We dive into groundbreaking new research that's offering a glimpse into the inner workings of large language models. Forget the idea that these powerful AI systems simply predict the next word – we explore how a new technique called circuit tracing, used by Anthropic on their Claude 3.5 Haiku model, is revealing surprisingly complex and sometimes "inhuman" reasoning. Discover how Claude handles tasks like translation with language-neutral circuits, performs math in unconventional ways, and even shows foresight when writing poetry. Join us as we unpack these fascinating findings and what they mean for our understanding of artificial intelligence, even as researchers admit this is just the beginning of a long journey to truly understand these enigmatic technologies.