previous generation of model is teaching the next generation of model.

23:36And as these models get smarter, the data that they can create can be really quite interesting and helpful. And we saw that with this model. We expect to see that more with every future model. This gets to the AGI question, which is at what point do you think it becomes where the model starts to teach itself? That's definitely happening now to some degree. Definitely part of the reason GPT-5 is better than previous models is those previous models helping to teach this new one. And so where do you think we are on this AGI path then? What do you what's your personal definition of AGI? And I'll answer.

24:14Oh, that's a good question. Well, what is your personal definition of AGI? I have many, which so I think it's not a super useful term. If I think one definition that people like is, you know, when it's doing like a significant amount of that work in the world, although that keeps changing because people do new jobs. Another that some people use is when it can do like really good AI research on its own. I think the point of all of this is it doesn't really matter. And it's just this continuing exponential model capability that we'll rely on for more and more things. Maybe another one that people like is when it discovers like an important new mathematical theorem and an important new piece of science.