I think you're on the right track with this thought: the obvious use case for models like this is their ability to classify data based on their training. Like almost everyone has immediately thought "AI moderator" as a use case - but the most obvious use is for the AI to moderate it's own training data for the next version.
Once they can do that and produce a productively improved model, then that's really the start of self-improvement.
Once they can do that and produce a productively improved model, then that's really the start of self-improvement.