[CITATION NEEDED] 002

​Ever wondered what really goes on under the hood of cutting-edge AI research? [Citation Needed] is where we break down dense, technical AI research papers and uncover the ideas shaping the future - together.

​Paper for this edition of [Citation Needed] is: "On the Biology of a Large Language Model" (Lindsey et al., 2025) – https://transformer-circuits.pub/2025/attribution-graphs/biology.html

​Authors: Jack Lindsey, Emmanuel Ameisen, Adam Pearce, Joshua Batson, Chris Olah, et al. (Anthropic Interpretability Team)

​We use LLMs every day, but no one (not even the labs that build them) can fully explain what's happening inside. Anthropic's interpretability team set out to change that. Borrowing from neuroscience, they built what they call a "microscope" for Claude 3.5 Haiku, then traced the actual circuits the model uses to do things like solve math, write poetry, and refuse harmful requests.

​The findings are stranger than you'd expect. Claude plans rhymes several words before it writes them. It does arithmetic through pathways no human would ever teach. It sometimes fabricates its own chain of thought after the answer has already been decided. And it has language-independent internal representations, meaning the model "thinks" before it picks which language to answer in.

​If you've ever wondered whether LLMs are stochastic parrots or something stranger, this is the paper that gives you a literal look inside.

​Some guardrails for the event:

• ​Pre-reading the paper is mandatory! You can use CiteCat to "talk to the paper" with AI beforehand ask it anything, clarify tough parts, and come ready to contribute.

• ​Everyone must come with one key question you have about the paper or its contents. We'll explore this question together. The question has to be related to the paper.