Vibeleaderboard
← All Intel
Intel / video

OpenAI Sora 2 Team: How Generative Video Will Unlock Creativity and World Models

Source
youtube.com
Author
Sequoia Capital
Date

Context

In a Sequoia Capital podcast, Bill Peebles, who heads the Sora team at OpenAI, explains how its video models generate. A diffusion takes video, adds a lot of noise, and trains a network to predict that noise. Generation then removes noise step by step, working on the whole clip at once instead of one piece after another. Peebles says this helps with quality drifting over time.

The model treats video as small blocks covering both space and time, called spacetime patches. Because every patch can attend to every other, Peebles says, properties like object permanence can emerge. Peebles also describes a Sora 2 behavior: when a prompted basketball shot should miss, the ball rebounds off the backboard instead of being forced into the hoop. Peebles presents this as the model respecting physics, though it is the team's own claim.

Terms in this piece · Glossary
  • transformer — The neural network architecture behind modern AI models, built on attention — letting every word directly consider every other word in parallel.
  • token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
Read the source www.youtube.com
More from Sequoia Capital
Recommended reads
Comments

Checking sign-in…

Loading comments…