Mateusz Baran

Yes, exactly, I don't want to rush 1.0 but just have a place to discuss what breaking changes are planned.

There is already a Julia package that has combined function and gradient calculation figured out: https://github.com/JuliaNLSolvers/NLSolversBase.jl .

When you calculate gradient using AD then you get cost "for free" as a part of the output of AD -- so it's nice to be able to use that.

I think #150 is supposed to resolve this.

I think there are no easy solutions here. One idea would be defining environment variables in CI which are read by runtests.jl and enable/disable selected tests. So some different CI...

I'm fine with one hour of testing but it's near the upper limit of what is acceptable. And there are some related issues, for example if we wanted to split...

Sure, I think we can keep this open.

Cool, that would be a great speedup :+1: .

I'd prefer to have a trait instead of an abstract type like `AbstractRiemannianManifold` but there should definitely be a way to distinguish Riemannian and pseudo-Riemannian manifolds.

Great, I'd definitely like to see support for ML on manifold-valued data :+1: . I can help with this as well :slightly_smiling_face: . Regarding that vectorization, there are a few...