From one perspective, the output of your Harry Potter neural net is better than the output of a travesty generator I programmed back in the 90s, which analyzed and then reproduced the frequency distribution of n-grams—three-letter combinations, four-letter, or whatever you set it for—in a given input text. (Sorry—I didn’t save any examples.) The output of this new method makes sense more often, it captures more of the tone of the original, and it recreates a number of structural features, such as the back-and-forth of a conversation, better. But from another angle, the two methods are similar. In both, the sentences are often fragmented; you get a certain cracked or surrealistic logic; you have no sense of a real story. Neural Potter (Harry Net?) is a fun step along the way to a genuine parody, I’ll grant you that, but since I’m unsure what the broader purpose is, I’m unsure what to make of these results.