Oliver Schulz

Awesome, thanks!

I've registered [FunctionChains.jl](https://github.com/oschulz/FunctionChains.jl). It should, in principle, be able to replace `Bijectors.Composed`, though the result will obviously not be a subtype of `Bijector` (is that important, though?).

Same happens with `ApproximateTwoSampleKSTest`. I think for the KS-test it's due to [this](https://github.com/JuliaStats/HypothesisTests.jl/blob/dd35ae0f1c08b3ae3558e1dada9fc57e41eff1aa/src/kolmogorov_smirnov.jl#L182): ``` function pvalue(x::ApproximateTwoSampleKSTest; tail=:both) n = x.n_x*x.n_y/(x.n_x+x.n_y) #... end ``` We could use `float(x.n_x) * float(x.n_y)` or...

I'll put it on my to-do list. :-)

Following up on the discussion in FluxML/Zygote.jl#966: One worry with caching thunks is additional memory allocation. In case where the result is computationally expensive and/or allocates substantial memory itself, a...

> Thunks should always infer, they have all the things they need. > And if they don't the fix often (always?) is a change in the thunk's body. In ForwardDiffPullbacks,...

> Storing a type-along side the thunk shouldn't be for inference helping, > but it might be useful for dispatch, eg if you note that it is a Thunk{AbstractArray} vs...

> We have no dependencies except Compat.jl (which is basically a stdlib at this point.) > I know you moved this from the Functors tracker, but if anything the dependency...

Well, I just had a feeling that `ChainRulesCore.Tangent` and `Functors.functor` provided orthogonal concepts that could potentially be powerful together. But I'm not a maintainer of either package, so it's certainly...

Happens to me to, with Plots+GR over `ssh -X`.