This is extremely cool but I wouldn't call it that sophisticated, it's a first draft of the approach. I feel like you could do this with maybe a third of the tokens, for starters
something about the tone of the prompt really gets me
like, really trying to help a nascent ai grow into it’s abilities by priming it with its own behavioral patterns and faults… and it’s working???
"Sounds like it's going to be hard to get GPT-3 to compensate for its known shortcomings."
"Actually it's going to be super easy, barely an inconvenience."