This summer I presented the same work, on the geometry of ideology in language models, at two conferences with different audiences.
At IC2S2 in Burlington, Vermont (July 28–31), social science was well represented, but mechanistic interpretability was not.
At NEMI, the New England Mechanistic Interpretability workshop at Boston University (August 14), it was the other way around. Most of the room came from computer science, and few social scientists attended. Yet a field that asks what models represent needs the conceptual work social scientists do. In his keynote, “Why I Don’t Work on Mechanistic Interpretability (plus what I’d do if I were),” Stephen Casper mentioned Tim Miller’s Explanation in Artificial Intelligence: Insights from the Social Sciences.
I wrote more about measuring concepts inside and outside of LLMs in (Re)reading Sartori in another computer revolution.