concept also: world models, generative world model
World model
A model that learns an internal, navigable simulation of an environment — predicting how a scene evolves as you act in it, rather than just generating a single frame.
World models matter because they give an agent somewhere to practise. Instead of training only on logged data, a system can roll out imagined futures inside the model, try things, and learn from outcomes — cheaply and safely.
The hard part is consistency: objects must persist, physics must roughly hold, and the world must stay coherent over long interactions. Getting that right is what separates a tech demo from a usable training ground.