To me, as soon as it was obvious that training had distilled and connected abstract concepts of increasing generality, it was only a matter of time. Almost any "new" idea can be decomposed into a combination of old component concepts.
It was evident in GPT-3.5