Julia is interesting because it's performant enough to implement these things directly--Julia's machine learning frameworks are written in pure Julia (with the exception of the calls to CUDA libraries for GPU) and can achieve near-native performance, so solving the "two language problem" is part of its value proposition. I do wonder how valuable it is for the whole stack to be written in one language, and whether that will blur the lines between the software engineers who implement data science packages and the data scientists who use them.
https://juliagpu.org/ https://juliagpu.org/2020-09-28-gemmkernels/
> Can we write and execute all our code in a Jupyter Notebook? Yes! Should we? Probably not. However, I enjoy the workflow, and making this an enjoyable process is important to me.