Note

Advanced AI Sycophancy

I am not the first person to notice this behavior in frontier models. I’ve noticed it myself when workshopping drafts for this blog. Sometimes I’ll have an argument that goes A->B->C, and the model will suggest I reorder as B->A->C. If I try that and feed it into a new instance of the same model, it’ll sometimes say “that’s great, but I suggest ordering it as A->B->C”, and so on forever. It really does seem as if the model is trying hard to give me some kind of superficial pushback that I can either smugly ignore or happily accept.

Sean Goedecke’s article.