“OpenAI, DeepMind, and Anthropic recently co-published a paper exploring how observing a model's thinking process can be used as a tool for alignment research.”