This seems to be the trend in every computational science field. Take something compute intensive, apply DL magic and get results in a fraction of the CPU time

This seems to be the trend in every computational science field. Take something compute intensive, apply DL magic and get results in a fraction of the CPU time as the old approach. Similarly on the commercial side, plenty of opportunity to solve existing problems with DL. So while it may be true that DL research progress has slowed, still plenty to do in applying existing DL.