Interesting - I wonder what's the ratio of true vs. false knowledge in the corpus that GPT-3 is trained on. Perhaps someone has performed a study on this for a specific domain?