Will Tebbutt

Not ideal, but at least there's a precedent 😂

Note: @simeonschaub is essentially trying to do this in https://github.com/FluxML/Zygote.jl/pull/909

I think `ldiv` would actually be okay. I suspect that we would at least require a tenary function in which the sensitivities w.r.t. two of the arguments share some computation,...

I think this sounds reasonable, but I'm having trouble saying for sure without a concrete example. Any chance that you could concoct one?

@simeonschaub what's the status of this? Has it been completely superceded by the other complex numbers stuff?

> It was a hard-choice, and we are not revisiting it here. I'm not suggesting to revisit it here, I was just trying to see if I've missed an option....

Is this what we want? Specifically, I don't see why someone shouldn't make a new type `T` "primitive" in our system (by doing stuff with `getfield` or whatever), and use...

> In julia if you have isprimativetype(T) == true then you can't do getfield, primative types (and Arrays (but not other AbstractArrays)) do not have fields. Sorry, I'm talking about...

> there is no relevance. I disagree. If you want to answer the question "should we allow `Tangent{P}` for some Julia primitive `P`?", you have to know whether or not...

> I would argue it is 1. It is 1. Because they are not objects with fields, and thus can't be represented with Tangent{T}. Because Tangent{T} operations are not defined...