Per the embedded video (also used on the official site), they used textgenrnn (https://github.com/minimaxir/textgenrnn) to train the text-generating AI, which I made. Although I really wish they said that explicitly: not as a means of attribution, but to make the text generation more accessible/transparent which is why I made textgenrnn in the first place.

GPT-2 would likely have much more interesting results on such longer input text. If anyone from AKQA reads this, I'm open to help setting that up.