These models only really work on languages that have a massive corpus of open source text to steal from.

There are orders of magnitude less Apex than there is JavaScript or Python. There are billions of problem permutations on complex codebases in the big languages that the ML models simply haven't seen in something as niche and unpopular as Apex.

So even if trained on Salesforce, its solutions would be far less stellar than that of the languages with huge, talented dev pools. Or it might spit out something entirely fake, like when I asked it for something complex in Apex just now and it spat out pristine Java instead lol