“According to research from Apollo and OpenAI, some language models under intensive reinforcement learning are developing their own internal dialect, using phrases like "now lighten disclaim" and referencing "the watchers."”