World Models

Expert

AI systems that learn an internal representation of the physical world to predict, simulate, and reason about reality.

Last updated: Sep 13, 2026

How do World Models work?

World Models combine various techniques to model physical reality. The core idea: compress sensory input into a compact latent space, learn dynamics in that space, then decode predictions back into observable outputs.

A learned world model estimates dynamics from data. A physics simulator evaluates specified physical equations. A generative video model can produce plausible frames without accurately predicting every action's consequences. These tools can be combined, but their claims need different tests.

Collect observations and actions. Measurement noise and missing information matter.

Prediction error grows with the rollout

Worked dynamics example: x(t) = v₀·t + ½a·t². The reference acceleration and the model's estimated acceleration differ. This is an analytic toy world, not measured performance of Genie or a learned model.

Position (m): ±45.0010 Time (s)
Reference: 45.00 mPrediction: 40.00 mError: 5.00 m

Changing the estimate changes every predicted state. Even a small acceleration error accumulates quadratically here. Real world models also face uncertain observations, actions outside training data and compounding errors.

Examples and boundaries

Examples are versioned research references, not a ranking of currently available products.

Genie 3 / Project Genie

Genie 3 was announced on 5 August 2025. Project Genie is a product experiment using it. Interactive world generation has limits in action control, consistency and rollout duration.

DeepMind (2025)

Genesis

Genesis is a physics simulation platform. Its speed claims depend on a specific simulator, task and hardware; they are not performance guarantees for learned world models.

Genesis
Ha & Schmidhuber: World Models (2018)Genie