Seth Axen

> Also, regarding the above plan, @sethaxen what are your thoughts on it in light of our adjoint-adventures the other day? I agree in general with your comments. `to_vec`'s current...

> Also, regarding the above plan, @sethaxen what are your thoughts on it in light of our adjoint-adventures the other day? I'm nervous that if I start changing things, things...

> It's entirely possible that people are using this package to compute gradients, so losing that functionality would make for a bad user experience. Manifolds.jl's default "AD" backend is FiniteDifferences,...

Here's another example multi-dataset netCDF4: [tmp.zip](https://github.com/JuliaIO/JLD2.jl/files/9282992/tmp.zip). The latest commits in that PR seem to be able to load the names of the groups and the names of the arrays in...

Indeed, it seems to load perfectly! I don't know much about how netCDF 4 relates to HDF5, but I was a bit surprised by comparing the attributes ussing JLD2 vs...

> at least the name datasets give their respective dimension id. Are you referring to `_Netcdf4Dimid`? What can that be used for? I found another netCDF file that errors on...

Related: https://github.com/JuliaDiff/ChainRules.jl/issues/237

See also this issue about why Zygote doesn't do this (tl/dr Zygote basically treats all reals as embedded in the complex numbers): https://github.com/FluxML/Zygote.jl/issues/342 and an update here: https://github.com/FluxML/Zygote.jl/issues/472 Also ccing...

Maybe also get so it can pretend to be something with keys?

The current tests fail because 1. `zero(::NotImplemented)` throws a `NotImplementedException` 2. `zero(NoTangent())` is a `ZeroTangent()`, so this change breaks inferrability. This causes the ChainRules integration tests to fail for `copysign`...