Comment by markerz

The models are likely to check certain places. So either their training overrepresents it or the models were RLHF'ed to go there. I don't think it's that far away from checking StackOverflow for some bug, or reading Wikipedia for some trivia.