AI company Anthropic has published research providing a new perspective on how large language models process information, using 'mechanistic interpretability.' This method aims to understand the 'why' behind AI outputs, though its interpretation remains complex.