Ask HN: When to not use deep learning?

Given the "unreasonable effectiveness" of recurrent neural networks, and the increasing ease of their application to a given learning problem, I'm often left with the following uneasy feeling when reading through academic papers covering more specialized approaches: "Hmm, this was written in 200X... am I wasting my time reading this because deep learning could blow this technique out of the water?"

I want to develop my "smell" for judging whether reading through a paper for a non-RNN model is a good use of my time. Does anyone have any heuristics, red flags, smells, entire domains, etc., that they use to put things immediately in the category of "nah, this is smoked by deep learning techniques"?