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.
A world model is a network that learns to simulate an environment: given the current state and an action, it predicts the next state. Train it on enough video or experience and it builds an internal, navigable model of how a place behaves — letting an agent imagine and practise futures instead of only reacting to logged data. The hard part is keeping that simulation consistent over long interactions.
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.