World models offer a testing ground for physical AI, but validation remains a constraint

The article describes growing momentum behind world models for predicting physical consequences and comparing actions. Laboratory manipulation results and driving-simulation examples provide limited evidence of utility, while infrastructure applications remain hypothetical.

World models could help operators test actions before committing equipment or changing physical systems, reducing the cost and risk of experimentation. Their outputs still require independent real-world validation and safe fallbacks; conventional forecasting, optimization or simulation may remain preferable.

World models learn from images, video, sensors and action–outcome data to predict environmental changes, allowing planners to compare possible actions. The World Economic Forum describes growing interest in this approach as language-centred AI encounters limits in modelling physical consequences. Examples include Meta’s V-JEPA 2, DeepMind’s Genie 3 and Waymo’s camera-and-lidar simulations. Meta reports 65–80% success on selected laboratory manipulation tasks, but these results do not establish broader deployment readiness; infrastructure-planning applications remain hypothetical. The near-term opportunity is cheaper, safer experimentation rather than demonstrated operational autonomy. Visually convincing simulations can still misrepresent physics, and shared training-and-testing environments can conceal weaknesses. For infrastructure and mobility operators, the key distinction is between predicting a scene, choosing an effective action and performing reliably after deployment. The article calls for separate evaluation of each, supported by real-world comparisons, independent evidence and monitoring.

Evidence supplied by World Economic Forum.