[Reading Group] How LLMs Count

What's really happening inside a language model when it solves a problem? This week we'll be discussing a recent technical paper that attempts to answer this question by studying a deceptively simple task: breaking text into lines. Some researchers at Anthropic (the creators of Claude) discovered that their models solve this task by manipulating geometric structures called manifolds in a way that looks suspiciously similar to the biological structures found in animal brains.

https://transformer-circuits.pub/2025/linebreaks/index.html

The paper reveals how attention layers reshape information geometrically to solve tasks. It builds on previous interpretability research, but you don't need a background in neuroscience or advanced understanding of math to participate and get something out of the event!

We expect participants to be read up on the main ideas introduced in the paper, but not necessarily the intricate details. Come ready to discuss and learn more about how this alien intelligence achieves tasks, and what this tells us about how models think more broadly.
You can find a shorter overview of the work from a lead author here: https://x.com/wesg52/status/1980680563582538099

📆 February 18th, 18:00
📍 Sveavägen 76 (EA Sweden Office)
🍌 Snacks will be provided!

Looking forward to seeing you there! Ring "Mejsla" when you arrive and we'll let you in, then we're up on the third floor.

(Note that we also post our events on Facebook, where we tend to have more people signing up, so the Meetup list of attendees is not indicative of the total number of participants.)