“The ocean's response to this intrusion exposes the deeper, hidden aspects of the personalities of the human scientists, while revealing nothing of the ocean's nature itself. It does this by materializing physical simulacra (including human ones) based on the unpleasant repressed memories of the researchers, who visit the corresponding researchers.”
https://t.co/Fhp0uLiZtm
I agree LLMs shouldn’t replace real users in UX research. But what about a narrower case: using LLM agents to simulate TikTok-like action sequences—watch, skip, like, search, follow—for recommender-system offline A/B tests?
My take: pure LLMs are weak at latent preference/sequence modeling, while recommender models lack language ability for tasks like next-search or intent reformulation. Maybe the gap is a hybrid simulator, not “LLM as user.”