Running an experiment to see if LLMs will deliver you worse outputs if the model perceives you to be a crackpot. Less advanced models seem to respond the exact same way to the user, regardless of the user's indicated status. Haven't tested GPT 5.5 or Opus 4.7 yet though. (Joe Weisenthal)