(E.g. as they are able to learn code grammar.)
When I had the idea I also toyed around with word-rnn and similar RNN libraries. The results I got were pretty good, but training was extremely resource consuming. Cuda gave a 8x boost, but still training one of the smaller sites took 20 minutes on my simple graphics card while creating the Markov chain is done in 2 minutes.
I also have absolutely no experience with Machine Learning and just the setup was already quite an experience. So I stuck with what I know and stayed "the traditional way".
For smaller ones (and smaller datasets) a few hours of GPU is considered fast (at ~1$/h it is not much!). If you want to use it online, there is https://neptune.ml/ (no setup, $5 free credit for computing; full disclosure: created my colleagues).
In any case, I would be excited to see it on some site with code (like SO) or formulae (math or stats). Especially as I am a big fan of StackExchange and analysis (vide http://p.migdal.pl/tagoverflow/ :)).
You may have seen, that you can filter by site [1]. Code quite ruined my chains as it didn't appear in blocks, but rather everywhere, so I went the easy way and filtered all code blocks.
I didn't filter math as I couldn't find a proper way to do it, but you can see that it gets quite messy [2]