Lenny's Podcast · Sunday, October 18, 2026
Emmanuel Ameisen explains that Anthropic's interpretability team studies how LLMs like Claude operate by looking inside their intermediate numerical states. They discovered that models form robust internal representations for complex concepts across different contexts, such as recognizing the concept of an eye regardless of whether it appears in a photo, drawing, or diagram.
“they sort of represent, you know, these concepts. uh, uniquely across contexts.”
“if you pass a prompt that has a photo of an eye, or you talk about eyes in any language, or you draw a little SVG with an eye in it, and you look at the intermediate results inside the model, the same direction lights up”