This isn't true. A lot of the reason Julia would be so good for ML is that you can write your own high level kernels in Julia that compile down directly to devi

This isn't true. A lot of the reason Julia would be so good for ML is that you can write your own high level kernels in Julia that compile down directly to device code with CUDA.jl. The reason Python is in a lot of ways a really bad fit for ML is that you can't write fast python. You instead have to call a C/C++ library.