Neural networks generalize, otherwise they would not be as powerful as they are today (and I don't know how you can deny that). If your neural network does not generalize then the model is overfitted.
They don't generalize *well. "Deep Learning" as it is done currently, is very limited in its ability to generalize to out of distribution tasks. This is a major area of discussion in research.
The type of generalization necessary to perform what the parent was talking about for instance (synthesizing schematics) is (currently) not possible.
Lots of smart people have been trying to get it to have capabilities anywhere close to what you’re describing for years now, to no luck.
We’re safe for now.