It is sort of frustrating that all these c++ numeric libraries build on top of different linear algebra packages. That makes it more challenging not only to int

It is sort of frustrating that all these c++ numeric libraries build on top of different linear algebra packages. That makes it more challenging not only to integrate an external autodiff tool, but even just to use in your code. Ensmallen appears to build on top of Armadillo. There is autodiff.github.io, but it builds on top of Eigen, as does Ceres and Sophus. Boost has some nice integration routines, but a lot of them only work with boost vectors/matrices. Contrast this with the python ecosystem, where basically everybody uses numpy. That makes it so nice to be able to compose different scientific libraries together. I'm sort curious how python didn't end up with 5 different competing array libraries.