There's excellent new reporting out today on OpenAI's sycophancy crisis: how early the risks were known, and the safety tools OpenAI wasn't using I wrote a short post, highlighting the new facts I learned and mixing in a few reflections