If you are the sort of person who needs a trigger warning on somewhat sexual or profane language, please do not read this.
A couple observations about the process. This was a one-shot prompt. I edited the first completion to my liking, then resubmitted. This process went on for several rounds, after which I did another couple editorial passes. One paragraph in particular, the riff on filmmaker Stan Brakhage, is entirely machine generated and untouched by me. I tweaked and massaged various other sentences generated by the AI.
I’m coming to see GPT-3 as truly a writing instrument. One big potential mistake I see is in thinking of the AI as a labor-saving device. I actually don’t want the AI to save me time or effort. I intend to apply the same level of attention to machine-generated prose as I do to human-generated prose. The fun part is actually making the thing.
Second, I think there’s great potential in establishing an absurd but defined premise. I’ve always liked the stories of Aimee Bender, who starts with a strange premise, like a family of pumpkin-headed people who have a child with a head that’s a steam iron. Then she tries to logically work out how such a family would actually operate, what problems they’d run into, etc. I see GPT-3 doing the same. Start with something utterly outlandish, then work out the nuts and bolts implications of that premise.
Another thing I’m observing. I like the dynamic of GPT-3 trying and not quite succeeding to imitate my voice, followed by my attempting to imitate GPT-3’s imitation of me. This is incredibly exciting.