or the model get trained from teaching code which naturally contains lots of comments.

the dev is just lazy to not include them anymore, wheres the model doesn't really need to be lazy, as paid by the token