This is a neat observation, training has been optimized to hell and inference is just beginning.