The thing is I don't trust this explanation for a second especially as it applies to non-paying customers; they could have just as easily trained a generic ML algorithm on a publicly available data set, like I don't know, the public stuff on github.

Moreover, they could have trained their suggestions to actually be useful before throwing this out there as a feature set they thought people would want to use.

Plus then it'd make sense for people to open up their code, as a "local dictionary" of sorts that could be prioritized over generic suggestions. But at least then it would have had demonstrated value.