It's not actually an AI written chapter. This is false representation. AI generated text still doesn't know how to make a story stay on topic and slowly progress like this, while somewhat still make sense for the characters (even as poor as it does in this)
The relative narrative coherence and the use of the term “predictive keyboards” instead of Markov-somethings has me fairly convinced this was actually just written by a human.
I think "used predictive keyboards" leaves enough wiggle room that there can be a human looking at the handful of suggestions and navigating through them, while letting the autosuggestions work their surreal magic. It seems a powerful combo to me. I love me a good Markov chain composition, but yeah, they really lack in the coherence department, which means they are whiffing on a level of good humor.
>while letting the autosuggestions work their surreal magic
I think that (unfortunately) gives way too much to the autosuggestions, and that the humans are doing the vast majority of the creative work.
My predictive keyboard on android offers 3 words, and if I try, I can delay the trap of falling into a loop and get a small amount of diversity in my sentences. If I had 18 words to work with, and they were specifically trained on the content of Harry Potter, that's a lot of room to exercise the kind of creativity that is still (for now) unique to humans.
The sentences were generated using predictive keyboards, and then the sentences were arranged into a story by a human editor, according to an article I read about it.
Does anyone know what technique they're using to generate this? I've seen similar things done with LSTMs, but the results generally aren't nearly as clean or comprehensible as the text here.
You can try it yourself. It gives you 18 choices for each word, basing the choices on the predictive engine and previous words. It's basically written by a human.
We had an argument on IRC about whether it was possible that this was written by a char-RNN or a word-RNN and concluded it wasn't, because it maintains world state too well, never misspells things, an some things should not ever appear because they don't appear in the books' text (like nowhere inside the novels do they ever refer to the titles, of course, so how would a NN learn the naming scheme 'Harry Potter and the X of Y'?). Not to mention the narrative arc! The predictive keyboard explained all this, plus the crowdsourcing & editing it all into order. It's probably better to think of it as a kind of souped-up Mad Libs than augmentation or AI.
It's perhaps less interesting as technology than pure machine-generated text, but I find it fascinating creatively. In practice it's a collaboration where the machine provides absurdity, and the human writer shapes it into a form that humans can easily read and enjoy. Traditional machine-generated text often has too many sharp edges to work as surrealist humor.
My daughter, who read many Rainbow Magic books a few years ago, suggested that a predictive system like this could probably be used to make realistic Rainbow Magic books.
