Re the new Chinese law, you say: “The model should refuse to answer at least 95% of questions that would violate the law, while answering at least 95% of questions that are not illegal.”
Could you clarify whether illegality here refers to the question or the (potential) AI-generated response? I would assume that it relates to the response rather than the question but your statement seems to indicate the former.
Thanks for this—I really enjoy your newsletter!
Re the new Chinese law, you say: “The model should refuse to answer at least 95% of questions that would violate the law, while answering at least 95% of questions that are not illegal.”
Could you clarify whether illegality here refers to the question or the (potential) AI-generated response? I would assume that it relates to the response rather than the question but your statement seems to indicate the former.
The related twitter thread (where I assume you got the info from?) seems unclear to me.