My hunch is that this effectively creates a differentiable minimax “search” “tree” that can be backpropagated through. Not a tree — a dag really — and not search, but learning. :)