There was a post [1] from the creator about why he's not using RNNs, namely: 1. It requires a GPU and a ton more computational power, whereas a Markov chain is

There was a post [1] from the creator about why he's not using RNNs, namely: 1. It requires a GPU and a ton more computational power, whereas a Markov chain is something a normal server can do with a few spare cycles. (I believe he's using one of Reddit's servers.) 2. char-rnn, the most accessible implementation, doesn't seem to support online updating I look forward to when RNN tools are convenient to use, or when there's a computationally simpler version of the same idea, like GloVe is to word2vec. [1] https://www.reddit.com/r/SubredditSimMeta/comments/3goxte/what_if_subreddit_simulator_used_recurrent_neural/ https://www.reddit.com/r/SubredditSimMeta/comments/3goxte/wh...