It really is that slow. You're completely correct that the speed-sensitive parts are written in lower-level libraries, but another way to phrase that is "Pytho

It really is that slow. You're completely correct that the speed-sensitive parts are written in lower-level libraries, but another way to phrase that is "Python can go really fast, as long as you don't use Python." But this also means ML is effectively hamstrung into only using methods that already exist and have been coded in C++, since anything in Python would be too slow to compete. There's lots of languages that make good tradeoffs between performance and usability. Python is not one of those languages. It is, at best, only slightly harder to use than Julia, yet orders-of-magnitude slower.