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

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

If prefixed uncertainty words appear to elicit some degree of uncertainty/meta-reasoning from GPT-3, perhaps explicit probabilities can quantify it? Revisiting the questions from before, I try prefixing probabilities to answers. While GPT-3 imitates the pattern with no problem, it continues to make errors on questions it seems to’ve solved before without as much trouble (like the weight comparison questions), and the probabilities don’t have any clear connection to the correctness of the answers, so it looks like the numbers still don’t work even when using the prefixing trick.