Exactly, understanding something requires holding a model (or several) of the thing in the mind and be able to work with that model to some extent. That’s why w

Exactly, understanding something requires holding a model (or several) of the thing in the mind and be able to work with that model to some extent. That’s why we dont see statistical tricks as understanding - as soon as trivial errors show that operation in those models is not feasible, we reject it. And that’s why winograd works: it establishes a simmetry that requires a base semantic model outside of the provided text.