I also think S4TF's approach has been really neat here: there's a really nice python interop so you still have access to the entire body of work and tools in python available, and it basically feels like writing native swift code.
I primarily write C++ and Python and never feel like this.
But in general, c++ to python is an awkward comparison because c++ is extremely low-level by comparison. So you do get static typing, which is nice, but you also get a lot of other issues to deal with - namely manual memory management - so with either python or c++ you have a high probability you will have code which fails at runtime. It would just be the case that in c++ most of the time it would be because you handled memory incorrectly, whereas with python it would more often be because you made a type-related error. Swift is much more comparable with Python in terms of reducing cognitive load by obviating low-level details from the programmer.
I prefer the landscape for ML not to be fragmented based on non-essential qualities like the language that is used.
How great is it that researchers publish code in the same language, and everybody can use that code immediately without having to learn language X and/or porting the implementation?