You're about to train a 40-layer ReLU MLP with no normalization layers, and a failed run costs a day of compute.