Interview
OpenAI Sora 2 Team: How Generative Video Will Unlock Creativity and World Models
- Sora's Strategic Philosophy: OpenAI is pursuing an "iterative deployment" strategy to co-evolve society with video technology, aiming to establish "rules of the road" before long-term capabilities (such as autonomous "copies of yourself" performing tasks in the physical world) become widespread.
- Technical Architecture (DITs): Sora utilizes Diffusion Transformers (DITs) rather than autoregressive models, generating video by progressively removing noise from a signal rather than predicting tokens sequentially; this allows the model to generate the entire video simultaneously.
- Space-Time Tokens: The fundamental building block for Sora is the "space-time token" (or patch), a voxel-like unit representing both spatial dimensions and temporal locale, which enables the model to maintain global context and emergent properties like object permanence.
- Sora 2 vs. Sora 1 Improvements: Sora 2 represents a "step function" improvement beyond simple scaling, specifically enhancing the model's ability to respect physical laws (e.g., a basketball missing a hoop and rebounding realistically) rather than hallucinating success to satisfy the prompt.
- Implicit Agents: The "agent-like" behavior in Sora 2 is largely an emergent property of scale and robust internal world modeling rather than explicit engineering of agent logic.
- Data Strategy: The team is not exhausting pre-training data; video data is considered a massive, untapped resource where "intelligence per bit" is lower than text, but total volume is significantly higher and likely infinite for the foreseeable future.
- Simulated Science: OpenAI anticipates Sora 3 or 4 could achieve a "GPT-4 level" breakthrough enabling robust scientific discovery, such as running biological or fluid dynamics experiments (e.g., turbulence theory) entirely within the simulator without wet labs.
- Product Launch Metrics: Since launch, Sora has generated approximately 7 million videos daily, with a user base that is surprisingly diverse beyond the typical "AI film" niche.
- Creation vs. Consumption Incentives: The platform's ranking algorithm is explicitly optimized to prioritize "creation over consumption," contrasting with Instagram's historical shift toward optimized consumption, aiming to keep users in a creative, remixing mindset.
- Viral Feature (Cameos): The "Cameo" feature (allowing users to insert themselves or others into generated scenes) was an emergent product success that drove high retention; internal data showed ~70% of returning users create content and ~30% post to the feed within days of onboarding.
- API and Developer Strategy: A separate API is being rolled out to support "long-tail" enterprise use cases (e.g., CAD integration, filmmaking tools) without forcing all users into the social app interface.
- Future of Media: The team predicts feature-length filmmaking will become economically viable for individuals within the near future, though the format will likely evolve into a "new medium" rather than a direct replica of traditional cinema.
- Intellectual Property (IP) Monetization: OpenAI is actively developing a new economic framework to compensate rights holders when users generate content using specific IP (e.g., character cameos), ensuring monetization flows back to original creators while allowing user customization.
- Emergent IP: The team has observed unexpected emergent behaviors, such as a "walking clock" generated from a simple prompt, highlighting the creative unpredictability of the model.
- Long-Term Vision: The ultimate goal is a "digital clone" platform where users interact with persistent versions of themselves and others, functioning as a "mini alternate reality" for knowledge work and social interaction.
- Scientific Discovery Timeline: The team expects physical phenomena represented by visual data (fluid dynamics, classical physics) to be the first solvable by Sora, while non-visual disciplines (quantum mechanics) will likely be simulated last or never.