The big stuff is mostly done in C/C++ (because Python is slow) and bound to, so you could end up with basically the same bindings for JS/TS.
The incredibly long time it took to port very popular libraries from Python 2, to Python 3, demonstrates how unrealistic this thought is.
I've tried using both tensorflow and pytorch via their C/C++ apis. It's possible, but just enough of the convenience logic is in python that it's not very practical.
Incidentally if you're going to do it, I recommend pytorch over tensorflow.