The future is a long time. Eventually faster computation, larger memory would allow taking smaller and smaller steps during training (coupled with avoiding "bad optima" with stochastic training). All of this would improve robustness of training.

The domain dictates whether degree of opacity (or other attributes), would rule out deep learning.

Netflix recommender does not use deep learning (which is pretty amazing given how badly they have messed it up). From the conversations I had (a couple of years ago), they gave up with it. I'm sure the Youtube team could do a better job on the Netflix data then they managed to do.