I agree that Flux.jl is very cool but you can't use it to train a SOTA LLM, can you? The Julia ecosystem does not have the resources to even attempt building a viable alternative to Pytorch.

I'm still curious what concrete things you mean that Julia is missing. Maybe I'm so stuck doing numerical analysis, can't think of anything else...

>I'm still curious what concrete things you mean that Julia is missing.

Basically everything not related to numerical analysis. The problem isn't that there isn't an option which does work, but that those options aren't fully developed and mature to actually use for a commercial project.

E.g. look at GUI libraries, the native ones are just wrappers around other libraries and poorly documented. There is a single web framework, but is that supported enough that you would bet your entire company on it being continually supported?

If you are doing anything which isn't related to numerical analysis, Julia isn't your first choice, it also isn't your second or third choice. And I think that is fine, clearly Julia wants to be a good at a particular thing and has achieved that.

> and just auto differentiate the whole thing automatically

Not so simple as it sounds. A lot of speed advantages of Julia comes from doing in place ops on arrays and Zygote does not work with in place updates. So you end up copying and then at that point Python + PyTorch becomes a superior choice as Python is the de facto the universal language.