An attempt to steelman OpenAI's alignment plan — LessWrong