Mohamed Tarek

Note that you can pass in the initial state here https://github.com/JuliaNLSolvers/Optim.jl/blob/master/src/multivariate/optimize/optimize.jl#L34 as a kwarg. This can be used query the (L-)BFGS approx of the inverse Hessian after the optimization. I...

I like the always threaded option

There seem to be a lot of suggestions here on how to signal to the compiler to infer the output type of a function. Perhaps NOT specifying any output type...

>Wouldn't not specifying a type with const or var constitute the same thing? Exactly, `const a: anytype = f(..)` is not more readable than `const a = f(..)`, it's less...

My hope is that the following would work. ```zig fn add(a: u32, b: u32) { return a + b; } const c = add(2, 2); ``` In this case, the...

Moving some of my discussion points over from Slack and polishing them a bit. ChainRules projects the function's input's co-tangent to match the input's type. This can result in loss...

I think the example above is an argument against projecting at all in the complex case, in that projecting complex numbers to real numbers leads to loss of information that...

Projecting `nothing` and structural no tangents is fine. Your example is not surprising to me if they return different types because one function promotes types to the complex domain and...

You can put that in the objective function. Make it a callable struct, add a counter field and then define the callback behaviour you want.

Beautiful writeup! I don't think special-casing broadcasting for the user-facing API is the right approach here. `map`, `mapreduce` and other higher order functions are all supported by CUDA.jl for any...