My personal passion is linguistics, and I find this whole LSTM or probablistic grammar stuff ridiculous. It would be nice to see some research on connecting knowledge models rather than just trying to mimick speech like a parrot.