Near-mode thinking on AI — LessWrong