Seven Million Videos a Day, and Most Return Visits Are Spent Creating

Sequoia's Konstantine Buhler and Sonya Huang sat down with the three people running OpenAI's Sora. Bill Peebles heads the team. His path is the academic one: undergraduate research on video generation, more of the same at Berkeley, and Sora on day one at OpenAI. He is also the inventor of the diffusion transformer, the backbone now sitting under most competitor video models in the United States and China.

Thomas Dimson's path bent. He spent about seven years at Instagram when it was a company of roughly forty people, building early machine-learning and recommender systems — including the original ranking algorithm. He left, built Minecraft in the browser (and a product called Global Illumination that he still believes in), and OpenAI acquired the team because it looked "cracked" at product. Inside the lab he bounced across products and post-training research before landing on Sora.

Rohan Sahai has been at OpenAI about two and a half years. He started as an individual contributor on ChatGPT, saw the video-generation research, got "Sora-pilled," and now leads the product team. The conversation they came to have is larger than a generation tool. It runs from spacetime tokens and internal world simulators to a social product optimized for creation instead of consumption, and to a longer bet that society should co-evolve with simulation technology rather than meet it as a surprise.

Most people know autoregressive transformers: emit a token, condition on everything so far, repeat. Diffusion transformers use a different objective. Take a signal such as video, add a great deal of noise, and train a network to predict the noise you added. Generation is iterative denoising, not next-token sampling. In Sora 1, OpenAI popularized that recipe for video. Because the model produces the whole clip simultaneously, it sidesteps a failure that plagued earlier frame-by-frame systems: quality that drifts or collapses as the clip runs on. That inductive bias, Peebles says, is why DiTs proliferated across video stacks.

The building block is a spacetime patch — Peebles also calls it a spacetime token, and treats the two as synonymous. Language's atom is a character. Vision's atom is a little cuboid with spatial X and Y plus a temporal locale, almost voxel by voxel. In the classic DiT, every patch talks to every other patch. Full global context at every position in spacetime is a powerful property for a network to have, and it is how object permanence falls out. A ball that disappears behind a sofa is still that ball when it reappears, because the model can see the entire clip at once.

Attention is the communication layer. The Sora 1 blog post, *Video Generation Models as World Simulators*, already argued that if you patchify video into spacetime tokens and use attention properly, information transfers across the whole sequence in one shot. That is not a metaphor they adopted later. It is the frame they have been using since the first model worked.

Sonya asked whether Sora 2 is just Sora 1 plus scale. Peebles said the team spent the interval on core generative-modeling research, working from first principles: models that are extremely good at physics, and that feel intelligent in a way prior video systems did not. Older models broke on complicated physical sequences — spiking gymnastics, riding a dragon. Sora 2 fails differently. Give it a basketball star shooting a free throw. If the simulated player misses, the model does not magically guide the ball into the hoop to over-respect the prompt. Most of the time the ball rebounds off the backboard. That is a distinction between model failure and agent failure: the person Sora is implicitly simulating missed, and the laws of physics still held. He had not seen that semantic failure case in prior video models.

The agents, though, are mostly implicit from scale. Object permanence emerged in Sora 1 pretraining once a critical FLOPs threshold was hit. Push the next frontier and you see agents act more intelligently and physics get respected in ways they are not at lower compute. Internally they treat Sora 1 as video's GPT-1 moment, the first time the modality really worked, and Sora 2 as GPT-3.5: it crosses a usability barrier and mass adoption begins. A GPT-4-class jump is still required before the simulator is robust enough for science. He watches GPT-5 Pro improve convex-optimization lower bounds on Twitter and expects Sora to do the same for the physical sciences. One day a world simulator that has generalized the laws of physics well enough might let you skip the wet lab and run biology inside Sora. That system does not exist yet.

Spacetime patches are a simple, reusable representation, which is the point. Language models trained on BPE characters still grow internal world models because next-token prediction forces them to simulate. Video makes the same demand more explicitly: the data are raw observational bits of reality, whether footage of a conversation, anime, or cartoons. One network on that mix is under pressure to grok generalizable properties in a data-efficient way. The team does think about the mix. Anime is in the data because people love generating it, even though a flying dragon is probably not how you grok aerodynamics. Simplified primitives — people still locomote through scenes — may help; whether pretraining on sketches or other simplified visuals makes concepts cheaper to learn is, he says, an open scientific question.

Intelligence per bit is lower in video than in text, but the integral is much larger. Peebles finds it hard to imagine ever fully running out of video data. On embodiment, every extra 10× of compute has had remarkable things fall out with limited changes to the recipe. Physical agency would probably help collisions; he would be surprised if it made the model worse. Video-only is still striking enough that he would not be surprised if it were AGI-complete for a general-purpose world simulator. A basketball-game simulator might need only video and maybe audio. Extra modalities are unlikely to hurt; whether they add much beyond mastery of video and audio is another open question. In the lightning round, the first world-model scientific discovery is, in Peebles's bet, a better theory of turbulence; Dimson guesses a fluid-dynamics problem such as Navier-Stokes; Sahai wants whatever yields to huge numbers of simulated iterations. The last things to simulate, Peebles says, are phenomena video represents poorly — high-speed particle collisions, quantum mechanics, anything without native footage. Sahai adds smell, which almost nobody is working on.

The product story is not linear. There has been a Sora product team from the start, with Sahai leading it in the Sora 1 days, aimed mostly at a prosumer demographic. Videos without sound, Dimson says, are a different environment. In parallel, OpenAI was exploring social applications of AI. Most prototypes were quite bad. The magic showed up before ImageGen was released, in an internal social context: people took an image and built a chain of remixes. A duck, then the duck on someone's head, then everything upside down and smoking a cigarette. You cannot really do that on ordinary social media. The barrier is not the idea. It is the camera, the setup, the edit.

They tried to productize that behavior, at first away from Sora, while the research was still showing only signs of life. Then the research crossed into something people would clearly want, and they locked in around the week of July 4. Dimson disappeared into a two- or three-month sprint. The first build had no magical features — just a native video environment with audio, full screen, and quick generations. Coming from Instagram, he knew it is almost impossible to get people to create there, even though creation is the most valuable action. Remix from ImageGen still applied. Cameo was on the brainstorm list. Dimson did not think it would work. He is, he says, bad at predicting technology, and it was not obvious you could take a likeness and imagine it into video.

Early prototypes included people reacting in a corner of the frame. Then cameos started working. Internally, the day they shipped it, the feed became entirely cameos. Everything the team generated was of each other. A week later it was still all they did. At first they worried: is this good? Does anyone care about anything except cameos? They concluded it was actually good. A lot of AI video is beautiful, complicated, static, and missing a human touch. Cameo put people back in. ImageGen had already shown the same lesson with Ghibli selfies and photos with idols. Put yourself in the scene with your friends, then riff on Rohan doing something, or on Bill wrapped as an action-figure package — a clip that has been remixed thousands of times.

Dimson has something like 11,000 publicly posted generations. Sahai is a little under that. Konstantine's mother, they joked, has used Thomas's cameo for about 10,000 of them. The user mix surprised Sahai. He expected the Twitter AI crowd to dominate the feed; they dominate the press cycle he sees, not actual use. The firehose is about seven million generations a day, across a huge range of ages: motivational self-inserts, friends memeing, people cameoing public figures who opened their likeness. Hitting the top of the App Store pulled in people who were just browsing, far beyond the niche AI-film early adopters. A platform of all AI-generated content, Dimson later said, is strangely more social than a lot of social networks. You remix your friends, tag them, have an anime fight. That was not obvious in the sprint, and it is why they would not have launched a feed of AI content that did not feel human.

Dimson still sometimes has to defend the original Instagram ranking change. It was not a random decision to optimize for ads. A chronological feed guarantees every poster the top slot with all of their followers, which rewards posting constantly. As the power law got more head-heavy, high-volume accounts crowded out people you actually care about. Follow National Geographic and they post twenty times a day; your friend posts one coffee photo you never scroll down to. The solutions are unfollow, or permute the feed. Instagram tested the second path. Early results were unambiguous: people saw more of who they cared about, the experience improved, and creation actually went up because the content felt more accessible. Over time the thing can go astray. Open the firehose of unconnected content, add ad pressure, and the company incentive is blind consumption — cheaper content, more scrolling, a more mindless mode that can make people create less.

Sora's stated intent is the opposite, because the technology makes everybody a creator. The feed should be optimized to inspire you to make something. When other people hear "inspiration" they picture an elegant, beautiful scene. Dimson pictures meme culture: something funny enough that you have to riff on it. That is a different brain mode while browsing. Incentives come first. They will not get every edge case right; a recommender with no history, on a platform still evolving, is a living system. Sahai said the intent is working. Almost 100% of people who get past the invite code create on day one. When they come back, about 70% of the time they are creating. About 30% post into the feed, not just generate for themselves.

Beyond ranking, they try to knock people out of a consumption flow state. Sahai cited an article on the curvilinear nature of casinos: no windows, walk in a circle, never make a decision. Sora tries not to do that. An in-feed unit might note that you just watched a few videos in a domain and ask whether you want to create one. They actually shipped things like that. Dimson was never that afraid of competitors during the sprint, because they were stacking decisions that are obvious only in retrospect: cameos, then multiplayer cameos rather than a one-player onboarding tunnel, then the anime-fight version of hanging out. Each layer was the human element. That, not a prettier generator, is the product.

The app and the API currently expose the model in the same state. The API exists so OpenAI does not have to build a thousand interfaces for long-tail uses while still chasing a ChatGPT-scale consumer audience. In the Sora 1 days the team talked to studios that wanted the model in a specific part of their stack, in a specific way. So far that looks like filmmaking tools and at least one company doing something with CAD. Gaming is the use Dimson cannot stop thinking about. He would never bet against people finding a creative way around latency. The games that excite him are not the obvious real-time clones. Infinite Craft, a web game that starts with fire, water, earth, and air and lets an LLM derive the crafting tree — volcano, underwater volcano, eventually Godzilla — is a process of discovery. His philosophical view of language and video models is the same: everything is in the weights, and the prompt is a secret code that unlocks it. The best moments in games were always discoveries the designer enabled but did not place. Early GPT-3 text games had that DNA and then went off the rails; Sora, he thinks, still has some of that exploratory feel.

On filmmaking, Peebles wants both the long tail and the head. Creatives who went all-in on DALL-E 1 and 2 helped steer the lab, and Sora will keep building for power users. At its best, though, AI is a democratizing tool. When a beautiful anime prompt hits the top of the feed, anyone can remix it and learn from people who already know how to get the most out of the model. Feature-length work will arrive in stages. Daniel Freiden on the Sora team made the launch video by himself in days, and has kept putting short stories on OpenAI's Twitter. Filmmaking latency is already collapsing. Letting any kid at home do the same is an economics problem: video is the most compute-intensive modality, the app is free today, and paid access is how they will scale it. There will be a lot of bad movies. There is also, he thinks, a great director still in a parents' house in high school, who never had the equipment or the crew.

Sahai does not think the equivalent of a feature film will look like today's features. When recording cameras arrived, the first thing people did was film plays — the least interesting use, a way to skip traveling theater — until someone realized you could shoot in the world. Sora is in that first inning, especially as latency and length constraints change. Konstantine offered a film-history rhyme: an early motion picture settled a bet about whether a galloping horse lifts all four legs, a scientific question the new medium could answer. Dimson's recommender-systems lesson is that content is also a social phenomenon. When anyone can create, attention becomes the scarce resource, which should elevate quality — the inverse of today's world, where creation is scarce and consumption is not.

Rights holders are in active conversation. Peebles says the pitch is a new frontier: every kid using beloved IP in a personal, custom way, with monetization flowing back. The details of character cameos are being scoped from scratch, a new creator economy they are inventing with open minds and fast iteration. Nearer-term, pet cameos are a committed feature. Bill's dog Rocket was an internal test. Dimson recorded two seconds of a clock his father got from Veritas, told Sora it was a walking clock from an old *Simpsons* earworm, and generated a talking clock that played opposite Bill. "Who needs Pokémon when you can have a walking clock?" Emerging IP, in other words, does not have to come from a studio.

The app looks like a social network. Internally they view Cameo as the lowest-bandwidth way to give Sora information about you: appearance, voice. That bandwidth should increase until the model understands your relationships, how you have grown up, and can function as a digital clone. The Sora app becomes a mini alternate reality running on your phone. Versions of you go off, interact with other people's clones, do knowledge work, and report back. Entertainment is the on-ramp; the long-term object is a platform aligned with world-simulation capabilities. Sonya summarized: so you are building the multiverse. Peebles said actually kind of, yes.

That is why OpenAI, across the board, cares about iterative deployment. Peebles does not want to drop bombshells on the world when a research breakthrough lands. He wants society to co-evolve with the technology. Sora 2 is the GPT-3.5 moment for video: make the world aware of what is possible now, and start getting comfortable with the rules of the road for a future in which copies of you run around in the ether doing tasks. Asked for a probability that we are simulated, he put himself at a solid 60%, more likely than not. Dimson was low; Sahai was with Peebles. Theoretical limits include whether you can simulate a GPU cluster inside Sora, and well-defined bounds on how much computation you can run in a system given the compute you are actually spending. He has not thought deeply enough, he said, but the existential questions are real — which, Sonya noted, is why his simulation prior is so high.

Lightning-round favorites sketched the culture already forming. Dimson's first night with cameos open filled with Chinese factory-tour remixes; he is the chili-factory ribbon-cutting guy, zero likes, one like from himself. Sahai likes wholesome friend videos that never go mega-viral, and a Mark Cuban-and-Jorts clip. Peebles would put Sam Altman's GPU-themed K-pop dance on a Spotify playlist if the full song existed. On whether a fully AI work wins an Oscar, Grammy, or Emmy first, they leaned Oscar for a short, and wondered whether it has already happened without anyone noticing. Konstantine wants the long tail of history — stories locked up by production cost — unlocked the way *The Last Duel* eventually was. The through-line of the hour is the same as the opening clip: do not wait for the research bomb. Put a world simulator in people's hands while it is still a cameo and a remix, and let the rules get written in public.