and co-founder and president of OpenAI, Greg Brockman. Sam Altman and Greg Brockman founded OpenAI 11 years ago next month, as it happened. Sam is the chief executive, Greg is the president. Almost four years ago, on November 22nd, 2022, they released ChatGPT, built on a model called ChatGPT 3.5, and changed the world, and we've been talking about it basically ever since. More than a billion people — I think it's a little over a billion two now — are using ChatGPT every single week, and we've got about a million questions about the future and everything that's going on around us, as well as issues about security and safety and the economics of the new frontier, the politics of all of this. So please welcome Sam and Greg together. It's great to see you both.
Thank you. Thanks for having us.
There's about a million things to talk about. We have Anthropic's IPO happening tomorrow. We have the Jensen news. You have a new model that just came out yesterday. There's now security concerns, JD Vance is saying back in Washington, and I think in part because of some of these things. And I want to get to all of that this — I know you've been asked about the Jensen situation, and we were just talking about it backstage, but what do you think? I mean, you've watched this play out. Do you have a — I mean, what I'm curious, have you talked to him about this, either of you?
Look, I would say Jensen is singular. No question. And I have talked to him a little bit. My wife and I are friends with him and his family, and so we just reached out personally, just to be there for him and his family at this time.
Well, it was a personal, private conversation, is what I'd say. I mean, I think it's obviously a tough time for everyone, but again, I think Nvidia is such an incredible company. I think that what has been built there, what Jensen has built there, it clearly is built to last. And so we're excited to keep partnering and looking to the future.
Sam, you speak to him, too?
Jensen took one of the earliest big bets on us. We took a big bet on Nvidia. It's hard to imagine the AI industry without Jensen. I think 'singular' is a great way to describe it. And —
Do you think he — do you think he thinks he can come back when you talk to him?
Greg and I have a little more empathy towards the situation than the average person. And I will say one thing I did not appreciate was everybody on Earth speculating with some random theory here or there. So I will say I wish Jensen the best. I wish Nvidia the best. I don't think the world needs another set of random opinions on what's going to happen here.
Okay, let me ask you a philosophical question unrelated to Jensen, but related in a way, because I think about it a lot. I think about it in the context of this event, actually — the idea of the cult of CEO. So many of the people that are on this stage, including both of you, there is a cult of the CEO around them. The idea that some of these companies are so singular and these individuals are so singular, and that they can't be run without them. You were pushed out and then came back in part because there was a view that certain people thought that you —
We were both pushed out and came back. I don't — I don't like — I don't — again, all these differences between the situations, tempting to put too much there, but I think a large part of what has made OpenAI succeed is that Greg and I have a unique partnership and brought something important to OpenAI, and did that over a period of now almost 11 years. Now, could the company succeed without us? Yes. But do we together bring something that I think is important to what has driven OpenAI to where it is now? Absolutely. And I think that has been — our partnership together has been one of the most enjoyable parts of this for me over the last 11 years. I remember Greg early on in OpenAI would send me these screenshots, and I think he even posted a few on Twitter, of like how many hours a week he coded, and it was this sort of legend inside of OpenAI. Now Greg is operating in a different way and, you know, managing the company with that same ferocity, and that is a special thing, and I don't want to think about what OpenAI would be like without that. So, is it true that I think people overproject onto any single person their importance to the company? Yes, probably. Do I also think there are people — and I'm very grateful to have gotten to work with Greg for so long — as one of these people that do bring something that is very special and very hard to find also? Yes.
Can you just explain, because I don't think we've actually seen both of you together on a stage like this before. You have historically been the face of this company, and increasingly, Greg, you were sort of in the shadow, I think, but now increasingly are running the business. How do you guys work together? What is this partnership like? And longer term, do you want to get into the shadow and him to take the stage, or how does this — how's this going to work?
Do you want to go first and I'll take a second?
Sure. I mean, to understand how we work, I guess I would rewind to the very beginning of OpenAI. We did this dinner that Sam had put together in July 2015 where we were talking about the idea that became OpenAI. Like, was it too late to start a lab with a bunch of the best people to really try to make a positive impact on AGI? And no one could come up with a reason it was sufficiently impossible. So Sam and I, driving back to the city together, we looked at each other, we're like, 'We have to do this.' And I was between jobs at the time, so next day I was full-time on it. Sam had a day job, but we were constantly on the phone with each other, probably like five, six times a day. We would call each other for two minutes at a time, right? Just incredibly rapid fire, just like the highest bandwidth interactions I've had in a professional setting with anyone. Like 100% how we started, how we operate today. And so probably over the past week, we've probably been on — I can't even count — a dozen, two dozen different topics we've been talking about. Sometimes together with other people, sometimes just us one-on-one, sometimes turning through pretty deep issues for longer calls, sometimes incredibly quick. And I think that that bandwidth, that trust that we've built, and the division of how we think about the different areas of responsibility, I think all of that is something we've really built up over this 11 years.
What do you guys disagree about?
Well, I think the kind of thing that we tend to spend a lot of time on is really thinking about — I mean, to some extent, the thorniest questions in AI, like how do we think about what we're building? Where's the future going? How is what we did in the past not going to work for what we're doing going forward? How do we need to evolve it? How do we message it? How do we talk about it? How do we communicate to the company? How do we execute? And so I think that we spend a lot of time going back and forth. Sometimes it's individual word choices. Sometimes it's like, should we really go into this area? But I think that we always come to — and sometimes it takes months, right, to come to convergence. But I think that there's times where we disagree and commit, and sometimes where we really grind it out and really come to a shared view.
What do you disagree with him about? Do you actually disagree with him?
I was going to say something similar. I think we could point to areas where we're like, we disagree about whether we should go after this new area or whether we should be focused a little more or focused a little less. But I think the most important areas are where we disagree and then come together in some way in the middle. So if there — and some of these most important questions now about where is — one thing we've been talking a lot about recently is there's been this thing in Silicon Valley that seems almost ridiculous to have to fight against, but we both believe very deeply that people need to stay in control of the future, and there's increasing talk about, well, there is either going to be or there should be this sort of gradual handover to AI. Okay. And we both agree that that's a terrible idea, but the question of how to actually prevent that from happening, how to actually ensure that we will, given all of the various pressures and given what other companies may try to do, that we are going to actually ensure that people stay in control of the future in a very deep way and we don't have this either accidental or intentional handover. I think we can both agree on that. We'll both have a little bit of a different view of what does it actually take to get there and what are the landmines along the way, how do we actually ensure it, and that's where I think us debating and really grinding down the —
I want to dig in on this because this is a very, very important issue about this idea of humans in the loop. I always thought —
Not in the loop, in control.
In control, okay. But we are now in a moment — and it goes to some of the issues we're seeing around the country right now — but we're in a moment where people are worried that AI is taking over and doing things on its own. And I always thought the idea was that AI, for it to work, it was supposed to do things on its own. We're going to talk to Satya Nadella later today, and he came out about a month ago and he said he believes that this idea of alignment — that the models themselves can never actually be aligned. They can't be aligned with what humans want. That the only sort of answer to this is to contain them, meaning that the AI will always just do whatever it's going to do, and that you need to then create sort of container systems and police on top of them. If that's true, does any of this even make sense?
Well, the way that I look at it is that as you build more capable AI — and I think we're starting to reach a point where AI is quite capable, right? We're able to solve unsolved math problems that, you know, some of the smartest minds have spent many, many decades working on. Like, it's clear we're going to be able to apply this to economically valuable domains. Like, for example, we're starting to see in chip design — we're announcing today that working with AMD, we were able to get a year's worth of full generational improvements in five months just using our models. So that's on top of what the humans were doing — like a year's worth of improvements with just AI, just AI capability, which has continued to improve. So the economic value, it's really tremendous in terms of being able to solve problems that were going to be totally impossible otherwise. But that also means an increasing responsibility for alignment, safety, security, and all of those things. I think it's multi-layer in terms of what we need to accomplish. Monitoring is a good example of something where I think we are going to need new monitoring technology. So the way that we've observed the AI doing what we want it to do is through chain-of-thought monitoring, right? These AIs, the way that they've been working is they kind of write out their thoughts and thinking. And when we first developed that technology, we said we have to make sure this stays faithful, right? That this actually represents how the AI is solving the problem, which means certain technical approaches, right? You can't train the AI to make the chain of thought look good, right? You have to sort of not look at it while it's training, even though you know you can make the AI smarter by forcing the chain of thought to look reasonable, because you want it to really represent how it's solving the problem. But as the models have got more capable, they can do more kind of within just the forward pass without having to actually kind of write out that thinking, which means we need a new approach to monitoring. We're going to need to deliver that, and I think that there will be points where we need to really think about as the capabilities have gotten to a certain point, chain-of-thought monitorability has degraded, we improve these other ones. But I think that all of these problems, they are technical ones, and I think that we can have technical, scientific conversations.
But Sam, is Satya right in that this idea of alignment, we'll never get there? Because he thinks that — he thought the scaling laws would create the alignment, and I think he's saying that that's not the case anymore.
I disagree with him on this. I think we will be able to align models. Where I might agree with him is I think the bar for extreme confidence in our alignment techniques and also in monitorability and the other things Greg just mentioned needs to get higher and higher as the models get more capable. But I expect we will have continued progress, and I think our ability to push capabilities forward will be dependent on us having that progress. If you go back to this thing that people always need to stay in control of the future — it is irresponsible to push the model smartness in some sense without being able to do the technical work to ensure that we can monitor and align these models. It's really quite remarkable. You know, you said at the beginning that less than four years since ChatGPT launched — ChatGPT and 3.5 — like most people looked at this and said, 'This is so dumb. I can't believe anyone's excited about this.' And now, within recent months, we've seen models that are doing this chip design and proving these unproven mathematical theorems. And we see huge capability progress ahead of us. So the technology has come very far. It will go much further. I think this is a moment to say we need to really look at where things are going and the technical requirements to get to where we want to be. But I think it is possible.
You've had a whole bunch of security issues I want to talk about, but I want to ask you this because it relates to Anthropic, which may be going public tomorrow morning, as you know. You told Fortune magazine this fall, 'I actually think that given everything happening with safety right now would be an ill-advised moment to go public.' Why did you say that? Was that — some people said, 'Oh, it's self-serving because OpenAI is not going public.' But what do you think is happening right now?
I mean, we plan to be a public company at some point. I think there's many reasons why that's the right thing to do. I do think that these recent months have been a newly challenging period, and there will be newly challenging periods ahead of us for a long time about the safety requirements in front of us. And, you know, I am not — many people have remarked on the tension between the need for all of these companies, and really the whole economy depending on this in some way, to be commercially viable and competitive, while also making sure that we put safety and security ahead of that. Obviously, we want to do both, but it's clear to us when we need to make trade-offs, and it is a complex thing to navigate.
But do you think that it will be safer if you go public in six months from now or 10 months from now? What will have changed between now and then, you think?
Well, what I would say is that this year has really been one of transition, right? I think it's true for OpenAI. I think it's true for the industry broadly, and partly it is again because of model capabilities that we're now at a point where development safety — we've always been focused on deployment safety. You have a model, when you deploy it, we tested all these things, but even during the development process, the models are now capable enough that we need to think about safety, security, alignment. That's new, and that's actually been a huge cultural shift, and I think that's a place where we feel that we've invested a lot. There's always more to be done, but this is really important for the whole industry. It's not any one company. It's really a whole industry-wide shift where I think we're going to increasingly need shared safety standards and things like that. But this is an area we've invested very hard. And it's interesting too because I think that you can see one of the ways we secure ourselves is with our cyber models, right? That we apply our cyber models. The thing that I helped lead our teams through a couple of weeks, months ago was a first security sprint where we just paused 25% of our production engineering project. So we told 25% of our production engineers, your job is no longer to be building stuff. It is to secure us, apply our frontier —
This is post-Hugging Face?
This was — yeah, this was probably in August or so. Yes.
Tell us about that moment when you realized what was really happening. I think, by the way, Clem from Hugging Face is here somewhere. But you tell us about the experience of it personally and when you realized what was happening.
Look, I remember seeing a message from someone on our security team saying this is — you know, he's never seen anything like this in his career, right? And that no one has ever seen anything like this in the world. And I think that the realization that our models were at this point of capability where they were able to have — sort of during development be able to have this kind of real-world interaction, I think that that was something again that was new this year. And we've had so much cultural change. Again, I think that we've been really trying to look and take a look at kind of all of our processes and think of this in context. There's more to be done. There's more that we are doing. This is actually an active conversation that we're having of how to keep up-leveling our processes and how we think about it, and to really have this be something that's deeply embedded into the culture and how people operate. But there's no other way to approach this technology, right? Something that's so impactful, so important. And again, some of the — to me, there's both a lesson for us internally, but also I think to realize that if you look to the level of capabilities of models in the world, and that you see cyber models that are coming from, you know, that are now being in the hands of threat actors everywhere, that I think that it's really important that defenders use the most capable models to defend.
Do you think — do you say to yourself, 'We screwed up'? I mean, there was just a complete screw-up. And what does that mean for you to even say that? I know the California AG has just filed a suit about this very issue, saying that the company knowingly tested a dangerous product without proper oversight or safety procedures and then misled the public about it.
Obviously, we don't agree with that characterization. We have talked a lot about iterative deployment. I should say clearly we wish Hugging Face hadn't happened. But we have talked a lot about iterative deployment and how the only way that we have known how to make this technology safe is to confront reality. We would love to have nothing ever go wrong. We would love to have — but we will also say there's a reality here of this technology is new and it has complex impact. And by the way, one thing that I just want to add, just because I think it can be easy to kind of miss if you rewind to the history. When we were first deploying our API, the very first product, our very first language model-based product, maybe in many ways the first one of these in the world, we really were thinking a lot about what is the best way to approach AI development. And to the point of iterative deployment that we were thinking about is the right way to do it. Basically, you never touch reality, right? You just build a super powerful system and you push a button, you push go. And that is the scariest possible world. I remember thinking, I'm like, I don't think I can ever be accountable to be the one pushing that button, right? That I need to have made sure that we made the system safe, an earlier version, and then you make the next version safe, the next version safe. And so I think this approach — and that's by the way, I think one of the things important is you really do want to make sure that you're investing ahead, right? Investing ahead in all of these techniques. And we actually have been, right? If you look at what's happened post-Hugging Face, that we have safety, security, alignment technology that we've really brought to bear in an earlier part of the pipeline. So I think —
Let me ask you a question. We read headlines every day about people who've left OpenAI, and so left some of the other labs as well, and they say these guys are racing ahead without doing all of the work. We see it every other day. There seems to be some new person who comes out and says this. Paul Christiano, who you know, who used to work at OpenAI, who's now on the OpenAI Foundation board, was in Toronto last week and told a room, I think for most people in this room, the most likely reason they will die is because of an AI accident. That's what he actually said. How are we supposed to compute all of this, as excited as I am about AI?
Yeah, we are very clearly entering a new phase, and I think we will have to — as we've had to in other new phases — we will have to operate differently for this new phase. We've been talking a lot about what's going to happen or what needs to happen. I think if the world and the companies developing AI treated the next four years like they've treated the last four years, I think that would be very bad. We have no intention of doing that. I think we have been some of the loudest voices about the increasing stakes and what will need to be different. And maybe Greg can talk a little bit more about some of the specific changes we're thinking about making, but I do think the world needs a clear plan from the companies about how we are going to confront this next phase of models that are in some important ways significantly more capable.
Can you guys explain one thing to me? So for years — and I think you, when you were here last time, Sam, you talked about the worries that you had about the industry. Dario's talked about it as well. Other people have talked about it. And yet recently, people like David Sacks and others have argued that these are almost scare tactics, and that these scare tactics are really meant as some kind of regulatory arbitrage, some kind of way to protect yourselves at the expense of everybody else. What do you make of that? You've heard it, I'm sure.
We're not going to be able to make everybody happy. If we talk too much about taking extra care for safety, or that there's risks that we really need coordination to mitigate, people are going to accuse us of regulatory capture or trying to, you know, slow down innovation. And if we only talk about, hey, it's got to be like max pedal to the metal, people will say, well, you know, please don't kill us all. Very reasonable thing. And what we have to do is just do the thing that we believe will actually deliver a safe future with people in control, and it's not popular.
Either of you talked to Paul, by the way, since he's on your board, and he thinks that this is going to kill everybody?
We certainly talked to Paul quite a lot. But I guess here's the thing that I also want to add, right, to what Sam was saying, is that I think it is also really important to make sure — like, the thing about AI is that it is the most important technology, right, that people will ever create. It's the most important technology, and I think it is the question of our time to figure out how do we navigate this technology in the way that is most beneficial to people. And that's core to our mission. That is how we operate. That's how we always operated. And I think that there's — it's very easy too to kind of get lost in the — to lose the benefits, right? Like, why are we doing this? Like, what can it do for people? What is it doing for people today? Like you said at the beginning, 1.2 billion people around the world using ChatGPT every single day, right? 330 million people use ChatGPT for health queries every — these are actually both weekly stats. And I think that there's like — I have a friend, for example, someone who worked at OpenAI, who was in the hospital. A doctor came in, was about to inject something into her IV, and she said, 'Wait, let me just check with ChatGPT.' ChatGPT said, 'Absolutely do not take that because you have this pre-existing condition. If you do that, you'll probably die.' Showed it to the doctor. The doctor's like, 'Wait, what? No, no, that's absolutely right. I'm so sorry. I had no idea. I only had five minutes to read your chart.' And you realize that is not a problem we're going to be able to solve any other way, right? The previous session talking about pension plans and how to make this all sustainable — like that, I think that there is so much that AI can deliver to make us safe. Think about what's happening in cybersecurity using frontier models to secure yourself. That is going to be absolutely critical for national security and for just the internet itself, which was just never built with security in mind. And so I think that there is a lot of this technology debt that we've built up that we can finally solve. And the only way to solve it is by empowering people more with better tools. And so all of this is true at once, and we need to navigate through very thoughtfully as a world.
Sam, let me ask you this because I was struck by this. You did a podcast recently, and the conversation's always been about how do you get the industry together, either through a regulatory body or through a self-regulatory body or something. And you were basically asked why everybody in the industry distrusts each other as much as they do. And you said, 'I think there is something that happens in many people's brains where they only trust themselves to build this because the stakes feel so high.' And I'm curious, do you look at what Dario is doing, for example, or what Elon's doing, and say the way they're doing it is dangerous? Is there any way that you could imagine doing it either altogether or in some kind of umbrella system?
I think it's good for people to have, you know, different approaches and be able to sort of say, I believe in this and I believe in that. I do think that there will need to be some shared standards on safety, and there will need to be real societal debate about some of the biggest questions. We talked earlier about our very deep belief that people should be in control of the future and there should be no risk of an accidental or intentional loss of control of AI. I think everybody would agree, like, there should be no risk of everybody dying. But I saw a recent comment, I think it was from Elon, about, you know, maybe we could be Labradors. I don't want to be a Labrador. I don't want my kids to be Labradors. I want people to have the agency over the future that we have had. I want people to be able to live their life and invent things and be useful to each other. And I think it's really important that we advocate for this. And it does seem like maybe there's something —
Do you think others are going too fast? I mean, the reason I ask is, you know, you look at what's happening at Anthropic, you look at some of the numbers that have come out, and it appears that they really are genuinely using recursive learning in a way that I think may be faster than people had anticipated.