> It's the output model that eventually makes it into the back-end ¯\_(ツ)_/¯.
Yes, the model itself is part of the backend, not the code used to generate the model. But the website explicitly mentions "machine-learning __models__" as an example of something that's "not backend".
> Back-end has always meant the CRUD/API portions of a CRUD app.
This is not true. Compilers have backends. Speech processing pipelines have backends. The dragon book, which pre-dates the advent of the WWW, distinguished frontend from backend (and in those words).
Web-based CRUD apps inherited the terminology.