When AI Learns the World

Language models learn patterns in text. So-called world models aim to represent states, actions and changes in environments that may be visual, physical or simulated. This is still a heterogeneous family of approaches: a shared label does not mean that systems share objectives, data or methods.

Google DeepMind describes Genie 3 as a model capable of generating interactive environments. Waymo has described a driving-specific world model, built on Genie 3 and adapted to a domain where rare or difficult situations need to be explored without collecting them all on public roads.

Other tools, including NVIDIA Cosmos, focus on models and synthetic data for developing physical AI. AMI Labs has announced a research direction inspired by JEPA, which seeks to predict abstract representations of states rather than directly producing realistic images. These are different examples; it would be inaccurate to treat them as equivalent variants of a standardised technology.

This distinction matters because a convincing scene is not necessarily a useful prediction. Simulation is valuable when it represents the variables relevant to a decision, reflects the limits of its domain and is compared with real data. Visual realism is one property, not a complete measure of accuracy.

Imagine a luxury house choosing the assortment for a boutique. A general simulator may help represent spaces and footfall; it does not automatically know the preferences of that house’s customers, the effects of a private event or how a past choice changed later demand.

Specific forecasts require data linking decisions, conditions and outcomes: which assortment was selected, in which market, with what footfall and what happened afterwards. Sales data without context record outcomes but may not explain the mechanism.

The opportunity for a company does not necessarily lie in building a general-purpose world model. It may be more practical: recording decisions and outcomes consistently enough to evaluate simulations and forecasts. The issue is how representative the data are and who governs access to them. A historical archive does not automatically become a reliable simulator.

The meaning of “world model” is evolving. This paper describes public examples and considers their implications for business; it does not compare the performance of the systems cited.