I am similarly very bullish on this field, even for what I once thought were incredible claims that it can eventually understand the semantics behind documents, do question and answering, and even reason.

What blew my mind was this lecture, given by Hinton https://drive.google.com/file/d/0B8i61jl8OE3XdHRCSkV1VFNqTWc...

The main idea is that reasoning is just a sequence of 'thought vectors' that can be encoded within a recurrent neural network.

Sounded almost outlandish to me until I watched Richard Socher's lectures and started to understand that words can be represented as vectors, and that these vectors can be then encoded into new vectors of even higher representation. 'Thought vectors' may not be so outlandish after all.