World Model

A world model learns representations of an environment and predicts how it may change. When conditioned on actions, it can predict possible consequences of those actions. The 2018 World Models paper ↗ studied compressed spatial-temporal representations and policies trained within a learned environment.^[1] Its “world” is a modeled game, scene or task environment, rather than the entire universe.

Multimodal AI concerns the inputs a system combines; a world model concerns environmental change; an AI Agent concerns action selection. One system can combine these roles without making them interchangeable.

The DreamerV3 method ↗ distinguishes compressed observations, predicted future states and rewards, and an action-selecting policy.^[2] Other world models need not share its architecture. The following is a simplified explanation.

Represent the current observation as state z and predict a distribution over subsequent states given action a. Several outcomes can remain possible. A policy uses these predictions to choose an action, then new observations reveal prediction errors. Planning can use latent states without rendering every imagined frame, so defining world models solely as video generators excludes important designs.

Sources were checked on 20 October 2026. Google DeepMind's Genie 3 model page ↗ describes interactive environment generation alongside limitations in action range, location accuracy and interaction duration.^[3] Distinguish the Genie model from the Project Genie prototype.

The 19 May 2026 announcement ↗ introduced Street View grounding and expanded access.^[4] The January launch's US-only notice should therefore not be presented as the entire current access policy. Starting from real imagery does not establish that every subsequent generated scene is geometrically or physically validated. Recheck official regional and account conditions when using the service.

In this educational example, route A takes a predicted eight seconds with a predicted contact probability of 0.20; B takes eleven seconds with probability 0.02. Minimizing time favors A, while a hypothetical constraint of at most 0.05 leaves only B. These numbers are neither product measurements nor safety thresholds.

Calibration matters: an unvalidated prediction of 0.02 does not establish a trustworthy real frequency. An unfamiliar transparent obstacle could invalidate both forecasts. Physical AI requires separate validation of observations, control and stopping systems on the actual device.

Evaluate appearance, responsiveness to actions, object consistency, long-horizon error and downstream task completion separately. Attractive video that ignores a right-turn command from the same starting state may be an inadequate action model for that use. This is ORBiS Tree's evaluation framework, not an official product scorecard.

Match initial states, action sequences, horizons, available observations and compute budgets. Test conditions excluded from training and retain failures. Benchmark Data Contamination addresses evaluation overlap; Data Provenance helps distinguish generated material from observations.

Correlations learned from data need not describe causal effects under intervention. Repeated prediction can compound small errors, and actions rewarded inside a model may fail outside it. Uncertainty can also be underestimated under unfamiliar conditions.

Report training in generated environments, testing in those environments and independent real-environment evaluation separately. Record environments and exclusions in Safety Evaluation and intended uses in a Model Card. A recent demonstration alone does not establish general physical understanding or field readiness.

Read AI Agent for action selection, Physical AI for interaction with reality, and Multimodal AI for input integration. Automated AI Researcher addresses research automation; Recursive Self-Improvement addresses repeated changes to the improving system itself.