One of these is the Adam optimizer performing actual gradient descent. The other is... ^weeeee

source twitter