It depends on the dataset. Char-RNNs do pretty well compared to Markov chains when used to generate smaller texts. (which is why I've become less of a fan of the Shakespeare example the original Karpathy post used).