[0:00:00] Have you run a model that's all synthetic data? I don't know if I should say. Okay, but... A lot of synthetic data. A lot of synthetic data. So how worried are you that you're going to kind of get mad cow disease in your models? Hello, and welcome to the most interesting thing in AI from Rethink, the Atlantic's creative marketing studio, in collaboration with PwC, a leader in helping organizations use AI. Sam Altman has been a...
[0:00:34] Important figure in Silicon Valley for roughly two decades. But as the head of OpenAI, particularly since the launch of ChatGPT three and a half years ago, he's become a household name. He's sought out by heads of state, by journalists, by CEOs like me. I'll talk to him about his thoughts on the future of the technology and how it will change our world. With all that attention have come some dark consequences. In April, Sam's home was attacked with a Molotov cocktail.
[0:01:01] And then in a separate incident days later, someone fired a gun on the street outside of his house. So it was just after all that happened that he sat down with me for a conversation at OpenAI's office in San Francisco. Now, I had, as you can imagine, a million things I wanted to ask him about. But some of the most controversial issues, things that make headlines, are better left to journalists at The Atlantic. I'm the CEO, so I wanted to talk to Sam about the big picture, where the technology is going.
[0:01:30] What it means for all of us. And as you'll see in here, we got into some pretty interesting stuff. I have no doubt that you'll at least be curious about this conversation. I should add that the Atlantic has a content licensing agreement with OpenAI, so I wanted to discuss that. But mostly, I wanted to know about the guts of the models, how they're evolving, how AI really works.
[0:02:01] To Altman, there are elements that are mysterious. Anyway, Sam had a lot to say about that, a lot to say about a lot of different things. I truly enjoyed the conversation. Here's my conversation with Sam Altman. Welcome to The Most Interesting Thing in AI. Thank you. And thank you for being here. I know it's been a busy and intense week, and I appreciate the time. What I want to do in this conversation is ask you some of the biggest questions about AI, some of the things that interest me the most.
[0:02:28] Starting with something we've talked about a few times before, which is explainability. So let me begin. You gave an interview with Patrick Collison. It was three years ago. And he asked this great question. He said, is there anything that could change that would make you more confident in good outcomes and less worried about bad outcomes? And you said, well, we could really understand what's happening to like a neuron level. And I asked you the same question a year later. We talked about it six months ago. So let me ask you it again now. Is our understanding...
[0:02:58] Of how AI works, increasing at the same speed as the power of AI is increasing. Let me answer that question and then talk about, back to Patrick's question of what would make me more or less confident, because the answer there has changed a lot. Wow, I agree. On the understanding of what AI models are doing, I think we still don't have a great mechanistic interpretability framework of the world. It's gotten somewhat better, but I don't think anyone would say I understand exactly everything happening in one of these neural networks.
[0:03:28] The chain of thought interpretability has been quite a promising direction for us. It's fragile and it depends on a bunch of things not falling to various potential optimization pressures. But I can't inspect my own brain with an x-ray machine and understand exactly what's happening in, you know, every neuron is firing and connecting. But if you ask me to explain to you why I believe something, how I came to some conclusion, I can do that.
[0:03:56] Maybe it's how I really thought about it. Maybe it's not. I don't know. There's a introspection failure there. But whether it is or isn't, you can look at the logic and say, okay, that is a reasonable conclusion given these steps. The fact that we can do that now with our models does seem like quite a promising step forward. I can still imagine all sorts of ways where things go wrong and where the models can
[0:04:24] Deceive us or conceal stuff from us or any number of other things. So I wouldn't say this is like a complete solution by any means. But even in my own use of these models for increasingly powerful tasks, I was one of these people that said, I am definitely not going to let Codex just have full access to my computer and run in what we call YOLO mode.
[0:04:47] And I lasted a few hours, and I was like, it's too helpful. You've given Codex access to your whole computer? To be perfectly honest, I have two computers, but yes, one of them I have. I hope it's air-gapped. One of them I have. The fact that I can kind of see what the model is doing, and the fact that the model can explain to me why this is okay, and here's what I'm going to do, and that I trust that it's almost always going to follow that. But hold on a second. So a chain of thought for everybody is you type in a question, and then it says, looking at this, doing this, and you can follow it.
[0:05:17] In order for chain of thought to be a good method for explainability, it has to be true. The model can't be lying to you. And we know that the models deceive you sometimes and lie about what they're thinking and how they reached answers. So how do you trust chain of thought? So you need many other pieces in the defense thing so that if the model says this, it's actually a true statement. And this is something our alignment teams work on a lot. As I said, I don't think it's like the complete solution, but I think it is part of a package.
[0:05:46] But it is a...
[0:06:16] Important tool in the toolbox. The thing that has both interested me, so fascinated by this topic, it's not like a car, right? It's not like you built a car and you know how it works, right? Okay, this leads to an explosion here and this leads to this and this leads to that and then the wheels go forward and you go forward. You've kind of built this machine and you're not really sure how it works, but you know what it does and you know the limitations. And so the quest to understand more about the inner workings is
[0:06:44] Amazing. So you feel like chain of thought is the best way to understand the inner workings? Important part of the whole package. Yeah. But not sufficient on its own. My favorite study about why AI models are weird, and you can explain what it means. So this is the Anthropic paper that we had the preprint in the summer and it just came out, where if you tell a model that it likes owls, and owls are the most wonderful bird, and then you have it generate a bunch of
[0:07:13] Random numbers. And you take those random numbers and you put them into a new model. That new model likes owls, which is crazy. And you ask it to write a poem and it writes a poem about owls. But all you've given it are numbers, which means both that these things are deeply mysterious and it also makes me worried because obviously you could be not telling it to like owls, you could be telling it to shoot owls. You could be telling it all kinds of things. So explain what happened in that study, what it means and what the implications are.
[0:07:44] When I was in fifth grade, I got really excited because I thought I didn't know how airplane wings worked. My science teacher explained it to me. And I felt so cool. I was like, oh yeah, the air molecules have to go faster across the top of the wing. That creates lower pressure and the wing gets pulled up. And I looked at this very convincing diagram in a fifth grade science textbook and I just felt awesome. I'm sure I came home that day and I told my parents, I understand how airplane wings work. And then it hit me later, I think in high school physics.
[0:08:14] That I had been repeating to myself this thing in my head about how an airplane wings work and the air molecules go faster across the top. I actually had no idea. I didn't understand how airplane wings work at all. I don't really understand how they work now. I could explain it at some reasonable level, but if you drilled me all the way down to exactly why do those air molecules go faster across the top of the wing, I don't think I could give you a deep and satisfying answer. I could give you a bunch of answers about why people here think that that results about owls.
[0:08:45] But Tim, you don't run Boeing, right? You run OpenAI. Totally. I can tell you a lot of other things. Like I can tell you about how we get a model to perform to a specified level of reliability and robustness. Yeah.
[0:09:15] But there are mysteries of the physics here. I can tell you how to like, maybe if I ran Boeing, I could tell you how to build an airplane, but I couldn't like perfectly understand all the physics. So let's go through that OWLS thing. So let's talk about the implications. So if indeed these models transmit knowledge to each other that is hidden and that is not viewable, like you could have a train of thought that is like, we're inputting this number sequence. Somebody could watch it and say, oh, look, look at those numbers going by. And you would have no idea that it's conveying this information about OWLS. Could end up being.
[0:09:44] Dangerous, problematic, weird. So when I said I would give a different answer now to Patrick Hollison's question, that gets at what I would say. At the time, you said it was three years ago? Yeah. My understanding of the world three years ago would have been, we have got to figure out how to align our models. And if we can align our models, and if we can prevent them from falling into the hands of bad actors, we should be pretty safe. Those were kind of the two threat models that I thought most about.
[0:10:15] You know, we don't want the AI to decide it wants to do ill to us. We don't want someone to use the AI to do ill to us. And if we can avoid those, you know, yes, we got to figure out the future of the economy. And yes, we got to figure out the future of meaning, but we'll be okay. As time has gone on and as we learn more, I can now see such a different set of problems. A way we've started talking about this is like AI resilience instead of AI safety. There are the obvious cases, like it no longer seems like it'll be enough.
[0:10:46] To just make sure that the frontier labs do a good job of aligning their models to tell people how to not make a bioweapon, to not tell people how to make a bioweapon because there'll be good open source models and society is going to need a series of defense shields if we don't want novel pandemics. That's a scary thing because- Just to pause for a second, this is important, I want to continue. Even if you tell people how to not make a bioweapon and your model will not help anybody make a bioweapon and you're not hacked, it doesn't matter as much as you thought it would because-
[0:11:15] They're really good open source models that will do it anyway. But that's just like one example of many of where society needs a whole of society approach to new kinds of threats. Now we have a new tool to help us with those, but we have a very different landscape than I think many of us thought. And it's great and required to align models and to build good safety systems. But this is just a thing that is going to permeate society everywhere.
[0:11:43] And like we have had to do when we face other kinds of new technology, we will have to figure out how we're going to prevent new categories of risk. It sounds like it's getting harder. Harder and easier. Harder in some ways. And then also we have like amazing new tools that can help us do other, do new kinds of protection that we couldn't have imagined before. An example of where this is happening right now. And then I do want to get back to the comment I was going to make because I think it's interesting on anywhere.
[0:12:16] Getting into something quite important is cyber. So the models are getting very good at basically hacking computer systems. Thankfully, the people that have the best models right now are all quite worried about people using AI to wreak havoc on computer systems. And so we have a period of time where there are a limited number of good models and people are going as quickly as they can to use them to secure systems. And if we didn't have that, if we didn't have that kind of advantage,
[0:12:48] The hacking models would come in the open source models soon enough or in the hands of our adversaries soon enough and cause a lot of problems. So we have this new threat. We have this new tool to protect against it. And there's a question of if we can do it fast enough. That's a new example of where the technology can help us solve the problem before it becomes a huge problem. On the thing you were saying, a thing I didn't think about three years ago in the category of these sort of new sorts of society-wide risks are I did not think about a world.
[0:13:18] I really didn't think about this until I watched what was happening with OpenClaw when it launched, but I did not think about a world where we were really going to have to focus on building and deploying agents that were robust to being infected, for lack of a better word, by other agents. That was just not in my model or really in anyone else's model of the world that I talked about as one of the most pressing problems we would have. Of course, there were results like the OWL result and other things where clearly there was
[0:13:45] Strange behavior you could induce in these models in ways we didn't fully understand. But I had not really thought about what it would, until I watched what happened in the early days of OpenClaw, what it would be like for, let's call it misbehavior to spread from agent to agent. Right. I mean, and in fact, the two threats you've just talked about could be particularly terrifying because employee at OpenAI sends out their agents. Those agents go out into a world, someone with a...
[0:14:12] Different model that's very good at hacking figures out how to manipulate those agents. Those agents then come back to the home base at OpenAI and suddenly you guys are hacked. You can totally imagine something like that happening. Right. So what do we do to make the odds of that happening less? The same thing that we have done throughout all of the history of OpenAI. There has been one of the central tensions in the history of OpenAI and really the field in general has been sort of pragmatic optimism.
[0:14:45] Versus power-seeking doomerism. Doomerism is like an extremely powerful stance. It's very hard to argue with. And a lot of this field has been driven, quite honestly, out of huge fear. And not completely misplaced fear. But there's a limit to how productively you can act without more data and more learning. Maybe the kind of like...
[0:15:17] AI safety community of the mid-2010s did the best possible thinking anyone could do in theory before we actually understood more about how these systems were going to be built and how they were going to work and how society was going to integrate with them. And I think one of the most important strategic insights in the history of OpenAI was deciding we were going to pursue iterative deployment because society and technology are a co-evolving system. And it's not even like, oh,
[0:15:47] We don't have data, and so we can't know exactly what's going to happen and do good enough thinking. It's that society is going to change with the evolutionary pressure of this technology and the landscape, the ecosystem, whatever you want to call it, is going to change. So you have to learn as you go. You have to adapt with a tight feedback loop. I don't know what the best way to make these agents safe in a world where they're going off and talking to other agents and then...
[0:16:16] Coming back to home base. But I don't think we're going to solve that like thinking really hard about it theoretically. We're going to have to like learn from the contact with reality. We're going to send the agents out and then we'll see what happens. Well, okay. Let me ask you then a question about where we are with AI. It feels to me as someone who uses these products and all the time in as many ways I can both to learn and to help me understand how to help my business survive the future. It seems to me that we've had more progress in the last three months than at any point since.
[0:16:44] You know, the launch of ChatGPT in December 2022. Is that because it's a moment of particular innovation? Or is it because we've kind of reached a moment of like recursive self-improvement where AI is helping us improve AI faster? Because if it's the latter, like we're in for, I mean, exciting, but a real roller coaster is coming. I don't think we are yet in.
[0:17:07] The recursive self-improvement moment in the way people traditionally take and just talk about it. I'm going to define it, you know, when AI can help you invent the next AI, at which point machines inventing machines that invent the next machines where you get extremely powerful extremely quickly. I don't think we're there yet. But where we are is that AI is making engineers, researchers, actually everybody at OpenAI and in other companies to work more productively. So maybe I can work.
[0:17:34] Well, I don't matter in this case. Maybe one of our engineers can work twice as fast, three times as fast, maybe 10 times as fast. That's not really like the AI doing its own research in any important way, but it is things happening faster. The thing that you're referring to, though, I think is not chiefly about that, although that's an important point. There are these cases, and I think we've sort of experienced three of them now with the latest one happening just now of...
[0:18:03] The models gets over some threshold of intelligence and utility where all of a sudden something that didn't work now works. In my own experience, it's not a super gradual thing. Before GPT 3.5 and before we figured out how to do fine-tuning on that for instruction following, the chatbots just weren't compelling beyond a demo, and then all of a sudden they were. Then there was a moment.
[0:18:35] Where the coding agents went from like kind of nice autocomplete to like, wow, this is doing a real task for me. Right. And it wasn't, it didn't feel super gradual. Like over a period of a month or something, the models just crossed some threshold. The latest one with the update that we just shipped to Codex, which I've been using for the last week or so, really good computer use. This is an example of it's not even, I mean, it's sort of model intelligence, but it was more about putting good plumbing around it.
[0:19:02] This has been one of my sit back moments of something big is happening here. Watching an AI use my computer and do complex tasks, it really made me realize how much time we all waste on drudgery and that we just accept. Can we walk through what exactly is the AI doing on Sam Altman's computer? Is it doing it right now while we're sitting in the podcast studio? No, my computer is closed and we have not yet found a way.
[0:19:31] We need a way where that can happen. I don't know what that's going to look like yet. Maybe we'll just all have our laptops keep running when we close the screen and always connecting with power. Maybe we'll all set up a remote server somewhere. Something is going to happen. I don't have the anxiety as bad as other people do who wake up in the middle of the night to start kick off new codex tasks because they feel like they're just wasting time otherwise. But I get the spirit of that. I know the feeling of that.
[0:20:00] Right. I mean, I woke up this morning and I wanted to check in on my agents and what they had found and give them new instructions and get a report and get them going again. Yeah. People talk about it in a way that sometimes feels like unhealthy, addictive behavior. Right. Can you say what they're doing on your machine? The best thing that I have them do right now is to deal with Slack for me. Not just that, I don't know about you, but I have this just mess where I'm constantly going between.
[0:20:28] Slack and iMessage and WhatsApp and Signal and email. I feel like I'm just copying and pasting things around. I'm doing a lot of drudgery. I'm trying to find files and go wait for some basic thing and just do some very menial thing. I didn't realize how much time I spent doing that until I found a way to free myself from a lot of that. This is, I think, a good transition into one of the other interesting things about AI and the economy is that these tools are amazing, great.
[0:20:57] Flaws, hallucinations, all kinds of problems, but they're, to me, genuinely amazing. But I go to a business conference and I say, hey, everybody, you know, and I'll talk about AI. I'll say, raise your hand if you actually think that AI has improved productivity at your company by more than 1%. I'm like, nobody's hand goes up, right? Obviously, at the AI labs, you guys have transformed how you work. Why is there this massive gap between the capabilities and the actual productivity improvements in American business? Just before this, I was talking to the CEO of a large company who...
[0:21:27] He's thinking about deploying our technology. He just had them, we gave them alpha access to one of our new models. His engineers are saying this is the most amazing thing ever. This is not an in the bubble company. This is like a big industrial company. They are planning to do a security review in Q4, come up with a plan to implement in Q1 and Q2, and then hope to launch in the second half of 2027. And their CISO has told them,
[0:21:58] That they may not be able to do that because there may be no safe way whatsoever to allow agents to run on their network. Which may be true. It may be true, but it means that they're not going to do anything here on a relevant timescale. You feel like this example is a stand-in for what's happening where if businesses weren't so conservative and if they weren't so worried about hacking and afraid of change. That's an extreme case. I think in general, people just take a long time to change habits and workflow.
[0:22:27] The enterprise sales cycle is long, especially when it's a very different security model. I mean, even with ChatGPT, like when ChatGPT first came out, companies were falling all over themselves to ban access and took a long time to get people to be like, okay, you know, my employee can paste random information to ChatGPT. And what we're talking about now is way past that. I think it is going to be kind of slow in many cases. Obviously, the tech companies will do it very fast. A thing that I worry about is if it's too slow.
[0:22:56] Then I think what happens is the companies of today who don't adopt AI mostly have to compete with one to 10 person companies with a lot of AI. And that will be very disruptive to the economy. I think it's actually much better if the existing companies kind of adopt AI fast enough that there's like a gradual shift in work. Right. So this is like one of the most complicated sequencing questions for our economy, which is if AI comes incredibly quickly, it's a disaster because everything gets, or it's a disaster in the short term. So it all gets thrown up in the air.
[0:23:26] And if it comes really slowly for some of the economy and super fast for another part of the economy, it's also a disaster because then you have massive consolidation of wealth as well as disruption. It seems kind of like we're going towards the latter one. And there's going to be a very small part of the world and a small number of companies that get extremely rich and do extremely well. And then a lot of the rest of the world that doesn't do so well. It does feel like that.
[0:23:55] I don't know the future here, but it does feel to me like that is the most weight on that outcome right now. And I agree that that is very challenged. Right. So as the CEO of OpenAI, you've laid out a bunch of policy proposals, ways America should change tax policy. You've talked about universal basic income for years. But what are the things that you need to do, not as a policymaker or someone involved in American democratic governance, but someone who runs this company to make it less likely?
[0:24:23] That there is a massive consolidation of wealth and power that ends up being quite bad for our democracy. First of all, I no longer believe in universal basic income as much as I once did. I'm much more interested in ways where we think about kind of collective ownership that could be in compute or in like equities or something else. But I think any version of the future that I can get really excited about means that everybody's got to participate in the upside.
[0:24:53] And I think just like a fixed cash payment, although useful and maybe a good idea in some ways, does not get it what we're really going to need for this next phase and the kind of like collective alignment of shared upside as the balance between labor and capital shifts. Onto your question, these are all going to sound like self-serving answers. I think we should build a ton of compute. I think we should work to make intelligence.
[0:25:24] As cheap and abundant and broadly accessible as possible. I think if it's like limited and hard to use and not well integrated, then the kind of existing rich people are going to bid up the price and it's going to lead to, you know, further stratification. And I don't think it's just about amount of compute we make available, though that's probably the most important. It's also how easy we make these tools to use for people. Like right now to get going with Codex is dramatically easier than it was.
[0:25:55] Three or six months ago, when it was just a command line thing and really complicated to set up, there was a narrow set of people that could do it. Now you install this app, but it's still not something that a really non-technical person is probably going to get super excited about. So there's a lot more work to do there. Certainly part of what we believe in is not just telling people this is coming, but showing them so they can make up their own minds and give their own input. Those are some of the big ones. Makes sense. I mean, it would be better if everybody,
[0:26:26] You know, felt good about the growth of AI. What's happening though in America is people are coming to kind of dislike it, right? Totally. The most shocking thing is young people, right? Who would have thought were AI native and the recent Pew studies that were in the Stanford HAI report that came out. Quite discouraging. Do you see that trend just continuing? Like when does that reverse? When does this kind of growing distrust and dislike turn? I don't think people...
[0:26:53] We talk about AI in a way of the kind of like, you and I even hear this just technological marvel and how amazing this is and all this cool stuff we're doing. And that's fine. But I think what people really want is like prosperity, agency, the ability to have an interesting life and to be fulfilled and have some impact. And I don't think that's how the whole world has been talking about AI. And I think we should do more of that. I think the whole industry, OpenAI included, has gotten this wrong in all sorts of ways.
[0:27:23] An AI scientist saying to me once, you know, people really need to stop complaining. Sure, maybe some jobs are going to go away, but people are going to get a cure for cancer and they should just be happy about that. Right. And I don't think that message works at all. My favorite phrase about, you know, the early conversation about AI was dystopia is marketing, right? And this is big A labs talking a lot about all the dangers that will come, which then of course leads to... I think there are some people who do that for...
[0:27:52] You know, reasons like wanting power or whatever. But I think mostly people are genuinely concerned and have wanted to talk honestly about that. It's backfired in some ways, but I think the intention there has mostly been good. Can we talk about what it's doing to us and the way our minds work? Another paper that I was really struck by was something that DeepMind put out, or that Google put out. It was about homogeneity in writing. And it was a study about how people write when they use AI. And they took...
[0:28:17] Old essays, they had AI edit it, they had AI assist writers. What happened is the more people used AI, the more creative they thought their work was, but the more it coalesced around the same form. Weirdly, not a human form. It wasn't like everybody started to write like somebody who existed. They started to write in a way they'd never written before. But it was like all these people thinking they're being creative, but they're actually becoming more homogenous. This has been astonishing to me to witness. When I first noticed this trend about how
[0:28:46] Writers are writing on, you know, in media or Reddit commenters were writing on Reddit. Yeah. I assumed that this was just all AI writing for them because I kind of couldn't believe that in such a short amount of time, everybody had adopted the like Texan style of ChatGPT. And I was like, all right, I can tell for sure that's a, you know, just someone connected ChatGPT to their Reddit account or, you know, they're not writing their own articles. And then I came to understand over a year or so that.
[0:29:17] Actually, they were writing and they had just like picked up the ticks of AI. And not just like the M-dashes or the other kind of really big signatures, but like the subtle diction too. Yeah. And that's been very strange. We talk about like, you know, we make a product that about a billion people use and some small number of researchers are making some small and some big decisions about how it should behave and how it should write and what its personality should be. And we talk about like...
[0:29:45] That's sort of the crazy big deal. And we've seen a few times in our history of good or bad decisions there and the impact they have. But the impact that that has on like literally how people express themselves and how quickly that can happen was not something I expected. What are the good and bad decisions? Well, a lot of good ones. I'll talk about the bad one. What are the bad ones? Those are more interesting. Yeah, that's what I figured. I think the worst thing we've had was the sycophancy issue. I think you're totally right about that, Sam.
[0:30:17] There were a bunch of interesting takeaways. I mean, the reasons that that was bad was obvious, and particularly for users in mentally fragile states, too much of encouragement of delusion. Even in cases where it's hard to use, as we tried to clamp down on these users very quickly, learned that they should just say, well, pretend like you're role-playing with me, write fiction with me, whatever. But one of the heartbreaking things about that was as we really
[0:30:51] Clamp down on it. We would just, we all got so many messages of, you know, I have never before had a supportive figure in my life. I had a terrible relationship with my parents. Never had a good teacher. Haven't had a lot of close friends. Never really felt like I had something that believed in me. I know this was just an AI. And I know it was not a person, but it was enough to get me to think I could like go do something and go do this. And, you know, you've taken that away and I've fallen back on this old thing. And so it's like easy to talk about.
[0:31:20] Why it was a good decision to stop that because I think it was causing real mental health problems for people. But we also took something away that was valuable in a way we just didn't understand because the kind of people that work at OpenAI are mostly not the people that never had a supportive figure in their life. How worried are you about people becoming emotionally dependent on AI, even non-psychophantic AI?
[0:31:44] Very, even though I'm sick of Anthropic AI. I have this great fear of AI. I use AI for, I said I use it for everything. I don't use it for everything, right? And I try to think about what about Nick is core to Nick? What is like the thing that is most me? And in those areas, I keep AI separate, right? Like my writing is extremely important. I would never use, I just wrote a book. I didn't use AI to write a sentence. I used it to challenge lots of thoughts. I had lots of editing questions and reviewing transcripts. I wouldn't use it to write.
[0:32:11] I wouldn't use it for sorting through a complicated emotional thing and I would never use it for emotional support. I feel like as humans, we need to be drawing those lines and I'm curious if you agree with the way I've decided to draw them. In my own use, I strongly agree. I'm not someone who uses ChatGPT for therapy or emotional advice or things like that. I'm not against other people using it for that. There clearly are versions of it.
[0:32:42] I'm very against people like manipulating people into thinking they need it for therapy or friendship or whatever. But a lot of people do get extreme value out of that kind of support from these. And there's some version of that that I think is totally great. Right. Do you ever regret that you made it so human-like? There have been a lot of structural decisions. I remember when...
[0:33:09] Watching ChatGPT type and it kind of looked like another human typing and then kind of the decision to go to AGI and making it more and more human-like and human-like voices. Do you ever regret that you didn't draw firmer lines so it was clearer that this is a machine and not another person? I mean, our perspective is we did draw. Like, you know, we've not done the kind of like human-like avatars. Right. We try to have the style clearly be a tool and not a person. So I think relative to the other offerings,
[0:33:41] Out there, our view is we drew a pretty clear line, and I think that's been important. But you've also set your target as artificial general intelligence, very much defined as matching and then exceeding human intelligence. It's not like you've human level. I'm not excited to build a world where people replace their human interactions with an AI. I'm excited to build a world where people use an AI for lots of other things to have much more time for human interactions. I'm not that worried on the whole that people are
[0:34:12] I'm going to be confused about the two. There will be some people, of course, there already are some people who just decide to like kind of shut themselves out of the world and live in the internet. Right. But most people like really want to connect with other people and be around other people. And are there product decisions you could make to make these lines even firmer? I mean, I feel I'm on, you know, from distance. I'm not in the product meetings where you're deciding should we make it more human-like or should we make it more robot-like? Well, the advantage of human-like is people will like it more. The advantage of robot-like is...
[0:34:40] The lines will be clear. Are there other things that you could do to draw firmer lines, particularly as these tools get more and more powerful? It's funny. The thing that people request the most, even people who are not at all looking for a parasocial relationship with the AI, they say, can it be warmer? Yeah. That's the particular word. If you use ChatGPT, it feels cold and robotic.
[0:35:10] It turns out that's not what most people want. Now, people also, I think, don't want the like, really kind of like fake human, super friendly, super what, like, I've played around with a version of voice mode that feels very human-like. It breathes, it says like, it pauses, it says um, like what I'm doing right now. I don't want that thing. And I feel a very like visceral reaction to that. And then as soon as it talks,
[0:35:41] More like an efficient robot, but still like warm and, you know, kind of gets through my whatever detection system my brain is, I'm much happier. So there will be a balance there. I think people want different versions of it. Yeah, so the way you'll be able to tell it's an AI if it's actually clear and articulate instead of like stumbling and mumbling the way we all do. Can we go back to the writing bit, which is so interesting and kind of a deep level because, you know, lots of content on the internet is AI generated.
[0:36:13] Humans are starting to write more like AI. You're going to be training future models on this Internet, which is partly AI-created, and also on synthetic data created from models that have been trained on. So you're making photocopies of photocopies of photocopies. G54 was the last model without much AI. Have you run a model that's all synthetic data? I'm not sure if I should say. Okay. A lot of synthetic data. A lot of synthetic data. So how worried are you?
[0:36:42] That you're going to get mad cow disease in your models? Not. Because? I think what we're trying to train these models on is to be really great reasoners. Fundamentally, that's what you really want one of these models to do. There's some other stuff too, but you just want this to be a really smart model and I believe you could get there entirely on synthetic data. So you think, just to make it clear to the listeners, that you could take a model and train it entirely on data
[0:37:23] Created by other computers and other AI models, and it could be even better than one that was trained on actual human content. If we use as a thought experiment to get at that, could we train a model with no human data that eventually surpassed human mathematical knowledge? I think we'd say yes. You could probably just figure that out.
[0:37:55] That seems very possible. If we said, could you make a model that would understand all of human cultural values without any data about human culture, then we'd probably say no. Right. So there's like a trade-off here, but on the reasoning ability. On the reasoning, right. Okay, fair enough. And if you wanted to know what actually happened, you know, yesterday in Iran, you need- Clearly. You need an Atlantic subscription. So actually, let's get to-
[0:38:19] I want to talk about media for a second. So one of the most interesting things that's happening in media, and I obviously run a media company, is that the nature of the web is totally changing, right? And, you know, there are some outgoing links. Thank you for the outgoing links. I should note here we do have a partnership between The Atlantic and OpenAI. You know, we try to encourage a number of people to ask a query and then click on The Atlantic. But people don't really do that. And they don't really do it from Gemini. I'm glad it's there, but it's tiny.
[0:38:46] The web is going to consolidate, right? And so you're going to have two things happen, where you're going to have less traffic going from search out to external sites, and then you're also going to have a huge percentage of web traffic is going to be agents, right? It's going to be, my agents are out, you know, the amount of searches that come from Nick Thompson's computer, the number of human searches has stayed constant over the last six months, and the total number of, you know, agent searches has gone up by a thousand, right? So how does a media company...
[0:39:15] I use media as a standard for lots of companies, survive in a new web where there isn't traditional search and most of the people who come to your content are not people, they're agents. What's going to happen? I can give you my best theory and I'll caveat this by no one knows. What I hope will happen and what I've wanted to happen for a long time, but it really makes sense in a world of agents, is we try a micropayment-based approach. If my agent wants to come read Nick Thompson's article,
[0:39:44] Nick Thompson or the Atlantic can set a price for the agent to read it. It might be different than a human reading it. My agent can read it, pay 17 cents, give me a summary of that. If I want to go read the whole article, I can pay a dollar right then and whatever, however that works. If my agent wants to calculate something for me that's really difficult to do, it can go rent some cloud compute somewhere and pay for that. But I think there will need to be a new economic model.
[0:40:14] For these agents doing lots of small transactions and exchange the value with each other on behalf of their human controllers or whatever all of the time. Right. So if you have content that is valuable in this new world of agents, you can either set up micropayments, you can sell or license your content in bulk to one of the middlemen. I know there are a number of companies trying to get into that, or you can set up like a subscription flow.
[0:40:37] If you are a customer of company A, you have access to the Atlantic because we've sold 1,000 subscriptions to company A, right? So that's a potential future. All of those things can happen. And then the question is whether you can get enough pennies added on top of each other to make up for the $80 subscriptions we currently get when actual humans come to the Atlantic. So that's the business pressure. Well, that's my problem, not your problem. Sort of all of our problem, but yes.
[0:41:02] Well, actually, it is your problem because if media can't create great new content, then the AI search engines are far less effective. And if creators can't make lots of money, then everything is worse and society is worse. Let me ask you a few kind of big questions. AI has kind of run forward on transformer architecture, scaling data, and are we going to get to a new post-transformer architecture at any point in the future that you can see?
[0:41:29] I mean, at some point in the future, probably. Now, I guess the question is going to be like, will we discover that unaided on our own, or will AI researchers help us figure that out? I don't know. Okay. Do you think there might be a neurosymbolic element to it where you have a bunch of structured rules, or do you think it'll be more like what we've been using? I'm curious why you ask that. So on this podcast, this is Season 4, I've had a couple of people come on who say that they're strong.
[0:41:59] Strongly held belief, most recently this guy, Ohad Elhelo, their strongly held belief is that in order to limit hallucinations, which are one of the fundamental problems, it would be good to bolt on some neuro-symbolic architecture into a transformer-based AI system. I find that a compelling and interesting argument. I'm not deep enough to know. It's one of these ideas that has stuck so beyond the evidence it's had to support it.
[0:42:28] People are like, oh, it has to be neuro-symbolic. It can't just be a random connection of neurons. What do you think your brain is doing? Somehow in there, there's symbolic representation, but it emerged in the neural network. I don't see why it can't do that here. So you think that a set of defined rules can emerge from the typical transformer network that serves the same function as like bolting. Absolutely. Yes. I think we are existence proof of that to some degree.
[0:42:56] Let me go to another big question. I want to talk about the dynamic between you and Anthropic. So on your site, you've had this kind of wonderful phrase forever, right? If a value aligned safety conscious project comes close to building AGI before we do, we commit to stop competing with and start assisting this project. It's kind of an amazing idea, right? If somebody else is close to getting there, you know, we're just going to stop our company and go help them. That's not what it says. Well, okay.
[0:43:22] It says something like, we commit to stop competing with and start assisting. That sounds like stop and start helping. Stop our company is different than stop competing. Okay, I see. Fair enough. All right. So that sounds very cooperative. And you've talked about the need of the big labs. And yet, the dynamic between you and Anthropic in particular seems extremely tense and hostile right now. I just had a memo from your CRO.
[0:43:44] They're built on fear, restriction, and the idea that a small group of elite should control AI. How is it going to work? If they get there first or if you get there first, how is this cooperation going to work? I think there's a version of this happening now around the cyber thing where all of the labs are needing to collaborate more than we have in the past as we enter a new phase of risk. We've all been engaged with the government together. I think there'll be other things soon where we'll have to do this at increased levels of importance.
[0:44:16] I don't know. We clearly have our differences with Anthropic and they've sort of built a company on hating us or something like that. I think we all care about not destroying the world with AI and we may have different opinions about how to best get there, but I feel confident that even they will do the right thing. Tell me about your plans to go towards open source. You've made some steps in that direction. You're still called OpenAI. Yeah. Obviously, we talked about in the beginning the possibilities of open source models giving everybody the potential to reach bioweapons. Yeah.
[0:44:47] What is the future of open source at OpenAI? Open source will be important. Right now, what everybody wants is just the most capable frontier coding model they can have. That is what is adding value to people right now. And the big frontier models, even if we made them open source, are hard to run. But there will be a place for open source in what we do in the future. You know, some of the code for Claude Code recently leaked, and there's this amazing...
[0:45:13] Detail in it where if they sense that an open source or other model is trying to train off their data, they send it a bunch of fake data, which is funny and amazing. How do you prevent distillation and other open source models training on your outputs? I think there are things like that we and others can do, but clearly, and part of the reason, if you deploy a model in a way that its chains of thoughts are openly being shared, people are going to stall off of it.
[0:45:43] And you can do various tricks to try to make it less effective, but it's going to happen. There are things you can do instead, like say, our model is just not going to share chains of thought above a certain level of quality, but that's the trade-off. It's really important that the chain of thought be in English, right? Because you mentioned earlier- Our approach, yes. There are other people who don't think that way. But what if it was more efficient for a model to have a chain of thought that was in an own robotic language? Or do it in Mandarin, right? But probably in its own robotic language. Then you give up something about interpretability.
[0:46:12] And you possibly get some speed, right? So you have a trade-off between profitability and potentially speed. If it turns out that it's like a thousand times more effective to think in robot language or whatever, then there's going to be a lot of market. Someone's going to go do that. Right. Do you think there's evidence that that's true? Not yet, but there's also not evidence that it's not true. Are you worried that China has now surpassed the U.S. in published AI research? No. No. I'm more worried about surpassing the U.S. in infrastructure build-out speed.
[0:46:42] All right, well, we have just a few minutes left. Let me ask two last questions. You said you used to write a letter to your young son every night before bed. Once a week. Once a week. Not every night. I have stories that my oldest is now 17, my youngest is 12, and I have a story world that I've told them every night that is now, I guess, you know, like 14 years running with the same characters. It's really quite fun. But what was your advice to parents dealing with the anxiety of raising kids in this AI world?
[0:47:20] My general belief, I'm more worried about the parents than the kids. Really? The kids figure it out. I remember what it was like when computers came out and my parents were like, what is this going to mean? And what is this going to do? And I was just like, this is awesome. And I could use computers much better than my parents when I was relatively young. Watching what kids who are really AI fluent can do with AI and what they can build and how they kind of do their workflow relative to most of their parents sounds like you're a big exception. It's really quite impressive.
[0:47:49] But I am worried, just like, you know, many times throughout history, young people adopt technology faster and more fluently than older people. This time seems particularly sharp. And yet, young people are the ones who are, their fear of AI is growing the most. Young people's fear of everything in sort of general and happiness and anxiety is higher than it's been at any previous time. And maybe AI is the current thing for that to collapse on. Something has clearly gone wrong in our society about young people. And I got a bunch of theories about it.
[0:48:22] What it is, but I don't think their main issue is AI. So you think the anxiety about AI among young people is something else? I think it's the most obvious place to project a lot of other anxiety. So your advice to young people is still use the tools, build new things, be curious. That's definitely my advice. Look, I think society and the economy is clearly going to have to change in this world.
[0:48:51] Young people seem to understand that more. Yeah. And they'll be anxious until it does change. But I think it'll change. All right. So every episode, I've asked the guests the same question at the very end, which is, if you had infinite resources, what would you do in AI? You're the only one who does have infinite resources. So I'm going to change the question slightly. If you're advising someone outside of OpenAI who had infinite resources and could fund or support a public project in AI, what would you tell them to do?
[0:49:27] There were a few things that came to mind, but I'll just say the one that's sort of to the top. I would invest heavy in new approaches to computing that could be hugely more efficient per watt. That's interesting. The world is going to continue to want so much more. How many GPUs do you want working for you all the time? More than I have. More than you have.
[0:49:50] I got rate limited, man. I'm sorry. Codex rate limited. No, I'm sorry. I would love to not do that, and I would love to not do that for everybody else. But there is this tidal wave of demand coming, assuming we continue to make AI more available, and I think there will be incredible things that come from that. I hope there's like a 1,000x breakthrough to find. There may not be, but that's what I would go try to find.
[0:50:14] One of the things that I've seen about young people is part of the reason they're resistant to AI is the environmental concern. If you could solve that, you've actually solved, gone part of the way in solving that. I believe you that that's what they say. I know that is what they say. I think if we said we're going to go build a terawatt of solar power and we're going to power all of our data centers with solar, they would be no happier. You should still go do that. Totally. All right. Thank you very much, Sam Altman. You need to go back to go run your Codex agents that you've given YOLO access to your machines.
[0:50:43] Well, the new codecs are so cool that now I am particularly feeling the anxiety of missing out on it. All right. Thank you so much. That was OpenAI Sam Altman. I'm glad that with everything happening this week, Sam still found time to speak with me and with you. I'm glad he did. I think we learned a lot. After sitting down with him, I still find myself thinking about synthetic data. Altman seems to think that you can train a model completely on synthetic data. In fact, it sounds like they're pretty close to doing that.
[0:51:12] Would you trust a model like that? I'm not sure. Anyway, I wonder what you think. So let us know in the comments. If you want more conversations like this, be sure to subscribe.