An "Observatory" For a Shy Super AI? — LessWrong