Definition
World models are AI systems that learn spatial-temporal representations of environments, enabling agents to perceive, predict, and act.
Key Points
- Next-State-Prediction: orca-world-model (BAAI, July 2026) models abstract world-state transitions via unified latent space rather than isolated next-token/frame/action objectives
- Gameplay data provides interactive first-person training signal vs static datasets (general-intuition)
- Same model can power in-game agents and physical robots (sim-to-real)
- GAIA-2 video generative models from General Intuition research line
- Orca pre-training: 125K hours video, frozen Qwen3.5 backbone, lightweight readout heads for text/image/action