OpenAI Sora 2 Team: How Generative Video Will Unlock Creativity and World Models
OpenAI’s Bill Peebles, Thomas Dimson, and Rohan Sahai explain how Sora 2 crosses the GPT-3.5 usability threshold for video and where world-simulator science begins.
- Sora 2 is the GPT-3.5 moment for video; Sora 1 was GPT-1 — Sora 2 crosses the mass-usability threshold.
- Bill Peebles invented the diffusion transformer (DiT); virtually all competitor video models, including Chinese ones, now use DiTs.
- Sora 2’s novel failure mode: a missed basketball rebounds off the backboard instead of being magically guided in — physics overrides prompt intent.
- ~7 million generations per day; nearly 100% of new users create on day 1; 30% post to the public feed.
- Thomas Dimson wrote Instagram’s original ranking algo to fix chronological crowding by high-volume accounts, then watched it drift toward mindless consumption under ad pressure — Sora is deliberately designed against that.
- OpenAI is building an IP monetization layer so rights holders receive revenue when their characters are used in cameos.
- Long-term vision: Sora becomes a persistent alternate reality running digital clones that do knowledge work and report back to the physical world.
- First predicted world-model scientific discovery: something in fluid dynamics or turbulence — phenomena that lend themselves to simulation iterations.
2025-11-06 · Watch on YouTube
