## TL;DR
World models learn to simulate reality from video data — predicting future frames, understanding physics, and enabling agents to plan. OpenAI's Sora demonstrates emergent 3D understanding; UniSim creates interactive environments from internet video.

## Core Explanation
A world model is a generative model that predicts the future state of the world given current state and an action: s(t+1) = f(s(t), a). In video prediction: given past frames and text/action, generate future frames. These models learn physics implicitly from millions of hours of video — object permanence, collisions, lighting, camera motion.

## Detailed Analysis
Sora is a diffusion transformer that treats video as a sequence of spacetime patches, trained jointly on videos of variable resolutions and durations. Emergent capabilities: (1) 3D consistency — objects stay consistent as camera moves; (2) long-range coherence — objects remain visible after occlusion; (3) basic physics — fluid dynamics, rigid body motion. DreamerV3 (DeepMind) uses world models within RL agents that learn in latent state space.

## Further Reading
- DreamerV3: Mastering Diverse Domains (DeepMind)
- GAIA-1: World Model for Autonomous Driving (Wayve)
- "World Models" by Schmidhuber