Most people fall for the traditional AI fallacy that thought in the brain must somehow resemble lisp expressions. You can tell someone what thought you are having by producing a string of words that would normally give rise to that thought but this doesn't mean the thought is a string of symbols in some unambiguous internal language. The new recurrent network translation models make it clear that you can get a very long way by treating a thought as a big state vector. Jay McClelland was pushing this view several decades ago when computers were much too small to demonstrate its power.

Traditional AI researchers will be horrified by the view that thoughts are merely the hidden states of a recurrent net and even more horrified by the idea that reasoning is just sequences of such state vectors. That's why I think its currently very important to get our critics to state, in a clearly decideable way, what it is they think these nets won't be able to learn to do. Otherwise each advance of neural networks will be met by a new reason for why that advance does not really count. S