Double backwards

Double backwards is PyTorch's way of implementing higher order differentiation. Why might you want it? How does it work? What are some of the weird things that happen when you do this?