Comment by imh

It's a bit beyond me too, but here's my take from skimming: It's already known that you can do differential calculus in some pretty abstract spaces (not just functions from R^n to R^m). Automatic differentiation is a way of automatically keeping differentiability and computing derivatives when sticking other differentiable stuff together. This puts those ideas together, so you can automatically differentiate some more interesting things stuck together in some interesting ways related to programming. That was what I got from my rusty semi-layperson's skim, at least.