1) Deep Learning-based techniques outperform hand engineered natural language processing stacks in every measurable way.

2) There will be considerable progress made on the Winograd Schema challenge.

3) The (unconstrained) Turing test won't be "officially" passed until October 2017 though (it will actually be passed in July or August, but the news won't leak until October)[1].

4) Some people will continue to insist that Deep Learning is nothing different to what was being done in the 1990s. At some point someone will get frustrated enough with this to blog everything that is different now.

[1] Specific enough prediction?