Скачать или смотреть Why Video AI Fails at Physics (Even at Massive Scale)

Can video generation models actually understand the physical world — or are they just copying patterns?

This video breaks down a 2025 ICML research paper that puts modern video diffusion models to the test using pure physics simulations.

Researchers trained large-scale video models on millions of simulated videos governed by classical mechanics — including motion, collisions, and gravity — and evaluated them across:

• In-distribution generalization
• Out-of-distribution (OOD) scenarios
• Combinatorial physical reasoning

The result?

🔹 Near-perfect performance on familiar scenarios
🔹 Total failure on true out-of-distribution physics
🔹 “Case-based imitation” instead of rule learning

Even more surprising: the model prioritizes color over size, velocity, and shape — explaining why objects often change form or teleport in AI-generated videos.

This study challenges the idea that simply scaling video models will lead to true world models.

📄 Paper: How Far Is Video Generation From World Models? A Physical Law Perspective
🎥 Topic: Video diffusion, physics reasoning, world models, AI generalization

Subscribe for more 5-minute research breakdowns on cutting-edge AI.