generalization

Generalization is the difference between memorising and understanding. A model that generalises well can take something it learned in one setting and apply it somewhere it has never seen before. A model that generalises badly looks impressive on familiar ground and falls apart the moment the shape of the problem changes slightly.

It is one of the most argued-about questions in AI right now. Optimists point out that models trained mostly on text turned out to be decent at logic and planning, which nobody explicitly taught them. Sceptics counter that recent models are getting narrower rather than broader, spiking at coding and maths while standing still elsewhere, and that what looked like understanding was really very wide coverage of the training data. Nobody has settled it, which is worth remembering when a lab claims its model is generally smarter.