Paper Club: Looking inside Claude’s “Brain”

​Join us as we explore Anthropic’s groundbreaking research paper, “On the Biology of a Large Language Model.” This fascinating study applies neuroscience-inspired techniques to peek inside Claude 3.5 Haiku and understand how it actually “thinks.” The researchers built an “AI microscope” that traces computational pathways within the model, revealing surprising mechanisms behind its capabilities. Through these innovative methods, the team uncovered remarkable findings about Claude’s internal processes. They discovered that Claude plans ahead when writing poetry, considering potential rhyming words before beginning a line. They found evidence that Claude processes multiple languages using shared conceptual representations - essentially thinking in a “universal language.” The research also revealed how Claude performs mental calculations through parallel pathways, forms chains of medical reasoning, and maintains internal mechanisms to distinguish between what it knows versus what it doesn’t.

​Ever wondered what’s actually happening inside the “black box” of a language model? Join us to peek behind the curtain and discuss these fascinating insights into how modern AI systems really work! If you’re pressed for time, you can read this blog post before coming: https://www.anthropic.com/research/tracing-thoughts-language-model
If not, you can read the paper here to prepare yourself for the discussion: https://transformer-circuits.pub/2025/attribution-graphs/biology.html