Tim DaviesCEO Projekt Rising - The expert AI partner for SMEs

AI Learns Physics from TikTok - ByteDance's Seed Dance EXPLAINED!

6 Jul 2026 · 1 min

Watch on YouTube

0:50

In this segment, Tim Davies introduces ByteDance's SeedDance 2, a video‑creation model trained on millions of TikTok clips. He highlights that, despite being trained solely on visual input‑output pairs, the model exhibits emergent capabilities that resemble an understanding of real‑world physics: it can predict object trajectories, infer positions in non‑visual (3‑D) space, and anticipate camera motions such as spins. Davies emphasizes that these abilities arise without explicit physical supervision, suggesting that large‑scale video data can implicitly teach models to apply physical reasoning.

Chapters

We see the SeedDance 2 video creation model from ByteDance. Amazing. It's just been trained on TikTok videos. Yeah. Millions and millions of videos. And now all of a sudden it understands physics, real-world placement stuff. Does it understand physics or can it apply physics? Well, no, it's a very good point. It can apply physics and it hasn't actually needed to state the physical properties of it. Yeah, and you just see things like, um, placement in non-visual space, or camera spins, and it understands not just where the ball is but where the ball's gone to on the arc that it would have followed. And that's beyond most people's understanding, and yet the models have learned it from video, just from watching and doing effectively input-output analysis.