OK, I’ll bite. : )

The vast majority of coding data ingestion, mooel discovery, and training that we currently do will all go away.

The job will become much more interesting, because researchers will try and understand why certain architectures/training regimes are unable to perform certain tasks. Also, I think the architectures for some fundamental tasks like computer vision, and audio are going to become modular. This whole training models end to end is going to be verboten.