Ah, but we do use specialized hardware and compilers! (nevermind that they're sold by nvidia, nobody else is buying quadros by the truckload) And moreover, you don't do the differential equation in the code, you do the partial specialization by hand and throw that into your pytorch kernel.

The fact that we are using python, instead of fortran or matlab, really does show exactly how little hard math is being done by the computer, even in these fields.