Physical reasoning has been a blind spot for generative AI. A new paper from Google DeepMind shows their video model Veo 3 developed an understanding of physics, skills it wasn't trained for but learned on its own. Why this matters: Fei-Fei Li and Yann LeCun are both confident spatial intelligence is AI's next frontier.
Trained only to generate realistic videos, Veo 3 solves mazes, simulates how liquids behave, models light, edits videos from sketches, and predicts which objects will float or sink.
As these capabilities evolve, video models won't just generate realistic footage. They could reshape creative tools across "spatial" industries, transforming how game designers, filmmakers, VFX artists, industrial designers, architects, and engineers imagine and build.
Congratulations to the DeepMind team and Nick Matarese.