Someone correct me if I am wrong here, but these char-rnn based outputs are great to look at and experiment with - but they dont seem to be of any practical use. All that they are capable of showing is that an RNN can "remember" things. The biggest question is how do you make any use of these things ?

Is it possible to nudge the RNN in a particular direction - so that it produces something that we want ? Perhaps there is an answer to this in Alex Graves et al. Paper on Handwriting generation. A more thorough explanation or exploration in this direction would perhaps help ?.

Anyone really working on generating CSS ( conditioned! ) ?