Single-shot learning, but the goal is to produce an enormous corpus of colloquial responses, interpretations and extrapolations of the training example, which need not be “correct”. And some clever acronym for the technique like RelYGAN or TaNNACh
-
I can envision this for "extrapolation" but I don't see how you could get meaningfully unique "responses" without more training data. but maybe I'm not quite understanding what you mean?
Well just to be completely clear, this was a fever dream tweet proposing a science fiction idea that seemed cool. Perhaps you could get "responses" by running a pool of randomly seeded agents that punish each other for mimicking the original input poorly