Thoughts on sharing information about language model capabilities — LessWrong