OpenAI Sora 2 Team on Generative Video, World Models, and Co-Evolving Society
The OpenAI Sora team (Bill Peebles, Thomas Dimson, Rohan Sahai) sits down with Sequoia's Training Data to discuss the technology and product thinking behind Sora 2. Bill explains how diffusion transformers operate on 'space-time tokens'—cuboid patches that span both spatial and temporal dimensions—which gives Sora global attention over an entire video and enables emergent properties like object permanence. He frames Sora 1 as the 'GPT-1 moment' for video and Sora 2 as the 'GPT-3.5 moment,' with Sora 2 specifically trained to respect physics rather than bend reality to satisfy a prompt (e.g., a missed basketball bounces off the backboard instead of magically going in). He envisions a future 'GPT-4' leap for video that could enable scientific experimentation inside world simulators. The product team describes the rapid, ~3-month build of the Sora social app, emphasizing that the cameo feature—allowing users to insert their likeness into generated videos—was an internal accident that became the product's emotional core. Thomas Dimson draws on his Instagram ranking experience to argue that Sora's feed is explicitly optimized for creation rather than passive consumption, citing that ~100% of new users create on day one and ~70% continue creating on return. The team also discusses exposing Sora 2 via API to support long-tail use cases like filmmaking and CAD, and outlines an in-progress framework to monetize cameos featuring licensed IP for rights holders. Looking forward, the team describes cameo as a 'lowest bandwidth' way to give the model information about a user—a foundation they expect to grow into persistent digital clones operating in a Sora-mediated alternate reality. They explicitly invoke OpenAI's 'iterative deployment' philosophy, saying releasing Sora 2 now (rather than waiting) is meant to let society begin establishing norms around AI likenesses before the technology reaches agents that act in the physical world. Future roadmap items include pet and object cameos and longer-form narrative work.
Original sources
Podcast Training Data
Episode:OpenAI Sora 2 Team: How Generative Video Will Unlock Creativity and World Models ↗
YouTube Training Data
Video:OpenAI Sora 2 Team: How Generative Video Will Unlock Creativity and World Models ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.