World Model
A world model is a visionary concept in AI research: building an internal representation of how the real world operates — physics, spatial geometry, causality, social dynamics. AI with a world model can "imagine" the consequences of actions inside its head — a key step toward AGI and the core driver of the 2024-2025 video generation wave.
Two mainstream definitions
Narrow (embodied / RL view)
- Model is an internal simulator of the environment: given current state + action, predict the next state
- Used in model-based RL, planning, robotics, autonomous driving
- Representatives: Ha & Schmidhuber 2018 "World Models," JEPA (championed by Yann LeCun)
Broad (generative AI view)
- Model learns how the world works from massive video / image / physics data
- Expressed via generation: Sora / Gen-3 / Veo / Wan / Hailuo all head this way
- NVIDIA Cosmos, Google Genie, Meta V-JEPA 2 are explicit "World Foundation Models"
Why the sudden surge in 2024-2025
- Sora release (2024-02): made "video generation = world simulation" narrative plausible
- Video models were surprisingly good at physical consistency (objects don't suddenly disappear, lighting reflects correctly)
- Yann LeCun repeatedly emphasized LLMs lack world models; AGI must go the JEPA route
- NVIDIA positioned Cosmos as "the world foundation model for robotics / autonomous driving"
- Massive academic + industry investment made "world model" co-equal with "LLM" as a buzzword
Key techniques
Video generation = implicit world model
- Model learns 3D geometry / physics in latent space
- Not an explicit physics engine, but generated results obey physics
- The Sora "video as world simulator" paper is the core narrative
JEPA (Joint Embedding Predictive Architecture)
- Championed by LeCun: predict the embedding-space future state, not pixels
- Avoids generative models' computational waste
- Meta V-JEPA 2 is the open-source representative
Model-based RL
- DreamerV3, MuZero etc.: use learned world models for planning / tree search
- 10-100× more sample-efficient than model-free RL
Physics simulation integration
- NVIDIA Cosmos + Isaac Sim
- Google DeepMind Genie 2 (interactive environments)
- Combined with neural radiance fields (NeRF / 3D Gaussian Splatting)
Deployment scenarios
- Robotics / embodied AI: sim-to-real transfer, planning
- Autonomous driving: Wayve / Tesla world models are central
- Games / simulation: Genie 2, Decart real-time game generation
- Film / advertising: controllable video generation
- Scientific simulation: NVIDIA Earth-2, DeepMind weather models
Limitations
- True physical consistency: current models still violate basic physics (gravity, rigidity, momentum)
- Long-horizon consistency: videos beyond a few minutes often "drift" (outfit changes, scene contradictions)
- Correlation vs causation: models learn visual correlations, not necessarily real causality
- Interactivity: most video generation is offline; not real-time interactive yet