True that most excitement in the field has been in finding new architectures and building general purpose AI. However I would like to point out that people often underestimate potential of transfer learning. We have started solving real problems companies are facing right now because of lack of enough data to be able to train an accurate model themselves.