Kara Swisher: [Origins of scaling] …You had started it [OpenAI] as a nonprofit.
Sam Altman: I am not a natural fit for a CEO; being an investor, I think, suits me very well. I got convinced that AGI was going to happen and be the most important thing I could ever work on. I think it is going to transform our society in many ways, and I won’t pretend that as soon as we started OpenAI, I was sure it was going to work, but it became clear over the intervening years, and certainly by 2018–2019, that we had a real chance here.
K. Swisher: What was it that made you think that?
S. Altman: A number of things. It’s hard to point to just a single one, but by the time we made GPT-2, which was still weak in a lot of ways, you could look at the scaling laws and see what was going to happen. I was like, “Hmm. This can go very, very far.” I got super-excited about it. I’ve never stopped being super-excited about it.
Swisher: Was there something you saw that scaled or what was the …
Altman: It was looking at the data of how predictably better we could make the system with more compute, with more data.
K S: There had already been a lot of stuff going on at Google with DeepMind. They had bought that earlier, right around then [2014].
S A: Yeah. There had been a bunch of stuff, but somehow it wasn’t quite the trajectory that has turned out to be the one that really works.

S: …When people critiqued ChatGPT, you wish that you said, “Wait for GPT-4.” Now that it’s out, has it met expectations?
A: A lot of people seem really happy with it. There’s plenty of things it’s still bad at.
S: I meant your expectations.
A: Yeah. I’m proud of it. Again, a very long way to go, but as a step forward, I’m proud of it.
S: What are you proudest of?
A: Well, I enjoy using it, but more than that, it’s very gratifying to just go search for GPT-4 on Twitter and read what people are doing with it, the amazing discoveries people make of how to use it to be more productive, more effective, more creative, whatever they need. It’s nice. It’s nice to build something that’s useful for people.
S: You tweeted that at first glance, GPT-4 seems “more impressive than it actually is.” Why is that?
A: Well, I think that’s been an issue with every version of these systems, not particularly GPT-4. You find these flashes of brilliance before you find the problems. And so, a thing that someone used to say about GPT-4 that has really stuck with me is it is the world’s greatest demo creator. Because you can tolerate a lot of mistakes there, but if you need a lot of reliability for a production system, it’s not as good at that. GPT-4 makes fewer mistakes. It’s more reliable, it’s more robust, but there’s still a long way to go.

A: …These systems are trained to do something, which is to predict the next word in a sequence. And so, it’s trying to just complete a pattern, and given its training set, this is the most likely completion. That said, the decrease [in loss] from GPT-3 → 3.5 → 4, I think is very promising. We track this internally, and every week we’re able to get the number lower and lower and lower. I think it’ll require combinations of model scale, new ideas—
S: More data.
A: A lot of user feedback.
S: Model scale is more data.
A: Not necessarily more data, but more compute thrown at the problem. Human feedback—people flagging the errors for us, developing new techniques of the model—can tell when it’s about to go off the rails.
S: Real people just saying “This is a mistake.”
A: …Yeah. We pay experts to flag, to go through and label the data for us…Not just bounties, but we employ people. We have contractors; we work with external firms. We say we need experts in this area to help us go through and improve things. You don’t just want to rely totally on random users doing whatever, trying to troll you, or anything like that.
S: So humans, more compute. What else?
A: I think that there is going to be a big new algorithmic idea, a different way that we train or use or tweak these models, different architecture perhaps. So I think we’ll find that at some point.
S: Meaning what, for the non-techy?
A: Well, it could be a lot of things. You could say a different algorithm, but just some different idea of the way that we create or use these models that encourages, during training or inference time when you’re using it, that encourages the models to really ground themselves in truth, be able to cite sources. Microsoft has done some good work there. We’re working on some things.

S: [Openness] Talk about the next steps. How does this move forward?
A: I think we’re on this very long-term exponential, and I don’t mean that just for AI, although AI too—I mean that as cumulative, human, technological progress—and it’s very hard to calibrate that, and we keep adjusting our expectations. I think if we told you 5 years ago we’d have GPT-4 today, you’d maybe be impressed. But if we told you 4 months ago after you used ChatGPT that we’d have GPT-4 today, probably not that impressed. Yet it’s the same continued exponential, so maybe where we get to a year from now, you’re like, “Meh. It’s better, but the new iPhone’s always a little better too.” But if you look at where we’ll be in 10 years, then I think you’d be pretty impressed.

A: …Well, for us it’s the opposite. I mean, what we’ve said all along—and this is different than what most other AGI efforts have thought—is everybody needs to know about this. AGI should not be built in a secret lab with only the people who are privileged and smart enough to understand it. Part of the reason that we deploy this is, I think, we need the input of the world, and the world needs familiarity with what is in the process of happening, the ability to weigh in, to shape this together. We want that. We need that input, and people deserve it. So I think we’re not the secretive company. We’re quite the opposite. We put the most advanced AI in the world in an API that anybody can use. I don’t think that if we hadn’t started doing that a few years ago, Google or anybody else would be doing it now. They would just be using it secretly to make Google search better.

S: [Funding] But you are in competition. And let me go back to someone who was one of your original funders, Elon Musk. He’s been openly critical of OpenAI, especially as it’s gone to profits: “OpenAI was created as an open source (which is why I named it”Open” AI), nonprofit company to serve as a counterweight to Google, but now it has become a closed source, maximum-profit company effectively controlled by Microsoft. Not what I intended at all.” We’re talking about open source versus closed, but what about his critique that you’re too close to the big guys?
A: I mean, most of that is not true. And Elon knows that. We’re not controlled by Microsoft. Microsoft doesn’t even have a board seat on us, we are an independent company. We have an unusual structure where we can make very different decisions than what most companies do. I think a fair part of that is we don’t open-source everything anymore. We’ve been clear about why we think we were wrong there originally. We still do open-source a lot of stuff. Open sourcing CLIP was something that kicked off this whole generative image world. We recently open-sourced Whisper, we open-sourced tools, we’ll open-source more stuff in the future. But I don’t think it would be good right now for us to open-source GPT-4, for example. I think that would cause some degree of havoc in the world, or at least there’s a chance of that—we can’t be certain that it wouldn’t. And by putting it out behind an API, we are able to get many, not all, but many of the benefits we want of broad access to this society being able to understand the update and think about it. But when we find some of the scarier downsides, we’re able to then fix them, and we are going to.

S: How do you respond when he’s saying you’re a closed-source maximum-profit company? I’ll leave out the control by Microsoft, but in a strong partnership with Microsoft. Which was against what he said. I remember years ago, this was something he talked about a lot and was—
A: Was what part?
S: “Oh, we don’t want these big companies to run it. If they run it, we’re doomed.” He was much more dramatic than most people.
A: So we’re a capped-profit company. We invented this new thing where we started as a nonprofit—
S: Explain that. Explain what a capped profit is.
A: Our shareholders, who are our employees and our investors, can make a certain return. Their shares have a certain price that they can get to. But if OpenAI goes and becomes a multi-trillion-dollar company, almost all of that flows to the nonprofit that controls us.
S: What is the cap?
A: It continues to vary as we have to raise more money, but it’s much, much, much, and will remain much, smaller than any—…In terms of a number, I truly don’t know off the top of my head.
S: But it’s not substantial. The nonprofit gets a substantial chunk of the revenue.
A: Well, no, it gets everything over a certain amount. So if we’re not very successful, the nonprofit gets a little bit along the way, but it won’t get any appreciable amount. The goal of the cap profit is in the world where we do succeed at making AGI and we have a substantial lead over everybody else, it could become much more valuable, I think, than maybe any company out there today. That’s when you want almost all of it to flow to a nonprofit, I think.

S: I want to get back to what Elon was talking about. He was very adamant at the time and, again, overly dramatic, that Google and Microsoft and Amazon were going to kill us. I think he had those kinds of words, that there needed to be an alternative. What changed, in your estimation?
A: Of?
S: To change from that idea.
A: Oh, it was very simple. When we realized the level of capital we were going to need to do this, scaling turned out to be far more important than we thought, and we even thought it was going to be important then. And we tried for a while to find a path to that level of capital as a nonprofit. There was no one that was willing to do it. So we didn’t want to become a fully for-profit company. We wanted to find something that would let us get the access to and the power of capitalism to finance what we needed to do, but still be able to fulfill and be governed by the nonprofit mission. So having this nonprofit that governs the capped-profit LLC, given the playing field that we saw at the time, and I still think that we see now, was the way to get to the best of all worlds. In a really well-functioning society, I think this would’ve been a government project.
S: That’s correct. I was just going to make that point. The government would’ve been your funder.
A: We talked to them. They not just would have been our funder, but they would’ve started the project. We’ve done things like this before in this country. But the answer is not to just say, “Oh well, the government doesn’t do stuff like this anymore, so we’re just going to sit around and let other countries run by us and get an AGI and do whatever they want to us.” We’re going to look at what’s possible on this playing field.

S: Right. So Elon used to be the co-chair, and you have a lot of respect for him.
A: I do.
S: I’m sure you thought deeply about his critiques. Have you spoken to him directly? Was there a break, or what? You two were very close, as I recall.
A: We’ve spoken directly recently.
S: And what do you make of the critiques? When you hear them from him, I mean, he can be quite in your face about things.
A: He’s got his style.
S: Yeah.
A: To say a positive thing about Elon—
S: Yeah, I’d like you to.
A: …I think he really does care about a good future with AGI.
S: He does.
A: And…I mean, he’s a jerk, whatever else you want to say about him. He has a style that is not a style that I’d want to have for myself.
S: He’s changed.
A: But I think he does really care, and he is feeling very stressed about what the future’s going to look like—
S: For humanity.
A: For humanity.
S: When we did an interview at Tesla, he was like, “If this doesn’t work, we’re all doomed.” Which was sort of centered on his car, but nonetheless, he was correct. And this was something he talked about almost incessantly, the idea of either AI taking over and killing us, or maybe it doesn’t really care. Then he decided it was like anthills; do you remember that example?
A: I don’t remember the anthills part.
S: He said, “You know how, when we’re building a highway, anthills are there and we just go over them without thinking about it?” And then he said, “We’re like a cat, and maybe they’ll feed us and bell us, but they don’t really care about us.” It went on and on; it changed and iterated over time. But I think the critique that I would most agree with him on is that big companies would control this and there couldn’t be innovation in the space.
A: Well, I would say we’re evidence against that.
S: Except Microsoft, and that’s why I think—
A: They’re a big investor, but again, not even a board member. Like true, full independence from them.
S: So you think you are a startup in comparison with a giant partner?
A: Yeah, I mean, we’re a big start-up at this point.
S: And there’s no way to be a nonprofit that would work?
A: If someone wants to give us tens of billions of dollars of nonprofit capital, we can go make that work.
S: Yeah. Or the government, which they’re not.
A: We tried.

S: [GPT-3 political bias] Greg Brockman, your co-founder, said you guys made a mistake by creating AI with a quote, “Left-leaning political bias.” What do you think of the substance of those critiques?
A: Yeah. I think the reinforcement learning from human feedback on our first version of ChatGPT was pretty left-biased, but that is now no longer true. It’s just become an internet meme. There are some people who are intellectually honest about this. If you go look at GPT-4 and test it on…It’s relatively neutral. Not to say we don’t have more work to do. The main thing, though, is I don’t think you ever get two people agreeing that any one system is unbiased on every topic. And so giving users more control and also teaching people about how these systems work, that there is some randomness in a response, that the worst screenshot you see on Twitter is not representative of what these things do, I think is important.
S: So when you said it had a left-leaning bias, what did that mean to you? And of course they’ll run with that—they’ll run with that quite far.
A: People would give it these tests that score you on the political spectrum in America or whatever. And one would be all the way on the right, 10 would be all the way on the left. It would get like a 10 on all of those tests, the first version.
S: Why?
A: A number of reasons, but largely because of the reinforcement learning from human feedback stuff.

S: [Competitors] What do you think the most viable threat to OpenAI is? I hear you’re watching Claude very carefully. This is the bot from Anthropic, a company that’s founded by former OpenAI folks and backed by Google. Is that it? We’re recording this on Tuesday. Bard launched today; I’m sure you’ve been discussing it internally. Talk about those two to start.
A: I try to pay some attention to what’s happening with all these other things. It’s going to be an unbelievably competitive space. I think this is the first new technological platform in a long period of time. The thing I worry about the most is not any of those, because I think there’s room for a lot of people, and also I think we’ll just continue to offer the best product. The thing I worry about the most is that we’re somehow missing a better approach. Everyone’s chasing us right now on large language models, kind of trained in the same way. I don’t worry about them, I worry about the person who has some very different idea about how to make a more useful system.
S: But is there one that you’re watching more carefully?
A: Not especially.
S: Really? I kind of don’t believe you, but really?
A: The things that I pay the most attention to are not, like, language model, start-up number 217. It’s when I hear, “These are 3 smart people in a garage with some very different theory of how to build AGI.” And that’s when I pay attention.
S: Is there one that you’re paying attention to now?
A: There is one; I don’t want to say.
S: You really don’t want to say?
A: I really don’t want to say.

A: …What I am personally most excited about is helping us greatly expand our scientific knowledge. I am a believer that a lot of our forward progress comes from increasing scientific discovery over a long period of time.
S: In any area?
A: All of the areas. I think that’s just what’s driven humanity forward. And if these systems can help us in many different ways, to greatly increase the rate of scientific understanding, curing diseases is an obvious example. There’s so many other things we can do with—
S: AI has already moved in that direction—folding proteins [eg. AlphaFold2] and things like that.
A: So that’s the one that I’m personally most excited about. But there will be many other wonderful things too. You asked me what my one was and—
S: Is there one unusual thing that you think will be great, that you’ve seen already that you’re like, “That’s pretty cool?”
A: Using some of these new AI tutor-like applications is like, “I wish I had this when I was growing up. I could have learned so much, and so much better and faster.” And when I think about what kids today will be like by the time they’re finished with their formal education and how much smarter and more capable and better educated they can be than us today, I’m excited for that.

A: …[Regulation] I do think we need regulation and we need industry norms about this. We spent many, many months—and actually really the years that it’s taken us to get good at making these models—getting them ready before we put them out. It obviously became somewhat of an open secret in Silicon Valley that we had GPT-4 done for a long time and there were a lot of people who were like, “You have to release this now; you’re holding this back from society. This is your closed AI, whatever.” But we just wanted to take the time to get it right. There’s a lot to learn here, and it’s hard, and in fact, we try to release things to help people get it right, even competitors. I am nervous about the shortcuts that other companies now seem like they want to take.
S: …Yes. So this is what happens a lot of the time, even in well-regulated areas, which banks are compared to the internet. What sort of regulations does AI need in America? Lay them out. I know you’ve been meeting with regulators and lawmakers.
A: Yeah, I did a 3-day trip to Washington DC earlier this year.
S: You did. So tell me what you think the regulations were and what are you telling them, and do you find them savvy as a group? I think they’re savvier than people think.
A: Some of them are quite, quite exceptional. I think the thing that I would like to see happen immediately is just much more insight into what companies like ours are doing, companies that are training above a certain level of capability at a minimum. A thing that I think could happen now is the government should just have insight into the capabilities of our latest stuff, released or not, what our internal audit procedures and external audits we use look like, how we collect our data, how we’re red-teaming these systems, what we expect to happen, which we may be totally wrong about. We could hit a wall anytime, but our internal road-map documents, when we start a big training run, I think there could be government insight into that. And then if that can start now…I do think good regulation takes a long time to develop. It’s a real process. They can figure out how they want to have oversight.
S: Reid Hoffman has suggested a blue-ribbon panel so they learn, they learn up on this stuff, which—
A: Panels are fine. We could do that too, but what I mean is government auditors sitting in our buildings.

S: Congressman Ted Lieu said there needs to be an agency dedicated specifically to regulating AI. Is that a good idea?
A: I think there’s two things you want to do. This is way out of my area of expertise, but you’re asking, so I’ll try. I think people like us who are creating these very powerful systems that could become something properly called AGI at some point—
S: Explain what that is.
A: Artificial general intelligence, but what people mean is just above some threshold where it’s really good. Those efforts probably do need a new regulatory effort, and I think it needs to be a global regulatory body. And then people who are using AI, like we talked about, as a medical adviser, I think the FDA can give probably very great medical regulation, but they’ll have to update it for the inclusion of AI. But I would say creation of the systems and having something like an IAEA that regulates that is one thing, and then having existing industry regulators still do their regulation—
S: People do react badly to that, because the information bureaus, that’s always been a real problem in Washington. Who should head that agency in the U.S.?
A: I don’t know.
S: Okay. So one of the things that’s going to happen, though, is the less intelligent ones, of which there are many, are going to seize on things like they’ve done with TikTok, possibly deservedly, but other things. Like Snap released a chatbot powered by GPT that allegedly told a 15-year-old how to mask the smell of weed and alcohol, and a 13-year-old how to set the mood for sex with an adult. They’re going to seize on this stuff. And the question is, who’s liable if this is true, when a teen uses those instructions? And §230 doesn’t seem to cover generative AI. Is that a problem?
A: I think we will need a new law for use of this stuff, and I think the liability will need to have a few different frameworks. If someone is tweaking the models themselves, I think it’s going to have to be the last person who touches it has the liability, and that’s—
S: But it’s not full immunity that the platform’s getting—
A: I don’t think we should have full immunity. Now, that said, I understand why you want limits on it, why you do want companies to be able to experiment with this, you want users to be able to get the experience they want, but the idea of no one having any limits for generative AI, for AI in general, that feels super-wrong.