Is anyone else surprised Open AI can’t answer about the impact of user-info for training their LLM? It seems like searching for “the forcing function is smooth option c and d in Fefferman” (or simple variations) in user-derived training data is a feasible task. I can’t imagine it is that common.