It ought to be a property of a world model, but a language model models language. We use language capacities to describe counterfactuals, argue about facts, and imagine all the time. What assistants are lacking is any reasonable kind of world model. We need to marry the stochastic to the symbolic.