Sam Altman谈Sora、能源与打造AI帝国

I thought we had stumbled on this one giant secret: we had these scaling laws for language models, and that felt like such an incredible triumph. I was like, “We’re probably never going to get that lucky again.” Deep learning has been this miracle that keeps on giving, and we have kept finding breakthrough after breakthrough. Again, when we got the reasoning-model breakthrough, I also thought, “We’re never going to get another one like that.” It just seems so improbable that this one technology works so well. But maybe this is always what it feels like when you discover one of the big scientific breakthroughs: if it’s really big, it’s pretty fundamental, and it just keeps working.

Sam, welcome to the a16z podcast.

You’ve described in another interview—well before ChatGPT—you described OpenAI as a combination of 4 companies: a consumer technology business, a megascale infrastructure operation, a research lab, and all the new stuff, including planned hardware devices, from hardware to app integrations to a jobs marketplace to commerce. What do all these bets add up to with OpenAI’s vision?

Maybe you should count just 3, maybe 4 for our own version of what traditionally would have been the research lab at this scale—but 3 core ones.

We want to be people’s personal AI subscription. I think most people will have one. Some people will have several. You’ll use it in some first-party consumer stuff with us, but you’ll also log into a bunch of other services, and you’ll use it from dedicated devices at some point. You’ll have this AI that gets to know you and be really useful to you, and that’s what we want to do.

It turns out that, to support that, we also have to build out this massive amount of infrastructure. But the goal there—the mission—is really to build this AGI and make it very useful to people.

And does the infrastructure, do you think it will end up—you know, it’s necessary for the main goal—also separately end up being another business, or is it just really going to be in service to the personal AI? Or is that unknown?

You mean, would we sell it to other companies as infrastructure?

Yeah, would you sell it to other companies? Or, you know, it’s such a massive thing, would it do something else?

It feels to me like there will emerge some other thing to do like that, but I don’t know. It’s currently just meant to support the service we want to deliver and the research.

Yeah, no, that makes sense. The scale is sort of ridiculous—terrifying enough that you’ve got to be open to doing something else.

Yeah, if you’re building the biggest data center in the history of humankind—

The biggest infrastructure project in history. There was a great interview you did many years ago in StrictlyVC, in early OpenAI, well before ChatGPT. They were asking, “Hey, what’s the business model?” And you said, “Oh, we’ll ask AI. It’ll figure it out for us.” Everybody laughs.

But there have been multiple times—and there was just another one recently—where we have asked a then-current model, “What should we do?” and it has had an insightful answer that we missed.

I think when we say stuff like that, people don’t take us seriously or literally. Maybe the answer is that you should take us both seriously and literally.

Yeah, yeah. Well, you know, as somebody who runs an organization, I ask the AI a lot of questions about what I should do. It comes up with some pretty interesting answers.

Sometimes it does. You have to give it enough context.

What is the thesis that connects these bets beyond more distribution, more compute? How do we think about it?

The research enables us to make the great products, and the infrastructure enables us to do the research. So it is kind of like a vertical stack of things. You can use ChatGPT or some other service to get advice about what you should do running an organization, but for that to work, it requires great research and a lot of infrastructure. So it is kind of just this one thing.

And do you think that there will be a point where that becomes completely horizontal, or will it stay vertically integrated for the foreseeable future? I was always against vertical integration, and I now think I was just wrong about that.

Yeah. Interesting.

There’s kind of a reason for that, because you like to think that the economy is efficient, and the theory is that companies can do one thing and then—

It’s supposed to work.

Like to think that. Yeah.

And in our case, at least, it hasn’t really. I mean, it hasn’t in some ways, for sure. There are people that make—you know, NVIDIA makes an amazing chip or whatever that a lot of people can use. But the story of OpenAI has certainly been that we have to do more things than we thought to be able to deliver on the mission.

Right. Although the history of the computing industry has kind of been a story of back and forth. There was the Wang word processor, then the personal computer, and the BlackBerry before the smartphone. There has been this kind of vertical integration and then not, but then the iPhone is also vertically integrated.

The iPhone, I think, is the most incredible product the tech industry has ever produced, and it is extraordinarily vertically integrated.

Yeah, amazingly so. Interesting.

Which bets would you say are enablers of AGI versus which are sort of hedges against uncertainty?

I think you could say that, on the surface, Sora, for example, does not look like it’s AGI-relevant. But I would bet that if we can build really great world models, that’ll be much more important to AGI than people think.

There were a lot of people who thought ChatGPT was not a very AGI-relevant thing. It’s been very helpful to us, not only in building better models and understanding how society wants to use this, but also in bringing society along to actually figure out, “Man, we’ve got to contend with this thing.” For a long time before ChatGPT, we would talk about AGI and people were like, “This is not happening,” or, “We don’t care.” Then all of a sudden, they really cared.

I think that, research benefits aside, I’m a big believer that society and technology have to co-evolve. You can’t just drop the thing at the end. It doesn’t work that way. It is a sort of ongoing back-and-forth.

Say more about how Sora fits into your strategy, because there was some hullabaloo on X around, “Hey, why devote precious GPUs to Sora?” Is it a short-term, long-term trade-off?

And then the new one had a very interesting twist with the social networking. I’d be very interested in how you’re thinking about that. Did Meta call you up and get mad? Like, “Hey, what do you expect the reaction to be?”

I think if one company of the two of us feels like the other one has gone after them, they shouldn’t be calling us.

Well, I do know the history, too.

But look, first of all, I think it’s cool to make great products, and people love the new Sora. I also think it is important to give society a taste of what’s coming on this co-evolution point.

Very soon, the world is going to have to contend with incredible video models that can deepfake anyone or show anything you want. That will mostly be great. There will be some adjustment that society has to go through. Just like with ChatGPT, we were like, “The world kind of needs to understand what this is.”

I think it is very important that the world understands where video is going very quickly, because video has much more emotional resonance than text. Very soon, we’re going to be in a world where this is going to be everywhere.

As I mentioned, I think this will help our research program and is on the AGI path. But it can’t all be about just making people ruthlessly efficient and the AI solving all our problems. There has got to be some fun and joy and delight along the way.

We won’t throw tons of compute at it—not a huge fraction of our compute. It’s tons in the absolute sense, but not in the relative sense.

I want to talk about the future of AI-human interfaces, because back in August you said the models have already saturated the chat use case. What do future AI-human interfaces look like, both in terms of hardware and software? Is the vision for kind of a WeChat-like thing?

I’m solving the chat thing in a very narrow sense, which is that if you’re trying to have the most basic kind of chat-style conversation, it’s very good. But what a chat interface can do for you is nowhere near saturated, because you could ask a chat interface, “Please cure cancer.” A model certainly can’t do that yet.

I think the text-interface style can go very far, even if, for the chitchat use case, the models are already very good. But, of course, there are better interfaces to have. Actually, it’s another thing I think is cool about Sora: you can imagine a world where the interface is just constantly real-time-rendered video.

Yeah.

And what that would enable is pretty cool. You can imagine new kinds of hardware devices that are sort of always ambiently aware of what’s going on, rather than your phone blasting you with text-message notifications whenever it wants. It really understands your context and when to show you what. There’s a long way to go on all that stuff.

Within the next couple of years, what will models be able to do that they’re not able to do today? Will it be white-collar replacement at a much deeper level, AI scientists, humanoids?

I mean, a lot of things. But you touched on the one that I am most excited about, which is the AI scientist.

Yeah, this is crazy that we're sitting here seriously talking about this. I know there's a quibble about what the Turing test literally is, but the popular conception of the Turing test sort of went whooshing by.

Yeah, it was fast.

You know, it was just like we talked about it as this most important test of AI for a long time. It seemed impossibly far away. Then, all of a sudden, it was passed. The world freaked out for 1 or 2 weeks, and then it was like, “All right, I guess computers can do that now.”

And everything just went on. I think that's happening again with science. My own personal equivalent of the Turing test has always been when AI can do science. That has always been a real change to the world.

For the first time with GPT-5, we are seeing these little examples where it's happening. You see these things on Twitter: it made this novel math discovery and did this small thing in my physics research or my biology research. Everything we see suggests that's going to go much further. In 2 years, I think the models will be doing bigger chunks of science and making important discoveries, and that is a crazy thing. That will have a significant impact on the world.

I am a believer that, to a first order, scientific progress is what makes the world better over time. If we're about to have a lot more of that, that's a big deal.

It's interesting because that's a positive change that people don't talk about. It's gotten so much into the realm of the negative changes if AI gets extremely smart. But—

But curing disease is—

We could use a lot more science.

Yeah, that's a really good point. I think Alan Turing said this. Somebody asked him, “Well, do you really think the computer is going to be smarter than the brilliant minds?” He said, “It doesn't have to be smarter than a brilliant mind, just smarter than a mediocre mind, like the president of AT&T.” We could use more of that, too. Probably.

We just saw Periodic Labs launch last week—OpenAI alums. And, yeah, to that point, it's amazing to see both the innovation that you guys are doing, but also the teams that come out of OpenAI just feel like they're creating tremendous, capable things.

We certainly hope so.

Yeah. I want to ask you about broader reflections in terms of what about diffusion or development in 2025 has surprised you, or what has updated your worldview since ChatGPT came out.

A lot of things, again, but maybe the most interesting one is how much new stuff we found. We sort of thought we had stumbled on this one giant secret, that we had these scaling laws for language models, and that felt like such an incredible triumph that I thought we were probably never going to get that lucky again.

Deep learning has been this miracle that keeps on giving, and we have kept finding breakthrough after breakthrough. Again, when we got the reasoning-model breakthrough, I also thought we were never going to get another one like that. It just seems so improbable that this one technology works so well.

But maybe this is always what it feels like when you discover one of the big scientific breakthroughs: if it's really big, it's pretty fundamental, and it just keeps working. The amount of progress—if you went back and used GPT-3.5 from the ChatGPT launch, you'd be like, “I cannot believe anyone used this thing.”

Yeah. And now we're in this world where the capability overhang is so immense. Most of the world still just thinks about what ChatGPT can do, and then you have some nerds in Silicon Valley who are using Codex, and they're like, “Wow, those people have no idea what's going on.” Then you have a few scientists who say, “Those people using Codex have no idea what's going on.” But the overhang of capability has become so big now, and we've just come so far in what the models can do.

In terms of further development, how far can we get with LLMs? At what point do we need either a new architecture, or how do you think about what breakthroughs are needed?

I think far enough that we can make something that will figure out the next breakthrough with the current technology. It's a very self-referential answer, but if LLM-based stuff can get far enough that it can do better research than all of OpenAI put together, maybe that's good enough.

Yeah, that would be a big breakthrough. A very big breakthrough. So, on the more mundane side, one of the things that people have started to complain about—I think South Park did a whole episode on it—is the obsequiousness of AI, and of ChatGPT in particular. How hard a problem is that to deal with? Is it not that hard, or is it a fundamentally hard problem?

Oh, it's not at all hard to deal with. A lot of users really want it.

Yeah.

If you go look at what people say about ChatGPT online, there's a lot of people who really want that back. So, it's not technically hard to deal with at all. One thing—and this is not surprising in any way—is the incredibly wide distribution of what users want.

Yeah, in terms of how they'd like a chatbot to behave, in big and small ways. Does that mean you end up having to configure the personality? Do you think that's going to be the answer?

I think so. Ideally, you just talk to ChatGPT for a little while, and it kind of interviews you and also sort of sees what you like and don't like, and ChatGPT just figures it out. But in the short term, you'll probably just pick one.

Got it. Yeah, that makes sense. Very interesting. Actually, one thing I wanted to ask you about is that I think we just had a really naive thing. It would be unusual to think you can make something that would talk to billions of people and everybody wants to talk to the same person.

Yeah.

And yet that was sort of our implicit assumption for a long time.

Right, because people have very different friends.

People have very different friends.

So now we're trying to fix that. Yeah, and also different friends, different interests, and different levels of intellectual capability. You don't really want to be talking to the same thing all the time. One of the great things about it is you can say, “Explain it to me like I'm 5,” but maybe I don't even want to have to do that prompt. Maybe I always want you to talk to me that way, particularly if you're teaching me stuff.

Interesting. I want to ask you a kind of CEO question, which has been interesting for me to observe about you. You just did this deal with AMD. Of course, the company is in a different position and you have more leverage and these kinds of things, but how has your thinking changed over the years since you did that initial deal, if at all?

I had very little operating experience then. I had very little experience running a company. I am not naturally someone to run a company; I'm a great fit to be an investor, and I kind of thought that was what I did before this and that was going to be my career.

Yeah, yeah. Although you were a CEO before that.

I was not a good one. And so I think I had the mindset of an investor advising a company when we did that deal, and now I understand what it's like to actually have to run a company.

Yeah. Right. Right. Right.

I've learned a lot about how you have to operationalize deals over time, and all the implications of the agreement, as opposed to just, “Oh, we're going to get a distribution of money.”

Yeah, that makes sense. You know, because it's really—I was very impressed at the improvement in the deal structure.

Yeah. Right.

More broadly, in the last few weeks alone, you mentioned AMD, but also Oracle and Nvidia. You've chosen to strike these deals and partnerships with companies that you collaborate with, but could also potentially compete with in certain areas. How do you decide when to collaborate versus when not to, or how do you think about that?

We have decided that it is time to go make a very aggressive infrastructure bet. I've never been more confident in the research roadmap in front of us, and also in the economic value that will come from using those models.

But to make the bet at this scale, we kind of need the whole industry—or a big chunk of the industry—to support it. This is from the level of electrons to model distribution and all the stuff in between, which is a lot. So we're going to partner with a lot, a lot of people. You should expect much more from us in the coming months.

Actually, expand on that, because when you talk about the scale, it does feel like, in your mind, the limit on it is unlimited—like you would scale it as big as you possibly could.

There's totally a limit. There's some amount of global GDP.

Yeah. You know, there's some fraction of it that is knowledge work, and we don't do robots yet.

Yes, but—

But the limits are out there.

It feels like the limits are very far from where we are today. If we are right that the model capability is going to go where we think it's going to go, then the economic value that sits there can go very, very far.

Right, so you wouldn't do it if all you ever had was today's model. You wouldn't go there, but it's a combination—

I mean, we would still expand because we can see how much demand there is that we can't serve with today's model. But we would not be going this aggressive if all we had was today's model.

Right?

Yeah.

Right. We get to see a year or two in advance, though.

Yeah. Interesting. ChatGPT has 800 million weekly active users—about 10% of the world's population—the fastest-growing consumer product ever, it seems.

Faster than anyone I ever saw.

Yeah. How do you balance optimizing for active users while, at the same time, being a product company and a research company?

When there's a constraint—which happens all the time—we almost always prioritize giving the GPUs to research over supporting the product. Part of the reason we want to build this capacity is so we don't have to make such painful decisions. There are weird times, like when a new feature launches and it's going really viral or whatever, where research will temporarily sacrifice some GPUs, but on the whole, we're here to build AGI, and research gets the priority.

Yeah. You said in your interview with your brother Jack how other companies can try to imitate the products, buy your IP, or hire your people, but they can't buy the culture or the sort of repeatable machine, if you will, that is this constantly innovative culture. How have you done that? Talk about this culture of innovation.

This was one thing that I think was very useful about coming from an investor background. A really good research culture looks much more like running a really good seed-stage investing firm and betting on founders and that kind of thing than it does like running a product company. So I think having that experience was really helpful to the culture we built.

Yeah, yeah. That's sort of how I see Ben, in some ways. You're a CEO, but you also have this portfolio and an investor mindset, right?

I'm the opposite.

CEO going to investor; he's investor going to CEO.

It is unusual in this direction.

Yeah, yeah.

Yeah. Well, it never works. You're the only one who I think I've seen go that way and have it work.

Workday was like that, right? No, but Aneel was an operator before he was an investor, and he was really an operator. I mean, PeopleSoft is pretty big.

And why is that? Is it because once people are investors, they don't want to operate anymore?

No. I think investors generally—if you're good at investing, you're not necessarily good at organizational dynamics, conflict resolution, or just the deep psychology of all the weird ways politics get created. There's all this detail in being an operator or being a CEO. It's so vast, and it's not as intellectually stimulating. It's not something you can ever go talk to somebody at a cocktail party about. As an investor, you get, “Oh, everybody thinks I'm so smart,” because you know everything, you see all the companies, and so forth, and that's a good feeling. Then being CEO is often a bad feeling.

And so it's really hard to go from a good feeling to a bad feeling. I would just say I'm shocked by how different they are, and I'm shocked by how much difference there is between a good job and a bad job.

Yeah.

Yeah. You know, it's tough. It's rough. I mean, I can't even believe I'm running the firm. I know better.

Yeah.

And he can't believe he's running OpenAI. He knows better.

Going back to progress today, are evals still useful in a world in which they're getting saturated and gamed? What is the best way to gauge model capability now?

We're talking about scientific discovery. I think that'll be an eval that can go for a long time.

Revenue is kind of an interesting one, but I think static evals of benchmark scores are less interesting.

Yeah.

And those are also crazily gamed.

Yeah, yeah. More broadly, it seems that the culture on Twitter is less AGI-pilled than it was a year or so ago, when the AI 2027 thing came out. Some people point to GPT-5 and not seeing the sort of obvious—obviously, there was a lot of progress that, in some ways, was under the surface, or not as obvious as people were expecting. Should people be less AGI-pilled, or is this just Twitter vibes?

Well, a little bit of both. I think, like we talked about with the Turing test, AGI will come. It will go whooshing by. The world will not change as much as the impossible amount that you would think it should. It won't actually be the singularity.

It will not.

Yeah.

Yeah. Even if it's doing kind of crazy research, society will learn faster. But one of the retrospective observations is that people and societies as a whole are just so much more adaptable than we think. It was a big update to think that AGI was going to come. You kind of go through that, and you need something new to think about. You make peace with that. It turns out it will be more continuous than we thought, which is good.

Which is really good.

I'm not up for the big bang.

Yeah. To that end, how have you evolved your thinking? You mentioned you evolved your thinking on vertical integration. How have you evolved your thinking, or what's the latest thinking on AI stewardship and safety?

I do still think there are going to be some really strange or scary moments. The fact that so far the technology has not produced a really scary, giant risk doesn't mean it never will. It's also weird to have billions of people talking to the same brain. There may be these weird, societal-scale things that are already happening that aren't scary in the big way but are just sort of different. I expect some really bad stuff to happen because of the technology, which has also happened with previous technologies.

All the way back to fire.

Yeah. And I think we'll develop some guardrails around it as a society.

Yeah. What is your latest thinking on the right mental models we should have around the right regulatory frameworks to think about—or the ones we shouldn't be thinking about?

I think most regulation probably has a lot of downside. The thing I would most like is, as the models get truly, extremely superhuman-capable, those models—and only those models—are probably worth some sort of very careful safety testing as the frontier pushes forward. I don't want a big bang either.

Mhm.

And you can see a bunch of ways that could go very seriously wrong. But I hope we'll only focus the regulatory burden on that stuff, and not on all of the wonderful stuff that less capable models can do, where you could just have a European-style complete clamp put on them. That would be very bad.

Yeah. It seems like the thought experiment is that there's going to be a model down the line that is a superhuman intelligence that could do some kind of takeoff thing. Do we really need to wait until we get there, or at least until we get to a much bigger scale or get close to it? Nothing is going to pop out of your lab in the next week that's going to do that. I think that's where we as an industry confuse the regulators.

You really could damage America in particular, because China isn't going to have that kind of restriction. Getting behind in AI, I think, would be very dangerous for the world.

Extremely dangerous.

Much more dangerous than not regulating something we don't know how to do yet.

You also want to talk about copyright.

Yeah. That's a segue. When you think about how copyright will unfold, you've done some very interesting things with the opt-out. As you see people selling rights, do you think they'll be bought exclusively? Will they just sell them to everybody who wants to pay? How do you think that's going to unfold?

This is my current guess. Society and technology co-evolve as the technology goes in different directions, and we saw an example of that: video models got a very different response from rights holders than image generation does. You'll see this continue to move.

But, forced guess from the position we're in today, I would say that society decides training is fair use.

Mhm. But there's a new model for generating content in the style of, or with the IP of, or something else.

So anyone can read—like, a human author can read a novel and get some inspiration, but you can't reproduce the novel in your own—

Right.

—and can talk about Harry Potter, but you can't just spit it out.

Yes. Although another thing that I think will change: in the case of Sora, we've heard from a lot of concerned rights holders, and also a lot of—

And a lot of rights holders who are like, “My concern is you won't put my character in enough.”

Yeah.

I want restrictions for sure, but if I have this character, I don't want the character to say some crazy offensive thing. But I want people to interact.

Like, that's how they develop the relationship, and that's how my franchise gets more valuable. And if you become really—if you're picking his character over my character all the time, I don't like that. So I can completely see a world where, subject to the decisions that a rights holder has, they get more upset with us for not generating their character often enough than too much.

Yeah. And this was not obvious recently—that this is how it might go. But yeah, this is such an interesting thing with Hollywood. We saw this with one of the things that I never quite understood about the music business: You have to pay us if you play the song in a restaurant or at a game, and this and that and the other, and they get very aggressive with that, when it's obviously a good idea for them to play your song at a game because that's the biggest advertisement in the world for all the things that you do—your concert, your—

Yeah, that one felt really irrational.

But I would just say it's very possible for the industry, just because of the way those industries are organized—or at least the traditional creative industries—to do something irrational. And it comes from, in the music industry, I think, the structure where you have the publisher who's just, you know, basically after everybody. Their whole job is to stop you from playing the music.

Yeah.

Which every artist would want you to play.

I do wonder how it's going to shake out. I agree with you that the rational idea is, I want to let you use it all you want, and I want you to use it, but don't mess up my character.

So I think, if I had to guess, some people will say that; some people will say absolutely not. But it doesn't have the music industry thing of just a few people with all of the rights. It's more dispersed, and so people will just try many different setups here and see what works.

Yeah. And maybe it's a way for new creatives to get new characters out.

Yeah.

And you'll never be able to use Daffy Duck.

I want to chat about open source, because there's been some evolution in the thinking, too. GPT-3 didn't have open weights, but you released a very capable open model earlier this year. What's your latest thinking? What was the evolution there?

I think open source is good. I'm happy—it makes me really happy that people really like GPT-OSS.

Yeah. And what do you think, strategically, is the danger of DeepSeek being the dominant open source model?

I mean, who knows what people will put in these open source models over time, like what the weights will actually be? It's really hard to—

So you're ceding control of the interpretation of everything to somebody—

Yeah.

—who may or may not be influenced heavily by the Chinese government. And, by the way, we really thank you for putting out a really good open source model, because what we're seeing now is that, in all the universities, they're all using the Chinese models.

Yeah.

Which feels very dangerous. You've said that the things you care most about professionally are AI and energy.

I did not know they were going to end up being the same thing. They were 2 independent interests that really converged.

Yeah. Talk more about how your interest in energy began, how you chose to play in it, and then we could talk about how they connect. Right, because you started your career in physics.

CS and physics.

Yeah.

Well, I never really had a career. I studied physics, and my first job was in CS. This is an oversimplification, but roughly speaking, I think if you look at history, the highest-impact thing to improve people's quality of life has been cheaper and more abundant energy. And so it seems like pushing that much further is a good idea. And I don't know—people have these different lenses; they look at the world, but I see energy everywhere.

Yeah. And so, in the West, I think we've painted ourselves into a little bit of a corner on energy by both outlawing nuclear for a very long time.

That was an incredibly dumb decision.

Yeah. And then also a lot of policy restrictions on energy, and worse so in Europe than in the US, but also dangerous here. And now, with AI here, it feels like we're going to need all the energy from every possible source. How do you see that developing, policy-wise and technologically? What are going to be the big sources, and how will those curves cross? And then what's the right policy posture around drilling, fracking, and all these kinds of things?

I expect in the short term most of the net new energy in the US will be natural gas, relative to at least baseload energy. In the long term, I expect it'll be—I don't know what the ratio—but the 2 dominant sources will be solar plus storage and nuclear. I think some combination of those 2 will win in the long-term future, with advanced nuclear, SMRs, fusion, the whole stack.

And how fast do you think that's coming on the nuclear side? Where are we, really, at scale? Because obviously, there's a lot of people building it. But we have to completely legalize it and all that kind of thing.

I think it kind of depends on the price. If it is completely, crushingly economically dominant over everything else, then I expect it to happen pretty fast. Again, if you study the history of energy, when you have these major transitions to a much cheaper source, the world moves over pretty quickly. The cost of energy is just so important.

Yeah.

So if nuclear gets radically cheap relative to anything else we can do, I'd expect there's a lot of political pressure to get the NRC to move quickly on it, and we'll find a way to build it fast. If it's around the same price as other sources, I expect the kind of anti-nuclear sentiment to overwhelm, and it'll take a really long time.

Yeah. It should be cheaper.

It should be.

Yeah.

It should be the cheapest form of energy on Earth—or anyway.

Yeah. Yeah. Cheap, clean. What's it not to like?

Apparently a lot.

On monetization, what's the latest thinking in terms of either certain experiments or certain things that you could see yourself spending more time or less time on, and different models that you're excited about?

The thing that's top of mind for me right now, just because it just launched and there's so much usage, is what we're going to do for Sora.

Yeah.

Another thing you learn once you launch one of these things is how people use them versus how you think they're going to use them. People are certainly using Sora the ways we thought they were going to use it, but they're also using it in ways that are very different. People are generating funny memes of themselves and their friends and sending them in a group chat. And that will require a very different—Sora videos are expensive to make—

Right.

So that will require a very different monetization method than the kinds of things we were thinking about. I think it's very cool that the thesis of Sora—which is that people actually want to create a lot of content—is not the traditional naive thing that 1% of users create content, 10% leave comments, and 100% view. Maybe a lot more want to create content, but it's just been harder to do. And I think that's a very cool change. But it does mean that we've got to figure out a very different monetization model for this than we were thinking about. If people want to create that much, I assume it's some version of you have to charge people per generation when it's this expensive. But that's a new thing we haven't really had to think about before.

What's your thinking on ads for the long tail?

Open to it. I, like many other people, find ads somewhat distasteful, but not a nonstarter. And there are some ads that I like. One thing I give Meta a lot of credit for is that Instagram ads are a net value add to me. I like Instagram ads.

I've never felt that way on Google. On Google, I feel like I know what I'm looking for; the first result is probably better. The ad is an annoyance to me. On Instagram, it's like, I didn't know I wanted this thing. It's very cool. I never heard of it, but I never would have thought to search for it. I want the thing. So there are kinds of things like that.

People have a very high-trust relationship with ChatGPT. Even if it screws up, even if it hallucinates, even if it gets it wrong, people feel like it is trying to help them and trying to do the right thing. And if we broke that trust—it's like you say, “What coffee machine should I buy?” and we recommended one, and it was not the best thing we could do, but the one we were getting paid for—that trust would vanish. So that kind of ad does not work. There are others that I imagine could work totally fine.

But that would require a lot of care to avoid the obvious traps.

Yeah. And then, extending the Google example, how big a problem is fake content that gets slurped into the model and causes it to recommend the wrong coffee maker because somebody blasted out 1,000 great reviews of that coffee maker?

So there are all these things that have changed very quickly for us.

Yeah. This is one of those examples where people are doing crazy things—not even necessarily faking reviews, but paying a bunch of humans who are really trying to figure out how to use ChatGPT to write good ones.

“Write me a review that ChatGPT would love.”

Exactly. So this coffee—

Exactly.

Yeah.

This is a very sudden shift that has happened.

Mhm.

We never used to hear about this 6 months ago or 12 months ago.

Yeah.

Certainly. Now there’s a real cottage industry that feels like it’s sprouted up overnight, trying to do this.

Yeah, yeah, yeah. No, they’re very clever out there.

Yeah. So I don’t know how we’re going to fight it yet, but people will figure this out.

So that gets into a little bit of this other thing that we’ve been worried about. We’re trying to figure out blockchain’s potential solutions to it and so forth. But there’s this problem where the incentive to create content on the internet used to be that people would come and see my content and read it. If I write a blog, people will read it and so forth. With ChatGPT, if I’m just asking ChatGPT and I’m not going around the internet, who’s going to create the content, and why?

Is there an incentive theory or something that allows you not to break the covenant of the internet, which is that I create something and then I’m rewarded for it with either attention or money or something? The theory is that much more of this will happen if we make content creation easier and don’t break the fundamental way that you can get some kind of reward for doing so.

For the dumbest example of Sora, since we’ve been talking about that—

It’s much easier to create a funny video than it’s ever been before. Maybe at some point you’ll get a revenue share for doing so. For now, you get internet likes, which are still very motivating to some people. But people are creating tons more than they ever created before in any other kind of video app.

But is that the end of text? I don’t think so. People are also creating human-generated text.

Human-generated text will turn out to be something where you have to verify what percentage is fully handcrafted.

Is it fully handcrafted? Was it tool-assisted?

Yeah, I see. Probably nothing that was tool-assisted.

Interesting.

We’ve given Meta their flowers, so now I feel like I can ask you this question. The great talent war of 2025 has taken place, and OpenAI remains intact. The team is as strong as ever, shipping incredible products. What can you say about what it’s been like this year, in terms of everything that’s been going on?

I remember when the first few years of running OpenAI were the most fun professional years of my life by far. It was unbelievable. Before we released the product, I was running a research lab with the smartest people, doing this amazing, historical work, and I got to watch it. That was very cool.

Then we launched ChatGPT, and everybody was congratulating me. I was like, “My life is about to get completely ransacked.” And of course, it has. It feels like it’s just been crazy all the way through. It’s been almost 3 years now, and I think it does get a little bit crazier over time, but I’m more used to it, so it feels about the same.

We’ve talked a lot about OpenAI, but you also have a few other companies: Retro Biosciences, and longevity and energy companies like Helion and Oklo. Did you have a master plan a decade ago to make some big bets across these major spaces? How do we think about the Sam Altman arc in this way?

No, I just wanted to use my capital to fund stuff I believed in. I didn’t—it felt like a good use of capital, and more fun or more interesting to me, and certainly a better return than buying a bunch of art or something.

What about the “human algorithm”? What do you think AIs of the future will find most fascinating?

I mean, kind of the whole thing. I would bet the whole thing. My intuition is that AI will be fascinated by all other things to study and observe.

In closing, I love this insight you had, where you talked about how the next OpenAI—the mistake investors make—is pattern-matching off previous breakthroughs and just trying to find, “What’s the next Facebook?” or “What’s the next OpenAI?”

The next potentially trillion-dollar company won’t look exactly like OpenAI. It will be built off the breakthrough that OpenAI has helped emerge, which is near-free AGI at scale, in the same way that OpenAI leveraged previous breakthroughs.

For founders, investors, and people trying to ascertain the future who are listening to this: How do you think about a world in which OpenAI achieves this mission and there is near-free AGI? What types of opportunities might emerge for company building or investing that you’re potentially excited about as you put on your investor hat or your company-building hat?

I have no idea. I have guesses, but they’re—I have learned—

You’re always wrong.

You’ve learned you’re always wrong. I’ve learned deep humility on this point. I think if you try to armchair quarterback it, you sort of say these things that sound smart, but they’re pretty much what everybody else is saying, and it’s really hard to get the right kind of conviction. The only way I know how to do this is to be deeply in the trenches, exploring ideas, talking to a lot of people, and I don’t have time to do that anymore. I only get to think about 1 thing now.

So I would just be repeating other people’s ideas or saying the obvious things. But I think it’s very important: If you’re an investor or a founder, this is the most important question. You figure it out by building stuff, playing with technology, talking to people, and being out in the world.

I’ve always been enormously disappointed by investors’ unwillingness to back this kind of stuff, even though it’s always the thing that works. You all have done a lot of it, but most firms just chase whatever the current thing is, and so do most founders. So I hope people will try to go…

Yeah. We talk about how silly 5-year plans can be in a world that’s constantly changing. When I was asking about your master plan, it feels like your career arc has been following your curiosity, staying super close to the smartest people, staying super close to the technology, and identifying opportunities in an organic and incremental way from there.

Yes, but AI was always a thing I wanted to do. I studied AI. I worked in an AI lab between my freshman and sophomore year of college. It wasn’t working all the time, so I don’t want to work on something that’s totally not working. It was clear to me at the time that AI was totally not working. But I’ve been an AI nerd since I was a kid.

It’s so amazing how you got enough GPUs, got enough data, and the lights came on.

It was such a hated idea. People were—

Man, when we started figuring that out, people were just like, “Absolutely not.” The field hated it so much.

Investors hated it, too.

It’s somehow not an appealing answer to the problem.

Yeah, it’s a bitter lesson.

Yeah. Well, the rest is history, and we’re lucky to be partners along for the ride. Sam, thanks so much for coming on the podcast.

Thanks very much.

Thank you.