At the moment it is much more interesting when applied to more narrowly targeted tasks. These can include video captioning but for clearly specified tasks. The (too quick of a) tentative jump from well automated targeted tasks to generic task automation is the culprit here.

In my few years of practice of deep neural networks in industry, I can say they are very nice to work with as they fit a large number of tasks and yield excellent results for many of them.