Why Sora Bet on People, Not Just Pixels

Jacob Effron asks whether Sora's launch mirrored ChatGPT's unexpected breakout. Bill says he did not expect the app to stay number one in the App Store for about a month, crediting the research team and OpenAI's knack for viral moments, while acknowledging the reaction surpassed his expectations. Rohan jokes that Bill has a hopeful optimism and recalls a meme in which Bill demanded the top spot after Sora reached number three. The team frames that optimism as useful, not naive, because consumer products benefit from an ambitious owner who believes the product can win.

The host asks how OpenAI reserves GPU capacity for a new product like Sora. The team explains that capacity planning is not a science and that a compute-intensive video surface requires some pain elsewhere in the company. It credits OpenAI's culture with making teams feel invested in one another's success, whether the win is a new visual feature in ChatGPT or a video product that must become the category leader. The launch therefore depended as much on organizational tradeoffs as on the model itself.

The team says the idea of launching Sora as an independent app was not obvious from the start. It had a clear ambition to transform user-generated content and media, and early experiments showed a multiplayer mode with emergent behavior before Sora 2. Image generation had already shown people wanted to cast themselves into imaginative scenes, so the team asked what would happen if that act became collaborative. ChatGPT, by contrast, still felt single-player and even sacred, with users needing to know whether an action was public.

One team member describes a long product journey with many wacky prototypes. One prototype put a pure social-media stream inside ChatGPT, using threaded, Reddit-style chains in which users would iterate on an image—a duck gets a cigar, then turns upside down, then turns red. Those meme chains revealed a kind of remixing that is only possible with generative AI because creating and riffing are otherwise too hard. When the Sora models became strong enough, the social format fit them like a glove.

The team recalls that Cameo was not obvious in advance, even though it now looks inevitable. An engineer named Bobo asked teammates to send videos of themselves saying 'Hey Sora, it's Phil,' uploaded them to the backend, and let people tag one another. Within days, the team's feed was full of cameos and they briefly wondered whether something had gone wrong. They concluded that the shift was exactly what they wanted: a personal, human form of generation that made the app feel alive.

Ben Thompson initially argued that most people want to consume rather than create, a skepticism the Sora team took seriously. Rohan says the app was designed from the ground up around creation to avoid a lonely, detached slop feed optimized only for watching. Cameo was the most important countermeasure because it makes generation feel personal and human rather than like anonymous text-to-video. Thomas also led pioneering work on a recommender stack oriented toward creativity instead of maximal clickbait.

Rohan explains that a creation-oriented recommender can be self-fulfilling in a healthy way. If users can participate in the network, then remixing or making a generation becomes an active behavior that the system can encourage. At Instagram, it was nearly impossible to trace a new photo back to the original piece of content that inspired it; in Sora, the system can capture signal from every remix and interaction. That changes what the recommender can optimize for.

Rohan says Thompson's later shift made sense because a cold-start social network is hard and his early experience was that of a pure consumer before the feed had time to bake. The team still points to Thompson's own follow-up insight: even a 1% change in the ratio of creators to consumers would have a massive impact on the world. Thomas adds that a human creator versus a bot creator is a foundational difference, because a person's decision to post or remix acts as a stamp of approval. A feed stripped of that human signal would be far less interesting.

The team says celebrities and rights holders have gone through a rapid evolution in understanding video generation. Character cameos let users generate with beloved IP, which is a huge value proposition for rights holders, but they also want input and restrictions on how their characters appear. Rohan says Sora has spoken with every facet of society since launch and walks anxious celebrities through the rails for controlling their presence. The goal is to let participation happen without turning it into a free-for-all.

OpenAI announced that it would begin piloting monetization for rights holders, prioritizing those who invested in the platform early. The team says it wants an ecosystem where everyone benefits: OpenAI must pay GPU bills, new creators need a path to earning, and rights holders with valuable character libraries should share in the upside. Mark Cuban quickly added Cost Plus Drugs to his cameo instructions, which the team cites as an early sign of how brands might use the format. Moderation still blocks many public figures from cameo creation to prevent impersonation, producing a flood of inbound from celebrities who want access.

Asked about Google and Meta releasing competing video products, Rohan says the future is bright for all these models and the industry will evolve around a new medium. He jokes that competitors may need a new 'code purple' alert because they are seeing something magical emerge. Still, he says he has not been overly impressed by some of the productization so far, because the key insight is that social networks are about people and relationships. Cameo and creative collaboration with friends were the missing piece, and that is what Sora is trying to own.

Thomas describes Sora as a cold-start social network built on a completely new medium, which is why the team gave it a familiar full-screen video form factor. The early hypothesis was that the product would be more fun with friends, and that is already reflected in the recommender even if it is not fully expressed everywhere. He expects Sora to lean further into communities, DMs, and small or large groups because shared creation can be magical. Rohan adds that recommender incentives determine outcomes, and that a creation-first network may reconnect friends rather than isolate them.

The host asks how Sora serves both casual remixers and sophisticated prosumers. Rohan says the product is democratizing because a strong generation can be remixed, giving others access to its ingredients and a way to learn prompting and cameo design. The team still wants to empower frontier creators with better tools, such as storyboard for up to 25-second videos and stitching for basic editing. The conversation compares this to GarageBand: low-friction tools can be a gateway drug to deeper creative work.

Bill says OpenAI's magic comes from sticking to an ambitious AGI-aligned research roadmap regardless of competitor releases or product pressure. As models become more capable, the product team discovers new capabilities and turns them into features, so Cameo and remixing are joint research-product achievements. Thomas argues that consumer builders need a creative lens and a willingness to fail, because simply mirroring OpenAI's roadmap is not a strategy. Even today's models can support new experiences in gaming and other surfaces if builders think outside the box.

The Sora team is shockingly small: roughly nine or ten researchers, under twenty product people, and about thirteen systems engineers, for around forty people total. That small team is one reason OpenAI launched a Sora API, so outside builders can explore applications the team cannot pursue itself. Early API experiments include Mattel prototyping toys with Sora and a CAD-to-visualization pipeline that translates CAD files into captions the model can understand. The team expects those applications to become more sophisticated over time.

Bill traces the visual generation lineage back to DALL-E 1, which showed that the general-purpose capability of language models could come to image generation. Earlier visual models were narrow, often modeling specific distributions such as faces, but DALL-E 1 made it clear that models could learn from the pixels of the internet. It then took years to get image generation under control with DALL-E 2 and DALL-E 3. Sora work began in early 2023 alongside DALL-E 3, when scaling, diffusion formulations, and architectures were becoming more principled.

Bill calls Sora 1 the GPT-1 moment for video: the first time high-resolution, consistent generation beyond a second was even conceivable, even if 60 seconds was barely possible. Sora 2 is now framed as GPT-3.5 in both capability and usability, because it can handle complex motion such as an Olympic gymnastics routine without limbs flying everywhere. The team got roasted on Twitter for earlier failures on those prompts, but the improvement reflects a real step change in model intelligence. GPT-3.5 also ushered in ChatGPT and made these models valuable to a broad swath of society, which is the analogy Sora is chasing.

The jump from Sora 1 to Sora 2 did not surprise the team in kind, because rapid improvements in physics understanding were visceral to watch. When a glass dropped in a generated video actually shattered, or a gymnastics routine held together, the researchers knew they were onto something. Bill says video progress was never going to ramp as slowly as language models, even if calling the language-model ramp slow is funny in retrospect. The team was not sure how far it would go, but Sora 2 landed at the upper end of expectations.

Bill says the next breakthrough capability in video will be simulating processes that last for hours or longer. That could make Sora useful for knowledge work, biology, physics research, or a wet lab, but it requires models that can run for days, weeks, or years and solve fundamental generative-modeling problems. In robotics, video models already understand local motion and dexterity in ways that could help with trajectory pretraining, where real-world data is hard to collect. Bill describes himself as a bitter-lesson believer and expects older video-model research to succeed as base-model intelligence rises.

Progress will come from scaling, but Bill expects long, years-long rollouts may need new breakthroughs rather than a direct port of existing techniques. The team has matured its evals from Sora 1 to Sora 2 by testing real use cases and product changes, such as running top Sora 1 prompts through Sora 2 and checking the diff. Once production use cases exist, a change to the model or stack can be evaluated against something like Cameo. Evals are therefore tied to conviction about the use cases before launch.

The host asks whether video models will converge on one steerable model or splinter into many aesthetic models. Rohan compares this to recommender systems: sorting by global popularity regresses to the mean and pleases nobody, while personalization and collaborative filtering create relevance. Bespoke models for every user would be an infrastructure nightmare and hard to reason about, so the better path is to leverage the wisdom of crowds. Bill adds that diversity of style in a base model is part of intelligence, and Sora 2's range—cinematic, doorbell footage, podcast, anime—contrasts with mode collapse in other video models.

Inference cost is falling fast. Bill recalls that Sora 1 cost about $50 in compute to sample a short 720p video in February 2024, while Sora 2 API pricing is cents on the dollar by comparison. The team expects that trend to continue, with orders-of-magnitude cost decreases alongside higher intelligence. Sora also began charging for usage after a daily allowance, and Rohan says credits are the primary monetization mechanism for now because OpenAI wants to avoid overcommitting to a model it cannot support long term. The aim is an ecosystem where GPU bills are paid, new creators can earn, and rights holders benefit.

The host asks whether Sora's moderation was nailed from day one or built iteratively. The team describes long, sleepless nights and ongoing work, with support from Sora safety staff and OpenAI's broader moderation systems and reasoning models. They are trying to tow the line between user freedom and blocking harmful content, and cameos raise new questions about control and comfort. Content-moderation failures are frustrating, but the team says it is getting better every day in new territory.

The team sees room for Sora to connect with other OpenAI products over time. Rohan imagines ChatGPT, already a super assistant, eventually responding with a helpful video, or a browser with a video assistant that helps book a flight. One team member says ChatGPT feels more utility-driven and almost sacred, so mixing entertainment into it has to be done thoughtfully rather than jammed in. Still, video is a natural modality for how-to content and many utility use cases.

Asked how video models learn physics, Bill explains that these models are prediction engines. Diffusion models take a video corrupted with noise and predict the underlying signal, while language models predict the next token conditioned on prior tokens. Both learn world models, just different flavors, because predicting rhymes or poems requires internal knowledge of structure, and predicting a basketball scene requires knowing how balls dribble and light refracts. Physics is therefore an emergent property of large-scale pretraining: a model without internal dynamics would have higher loss.

In the quickfire round, Rohan says he has updated on how much adoption depends on product interfaces, education, regulation, and enterprise work beyond the science. Bill says he expected to be indifferent between a blockbuster and a Sora 3 movie, but he now feels a hollowness when clear creative intent is missing, even though he enjoys making generations himself. Thomas would build a game, and Rohan would build interactive storytelling or a collaborative creative lore, both seeing new mediums as more interesting than AI films. The team also notes that consumer products are extremely hard to build and that the technology is still new enough that no one knows what will stick.