Plan B in AI Safety approach — LessWrong