Full Interview: Sam Altman on Sora and the Future of OpenAI

0:00And we are joined by Sam Altman and Bill Peebles. Sam, Bill, how are you doing? What's going on? Hey, guys. Hey, guys. Great to see you. Congrats on all the progress. I've been enjoying Sora a ton. Personally, I've been enjoying making them. I had a ton of fun making the collab posts yesterday. And I was wondering... And prompting your cameo feature. John made it so that he always appears as a bodybuilder if anybody is cameoing him. So you guys got to experience that. And it led to some chaotic results. Do you have favorite Sora posts that you've been coming back to or that have stuck out to you as particularly creative uses?

0:43I mean, definitely all of the ones of me stealing GPUs or doing other crazy things to get GPUs have been funny. In the last few days, at least in my feed, there have been these very beautiful sort of fantastic scenes that are just not things that could have ever existed without something like Sora. or wouldn't have been easy to make. And watching people build those and watching the trends flow through that has been pretty awesome. What about you, Bill? Any favorite uses of Sora so far? Oh, man. Mark Cuban came on the platform a few days ago, and there have been some hilarious Shark Tank memes.

1:18Those are probably my favorite. Pitching some Sora features to Mark. Also, leveraging the prompting function to always include an ad for Cost Plus Drugs I thought was especially hilarious, considering he's been one of the most vocal opponents of advertising in AI. He is leveraging the feature to the max. Yeah, I think there are going to be all of these weird new dynamics that we see emerge that just weren't possible in previous kinds of video. And this is like a fun period because it's all going to be so different every few days. Man, I'm watching this Polymarket ticker go by and it's so tempting to say things to all these markets.

1:56Yeah, yeah, yeah. To be clear, don't worry about the ticker. I don't think we're including, I don't think any of those markets are being featured in the ticker. But yeah, again, this is the new world we're in. Yeah, you can move the market live on TBPN. But today, people I'm sure will be happy or disappointed. We're here to talk about Sora, of course. Yes, yes. So none of the other topics. But I mean, I also want to know about ads. Why no ads in Sora on day one? I feel like you've laid out a really great mental model for how you think about ads on Stratechery, on the Andreessen Horowitz podcast.

2:35I'm bought in. Is it a technical thing? Do you need scale? Do you need to think about it more? Why no ads on day one? This is like a 10-day old product, right? It's hard to get anything to work at all. We don't assume success. We got to go hard-earned success, and then we can think about monetization for it. But this is like, it's gone great so far. It's still very early, and there's still a lot of work to build something that a lot of people are going to love first. What about surprising capabilities of the model? You mentioned that you've seen some fantastical scenes. I'm interested to know about specific breakthroughs that you've noticed that Sora 2, the model, is particularly good at.

3:19I noticed one about reflections being great. Obviously, people love the cameos. But what has surprised you in terms of just like, technically, the model can do something now that it couldn't do before? This model is a huge leap forward in terms of physics IQ. So pretty much all past video generation models really struggled with prompts that involve like backflips, gymnastics routines, et cetera. And this is really the only model that exists today, which can reliably handle these kinds of really complicated dynamics. One of the big features that people have really loved on the app so far is the steerability of the model.

3:51So, you know, if you give it like a really simple text prompt, that's maybe even only a few words, this model is really good about kind of telling a coherent story with like a beginning, middle and end and doing this like automatically in a way that doesn't require like a lot of direct steering from the user. If you want to like go into, you know, a ton of detail about exactly how your prompt should be laid out and how the story should unfold, it supports that too. So it can kind of meet you wherever you're at in the creative process. But really, this model is just like so hyper steerable. And it's like just vastly higher physics IQ just makes it able to do things that were like not possible a few months ago.

4:23Is that all within the model? Or is there some sort of like reasoning step where you're hydrating or unpacking my prompt and writing a bigger prompt or breaking down the problem in some way? Can you share anything about that? Yeah, it's a good question. So, you know, the intelligence for these text conditional video models kind of lies both in the core model itself, like Sora and some amount of it also comes in through the text prompt. So, you know, however the user decides to kickstart a prompt, you can have like a language model under the hood, add some details in. But for example, you know, when it comes to things like, again, like doing these backflips or any kind of physical interactions, how refraction is modeled, you know, when you're pouring water into a glass, all of these details have to be captured by the core video model itself.

5:06So that's intelligence, which is really innate to Sora. And certainly you can supplement it with intelligence from a language model as well. But it's not necessarily a prerequisite to get kind of amazing results out of these things. Are there any areas on the physics where you think that the model falls down and you want to improve? I mean, we went through the era of like six fingers. It seems like reflections in water are solved. But someone was saying something about doors being hard. I haven't noticed that one personally, but a lot of the stuff's great. But what have you noticed is like the next version is going to be even better at?

5:34This is still very early. A thing Bill said that I appreciated is this is like the GPT 3.5 moment for video. And if you went back to use the actual GPT 3.5, you'd be like, okay, signs of great promise can do the occasional impressive thing. But it was really not until GPT-4 where these text models started providing real value for people. And we know how to go make the GPT-4 equivalent of video models, and we will do that. And then a lot of these things that are currently annoying, like doors, or once in a while something goes through something else it's not supposed to. In the same way that the world loved to complain for a brief period of time about where 3.5 fell down and, oh, it's never going to be useful, it's never going to do this, it's never going to do that, and then we were able to just keep making it better and better and better and better.

6:22The model physics IQ is certainly the best I've ever seen, but it is nowhere near as good as it will be in the future versions. and I think, I hope, we'll see a similar thing to what happened with the GPT text models, which is people will always demand more and better and they will always find new and better things to use it for and the world will just make ever more amazing videos. How quickly... Oh, sorry. We're so early on the tech curve for video here. Like GPT-1 really was Sora 1 for this modality and the progress we've made kind of in the last 18 months getting to this 3.5 moment, right?

7:00is really compressed compared to how long it took to go from GPT-1 to 3.5 in the language domain. So we're really expecting progress to continue to be meteoric here in the near future. How quickly do you expect the Cameo feature to be cloned? That feels like an equally important part of the models made a leap, but the product and the experience of creating these assets is wildly innovative. We saw stories get cloned. We saw, you know, algo video short form feeds get cloned. I expect many other platforms to be looking at this functionality and realizing that this might be the future. You guys certainly believe that it could be important.

7:47So how quickly do you think? We're actually totally okay with a world where we do the product innovation and everybody else copies. And I don't think it works for them as well as they think it does. A lot of people have tried copying ChatGPT. You can go look at some of our competitors' apps, and they even copy the mistakes. They even copy the design decisions we really wish we hadn't made. And maybe it's worked well for them. I guess I kind of hope it has, but it's been fine for us. I think the key to this is not any one innovation, but it's repeatedly putting them out again and again and being first to come up with them and put them into a cohesive offering and, you know, that's what we want to be good at.

8:27And if other people want to clone the stuff that works, we also sometimes clone stuff that works. That's fine. But mostly we want to be able to drive the innovation. And I think Bill and his team have done an incredible job of figuring out how people actually want to use these video models, what the models need to do. Really, they've approached this as a full stack problem from how do you train the video model to how do you make this enjoyable for users. but cameos are one out of many ideas they have from here on the journey to like the product that we hope to eventually build. And so if people take some inspiration from us and copy us along the way, I'm sure they will.

9:03It's fine. How do you think about the like popular claim that we want AI detection? I want AI content flagged. Is that a stated preference? That's not a revealed preference? Because personally, I don't want bad AI content, but I don't want bad human-made content either. I want great both, and I'm fine when someone comes up with something genius and they instantiate it with a video model. How do you think about it? I think that is the real thing is you don't want slop. You want great content. Different people, One man's slop is another man's treasure for sure. But what you care about is like good, original, thoughtful, new, helpful, whatever content.

9:51And whether that is generated entirely by a human or entirely by AI or what I expect will mostly happen in the future, which is tool-assisted, human-driven generation, I don't think you care that much if the content is great. there's a lot of like, you know, stuff that is technically written or drawn or filmed by a human, but is completely derivative and much less original than what an AI has generated. And I think that will be what people really care about long term. You just want great content. Now, I also do want some human connection with it. When I read a great book, first thing I want to do is read about the author that wrote it and what life experience went into that.

10:34I don't think that'll go away. But if they're using an AI as a tool to help them make the writing better, sign me up. That sounds great. Similarly, I would rather watch a video about someone I know than some random AI generated character, which is part of why I think this was cool to offer. One design decision the team made that I thought was really great, and I was actually pushing them in a different direction earlier on, and then I decided they were totally right, and I thanked them and dropped it was the fact that the feed is, um, AI only and not a mix of AI plus some uploaded videos, I think is a subtle, but extremely important design decision in how people are relating to this.

11:13Yeah. It was a very weird experience for me. I was, I was thinking about the collab posts that I was making announcing this interview. And my initial thing was like, well, I'm going to have to think of a script or I'm going to have to think of, you know, what I say, or I should record a piece of this and then I'll use it. And it was like, no, I just typed the prompt in and then I get the front facing videos. Remarkable. What are you, what kind of indicators are you guys looking at, uh, as Sora can transition from, uh, what it was the second it launched, which was a creative tool into something that's more of a consumpt, like a consumption platform, traditional, you know, social media platform.

11:49Like what, talk about kind of what you guys are pushing forward to, because obviously you're seeding the network with the tool, but it's certainly much harder to turn it into something that people are spending hours a day in, purely consuming content and not creating content. You know, we really wanted to design this from the ground up to be centered around creation. And a lot of the metrics that we've been focused on optimizing here are really aligned with making sure as many people as possible are actually like getting their hands on the Sora 2 model itself and, you know, able to create content with their friends and like for the rest of the world.

12:26One metric that we're really proud of with this launch so far is that 70 % of our users are actually creating content, even to this day, you know, a week and a half after launch. And that's like vastly higher than on any other social media platform. And I think it really speaks to just how fun creation can be with the right tool set, right? If you look at any of these kind of legacy platforms, there's just like so much friction from like getting off the feed and like into some creative flow state, right? You have to like put the phone down, you have to go get like a camcorder, start recording yourself, find your friends, like do a dance, etc.

13:00It's just like a lot of work, right? On Sora, like you can just pick up your phone, find like any video you like in the feed, remix it, you know, cameo any of your friends. And I think one insight that was not obvious to us at first, but we've kind of clearly seen as an emergent behavior of this product is just like, there's all these people out there who would not necessarily want to be like, you know, influencers or something or have like a big social media presence. But the fact that like all of their friends can just access their cameo, right, put them in all of these crazy situations, actually like kind of gets them into the playing field in a way that felt really high friction before.

13:33And so, you know, we're closing in on close to like 2 million weekly active users now. We're really excited that such a huge percentage of that user base to this day is like still creating with Sora. And we're going to continue pushing on that direction and making sure people have even more powerful tools in the future. Yeah. So 70 % of Sora users are creating content. The typical benchmark that people kind of quote randomly is like 1 % creation, 99 % consumption, something like that. And that certainly feels like my experience on Instagram. I post a photo every once in a while, but most of the time I'm just kind of scrolling.

14:06And I'm wondering if you think that that 1 % will be much higher on Sora in terms of actual time in the app, time prompting versus time scrolling. If you have any data, that'd be super interesting. But then also, does that make it more of like a competitor to video games than traditional social media because it's such a lean forward experience versus just lay back? What do you think? Yeah, it's a great question. We still need to study this more exactly how creation versus consumption habits kind of change over time for folks on the platform. It's still pretty early days. I do agree with your point, though, that I think over time, this is going to feel much more immersive in a way that video games kind of do.

14:47You have more agency when you're actually using the platform, not just kind of mindlessly scrolling a feed hours a day. And one interpretation of this product, which I think is kind of interesting, especially from the research perspective, is Cameo is in some ways the simplest way where you can kind of like inject yourself into the model, right? So it's a very low bandwidth communication channel right now. You know, you're only giving like a few seconds of video footage of like any given individual like into the app. But like over time, right, you can imagine like these models know more and more about your life.

15:19They really like deeply understand your friends, how you want to like show up in the world. And like over time, this can almost become like a little mini like alternate reality, right? So like you're not just generating like videos of yourself with your friends. Like you actually just have like digital copies of yourself running in the model on this Sora platform, interacting with other people with agency. And so I think over time, we're really going to see this platform evolve into something that feels kind of familiar today and just something that really leans into the full intelligence of Sora 2 in the future and really leverages all the world simulation capabilities that we're working on internally.

15:50Yeah, I would add to that that if you think of this spectrum of the kind of entertainment you can have in front of a computer, at one end you have like watch a two and a half hour movie and you hit play and then you lean back and you don't do anything at all. And then at the other end, you have like a very intense video game and you're like, you know, sweating and your heart's racing and it's like super, super active. AI is going to push things to be more in between there. So you'll have, maybe you're still watching that movie, but now you can like say something a few times throughout the course of it and it changes what happens as the movie plays out.

16:27Or with Sora, you're seeing this amazing new phenomenon where most users are creating in a world where traditionally only 1 % of them did. And so yes, you're like watching a video feed, but you're doing a little bit more. And it, at least for me, really changes how fun the whole thing is and how I feel about it. Then maybe you'll do what Bill said and you'll be way more actively participating in the Sora feed. And I think you're just going to see that continuum blur a lot more. Did you see Bandersnatch by any chance, Sam? Have you seen this Netflix? It's like a Netflix choose your own adventure.

16:57And it was a really cool idea, but ultimately people, it never really took off and became like something that they do again and again and again. And I'm wondering if it was because it was like not customizable enough or people just want to just sit back and see a director's vision. I don't know. I never heard of that, but it sounds cool. Yeah. How do you think, question for Sam, how do you think about allocating compute to Sora versus the rest of the business? I imagine Bill is constantly in your ear. Every other hour. But how are you thinking about it? You know, my real answer is I've entirely changed my focus of how I spend my days to just go get more compute rather than have to make the compute allocation decisions.

17:38I still do have to make some short-term compute allocation decisions. But I hope we are heading to a world where I am instead telling people you've got to find a way to use more compute. And we're going to be very aggressive here. It feels like you're doing a great job of bringing things within your control, within the supply chain. what is outside of your control at this point i mean most of it but i feel like you have great you have great partners all up and down the stack multiple partners in different parts of the chain like when i think about scaling up sora i i i feel like it's crazy to bet against you like you're gonna you're gonna get the chip so you're not gonna be gpu 4 try to buy like 10 gigawatts an hour for delivery next year it's not so easy It's funny.

18:24How are the conversations going with Hollywood? Oh, yeah. Oh, actually. Yeah, you take it. Yeah, I was going to say, we've been chatting actually with a few very notable folks in Hollywood over the last week. I think people's first reaction to this is very understandably going to involve a lot of trepidation and anxiety. When we've gotten to just sit in a room with these folks, though, and really explain what we're building, I've actually been pretty struck by like how excited folks in Hollywood are about this. You know, we were chatting with one actor recently who mentioned that, you know, on Twitter like a year ago, saw like a deepfake of her generated with one of these like open source models, which really had like a lot of nasty content created.

19:08And when we really like walked her through kind of all of our safety mitigations, right, how we're making sure that we have this like very well-defined model spec, which dictates the behavior that we allow on this platform and how we are really leaning into like full control of likeness, right? More so than any other platform. Like you have to come in through the Cameo process. You can't just like upload an image of yourself and just like generate a video of it, like any person. You have to come in through Cameo. I think it became clear that, you know, we're really setting the right standard here in terms of making sure people are in full control of their likeness in Hollywood.

19:41I think that's where like a lot of this anxiety comes from, right? It's this feeling that, you know, some random person can just kind of take videos or images of you and do whatever they want with them and create all of this like, like, like terrible content that's like outside of your purview. But we've really been like designing Sora from the ground up to put users in full control of their likeness end to end from the moment you sign into the app to, you know, needing cameo permissions to like access any of your friends generations. So, you know, I think we need to engage more with Hollywood and we're going to continue to do that.

20:11But once we really explain the story of Sora, you know, they're very receptive to it. Do you think there's a world of folks? To add something to that, like, you know, the team asked me before launch if they could put my cameo in there, open access. And I, of course, thought about it for a second and said absolutely yes. I had all these Hollywood celebrities then messaging me on the first day being like, you're absolutely crazy. This is insane. This is like the dumbest thing I've ever seen. And then by about the third day, they were like, hmm, that was really smart. You got like a lot of, you know, free publicity.

20:39Maybe we need to be doing that. And I think you're now seeing actual celebrities say, okay, I'm going to do this, and I expect a lot more of them will. Similar thing on other kinds of characters in IP. I can totally imagine a world where our problem in a year or six months or maybe even less is not that people don't want their cameos or their characters appearing, but they think we are not fairly having their characters or cameo appearing often enough. This may turn out to be a really big thing for fan connection. Now, it may be that kind of the previous generation of celebrities don't want to do this and the influencer celebrities all do.

21:18I don't know how that's going to go, but I bet this will be like a pretty deep kind of new connection. Yeah, it seems like it's been good for DiCaprio in the memes. Like he's not directly monetizing those when you show the champagne meme or him pointing at the TV. But like, you know, it builds his aura in some way. A friend of ours posted something yesterday. This is Jeremy Gaffan. He said, the reason we're so upset about slop is because it's obvious we're all going to love consuming it in two to three years. It's not going to be slop for long. Do you agree, Sam? I mean, some of it will be slop to some people and some of it won't.

21:53I remember there was a real reaction like this in the early GPT days where people were like, I can't believe anyone reads this. It's like total crap. It's full of hallucinations. You know, it's like, it's not useful to anyone. And then it became more useful to some people, but they said, I can't believe anybody like ever thinks this thing writes a beautiful sentence. That's insane. And then with GPT-5, you have authors saying like, wow, this is a useful tool. It sometimes like writes a beautiful sentence. Yeah. And I kind of think it'll follow a similar trajectory. What do you think about the fact that people feel at least, I don't know if they actually can, but it feels like you can still clock GPT-5 writing, you know, it's not this, it's that, the M dash.

22:31Like, will we still see these artifacts in three years in Sora 5 that people are like, oh, if you know, you know, you can tell, but most people can't? Yeah, it's like, what's the M - of video? Because I don't think it's like six fingers. No, no, definitely not. That's the typo, which doesn't happen anymore. Yeah, I think right now the M - is like this like slightly wired speech pattern in Sora where it likes to say a lot of words very quickly. You know, these generations definitely have like a style to them. I think analogously to GPT, we really want to give users a lot of control over exactly how their videos show up right on the platform.

23:10Like if you really want kind of like a very soothing experience, right, not a lot of shot changes going on. We want to give users the ability to generate that. We're going to continue to give more optionality to people. So, you know, there'll be some default kind of behaviors and quirks of Sora for sure. But we definitely want all power users to be able to be in full control. Random question. Where did the name Sora come from? Yeah, this is a fun one. So the original Sora came out in February 2024, the OG blog post. We did not have a name for it. I think like up to two days before we like revealed the model to the world.

23:46We just could not agree on the team what it should be. Did you at least have a code word or something? We just called it like video gen. Okay. And so at some ungodly hour, I like just started pumping a bunch of crazy ideas into chat GPT. And then like we basically ran out of like English words. So then we switched to like Japanese words. Wow. And then Sora came out. I was like, wow, that sounds really nice. It means sky, you know, linked with like imagination, like all the possibilities of creation. And so then we just like last minute ship Sora. So yeah, it was kind of a mad dash. Okay, speaking of Japanese stuff, Sam, you said you were looking for an Acura NSX a while back.

24:24It's kind of this throwback car. It's not a Waymo. What do you think the piece of content or format will be that remains loved in an age where everyone's taking the Waymo of video, the Sora video generation? What do you think is like the way to do for that? Well, first of all, I got that NSX, and it lived up to all of the childhood hype. I mean, just incredible. That's amazing. That car is so fantastic. That's great.