Ep 76: Sora Creators Bill Peebles, Rohan Sahai & Thomas Dimson on Their Unexpected Viral Success

This episode features the core team behind Sora, OpenAI's groundbreaking video generation platform that became the #1 app in the App Store. Bill Peebles (research lead), Rohan Sahai (product lead), and Thomas Dimson (engineering/product lead with Instagram background) discuss the unexpected viral success of Sora's launch, the product journey that led to the breakthrough "cameo" feature (putting yourself in AI-generated videos), and their philosophy of building a creator-first social network that prioritizes human creativity over passive consumption. They reveal the technical milestones in video generation, their small team size (under 50 people total at launch), navigation of content moderation challenges, early monetization strategy, and their ambitious vision for video models as world simulators that could eventually contribute to scientific breakthroughs by 2028. The conversation captures both the tactical product decisions and strategic philosophy that made Sora a cultural phenomenon.

• 0:00Speaker 00

  I feel like I have the main characters of the AI ecosystem.

• 0:02Speaker 01

  People are very creative. It's a social network. So social networks are about people and relationships.

• 0:07Speaker 02

  Even if you change the ratio creator to consumer by 1%, like that is a massive impact on the world.

• 0:12Speaker 00

  Celebrities and rights holders are going through like a very fast evolution of understanding this technology and like how to use it.

• 0:17Speaker 03

  Shag is posting bangers, by the way. I think the first scientific breakthrough that comes as a result of simulating some phenomenon in video models would be an insane milestone, right?

• 0:26Speaker 00

  A prediction of when that might be?

• 0:27Speaker 03

  I would be shocked if we are not making breakthroughs like this by.

• 0:31Speaker 00

  Sora has taken the world by storm this past month. It's number one in the App Store, has spawned endless hilarious videos. And I had the privilege today of speaking with Bill, Rohan and Thomas, three of the main folks at Sora behind this magic at OpenAI. It was just a really fun conversation. I got to ask them everything that I've been thinking about that Sora spawned. We talked about their reactions to Ben Thompson's take on Sora, as well as kind of the key product decisions that factored into how they thought about making it into a social experience. We talked about the state of video models, how they progressed over the past years and what's next, future milestones, including novel physics discoveries and their predictions for timelines on those. We also hit on how video models in general will fit into the broader OpenAI ecosystem, including ChatGPT over time.

• 1:12Speaker 00

  Just a ton of fun to get to talk with these three brilliant folks about all these things in the video ecosystem. I think folks will really enjoy it. Without further ado, here's the Sora team.

• 1:23Speaker 00

  Guys, I'm so excited to do this. Thank you so much for coming in here.

• 1:26Speaker 02

  Thanks for having us.

• 1:27Speaker 00

  I feel like I have the main characters of the AI ecosystem. Just like as a podcast host, a fun thing to get. So thank you, guys. I think a bunch of different things we'll want to hit on today. You know, maybe to start, I feel like there's been all these, you know, stories told of the ChatGPT launch and how it was kind of not expected to be anything big and out of nowhere. Turned into this just, you know, massively used consumer product. I'm curious whether like the same was true of Sora or just like, did you guys kind of expect this reaction? Any kind of fun stories from that first day?

• 1:53Speaker 02

  I mean, I didn't expect to be number one in the App Store for like a month. But it's, you know, one of those things you look at in retrospect, like the research team absolutely cooked. We. Have a knack at OpenAI of creating these viral moments really well. And so, you know, happy to have done it, but I think it surpassed my expectations.

• 2:12Speaker 01

  I think Bill also has a hopeless optimism about him.

• 2:15Speaker 02

  Yeah.

• 2:15Speaker 01

  It's not hopeless, actually.

• 2:17Speaker 02

  Not hopeless.

• 2:18Speaker 01

  I think it was like, there was a meme, I saw the first, it was a couple of days after launch, where it was like, number three in the App Store, and then Bill was like, demanding number one. I think Sam was putting you in the cage or something. It's like, until you were number one.

• 2:31Speaker 00

  So my most like Gen.

• 2:34Speaker 01

  2.5K.

• 2:35Speaker 02

  And. Wild it to life.

• 2:36Speaker 01

  You know what, I'm always, I've worked a lot of different social media products.

• 2:40Speaker 00

  You're up to Instagram for a while.

• 2:41Speaker 01

  Yeah, and some successful, some failures within that ecosystem. So I tend to be very grounded and realistic and all this sort of stuff. But I have to say, like, that optimism is actually very important. We were, we ended up being number one on the App Store, which was wildly crazy for my expectations. But it's delightful to see.

• 3:00Speaker 00

  So I imagine there's an aspect of this where you kind of have to reserve some sort of GPU capacity, right, based on what you expect. And so I always wonder, like, how in the world do you figure that out?

• 3:07Speaker 03

  You know, it's not really a science. It's a lot of, yeah. I think everyone feels compute constrained across the entire industry, not just within OpenAI. And certainly when we launch an entirely new product surface, right, that is as compute intensive as video, it requires some pain elsewhere in the company. I think one thing that's really nice about OpenAI, and like, to Sam's credit, like, everyone kind of feels invested in everybody's success across the board, right? Certainly when ChatGPT, you know, releases like amazing new visual generation features, we want to make sure those are great and like, people are having an awesome experience in chat. And likewise, the company really feels invested in Sora's success. People see that, you know, we need to take these new bets and it's important for the company's longevity, not just to have a super successful LM-facing consumer product, but

• 3:51Speaker 03

  also the winning one in video. And everyone's willing to chip in, you know, do your part in wartime.

• 3:56Speaker 00

  When did the idea to like, create Sora as an independent app, you know, come about? Was that always obvious from the beginning?

• 4:00Speaker 02

  It wasn't always obvious. I think there was a clear vision and there has been like. You know, a great level and of ambition in terms of what this technology can be and how it can transform UGC and media generally. And so then if you pair that with what we were actually experiencing with the model with like, hey, there's something very special in multiplayer mode. We're seeing this emergent behavior even before Sora 2 in terms of like, what is AI enabling? People can participate in trends extremely quickly and create these trends extremely quickly. Like, we've seen that with other video generation software in terms of like, this is flooding social media more and more. And then with ImageGen, when we launched ImageGen, it's amazing, it blew up in ChatGPT, but people were just putting

• 4:43Speaker 02

  themselves in these amazing scenes. And so something about seeing yourself, casting yourself in these videos, and we just wanted to take that one step further, like, do it with your friends. And we felt magic internally. And I think once we felt that, it was, it was a very easy decision. The one thing to add also is it's not like inconceivable that maybe these, the surfaces come closer together. But ChatGPT does feel very single player, and that's sacred in a way right now. Like, it can feel a little jarring if you, you know, you don't know when you do something in ChatGPT, is this public or not? So we also, we're sensitive to that as well.

• 5:18Speaker 01

  Yeah, there's actually, there's a long journey.

• 5:20Speaker 02

  Yeah.

• 5:20Speaker 01

  Of, you know, this year, for me at least, I was like, just trying to like, think about what social means in this kind of bigger context. And we had lots of wacky prototypes along the way. I'm very delighted it ended up in Sora, but it was not clear to me, like, product journey is always extremely.

• 5:36Speaker 00

  What was the strangest thing you threw out?

• 5:38Speaker 01

  I mean, we, we had a prototype of just a pure social media stream inside of ChatGPT.

• 5:44Speaker 02

  Yeah.

• 5:45Speaker 01

  And we're trying it out internally. One of the things we saw, so I was just like, it was like, think about social. Okay, the first thing that comes to mind, well, we have ChatGPT, why don't we just put some social features in it, see what it looks like? And when we started using it, we used it internally, it happened to be right before the ImageGen launch. And what we were seeing We had these like, it was a kind of like threaded sort of Reddit style. And we would see these very long chains where what would happen is like, somebody would put up an image and then somebody would be like, okay, it's smoking a cigar now. Like as I say, it's a duck. It's a duck with a cigar. Now it's upside down. Now it's red. And you're like, it was really cool to see it evolve. And I was like, this is these little meme chains. And that seemed like something very special. And I'm thinking about more sort of what Rohan

• 6:28Speaker 01

  was saying is like, that's actually something you can't really do without generative AI, because it's too hard to create. It's too hard to remix, to riff on something. And so that was like the wackiest, I think, one that we had. And again, through this long, snakey product journey, the Sora models were getting pretty good. We always wanted to do something really ambitious with them to really mass deploy them. And fortunately, it just kind of like fit like a glove in terms of, you know, ImageGen is very cool, but videos are even even cooler in some ways, especially in this context.

• 7:00Speaker 00

  And it sounds like from the inventory, you kind of knew cameos would be this like killer feature for for these powerful models.

• 7:05Speaker 02

  I don't know if we knew.

• 7:06Speaker 01

  Not really, no.

• 7:07Speaker 00

  Interesting.

• 7:08Speaker 02

  I mean, yeah, it's one of those things in retrospect. Yeah, it's obvious. But I think this idea was in the ether, you know, from from when the research team was actually building the model. It wasn't until we actually like had it, like one of our engineers on the team, Bobo, was like, everyone send me a video of you saying, hey, Sora, it's Thomas or Phil. Hey, Sora, bring me to life. We like in a Slack thread, literally, he uploaded all these in the backend, then you could just tag the person. We had like talked about this idea.

• 7:35Speaker 01

  Yeah.

• 7:35Speaker 02

  And then finally, like, everything changed that day. I think like, we all just started cameoing each other. We didn't have a verb for this at the time. We just started tagging each other. And like, it happened before we even noticed. It wasn't like, boom, this is it. Just like, in a couple of days, I feel like we were like, our feed is just cameos. Like, what's going on? And we were all using the app.

• 7:56Speaker 01

  We actually had a moment where we're like, is this bad?

• 7:58Speaker 02

  Yeah.

• 7:58Speaker 01

  Wait a second, like, did we just like, we had interesting stuff before.

• 8:01Speaker 02

  Yeah.

• 8:01Speaker 01

  Now it's just people like, and it's like, wait a second. No, that's actually amazing. That's what we need.

• 8:06Speaker 02

  Yeah.

• 8:07Speaker 00

  One of the coolest parts of building these products is you release it day one. And I'm sure people start immediately using it in like a million different ways that you never intended. Like, so it's been like the most surprising way you've seen people use or capabilities that you didn't know actually existed.

• 8:18Speaker 02

  Some people were like putting themselves in motivational situations, like envisioning themselves in situations they wanted to be in.

• 8:25Speaker 01

  Anytime I see a style transform, it's maybe a too small of a word for this, but like, you know, you take something, take a person. There's one review that I didn't, I couldn't believe this would work and ended up being one of the most remixed videos where I just had a kid opening a package for Christmas and it's like, "It's a Bill Peebles action video." And it looks like Bill Peebles.

• 8:43Speaker 03

  This is what's keeping me on the leaderboard. I'm like 25 or something.

• 8:48Speaker 01

  Bill Peebles action video.

• 8:50Speaker 03

  It's a tosses, Jen.

• 8:51Speaker 01

  It's, it's, I like, it's remarkable that the model understands that and understands the context of just like from a couple of numbers can put you in a completely unusual scene. I was amazed that that emerged from the model. I see those type of things in my feed all the time. Not exactly the same, but it's, it's always some, some riff on that that's very cool. And yeah, any, anything that's massive transformed, the claymation ones I love.

• 9:16Speaker 03

  Yeah.

• 9:16Speaker 01

  The video game ones, I think it's a little underexplored right now where you have like a LucasArts adventure, but it's actually Rohan or Bill or something. Those things, I just eat that up. This is the immediate like.

• 9:25Speaker 03

  See what the secret emoji is on those ones.

• 9:28Speaker 02

  Some people are just getting insane, like the Kaigumatsumara gens, like just like things, we didn't even get close to that level of like stylistic output. So just seeing the creativity of the world.

• 9:41Speaker 03

  Yeah. I think after we launched Storyboard, which lets you generate up to 25 seconds of video, that's when like the quality bar for me like really spiked. Like I'm actually surprised how in like one shot out of this model, you can just get unbelievably like cohesive stories. Like this is something that you cannot even like get with like 100 attempts on like Sora 1. This is really a new thing in Sora 2 to just really reflect over the step function increase in intelligence.

• 10:06Speaker 00

  You know, on TPN, I think you shared about like 70% of Sora users or creators, which is mind boggling. I'm sure it's, it's still super high. You know, one thing I thought was interesting around the discourse around Sora is when it was initially released, I think Ben Thompson wrote this piece and he was like, "I'm skeptical of Sora." Like, you know, most people don't want to create. They want to consume. That's been the behavior of every other product. And then I think he like issued to me a call point. He was like, "Okay, I was wrong. This is a cool product." Curious like what your reaction was to that and, and you know, you think this level of creation kind of persists.

• 10:34Speaker 03

  Yeah, I mean, we really designed the app from the ground up to be focused on the creation side. And this was really our core hypothesis going into it. Like, there's so many ways existing social media networks are like really amazing and people get like so much joy out of them. And that can come from the consumption side. At the same time, I do think probably the worst harms and like the worst case scenarios come from like the doom scrolling. And like, we really wanted to figure out how do we kind of avoid this very dystopian future of just like a slop feed that's optimized exactly to what you want to watch. It's very detached from your friends. It feels very lonely and isolating. And we kind of tackled this across the board. So like, really, I think the the single most important thing we did there was really with Cameo. Like Cameo is really what makes a gen feel personal to you and makes it feel very human in a way that just kind of like text to video

• 11:19Speaker 03

  or just like all these very simple ways of prompting these models maybe doesn't. And then we also did a lot of work, Thomas specifically, on the recommender side, right? So really, the way that these things really go off the rails if you're you're not careful or being kind of like a good faith actor is if you design this Rexus in a way, which is sort of like maximally clickbaity slash engaging and like kind of incentivizes the doom scrolling behavior. And Thomas did a lot of like very pioneering work to really rethink what like a modern Rexus stack looks like in this product oriented towards creativity over consumption.

• 11:49Speaker 01

  Yeah. And some of that is actually is kind of self fulfilling in a way or like at a circle. But the fact that you can participate in the network means that you can also encourage like that's now an optimization objective that can be in there. And I think it's a very healthy one. If you're actually making a decision to, oh, I'm going to remix this generation. That's like a very active behavior. It puts you in a very creative mode. It's not that whole just consumption oriented thing. And so that was the idea behind the Rexus is like, because it's so easy to create, we can actually encourage people to create in a way that you don't normally optimize for. Because, I mean, having worked at Instagram, the idea that you saw this photo or this video, and then you helped you opened your camera, and then you took a photo, and you have to attribute it back to the original

• 12:34Speaker 01

  photo you saw, it's impossible. Or it's very close to impossible. You can get a little bit of signal out of that. But in this world, you can get signal out of everything. So I think it's a it's a very nuanced, but kind of interesting sort of behavior that you can see.

• 12:49Speaker 02

  Yeah, and like on Ben Thompson, I think his arc, you know, I was crushed when he first, you know, I listened to Strategery every morning. And so I was crushed when he gave his first round of feedback. But his arc makes sense, right? Like, he is a consumer. Cold Start Network is a very hard problem. So the feed wasn't super dialed in. He's an early adopter, too. So early adopter without like a feed that has had any time to bake, it wasn't that interesting. So as a pure consumer and not a creator, I think he jumped in there. I think he had a quote about it just being like a Sam Altman advertisement. But the feed did feel like that for a little bit, to be fair. I mean, I thought it was funny. But but then I think when he had his kind of like, you know, he brought it back to his mea culpa, he had a good point, which is like, even if you change the ratio of creator to consumer by 1%,

• 13:34Speaker 02

  like, that is a massive impact on the world, right? And that's like such a different product. And I think that was like a good summary of why we think Sora is special and what he might have missed the first time around, because he is particularly like a pure consumer and not a creator in his everyday life.

• 13:50Speaker 01

  I do think it's also, it's, we played with this in some of our early prototypes off Sora. But the idea that there's a human creating it versus there's like a bot creating it is a very, very foundational difference. It's like kind of hard to appreciate. But if you took the Sora feed today, like maybe excluding Camino's, and you just stripped the person that posted it, it would be very uninteresting, in my opinion. Like, part of the essential thing is that somebody has looked at this and decided this is like, kind of like, they're putting their stamp of approval on it. They're also involved in the creative process. And I think that is, it's like, easy to think about. I've seen a few of these social networks that are completely AI-driven, but they last for like five minutes of interest. I get a like from a bot, and I'm like, what does that like me? I don't really understand.

• 14:31Speaker 01

  Whereas like, if Ron likes my content, like, oh, yeah, you know, I know what that means. Hey, you saw my thing. Like, did you like it? Like, yeah.

• 14:40Speaker 00

  Right. I think it's almost that so visceral about the fact that a human created it, because they're in the, they're in it, right?

• 14:44Speaker 01

  That's also true.

• 14:46Speaker 00

  Other tools. You know, one thing that's also been really interesting to see as, you know, Sora's been out, it's just like, I feel like celebrities and rights holders are going through like a very fast evolution of understanding this technology and like how to use it. Can you talk a little bit about like what that journey's been like over the past month? And, you know, where where most folks are today?

• 15:02Speaker 02

  Yeah, I mean, we've chatted with kind of every facet of society since launching this. And, you know, I think if you rewind the clock like a month ago,

• 15:08Speaker 03

  You know, most people in the world really did not know video gen was a thing. Certainly, they didn't know.

• 15:12Speaker 00

  You knew it was gonna be number one.

• 15:13Speaker 03

  Well, higher of bars. And Sora, I was very confident. You go through my gen history. Over time, you know, we've really like, just had the opportunity to sit down with these people and hear out both where they're really excited about this platform. I mean, there's a huge value proposition here, especially for rights holders. We launched character cameos yesterday, right? You can imagine if you have some IP that people love and any kid now can generate a gen with them in that that character, I mean, that's gonna be huge for these rights holders. And we're also hearing them out on their concerns, right? Like, making sure that they have a lot of input on how their characters show up. They want restrictions. They don't want it to just kind of be a free for all.

• 15:52Speaker 00

  Yeah, I think it's a pretty clever UI for setting the basic parameters around that. I'm sure that I'm sure the real celebrities have more specifics. For me, I felt good.

• 16:00Speaker 03

  Yeah, most people go through this cameo restrictions flow. And if there's like a celebrity who's particularly, you know, anxious about getting on the platform, we really step them through exactly like how they should think about setting these things. But when we kind of walk people through, you know, what we have on the docket and the features we've already integrated with character cameos, cameo restrictions, et cetera, there's honestly a lot of excitement around this. And, you know, just earlier today, we announced that we're starting to bring monetization to Sora. We imagine in the future.

• 16:25Speaker 00

  That was very nice of you to announce that, you know, 30 minutes before you come out.

• 16:29Speaker 03

  You know, you got the exclusive.

• 16:30Speaker 00

  I assume you were like holding it for days. Release it right before this episode.

• 16:34Speaker 03

  This exact moment. And, you know, we are going to start piloting new ways for rights holders to monetize their content. And we're gonna really prioritize folks who have invested in the platform from the beginning, have gotten on now, and we think amazing things are gonna happen from that.

• 16:48Speaker 00

  Have there been any specific, like, celebrities or rights holders that you're like, oh, they already get it?

• 16:52Speaker 02

  I mean, Mark Cuban immediately added costplusrhogs.com to his cameo instructions. I feel like he, you know, was first to figure out what could be like a, you know, an amazing industry in of itself in terms of how you think about, you know, brand advertisement and stuff like that. I mean, there's been a lot of people, like, we have a moderation filter that prevents you from creating a public figure. So when you go through the cameo flow, a lot of public figures actually get blocked. Because we don't want, you know, we're very sensitive to impersonation and that kind of stuff. So I think it's funny because we get a lot of inbound because they're like, we can't use your product. But the amount of inbound we got has been amazing. So like tons of people are actually trying this out,

• 17:34Speaker 02

  using it, having amazing time. Shaq incredibly creative. Like we talked to his team and they're just like, Shaq's just having a blast. Just use it. Just like generating. Like I don't, I don't even know if Shaq cares about like the name he's making for himself on this new network. I think he is just genuinely having a creative fun time.

• 17:50Speaker 03

  Shaq is posting bangers by the way.

• 17:53Speaker 02

  Yeah, go follow Shaq.

• 17:56Speaker 00

  He's just got to open it up though.

• 17:57Speaker 02

  Yeah.

• 17:59Speaker 00

  You've had both Google and Meta release, you know, similar kind of product, you know, products around the same time. You know, everyone's kind of racing forward on these video models. I wonder like what you make of those releases and kind of how you think this space evolves over time. Clearly there's a few folks with very good video models today.

• 18:12Speaker 01

  I think the future is very bright for all these models. Like I think there's, there's no world. I think Sora will stay on top because Bill's fine work. But you know, like this is obviously a new medium. It's a new creative expression. I think the, the industry will evolve completely and this is going to be a big part of that use case. I think Meta, Google, et cetera are right to be like, I'm sure there's a code red or somebody joked yesterday that'd be a code purple. Or you have to invent a new color for this one. And so like, I think that's, that's actually correct. Because they're seeing something emerge here that has been created with this magical technology. That said, I haven't been overly impressed with like some of the productization of the work so far.

• 18:53Speaker 01

  And I really think that goes back to our, our idea of like people embracing people. People are very creative. It's a social network. So social networks are about people and relationships. And the thing that really enabled us to move forward is this idea of cameo and bringing, bringing this like creative aspect to your life with your friends. And I think that was kind of missed. I'm not surprised. I'm not surprised post on like a model level, but I'm not totally surprised because it was just a, it was actually a very again, nonlinear, snaky journey to get to this point. It wasn't like. I think there were ideas of doing cameo, but there was also other ideas where like, what does remix mean in this context? Like, what is it? We had one that was always weird. I don't think, nobody else likes this one, but I always liked it, which was like you would,

• 19:36Speaker 01

  you take a video of yourself reacting to the AI video, kind of in the corner.

• 19:40Speaker 00

  Oh, yeah, I forgot.

• 19:40Speaker 01

  There was a reaction.

• 19:41Speaker 00

  We should bring that one back.

• 19:42Speaker 01

  I even generated like a dog reacting to myself or something. It was cool. But we were like, we were trying things. We tried a lot of different stuff. Even like the, the UI, it was early on, Joey had this idea for Image Gen where it was just a simple UI where you had your contact list and it was like, you just like pick and choose people to be in the gen. That kind of makes sense, but it was not, it's not obvious. Of course, it's obvious in retrospect. That's why it's code purple. But like, it's, it's not obvious when you go through this whole thing. There's a lot of micro decisions that are along the way that can make or break your product. With Instagram, I remember talking to Mikey, one of the founders there, and it was like, the square photos was like a limitation that they just imposed. I think probably because they didn't want to deal

• 20:25Speaker 01

  with like aspect ratio nonsense. And that was actually essential for its success. So, long story short, I think this space is going to get very, very competitive, very, very fast. I'm reasonably confident we're going to stay on top. Hope so. And I do think you need to embrace people in this whole process and I think empower them with creative, creative tooling. And that's what we're doing.

• 20:47Speaker 00

  And it seems on the product side, you're continuing to lean into this. I know, I think you guys have talked about like kind of focusing on communities and like other kind of ways in the product. Talk a little bit about like how you kind of see yourselves leaning into this even further over time.

• 20:57Speaker 01

  Yeah, it's, this is cold start social network, impossible. Hardest problem ever.

• 21:03Speaker 00

  Fun one to work on.

• 21:04Speaker 01

  Yeah, it's, it's fun for sure, especially with, oh yeah. And by the way, it's a completely new medium. Okay, let me think about this. So we didn't really know exactly how things would evolve. And we gave it a very familiar form factor. You know, it looks very similar to other full screen media player apps. It definitely has a different feel. But we did have this hypothesis, and I think it's kind of proven correct, that this is much more fun with your friends. And that aspect definitely is in the product. It's emphasized in the recommender system, but I don't know if it's completely emphasized in all the ways that it could be. So I see us leading into that a lot more over time. I think the public feed is also incredibly important. And it's a kind of a way of even giving you inspiration to be like, oh, actually, well, Shaq's here.

• 21:48Speaker 01

  It'd be pretty cool to be with Shaq or have a riff on Shaq's Marilyn Monroe meme or whatever he's doing today. But like, so I think that there's some product work to be thought through. And I'm also just excited about like, you know, when you think about like, okay, what does this technology enable? That is more fun with your friends. There's probably things that we haven't thought of yet that we can actually lean into. I can't say exactly what they are. Certainly, we're more emphasizing the DMs feature over time because we think that's like, can be a very magical moment. But you can imagine things that happen in small groups, kind of like being very exciting. Even even large groups, I think, could be a very exciting angle. Certainly, our OpenAI group before we launched was very cool. You know, it's a very dense, connected, large network that people are just having a lot of fun with.

• 22:31Speaker 01

  And I'd like to be able to support that in the product in the future as well.

• 22:35Speaker 00

  That's interesting. And one thing I'm struck by is just clearly like the way you put the feed algorithm, like basically determines a lot of this behavior in the ecosystem, right? And like what you encourage there.

• 22:43Speaker 01

  There's a famous quote, I'm trying to remember who did it, but it's like, show me the incentives, I'll show you the outcome. And so this has never been more true than in those large recommender systems. And certainly when I worked on Instagram, there was a very explicit decision to prioritize your friends. We chose a lot of things to not introduce unconnected content in the main feed. I worked on the explore feed, which is a secondary service, but they made that a very explicit decision. And it's kind of drifted over time for the reason that people are posting less and all that sort of stuff. But if you look at your Instagram feed, it's kind of kind of wacky, I'd say, right now. It's like pretty disconcerting. Same with X. And I think because we brought this barrier to creation down so much,

• 23:24Speaker 01

  there's a good chance that we can actually emphasize that. It's very funny that AI videos are what bring you more connected to your friends. It's a little bit unusual. Yeah, the hottest densely connected network is an AI gen feed. But I'm all here for it, yeah.

• 23:39Speaker 00

  Yeah, I thought it was very like, in some ways, very like heartening use case for AI, right? I think, you know, one. One thing I was struck by is just like, it's just fun in a way that like, I feel like so many of the things that have been built on top of other OpenAI products are just very serious of like, you know, it's like every founder out of college wants to go build like a very specific vertical like enterprise app. And it's like, there haven't been, there actually really haven't been that many consumer products that have been built. I think a lot of people were playing around with this like, how do you, you know, have AI generations and people in the same, you know, I feel like character AI was even riffing on some of this stuff in the early days. And I feel like you guys have kind of nailed the first product around it. And it's just cool. I mean, I don't know, it's like good, cool for the world to see.

• 24:16Speaker 02

  I mean, it wasn't like, there's a whole crazy flow you have to go through to get your cameo, right? Of like an audio challenge and moving your head around. There were many moments where Thomas and I were like, we're cooked. There's no way, there's no way people are getting through this flow.

• 24:27Speaker 01

  You're talking about funny launch stories. Yeah. Cause we were trying to dial in the cameos and we're like, there, I don't know, there's a bit of voodoo in that whole thing.

• 24:34Speaker 02

  Oh, yeah.

• 24:34Speaker 01

  A lot of voodoo. If you stand in this particular pole.

• 24:36Speaker 02

  Oh, yeah.

• 24:37Speaker 01

  The time of the day is 4:00 P. M.

• 24:38Speaker 02

  Literally.

• 24:39Speaker 01

  This is like, we all went to this pole, we recorded cameos ourselves. And somehow we ended up at a flow, which actually is the optimal flow, but is impossible, impossible to do where you have to say the words while you move your head. That was, that was actually the optimal.

• 24:53Speaker 02

  Yeah, you were like, say a sentence, move your head around like this.

• 24:57Speaker 01

  Quick, brown, fuck. It was insane. But I mean, it would actually produce a slightly better, better cameos. Anyway, that was, that was a funny launch moment. We had to dial that back. We're like, I don't think anyone's going to get through that one because we can't get through it.

• 25:08Speaker 02

  Yeah.

• 25:08Speaker 00

  Is this a secret pro tip that my cameos will be better if I, if I had recorded that?

• 25:12Speaker 03

  That background looks very promising, actually.

• 25:14Speaker 02

  Yeah.

• 25:14Speaker 03

  Yeah.

• 25:14Speaker 02

  Yeah.

• 25:15Speaker 03

  Good lighting.

• 25:15Speaker 02

  Lighting.

• 25:16Speaker 00

  A very private place and just roll my head.

• 25:20Speaker 02

  Yeah, exactly.

• 25:21Speaker 00

  One tension I mentioned you get dragged into immediately because you're focused on creators is like, there's just such a range of sophistication of folks. We were talking about this before. There's like the kind of pure consumer type of creators that just want to like be able to remix something as easily as possible. Then you have these like prosumer type use cases where they're like, you know, super sophisticated. And you guys have obviously introduced some like basic editing functionality. How do you like, think about this product surface area over time?

• 25:44Speaker 03

  I think one thing that's cool about this product is that at its best, it's very democratizing in terms of who can create and who can kind of level up their game to become a traditionally pro creator, right? If you see kind of any of these god tier gens from like, you know, people who have really mastered Sora, you can just remix it and have basically direct access, right, to all of the ingredients that went into that. And you can sort of like learn the ropes gradually, right? About how you actually kind of master like prompting Sora, how you master like designing your cameo, et cetera. I think an important part of this is continuing to Let people who are at the frontier of creativity, these pro consumers, continue to be empowered with like even better tools, right? So they can keep pushing forward. And we are launching more features that are really aimed towards that group specifically. You know, Storyboard was a big one.

• 26:26Speaker 02

  Yeah.

• 26:27Speaker 03

  We're starting to introduce very basic kind of like editing features into the app, like stitching, which went out earlier this week. And so over time, you know, my hope is everyone kind of levels up, right? You know, we're going to really empower the top creators to do their thing. But just by virtue of seeing this stuff in the feed, right? And having all these amazing remixing and editing tools, everyone can kind of like gradually like turn into these people. And it will really just like make the feed an amazing place to be.

• 26:49Speaker 02

  There's something beautiful about an end state where more people can be creative. And then also people, that can be a gateway drug for even going deeper. I think I had this experience growing up with like GarageBand. I think it's like a good analogy here where it was so accessible. It was crazy, you know, in the simplest, you could just drag loops, right? You're not even really sitting down and playing an instrument, but you're starting to get a sense of like, all right, what are the elements of creating? And you can actually create some interesting things, but then you can go deeper and you're like, oh, now, now I'm actually curious enough to have the MIDI keyboard and like learn guitar and record. And like, it's a gateway drug, but you, we got there by making the, you know, the barrier to entry much lower.

• 27:34Speaker 00

  I'm curious, I feel like every app builder right now is kind of thinking about like, what do I build a scaffolding around these models versus like, I just go to a beach for two years and like the models are going to get way better?

• 27:43Speaker 03

  Yeah.

• 27:43Speaker 00

  How do you guys think about that as like a holistic team of like the, the types of scaffolding you want to build around the shortcomings in the model versus ways it'll just get better and you know?

• 27:51Speaker 03

  Yeah. In general, I think a lot of OpenAI's magic comes from just having this incredibly ambitious AGI-aligned research roadmap and sticking to it no matter what happens, no matter what the competitors release, no matter what the product pressure is. And that's certainly our philosophy on Sora. You know, I think what's cool is as these models get more and more powerful, we kind of discover all these amazing capabilities that they have. And that, you know, keeps like Rohan and Thomas very busy. So, you know, something like Cameo, something like remixing, right? Is really not just a feature that we pioneer on research, but it's really a joint collaboration with the work the amazing guys on product are doing to figure out all of the crazy things you can do with these models.

• 28:29Speaker 01

  I do think it requires a very creative lens on what you do with it. Or a willingness to fail, you know, like, let's try the AI, you know, completely AI feed with no cameos. It's actually not that cool. But I can imagine a lot of things in gaming and other, other surfaces that even today's LLMs and video models could support in a very interesting way. And so I think it's just a little bit of thinking outside the box and not trying to mirror the exact roadmap of OpenAI in any way. Uh, but being like, okay, like, what's a spin on this that actually makes sense? I'm always consumer build, not talking about enterprises, of which there are a thousand for these models, of course, but like the consumer things is just like, what's creative? What's new? Let's embrace that. There's way more you can do with this. There's control. There's all kinds of exciting things. Yeah.

• 29:09Speaker 03

  Yeah. This is a big part of the reason we started launching a Sora API as well with this generation of models. There is so much stuff, as Thomas is alluding to, that you can build with these things. And we're a very small team on Sora.

• 29:19Speaker 00

  Right. Shockingly small.

• 29:20Speaker 03

  Yeah, shockingly.

• 29:21Speaker 00

  Yeah. You were under 20 when you released it and now are like 50 or so?

• 29:25Speaker 03

  We're at like, so there's like roughly nine or 10 researchers on Sora. I think product is at what, like?

• 29:30Speaker 00

  Under 20.

• 29:31Speaker 03

  Yeah. Um, and then we have a systems team, which is about like 13 or so folks. So it's, it's like 40 ish in total. It's pretty small. And so, you know, all of, uh, you know, these creators who reach out or, uh, you know, all of these, uh, very entrepreneurial folks who want to build these new applications, they now have the ability to do that with the Sora API.

• 29:49Speaker 00

  Have you seen anything cool already on the Sora API? Or like, what gets you guys excited about, uh, what people can do?

• 29:53Speaker 03

  Mattel has done some cool stuff. So they've been prototyping.

• 29:56Speaker 00

  Like Barbie Mattel?

• 29:57Speaker 03

  Yeah. They've been prototyping new toys with Sora, which is just super cool to see. Um, you know, this is, I think, you know, over time, the applications will just get increasingly sophisticated. It's been, I think, three weeks since we launched the API, but yeah, already people are doing some pretty wacky stuff.

• 30:12Speaker 02

  I've seen some like CAD, CAD to visualization pipelines where people are taking like a CAD file, translating that to some caption that Sora could understand so they could visualize their parts, which I'm like, that doesn't seem very precise. And they're like, explained why this is actually critical and something missing in the file. I'm like, whoa. Yeah. Amazing. You know, maybe take a step back and talk about the video model side. Like, well, let me just kind of contextualize for our listeners, like, how has the AI video space progressed over the past few years? Maybe some of like the different milestones that matter to you?

• 30:42Speaker 03

  So I'd say for a while, uh, basically there was no progress in video and a lot of it was really on the image generation side. So one of the most important early papers in image generation, um, was Dolly one, which came out of, uh, OpenAI several years back. Um, that was like early work that Aditya did. And that was really the first time we kind of saw this, you know, breakthrough general purpose capability that we were starting to see in the LMs come to the visual generation side. Before that point, you know, there were models that were very niche in terms of modeling, very specific narrow distributions like human faces. But there was really not a moment where it was clear these things could model kind of all of the pixels that exist on the internet. I think from that point on, it became clear where things were headed. It took a few years to really get image generation under full control. So,

• 31:27Speaker 03

  you know, Dolly two, Dolly three. We started working on Sora in early 2023. It was co-developed at the time that Dolly three was being worked on. And we kind of had all of the necessary ingredients, like make a big breakthrough at that point in time. We started to start to understand how scaling works. Diffusion models were starting to become much more principled, both in terms of the actual formulation of diffusion, as well as the architecture of them. And Sora one was really like the GPT one moment for video. That was the first time you could even think about doing kind of high res, consistent generation for anything more than one second. We could do 60 seconds barely with that model. And since then, you know, we've really just been trying to push the frontier of intelligence and usability. So we, for a long time, we were actually trying to figure out what number of GPT is Sora two.

• 32:11Speaker 00

  Yeah.

• 32:11Speaker 03

  I think it's, we're very clear now. It's GPT 3.5 in many ways, both in terms of the breakthrough in capabilities. This is the only model that can, you know, do an Olympics gymnastic routine and like not have things just like go crazy.

• 32:22Speaker 00

  Right. I love that. That's one of the evils.

• 32:24Speaker 03

  Yeah.

• 32:24Speaker 00

  Yeah.

• 32:24Speaker 03

  Olympia, it's funny, when we released Sora one, everyone just like went off on Twitter with like all these Olympics proms. You saw limbs flying everywhere. Yeah, we got, we got roasted. So clearly, you know, it's justified by the model intelligence boost, but also in terms of the usability side, right? GPT 3.5 ushered in Chat GPT. That was the first time that these models became really valuable, just a huge swath of society. And that's what we're seeing with Sora as well.

• 32:48Speaker 00

  Yeah. I guess like, did that jump from GPT one to 3.5 like surprise you? Or did it feel, was it like the people in the know were like, yeah, if we keep scaling the way we're scaling, this is going to happen.

• 32:57Speaker 03

  Yeah. I mean, we knew we were on to something, you know, when you again see just this like rapid improvement in how the model understands physics, right? It's very visceral to see, you know, these gymnastics routines start working or. For you, when a glass gets dropped, it actually shatters, right? You know, you know, you're on to something. We just see the videos, right? Like seeing really is believing. So we knew it was not going to be as slow of a ramp as it was for the LMS. I mean, it's funny to call that slow, but you know, we're really accelerating timelines here. So we knew that it would eclipse the pace of progress in LMS. We weren't sure by how much. And so I, you know, I think we weren't sure, will we end up okay, like three, 3.5? We ended up kind of being on the upper end of expectations.

• 33:38Speaker 00

  Yeah. And what in your mind is like the jump from 3.5 to four? What are the, the kind of like unsolved problems or things that you're, you know, that you think need to get better in these models?

• 33:45Speaker 03

  I think the next like breakthrough capability in video is really going to be simulation of processes that need to last for hours or even longer on end. So we're really excited about how Sora will eventually be used for knowledge work or even for biology, physics, research, et cetera. And if you think about what it takes to simulate a wet lab, right? We really need a lot of improvements on the modeling side. You need to be able to run these models for days, weeks, years on end. And that requires solving a lot of very fundamental problems with generative modeling.

• 34:16Speaker 00

  I feel like this is one of the big questions. I mean, maybe just take robotics as an example. I feel like there's this big question of like, how much simulation data alone gets you? And like, I think for a while, it felt like maybe it could get you a bunch around locomotion, but not like manipulation or some of the more sophisticated tasks. The video models have gotten a lot better since some of that early thinking. What's your latest, you know, bet on, you know, how far just using these video models gets us in some of these domains where it felt like we'd have to build up massive real world data collection operations?

• 34:40Speaker 03

  I think they're going to be instrumental for making progress here, right? And to your point, one of the really core issues with robotics, right? It's just hard to get a large data set of trajectories, which is useful for pre-training. And video, video models clearly, right? Like deeply understand things like locomotion, like dexterity related tasks. And so we're very bullish on the ability to repurpose models like Sora for these things. You know, I think I'm very bitter lesson pill as, you know, many people are at OpenAI. And I really believe that some of the older kind of lines of research, which were built on very early video models, but didn't necessarily bear fruit, will start to be successful just with these increases in base model intelligence.

• 35:17Speaker 00

  Are you going to basically need one of these like state of the art video models to even do work in like bio robotics, material sciences?

• 35:22Speaker 03

  I think it will increasingly be the case, yeah.

• 35:24Speaker 00

  So I guess like these, these next frontiers, it seems like for you are really around like longer running videos and the ability to kind of keep. Like true to physics in the world.

• 35:32Speaker 03

  Yeah, it's really in pursuit of this world simulation goal, right? Like, how do we make it so Sora is not just kind of like a video generation system, but really deeply understands, you know, every bit of reality and is useful for tasks that go well beyond just kind of entertainment? I think entertainment is an awesome thing that people get tons of utility out of with these models, as we're seeing with the Sora app right now. But I really think this is just phase one for video. And this is, this is kind of the simplest thing we can build right now with the current capability of models and their usefulness will just skyrocket over the coming years.

• 36:02Speaker 00

  Yeah. Is there an eval that you have in mind that you're like, God, when video models can do this, that'll be?

• 36:06Speaker 03

  Maybe GDP eval eventually, you know? It's a good question. Yeah. You know, I think the first scientific breakthrough that comes as a result of simulating some phenomenon in video models will be an insane milestone, right? I mean, that will just really open the floodgates. And we've been debating on the team, like, what will that thing be? High air of ours, of course. I think something related to classical physics is very likely. Video is just a good modality for understanding, you know, physical phenomenon that are well represented with like observational data, right? Things like turbulence, et cetera. So I think that will be an insane moment. I guess it's not necessarily an eval, but certainly like a milestone that will really mark kind of like the beginning of a new era.

• 36:45Speaker 00

  Yeah. And because I'm a shameless podcaster, I have to ask you, you have any prediction of when that might be?

• 36:49Speaker 03

  Oh, man. I'm really bullish on timelines. I would be shocked if, if we are not making breakthroughs like this by early 2028.

• 36:59Speaker 00

  Yeah. And is it like, I mean, is somebody L M world where it feels like we have some scaling laws that we just need to follow? Or are there like other, you know, I guess on the, the mix in the L M world of like data algorithmic breakthroughs and just more GPUs? Like, is it a similar way to think about things in the video world? Or how should people adjust their mental models?

• 37:15Speaker 03

  I mean, I think there's lots of axes along which you can make progress. Certainly scaling is a big one. You know, I think when you begin to think about how do we do these kinds of like years long simulation rollouts, that maybe requires some new breakthroughs and you cannot just kind of port the existing techniques directly over that. Maybe you can, and maybe it'll be successful. But when we really, you know, get to the point where we need to build this alternate reality, it remembers, you know, every, you know, character on your piece of paper right there, kind of the details of, of the fabric of your jacket. That seems like it might need something kind of new. So lots of green pastures to explore.

• 37:49Speaker 00

  How do you do evals on on video models now? Like, what? How do you know when they're getting better? And, you know?

• 37:55Speaker 03

  How do you guys do it on the product side?

• 37:57Speaker 00

  Do you have any go to go to things you try?

• 37:59Speaker 02

  Some survives, but but we actually, I mean, there's something, an area where we've matured a lot. I think especially from Sora one to Sora two, just learning how, you know, how high leverage good evals are. And then actually having a product now with with real use cases. And before we launched, having conviction in the in the use cases lets us build, you know, like, basic product evals, right? Like, hey, we're changing the model or some part of the stack in XYZ. Let's run, you know, the top Sora one prompts through Sora two and get a sense of, you know, the diff. And now that we actually have, like, production use cases, just like, like Cameo is a great one. Like, anytime we make a change, we want to understand how does it impact like one of these core use cases?

• 38:42Speaker 00

  Yeah, I'm curious with like with video models. Obviously, you could imagine if you like, you know, averaged out the preferences of everyone, you may end up with like a personality list model that like is is perfect to no one. And I wonder, over time, do you think there end up being like a bunch of different models with different aesthetics and different like use cases? Or does it kind of converge on a single model that you can kind of steer to your like way of liking things?

• 39:02Speaker 01

  I think there's actually analogy here to recommender systems. It fits pretty nicely, which is like when people think of recommender systems, not everyone necessarily thinks through their feed very often, but like, and let's just say I just tasked you with like designing a feed. Your first thing you might think of is just like, oh, let me sort by popularity or like sort by something like that. And actually, you find that that is exactly you're the following, this regression to the mean kind of thing where it's actually interesting to nobody because it's just just a trend that is only globally popular, which is not suiting your own personal preference. And so what the magic is you introduce personalization in different forms. And that can be, there's lots of different ways of doing this. But the obvious one is just look at your history and see what you're doing, going to do next.

• 39:45Speaker 01

  And that really does change the feeling and the results of relevance, all that sort of stuff in your feed. I can't imagine that we wouldn't have a similar phenomenon basically everywhere. I mean, we're seeing a lot in chat GPT around personalization being a very important thing, the way the model talks to you. I wouldn't necessarily, I mean, I'm not going to speak to video models directly, but in recommender systems, you learn very quickly that you don't want bespoke models for everybody.

• 40:10Speaker 02

  Right.

• 40:10Speaker 01

  It's an infrastructural nightmare. It's also like not really in the pursuit of AGI in many ways. You can kind of do it, but it doesn't really scale. It also gets very hard to reason about. And much better is to leverage the wisdom of the crowds where you find analogies between people. You do collaborative filtering and so it helps. Actually, the scale benefits you in this beautiful way where it's like, I know that this person is like videos that are similar in the past and I'm similar to that person. And so now therefore I'm going to like very likely enjoy this video as well or be inspired to be creative from this video.

• 40:40Speaker 02

  Yeah, I mean, I think models that perceive the world the way humans do like, if you measure intelligence in that way, diversity and understanding different stylistic behavior baked into the base model is is an important part of that intelligence. So that was the craziest thing about Sora two to me, not the craziest thing, but one of the most amazing things. There's definitely mode collapse and I think almost all the other video models out there today, even the ones that are great at like cinematic content and physics. Sora feels like it's great at that and great at like a ring, you know, a doorbell footage. It's great at like an interview, like a podcast.

• 41:22Speaker 02

  It's great at all these different kinds of things that feel very anime, very stylistically and different on sort of like, just like, I don't know, the the range of the model is amazing. And I, yeah, I only anticipate that we we lean into that.

• 41:36Speaker 00

  Yeah. When you guys have obviously spoiled people on the LM side where it's just like the the default expectation is whatever you release is like 100 times cheaper. And, you know, in in six, 12 months, like, should we expect a similar thing on the video side?

• 41:46Speaker 03

  100%. And even if you rewind the clock to February 2024, when we showed off Sora one to the world for the first time, that model cost him like $50 and compute to sample like a 720p, you know, short video from. And if you look at like our API pricing for Sora two, right? It's like cents on the dollar compared to that. So already we're seeing orders of magnitude decreases in cost with vastly higher model intelligence with this release. And that trend will definitely continue.

• 42:13Speaker 00

  You wanted to break big. News on this podcast. So 30 minutes before we started, you tweeted that you were going to introduce some pricing, which I guess entirely reasonable. I think when you get 30 a day and then you have to start paying a little bit for for them. And I don't think the internet rioted from what at least I saw. That seems like a natural first step in monetizing. I know you guys have talked before about like both, you know, monetization for the inference cost of running these models, but also monetization to figure out how to incentivize rights holders and all sorts of folks to get involved.

• 42:37Speaker 03

  Yeah, I mean, we really want to create an ecosystem where everyone is benefiting, right? Like, we need to pay the GPU bills for Sora. We want, you know, even new creators who are coming up on this platform, right? They didn't necessarily have a following on TikTok or IG. We want them to have a path to success in making money. And then we want the rights holders who have, you know, this incredible library of characters, who so many people will just love using to also benefit. So when we were thinking about the initial ways to monetize, you know, we're going into this learning a lot every day. We want to kind of take it slow and, you know, we want to make sure we're checking the boxes on the most important categories here, which is, you know, giving a path to monetization to all of these different people within our ecosystem. And we really viewed credits as being kind of the primary way to do that,

• 43:23Speaker 03

  without necessarily over committing to a model that we don't know we can support long term or just like need to pull the plug on eventually. So we're, we don't think this will necessarily be the final way that we ultimately monetize Sora. We're really open minded in this. We're trying to be as transparent as possible and we kind of like make these decisions and just hear what everybody across the board has to say about it. Because we want to land this in a spot, right? Where, you know, certainly not OpenAI is like the only player like making money off of this, but everybody is benefiting. I think that's really important to the success of the platform long term, right? Like, this should be somebody's full time job if they're a creator and they want to go viral on Sora. We should have a way to really make that possible for them.

• 43:59Speaker 00

  Any other like types of pricing models you guys have riffed around or thought about?

• 44:02Speaker 02

  Nothing that's like super tangible in the short term. But again, going back to, you know, how you can just totally rethink. Is it like branding and the ways you might want to bring something that you sponsor to the table now that you have generative video like, you know, right now, if you're scrolling an Instagram feed, you get a new video for an advertisement. But, you know, what if you're a creator and you're okay with all the inanimate objects in your video being certain brands or something like that and you can auction off two certain brands? Like there's, there's like some wacky ideas out there that I feel like are, um, yeah, we have, like.

• 44:40Speaker 01

  I think other thing that is really interesting about this platform and different, and I'm sure it'll grow over time. Um, like I'm using myself as an example. Like I made my cameo public as one of the first people, first early adopters that did because we were on the team. Uh, and I don't know, I think I have like 17,000 cameo appearances now. And like I, if I sum the view count over those cameo appearances.

• 45:01Speaker 02

  It's probably insane.

• 45:02Speaker 01

  I probably have never had more distribution in my entire life.

• 45:04Speaker 02

  Yeah.

• 45:04Speaker 01

  I mean, I'm a nobody.

• 45:07Speaker 00

  My follower count is like creeping up to my one on Twitter. It's quite crazy.

• 45:12Speaker 01

  That is crazy. That is crazy. But, and then you multiply that by the cameo count and the viewing that, like actually there's a crazy amount of reach for that one individual that you don't, it's like kind of an impossible level of reach on other platforms because they have to, you know, you have to create it yourself necessarily. But I actually like my, like, I like seeing what I do every day. I change my cameo instructions. I'm like, oh, I've got a cool shirt on now. Musical. And I was like, when would all the musical numbers? That's just me looking at my own. Like I'm also seeing what, you know, other people are doing. So I think, um, it's a very interesting thing. I don't think we have a great analogy for this type of, um, uh, format. There's a lot of new medium things. But it's just interesting to think that reach in a different way. It's like, it's not just what you post necessarily.

• 45:56Speaker 00

  And one thing I think that's cool about the product too is that it's like global. And I wonder if like you've seen any cool, like different variations and ways people use it like across, uh, across geographies.

• 46:05Speaker 01

  So we just did the launch to some Southeast Asian countries yesterday night.

• 46:10Speaker 02

  Yeah, literally last night.

• 46:11Speaker 01

  Hot off the press. It was hot off the press, so that's out there. We originally launched to US, Canada, and then opened it up to Korea and Japan. And there's definitely a huge different flavor of all kinds of

• 46:24Speaker 02

  Totally.

• 46:24Speaker 01

  creation, comfortableness with cameos, comfortable with the character cameos. It's been inspiring to see. Sometimes people complain about the cross-cultural content. I love it.

• 46:33Speaker 02

  Yeah.

• 46:34Speaker 01

  It's really, everything just has a different flavor. The Japanese creators that we all love and follow.

• 46:41Speaker 02

  Kaigumatsmara.

• 46:43Speaker 01

  There we go.

• 46:43Speaker 02

  Yeah.

• 46:43Speaker 01

  Kaigumatsmara. Yeah, just like, it's like, very aesthetic. It has a very different feel than anything I've seen from the US. So, yeah, I'm sure there's going to be more to come there as well. And I like it when I can learn something from the gens as well. And one thing I saw was somebody talking the Toronto accent. And I was, I'm from north of Toronto. I didn't know there was a Toronto accent. As soon as I saw that, I was like, that is exactly how everybody talks. And I immediately opened Wikipedia and there's a huge article on the Toronto dialect or whatever they call it. And I was like, I've just learned something. That's very interesting. Which actually was my experience with TikTok too. Like, I'm actually a big TikTok user and fan.

• 47:28Speaker 01

  And so many things I've learned from that where somebody is like, psychological things about myself, where, you know, they're describing some behavior, some attachment theory thing. I'm like, wait a minute, that's me. Let me go look that up on Wikipedia. So I think we'll see a lot more of that. And I think cross-cultural type of stuff is also there. Also, when you see a remix chain, it's so fun to see all the different countries that are participating in that. Always a different flavor.

• 47:52Speaker 02

  Totally.

• 47:53Speaker 01

  The rap is a different, you know.

• 47:54Speaker 02

  Yeah.

• 47:54Speaker 01

  It's slightly different.

• 47:56Speaker 02

  Bollywood rap, yeah.

• 47:56Speaker 01

  I need a translate button, but it's like, it's so close, yeah.

• 47:59Speaker 02

  The latest feed, I love, it's like my favorite spotlight into what people are doing on the app. It's just like, literally the stream of what's happening. One of the most, to riff off of what Tom's talking about in terms of like, learning things, just learning what What people want to, what scenes they want to see themselves in and their friends in. Like it's, it's like just so interesting, the kind of things people are memeing about. I'm like, I don't know, like it's mundane to one person, but then if you, if you're in that mode of like, what do people care and laugh about? There's so many hilarious things people obsessed with. I was talking to a user today that's obsessed with cranes. She was like, I just love visualizing myself on top of cranes.

• 48:39Speaker 02

  It's just me.

• 48:40Speaker 00

  I mean, obviously you guys have been super thoughtful on the moderation side too. And I don't know whether like, I think both encouraging like a lot of fun content, but also I think have done a pretty good job of like navigating those waters. Like, was that an iterative process or do you just like nail it from step one? Like, you know, talk a bit about that.

• 48:55Speaker 02

  Oh my God. We've had long sleepless nights, um, getting to the point where we are now and still, you know, still a lot of work. Shout out to amazing, both people working on safety on the Sora team and like a ton of work has happened at OpenAI on, on our moderation systems, moderation models. We have reasoning models that we've written about that are in our safety stacks, which are amazing. And so we really, it's a great example of using our technology to make our products better. Uh, and like we were talking about before, uh, we, we are able to do a lot more with a smaller team now on this like moderation and safety front. But thank you for saying it's in a good place. That's not what people on Twitter tell us, but.

• 49:33Speaker 00

  Twitter angry about something? I can't, I can't.

• 49:35Speaker 02

  There's, you know, the content moderation failures are, are understandably frustrating because you know, we're trying to toe this line between user freedom, but blocking all the bad stuff and we're getting, we're getting better every day. We're in new territory, especially with cameos. We want to be really sensitive to, um, you feeling like you're in control and you know, having you're comfortable on this network, right? So with that comes a lot of thought in the gargles. Yeah.

• 50:00Speaker 00

  Yeah. You've kind of talked about this being a new social platform. I think, you know, Sam's spoken very publicly about like that being interesting surface area for OpenAI to explore. You talked about maybe this integrating with chat GPT over time in some way. Like, how do you see other like angles of things that are being worked on in OpenAI? Like one day, you know, co, uh, you know, colliding with us.

• 50:19Speaker 02

  I mean, Bill mentioned, you know, going beyond entertainment. You can like chat GPT is, is like sort of your super assistant. That's literally what we called it early days. Why, why shouldn't it be able to respond with a, with a really helpful video? I think that'll be really killer, uh, once we, we get that kind of stuff in. But yeah, it's like, I think the world's our oyster with ways that like all of our products could interact. I mean, we just released a browser.

• 50:46Speaker 02

  I don't know, you can imagine using a browser and having a little video assistant or something talking to you on the side. Would your agent help me book this flight? I don't know, those crazy, wacky ideas I've heard out there.

• 50:55Speaker 01

  Yeah, and I think a lot of this stuff builds off each other. Like I say, the reasoning models, which, you know, in the pursuit of AGI, which are justified. Is, you know, I, my first use case wouldn't be a moderation stack.

• 51:06Speaker 02

  Right.

• 51:07Speaker 01

  It's like, wait a second, that actually is a perfect use case. And so a lot of this stuff builds on each other. The ecosystem itself, I do think like, chat GPT does feel a little different, a little bit sacred.

• 51:15Speaker 02

  Totally.

• 51:15Speaker 01

  Doesn't mean it can't change over time, or like, there's definitely things that we can do there. But, you know, it is, tends to be a more utility-driven use case. And, you know, mixing entertainment and a utility-driven use case doesn't always work. So I don't think it's just like, jam it in there and it will just magically work.

• 51:29Speaker 02

  Right.

• 51:29Speaker 01

  I think we went on to do that very thoughtfully. But certainly, as a, as a utility use case, these models are going to discover new modes of turbulence modeling. So there's clearly utility.

• 51:40Speaker 02

  Yeah.

• 51:41Speaker 01

  I heard, I heard 2028.

• 51:43Speaker 02

  Early 2028. Not even, not even, not even by the second.

• 51:47Speaker 01

  And so you should be able to ask chat GPT about that. And, you know, video as a modality is, you've seen on YouTube, like, half the videos on YouTube are like how-to videos. That's something that we want to see modeled. But in my case, you know, here's my toilet fix, figure out how I like adjust it and like fix this problem with it, or whatever, whatever it is. So I think that can be very naturally connected.

• 52:10Speaker 02

  Yeah. Before we move to our quickfire round, I, Bill, I did want to ask you, you know, a shameless question, which is, maybe what's the simplest way to explain how these models learn physics?

• 52:19Speaker 03

  It's a great question. At a high level, right, these models are always doing prediction tasks. So in the case of, for example, diffusion models, you're getting a video, which we've artificially added a lot of noise to. And the goal of the neural network is to predict the underlying signal that we kind of obscured with all of that noise. L M's also a prediction task, right? You're predicting the next token condition on all the prior tokens. We think actually both L M's and Sora learn world models. They learn different flavors of world models. But fundamentally, right? If you're doing this prediction task, whether it's predicting signal or whether it's predicting the next token, you need to develop some understanding of kind of the dynamics of the world, right? So in the case of L M's, if I'm generating like a rhyme or something, or maybe a poem, then it's very useful for me to have some kind of internal knowledge, right?

• 53:04Speaker 03

  About the structure of poems, about the linguistic elements of them that I just pick up from data, right? Because if I haven't Grok these things, then it's very hard for me to predict the next token. And the analogy holds for video as well, right? So if I'm a video model and the prompt is a guy playing basketball, I sure better have learned, right? How basketballs dribble, how light refracts through a scene, and all these just little components of physics in order to paint this broader picture of like the entire scene. And so it's basically an emergent property from lots of compute and lots of data, which is like a very, you know, banal thing to say these days. Everyone gets it. Their lesson works. But fundamentally, right? If you don't have these kind of internal models of how the world functions, then you will always be a worse model and thus have a higher loss than one that does.

• 53:47Speaker 03

  And so that is kind of how we think about the optimization pressure, which makes physics emergent from large-scale video pre-training.

• 53:53Speaker 00

  In the L M world, there's, you know, we went from just like the entire internet to now like there's, you know, incredibly valuable data that is like PhDs, you know, doing different things. Is there like an equivalent in the video world of like, oh, that is like a really complex, you know, piece of physics that like is great for the model to see or like?

• 54:06Speaker 03

  Right, right. I mean, if you take any like video researcher and show them a bunch of videos and ask them which is their favorite, chances are, you know, if there's a gymnastics routine going on, they're gonna be like, that's the one. That's the one. But I think it's a really interesting question and, you know, it's, it's somewhat hard to answer because the way intelligence manifests in videos, I feel is very different than the way it manifests in language. You could, for example, imagine there is a lecture video, right? Where someone is giving a, you know, a course on calculus. And that clearly has an intelligence which is of the flavor of text, right? Where it's teaching you about some like deep concept about the world, mathematics, physics, what have you. On the other hand, though, when you think about these gymnastics routines, right? There's no intellect. Actual intelligence there beyond kind of the planning involved of doing like a gymnastics

• 54:50Speaker 03

  routine, but there is so much interesting detail about the the small collisions that happen as part of that routine. Again, about all of the people in the background shuffling amongst themselves, like building a model that can actually simulate not just the gymnast, right, but every single person in the broader scene. And so, you know, we're still trying to figure out what kind of data really makes for an amazing video model. But there's a lot to it. That's very multimodal in a way, right? Multimodal in the sense of kind of all these different pockets of intelligence, which exist in video, but not necessarily in other modalities like text.

• 55:20Speaker 00

  Well, I always like to end our interviews with a standard quickfire round where I stuff some overly broad questions into the very end. And so maybe to start, I'm curious for each of you, one thing you've changed your mind on in the AI world in the last year?

• 55:31Speaker 02

  I think my timelines for some things have accelerated for other things have delayed. I think like we, this is a common one that people say, but I just think we, we really overestimate estimate consumers and adoption and how people learn this technology. And we could be way ahead of like the general world in terms of the actual science, but way behind in terms of like the way we build a product interface that's acceptable, accessible and how we teach the world about it. And then when you get into like actual enterprise use cases, there's like regulatory sort of just like, yeah, things you have to sort of battle through to get people to actually adopt this technology. So all of that, you know, working as like a

• 56:13Speaker 02

  product person, even on the consumer product, just seeing like the crazy amount of sort of work that happens beyond behind the scenes to make things possible. Yeah, there's just a lot to do for people to actually adopt these things.

• 56:25Speaker 03

  The aspect of AI I've updated most on is if you asked me a year ago,

• 56:31Speaker 03

  would I care if I watch, you know, some blockbuster movie versus say some movie generated by a model that's, you know, like Sora three caliber or something? I would say the Sora three movie would be just as good as the blockbuster one. I don't need any information about it, like assuming it's, you know, Grok's cinematic content and like interesting plots like that's sufficient. Interestingly, though, I feel like the more I've interacted with these models, I actually do feel a hollowness when there's not clear creative intent behind it, which is something I would not have expected to have felt a priori in part because, you know, like, you know, I enjoy making the models. So, you know, it feels like weird for me to have that opinion. But I really have been struck by how much more interesting and compelling.

• 57:17Speaker 03

  Generations feel, even with Cameo, of course, right, seeing the people you know in them feels amazing. Seeing your pet rocket, uh, my dog, who's now, as of yesterday, available in Sora.

• 57:28Speaker 03

  But even knowing someone, uh, I think Thomas was mentioning this earlier, even knowing someone just had the creative intent, right, to reject a sample, right? Or like really iterate on a prompt and the plot that they're actually using and the underlying generation has some kind of deeper connection to them emotionally, right? You actually do kind of feel that in a way that is very surprising to me. Uh, and I'm not sure exactly what triggered this change in me. Like I, again, I would have been like very convinced by these completely AI-generated feeds that our bot made or something, as long as the bot was good. But I really now like do not actually think even I would find that compelling, which is quite interesting to me. And I'm not sure exactly actually why this is the case, but there's some, you know, there's still some like morsel of humanity,

• 58:09Speaker 03

  which is like very important to communicate through these gens for, for content to actually feel meaningful in some way. Uh, so I've really updated on that.

• 58:16Speaker 00

  If you guys all, you know, left OpenAI tomorrow and had to build things on top of the Sora API, uh, what would you go build?

• 58:22Speaker 01

  Game.

• 58:24Speaker 00

  Like a game where you put yourself in a video?

• 58:27Speaker 01

  I'm sorry, I can't provide more details.

• 58:33Speaker 00

  Here's my very well thought out business plan.

• 58:40Speaker 01

  I think there's, um, I am just, I grew up in gaming. I learned to program in gaming. Uh, I can't imagine how much crazy opportunity there is in that space. Um, and the idea that a lot of the things are very difficult, at least for me, when I was building games, uh, was like, oh, I have to go and, source art or build the art myself. And that was a labor, labor intensive process. Learn the tools to do all that sort of stuff. Um, I don't think this is exactly what. Needs to manifest, but like what kind of gaming models are enabled by this generative technology in the same way that like we're talking about Cameo, is that something uniquely enabled here is like what kind of things are out there. So I'm excited about that and seeing what people can come up with in that space.

• 59:22Speaker 02

  Totally. My answer is like, I'll get a similar answer from a different angle, which is just that I think, I think the most, I mean, it probably comes through in the product it built and how we're talking today, but the most interesting thing about this technology is how it will transform and create new mediums rather than just, you know, like an AI film. I think it's like the least interesting concept to me. There's people doing cool things with film, but like completely new things are most interesting. And so I think interactive, which is like maybe gaming from Thomas's point of view, from my point of view, maybe it's interactive storytelling of some sort. And no one's really nailed this, like, you know, there's choose your own adventure books. I think those were pretty popular and like some movies like Bandersnatch on Netflix or kind of like, but I think pushing that

• 1:00:07Speaker 02

  idea like a lot farther of some like, you know, collaborative piece of, piece of creative lore or something like that.

• 1:00:17Speaker 00

  Why don't you think there have been more consumer product, like fun consumer products built on top of these models? Is it just that like they weren't good enough until now? Or?

• 1:00:24Speaker 02

  I think it's really, one, it's really frigging hard to build a consumer product. A good consumer product is like one of the hardest things combined with the technology being so new. And, you know, it's unclear what, what actually sticks. I think, I think that's actually still quite a hard combination of things to nail. Yeah.

• 1:00:43Speaker 00

  I mean, you've just done it twice at the top. So I guess still, still difficult. What about something you'd build on top of API?

• 1:00:49Speaker 02

  Oh, man.

• 1:00:50Speaker 00

  I know we can't take you away from the models, but if we have to.

• 1:00:52Speaker 02

  I would just go train new models, you know? I don't know, something probably more on the science side or robotics facing. I think there's all just a lot of cool new work to be done there.

• 1:01:03Speaker 00

  Well, I want to leave the last word to you guys. I usually say where can folks go to learn more, but I'd be genuinely shocked if any of our listeners didn't use Sora.

• 1:01:09Speaker 02

  Yeah.

• 1:01:10Speaker 00

  So maybe I don't know if there's specific research you want to point people to or like cool new product features. The mic is yours. Anywhere you want to point our listeners, please go ahead.

• 1:01:18Speaker 02

  We just launched character cameos yesterday, so you know, if you use Sora, you know, you can upload yourself, um, create videos with your friends. Now you can upload anything, you know, pets, inanimate objects. You can create, you know, your own IP from Sora and create a character out of that. I've got a pickle that's going viral, Pickleton, so. Go cameo Pickleton.

• 1:01:44Speaker 01

  I have a wild, wild one. Okay.

• 1:01:47Speaker 00

  It's an open mic, please.

• 1:01:48Speaker 01

  Open mic is, uh, in Disneyland Hong Kong.

• 1:01:53Speaker 00

  Good start, good start.

• 1:01:54Speaker 01

  There is a, uh, it's the Haunted Mansion, but it's like, I think it's called Magic Mansion. I don't know, they have a slight riff on it. And, uh, there's this lamp that it's a dark ride, so you just go through it. And there's this lamp that pours, like the whole concept is this monkey has this lamp. And when it pours out, the objects come to life. And so it pours out and then there'll be like an armor on the wall and the armor will start moving and talking to you. And, uh, we actually kind of just built that, which like you just record an object and then you say, come make it alive.

• 1:02:24Speaker 00

  Yeah.

• 1:02:24Speaker 01

  And so, uh, I think check out that ride is very cool, but also back to character cameos. It's like, actually it has that magical property.

• 1:02:32Speaker 00

  Yeah, I'm sold.

• 1:02:34Speaker 03

  Um, yeah, we've temporarily made it so you don't need invite code to get in. Come in with your friends. Sora is way more fun if you're in there with a group and, uh, yeah, check it out. Give us feedback.

• 1:02:43Speaker 00

  Yeah. Well, guys, thank you so much. This is a ton of fun. Um, I want to make sure I get you guys back to actually building Sora, uh, because the world will far more benefit from that. But, uh, seriously, this was a ton of fun.

• 1:02:52Speaker 02

  Thank you so much for having us. See you in Disneyland Hong Kong.