Locating style and sycophancy in AI responses
A user says single-token differences in responses make diffuse behaviors like style and sycophancy hard to locate.
TLDR
A user says many approaches to locating AI behaviors use single-token differences in responses, making diffuse behaviors such as style and sycophancy hard to pinpoint. They suggest talking with Aruna to learn how to use causal methods to locate more complex generative behaviors.
Combined views
837
1 Source, first seen ago