However, for some applications there's a vendor with a neural network (or whatever) trained on a billion inputs, you see it perform better than the open source equivalent trained on a million inputs, and you don't have a billion inputs to improve the open source version with.