Plot twist: the loss function for training is basically a benchmark