Black box knowledge — LessWrong