I suppose if someone successfully created an ML-powered optimizer, powerful enough to level the playing field with C++ as the author suggests, in a higher-level language then they could run it on itself and there would be no need for C++ in the loop. Training the first unoptimized version of the model might be expensive, but if Google can throw $1.4 million of capacity (https://twitter.com/eturner303/status/1223976313544773634) at training a chatbot...

The author is saying that the main reason to use something like C++ instead of Python or even Java is for speed. He assumes that optimizing for speed is a problem that can fit neatly under the machine learning domain.

If a machine can optimize performance better than humans, then it would not make sense to use C++, in the context of performance.

We have JIT compilers that can optimize "better" than static. People are still using the latter because there are other benefits that nobody from the ivory tower can beat.