This is why I said "pretty fast", not necessarily faster than C++. Using ML means trading a small percentage of C++'s raw speed for code that can be debugged, adapted, etc. with much less trouble. I have found this to be a more than reasonable trade-off in virtually all cases.