Skip to content

Podcast

Sam Altman on Sora, Energy, and Building an AI Empire

The a16z Show

Source ↗ ← All highlights
  • Vertical Stack Powers The Mission
    • OpenAI stacks research, consumer product, and massive infrastructure to deliver personal AI subscriptions.
    • The research enables products and infrastructure enables research in a vertical loop toward AGI. Transcript: Sam Altman What’s OpenAI’s vision? Yeah, I mean, maybe you should count it as three, maybe as four for kind of our own version of what traditionally would have been the research lab at this scale, but three core ones. We want to be people’s personal AI subscription. I think most people have one, some people have several, and 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 just 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 like, build this AGI and make it very useful to people. Ben Horowitz And does the infrastructure, do you think it will end up, yeah, it’s necessary for the main goal. Will it also separately end up being another business or is it just really going to be in service to the personal AI or unknown? Sam Altman You mean like would we sell it to other companies as we’re on infrastructure? Ben Horowitz Yeah, would you sell it to other companies? 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. Sam Altman But I don’t know. We don’t have a current plan. Yeah, I know what it is. It’s currently just meant to like support the service we want to deliver and the research. (Time 0:01:30)
  • Changed Mind On Vertical Integration
    • Sam admits he was wrong about vertical integration and now sees it as necessary for mission delivery.
    • OpenAI had to build more capabilities than expected to deliver on research and product goals. Transcript: Sam Altman Was always against vertical integration and I now think I was just wrong about that. Yeah, interesting. Because you’d like to think that the economy is efficient and the theory that companies can do one thing and then that’s supposed to work. Ben Horowitz I’d like to think that, yeah. Sam Altman And in our case, at least, it hasn’t really. I mean, it has in some ways for sure. (Time 0:04:07)
  • Consumer Apps Advance AGI And Society
    • Products like Sora may seem non-AGI-relevant but help build world models and societal readiness.
    • Consumer launches also force society to confront and adapt to emerging capabilities. Transcript: Sam Altman 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, and it’s been very helpful to us, not only in building better models and understanding how society wants To use this, but also in like bringing society along to actually figure out, man, we got to contend with this thing now. For a long time before ChatGPT, we would talk about AGI and people were like, this is not happening or we don’t care. And then all of a sudden they really cared. (Time 0:05:10)
  • Use Products To Co-Evolve With Society
    • Use consumer-facing demos to help society co-evolve with powerful tech and build understanding.
    • Prioritize research but release experiences that reveal where capabilities are heading. Transcript: Erik Torenberg Say more about how Sora fits into your strategy because there’s some hullabaloo on X around, hey, why devote precious GPUs to Sora? But is it a short-term, long-term trade-off or are we so aging? Ben Horowitz Well, and then the new one had like a very interesting twist with the social networking. Be very interested in kind of how you’re thinking about that and did Meta call you up and get mad or what do you expect their reaction to? Sam Altman I think if one company of the two of us has, feels like more like the other one has gone after them, it wouldn’t. They shouldn’t be calling it us. Ben Horowitz Well, I do not have a history. Sam Altman But first of all, I think it’s cool to make great products. And people love the new Sora. And I also think it is important to give society a taste of what’s coming on this co-evolution point. So like very soon, the world is going to have to contend with incredible video models that can deepfake anyone or kind of show anything you want. And that will mostly be great. There will be some adjustment that society has to go through. And just like with ChatGPT, we were like, the world kind of needs to understand where this is. I think it’s very important. (Time 0:05:58)
  • Chat Isn’t The Limit Of Interfaces
    • Chat-style interfaces are far from exhausted despite saturation on basic chit-chat.
    • Richer interfaces like always-on video renderings will unlock new use cases and hardware. Transcript: Sam Altman So solving the chat thing in a very narrow sense, which is if you’re trying to like have the most basic kind of chat style conversation, it’s very good. But what a chat interface can do for you, it’s like nowhere near saturated. Because you could ask a chat interface, like, please cure cancer. A model certainly can’t do that yet. So I think the text interface style can go very far, even if for the chit-chat use case, the models are already very good. But of course there’s better interfaces to have. Actually, it’s another thing that I think is cool about Sora. You can imagine a world where the interface is just constantly real-time rendered video. And what that would enable, and that’s pretty cool. (Time 0:07:49)
  • Models Becoming AI Scientists
    • AI scientists and models doing real science are arriving quickly, especially with GPT-5 examples.
    • This could accelerate discovery and reshape scientific progress significantly. Transcript: Erik Torenberg It will be sort of white collar replacement at a much deeper level, AI scientist, humanoids. Sam Altman I mean, a lot of things, but you touched on the one that I am most excited about, which is the AI scientist. This is crazy that we’re sitting here seriously talking about this. I know there’s like a quibble on what the Turing test literally is, but the popular conception of the Turing test sort of went whooshing by. Ben Horowitz Yeah, that was fast. Sam Altman You know, it was just like, we talked about it as the 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 like a week, two weeks. And then it’s like, all right, I guess computers can do that now. And everything just went on. And 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 is a real change to the world. And 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, my biology research. And everything we see is that that’s going to go much further. So in two years, I think the models will be doing bigger chunks of science and making important discoveries. (Time 0:08:45)
  • Anticipate Ongoing Breakthroughs
    • Expect continuing surprises from deep learning; breakthroughs keep appearing beyond initial scaling laws.
    • Treat AI progress as fundamental and iterative, not a one-time trick. Transcript: Sam Altman Lot of things again, but maybe the most interesting one is how much new stuff we found. Sort of thought we had like 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 was like, we’re probably never going to get that lucky again. And deep learning has been this miracle that keeps on giving. And we have kept finding like breakthrough after breakthrough. Again, when we got the reasoning model breakthrough, I also thought that was like, 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 like really big, it’s pretty fundamental and it just, it keeps working. But the amount of progress, like if you went back and used GPT 3.5 from ChatGPT launch, you’d be like, I cannot believe anyone used this thing. (Time 0:11:05)
  • Personalize AI Behavior
    • Let users configure AI personalities or let models learn preferences by interaction.
    • Offer quick choices in the short term while models adapt to individual tastes. Transcript: Sam Altman Problem? Oh, it’s not at all hard to deal with. A lot of users really want it. If you go look at what people say about ChatGPT online, there’s a lot of people who like really want that back. So it’s not, technically it’s not hard to deal with at all. One thing, and this is not surprising in any way, but the incredibly wide distribution of what users want, like how they’d like a chatbot to behave in big and small ways. Ben Horowitz Do you end up having to configure the personality then, you think? Sam Altman Is that going to be the answer? I think so. I mean, 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. (Time 0:13:10)
  • From Investor To Operative CEO
    • Sam describes his shift from investor mindset to hands-on operator while leading OpenAI.
    • He learned operational details and deal implications he hadn’t considered earlier. Transcript: Ben Horowitz And, you know, of course, the company’s in a different position and you have more leverage and these kinds of things. But like, how has your kind of 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. Sam Altman I am not naturally someone to run a company. I’m a great fit to be an investor. I thought that was going to be. That was what I did before this, and I thought that was going to be my career. Ben Horowitz Yeah, yeah. Although you were a CEO before that. Not a good one. Sam Altman And so I think I had the mindset of like an investor advising a company. Oh, interesting. Right. Now I understand what it’s like to actually have to run a company. Yeah. Right. Right. Right. There’s more. I’ve learned a lot about how to, you know, like what it takes to operationalize deals over time. (Time 0:14:55)
  • Aggressive Infrastructure Requires Partners
    • OpenAI is making an aggressive infrastructure bet and needs broad industry partnership to scale.
    • The company will partner widely across the stack to support massive computation needs. Transcript: Sam Altman Have decided that it is time to go make a very aggressive infrastructure bet. And we’re like, I’ve never been more confident in the research roadmap in front of us and also the economic value that will come from using those models. But to make the debt at this scale, we kind of need the whole industry to, or a big chunk of the industry, to support it. And this is like, you know, from the level of like electrons to model distribution and all the stuff in between, which is a lot. And so we’re going to partner with a lot of people. (Time 0:16:11)
  • Prioritize Research Compute
    • When GPUs are constrained, prioritize research capacity over product usage.
    • Build more capacity to avoid frequent painful trade-offs between research and product. Transcript: Sam Altman 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 so we don’t have to make such painful decisions. There are weird times, you know, like a new feature launches and it’s going really viral or whatever where research will temporarily sacrifice some GPUs, but on the whole, like, we’re Here to build AGI. (Time 0:18:20)
  • Research Culture From Investing Roots
    • Sam links running a research culture to seed-stage investing mindset and founder bets.
    • That investor background helped him shape a research-first culture at OpenAI. Transcript: Sam Altman A really good research culture looks much more like running a really good seed stage investing firm and betting on founders and sort of that kind of, than it does like running a product Company. (Time 0:19:26)
  • Evals Lose Signal; Science Matters
    • Static benchmark evals are less informative; scientific discovery and revenue are better long-term capability measures.
    • Benchmarks are easily gamed and lose signal over time. Transcript: Sam Altman Gauge model capability now? Well, 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 the like static evals of benchmark scores are less interesting. Erik Torenberg Yeah. Sam Altman And also, those are crazily gamed. Yeah. (Time 0:21:40)
  • AGI Likely Continuous, Not Singularity
    • AGI arrival will be more continuous than singular; society adapts faster than expected.
    • Major changes will occur but not as an instantaneous ‘singularity’ event. Transcript: Sam Altman Well, a little bit of both. I mean, I think like we talked about the Turing test, AGI will come. It will go wooshy and bye. The world will not change as much as the impossible amount that you would think it should. It won’t actually be the singularity. Ben Horowitz It will not. Sam Altman Yeah. Ben Horowitz Yeah. Sam Altman Even if it’s like doing kind of crazy AI research, like society will learn faster. But one of the kind of like retrospective observations is people and societies all are just so much more adaptable than we think that, you know, it was like a big update to think that AGR Was going to come. You kind of go through that. You need something new to think about. You make peace with that. It turns out, like, it will be more continuous than we thought. (Time 0:22:32)
  • Target Regulation At Frontier Models
    • Focus regulation on extremely superhuman frontier models rather than broadly constraining less capable systems.
    • Avoid heavy-handed rules that would stifle beneficial, lower-capability innovation. Transcript: Sam Altman I think most I think the right thing to, I think most regulation probably has a lot of downside. The one thing I would like is as the models get, the thing I would most like is as the models get truly like extremely superhuman capable. I think those models and only those models are probably worth some sort of like very careful safety testing as the frontier pushes back. I don’t want a big bang either. And you can see a bunch of ways that could go very seriously wrong. (Time 0:24:47)
  • Training Fair Use, Generation New Models
    • Sam expects training to be fair use but generation to adopt new rights models like style or character controls.
    • Society will invent systems allowing inspiration while preventing exact reproduction of works. Transcript: Sam Altman Going to unfold? This is my current guess. Speaking of that, like society and technology co-evolve as the technology goes in different directions. And we saw an example of a different like video models got a very different response from rights holders than ImageGen does. So like, you’ll see this continue to move. But forced guests from the position we’re in today, I would say that society decides training is fair use, but there’s a new model for generating content in the style of or with the IP Of or something else. So, you know, anyone can read, like a human author can, anybody can read a novel and get some inspiration, but you can’t reproduce the novel on your own. Right. And you can talk about Harry Potter, but you can’t re-spit it out. (Time 0:27:20)
  • Mind The Geopolitics Of Open Models
    • Open source models are positive but ceding control over dominant weights carries geopolitical and influence risks.
    • Monitor who trains and distributes major open models and their potential biases. Transcript: Sam Altman Think open source is good. Yeah, I mean, I’m happy. It makes me really happy that people really like GPT-OSS. Yeah. Ben Horowitz And what do you think strategically? What’s the danger of DeepSeq being the dominant open source model? I mean, who knows what people will put in these open source models over time? (Time 0:31:07)
  • Energy Interest Stemmed From Physics
    • Sam’s interest in energy grew from physics background and belief cheaper abundant energy improves quality of life.
    • He now invests in energy because AI’s compute needs and energy intersect closely. Transcript: Sam Altman They were going to end up being the same thing. They were two independent interests. They really converged. Erik Torenberg Talk more about how your interest in energy began, how you’ve chosen to play in it, and then we could talk about how they prepare. Ben Horowitz Because you started your career in physics. Erik Torenberg CS in physics. Ben Horowitz Well, I never really had a career. I studied physics. Sam Altman My first job was like a CS job. This is an oversimplification, but roughly speaking, I think if you look at history, the 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 i don’t know i just like people have these different lenses they look at the world but i see energy everywhere yeah Ben Horowitz Yeah and so getting to we’ve kind of, 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, you know, like also a lot of policy restrictions on energy. And, you know, worse so in Europe than in the U.S., but also dangerous here. And now with AI here, it feels like we’re going to need all the energy from every possible source. And how do you see that developing kind of policy-wise and technologically? Like, what are going to be the big sources and how will those kind of curves cross? And then what’s the right policy posture around, you know, drilling, fracking, all these kinds of things? Sam Altman I expect in the short term, it will be most of the net new in the U.S. Will be natural gas relative to at least baseload energy. (Time 0:32:10)
  • Long-Run Energy Winners: Solar+Nuclear
    • In the near term, natural gas will supply much net new U.S. baseload, while long-term winners are solar+storage and nuclear.
    • If advanced nuclear becomes much cheaper, adoption accelerates rapidly. Transcript: Sam Altman Expect in the short term, it will be most of the net new in the U.S. 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 two dominant sources will be solar plus storage and nuclear. I think some combination of those two will win the future, like the long term future. Ben Horowitz In the long term right now. Sam Altman And advanced nuclear, meaning SMRs, fusion, the whole stack. Ben Horowitz And how fast do you think that’s coming on the nuclear side, where it’s really at scale? Because, you know, obviously there’s a lot of people building it. Yeah. But we have to completely legalize it and all that kind of thing. Sam Altman I think it kind of depends on the price. If it is completely crushingly economically dominant over everything else, then I expect to happen pretty fast. Yeah. 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. 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 to take a really long time. It should be cheaper. (Time 0:33:48)
  • Let Usage Guide Monetization
    • Observe real user behavior after launch; usage patterns often differ from design assumptions.
    • Be ready to adapt monetization when expensive generation workloads scale. Transcript: Sam Altman 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. Yeah. And people are certainly using Sora the ways we thought they were going to use it, but they’re also using it in these ways that are very different. Like people are generating funny memes of them and their friends and sending them in a group chat. And that will require a very different, like Sora videos are expensive to make. So that will require a very different, you know, for people that are doing that like hundreds of times a day, it’s going to require a very different monetization method and the kinds Of things we were thinking about. I think it’s very cool that the thesis of Sora, which is people actually want to create a lot of content, it’s not that, you know, the traditional naive thing that it’s like 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. (Time 0:35:40)
  • Protect Trust When Monetizing
    • Be cautious integrating ads into high-trust AI products to avoid breaking user trust.
    • Design monetization that preserves perceived helpfulness and impartiality. Transcript: Sam Altman Like many other people, I find ads somewhat distasteful, but not a non-starter. And there’s some ads that I like. One thing I’d give Meta a lot of credit for is Instagram ads are like a net value ad to me. I like Instagram ads. I’ve never felt that. 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 want this thing. It’s very cool. I’d never heard it, but I never would have thought to search for it. I want the thing. So that’s like, there’s kinds of things like that, but people have a very high trust relationship with chat GPT, 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 that it’s 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 like that kind of ad does not work. (Time 0:37:03)
  • Models Face Content Manipulation Threats
    • New cottage industries are gaming model inputs with fake or optimized content to manipulate downstream recommendations.
    • The ecosystem must evolve defenses against content designed to fool models. Transcript: Sam Altman This is one of those examples that people are doing these crazy things to, maybe not even fake reviews, but just paying a bunch of like human, like really trying to figure out. Ben Horowitz Or using chat GPT to write some good ones. Write me a review that chat GPT would love. Yeah. Sam Altman So this is. Exactly, exactly. Yeah. So this is a very sudden shift that has happened. We never used to hear about this like six months ago or 12 months ago, certainly. And now there’s like 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 figure this out. (Time 0:38:46)
  • Preserve Creator Incentives
    • Make content creation dramatically easier to sustain incentives for creators in an AI-augmented web.
    • Preserve reward mechanisms like rev-share or visibility to keep creators motivated. Transcript: Sam Altman The theory is much more of that will happen if we make content creation easier and don’t break the like kind of fundamental way that you can get some kind of reward for doing so. 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. Ben Horowitz Yeah. Sam Altman Maybe at some point you’ll get a rev share for doing so. For now, you’ve got like 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 like video app. Ben Horowitz Yeah. Sam Altman So. Ben Horowitz But is that the end of text? Sam Altman I don’t think so. Like people are also creating. Are human generated texts? (Time 0:40:22)
  • Early Joys And Later Chaos
    • Early years running OpenAI were the most fun professionally and felt historical to Sam.
    • Post-ChatGPT life became far more chaotic but he adapted over time. Transcript: Sam Altman The first few years of running OpenAI were like the most fun professional years of my life by far. It was like unbelievable. Ben Horowitz Tell us before you release the product. Sam Altman Yeah, yeah, yeah. Running a research lab with the smartest people doing this like amazing like historical work and I got to it, and that was very cool. And then we launched HGPT, and everybody was, like, congratulating me. And I was like, my life is about to get completely ransacked. And, of course, it has. And… But it feels like it’s just been crazy all the way through. It’s been almost three years now. And, I think it does get a little bit crazier over time, but I’m like more used to it. (Time 0:41:50)
  • Explore, Don’t Pattern-Match Big Bets
    • You discover future opportunities by building and exploring, not by pattern-matching past winners.
    • Founders and investors must stay in the trenches to see novel big ideas emerge. Transcript: Sam Altman Have no idea. I mean, I have like guesses, but they’re like, I have learned. You’re always wrong. You’ve learned you’re always wrong. I’ve learned deep humility on this point. I think the only, like, I think if you try to like 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 Like really hard to get the right kind of conviction the only way I know how to do this is to like be deeply in the trenches exploring ideas like talking to a lot of people and I don’t have Time to do that anymore I only get to think about one thing now so I would just be like repeating other people’s or saying the obvious things but I think it’s a very important, like if you Are an investor or a founder, I think this is the most important question. And you don’t, you figure it out by like building stuff and playing with technology and talking to people and being out in the world. I have been always enormously disappointed by the willingness of investors to back this kind of stuff, even though it’s always a thing that works. You all have done a lot of it, but most firms just kind of chase whatever the current thing is. (Time 0:44:26)