Automatic Question Generation has come a long way in about 20 years. In that, it has followed the general trajectory of the whole field of computational linguistics. From linguistically inspired approaches, supported by statistical computations, the whole field, and AQG with it, have moved to deep learning neural approaches, and then increasingly relying on larger and larger models. This was in part enabled by the increasing computation power of available hardware. The phenomenal success of large language models was not an obvious outcome when considered even circa the year 2018.