Relatedly, while I wouldn't use either for many types of ML...I wouldn't use Python either. The only benefit to using Python in this space is it has libraries bound atop C/C++ implementations. Erlang/Elixir doesn't, that I know of, but that doesn't prevent it from being done if someone wanted to. In terms of actually building something from the ground up...well, it depends. It isn't fast...but it does make the concurrency part easy to model ( https://www.amazon.com/Handbook-Neuroevolution-Through-Erlan... for instance), and for learning/prototyping that might be what you're prioritizing for. Certainly, Erlang/Elixir have seen use in high speed trading and ad bidding platforms, things known for needing low latency.
Get dozens of book recommendations delivered straight to your inbox every Thursday.