Karpathy's work on large datasets for deep neural flow is conceiving of the "backward pass" as the preparation for initializing the mechanics for weight ranges, either as derivatives in -10/+10 statistic deviations.