“You don’t train a neural network. You let it struggle, fail, adapt, and repeat—until its failures look like ...