“GPT-3 Nonfiction § Umeshisms”, Gwern2020-06-19 (, )⁠:

Nonfiction writing by OpenAI’s GPT-3 model, testing logic, commonsense reasoning, anagrams, PDF/OCR cleaning, creative nonfiction, etc

Scott Aaronson popularized the genre of “Umeshisms”: quotes about how optimal choices typically involve some tradeoff and a non-zero error rate; they are useful for counteracting one-sided attitudes where errors are minimized without regard to opportunity cost or expected value.

I took some of his and some of mine to see if GPT-3 would understand the abstract idea & create more examples; many of its results are valid & one could easily screen its output to manufacture many Umeshisms, but it also gets the “direction” wrong enough that I’m not sure if it entirely grasps the abstract point… (As of 2023-01, ChatGPT appears to do much better.)