Thomas Sobieck 🐘This is the dumbest idea of all time.
PeterOh, FFS! Once more for the tech-bros ... "human-seeming" is NOT human! No researcher I've ever met is going to accept LLM-generated responses as representative of human responses.
@PeterLG I think it speaks to the misaligned incentives in research if anyone is actually thinking about doing this. It might get a paper published, but it is going to tell you exactly nothing about reality.
Yes, I'm really hoping it is a "possibilities" paper, rather than a "probabilities" one.
Syn Reitan@Sobieck I especially like the section where it claims that the IA could make studies MORE diverse somehow?? Are these people not aware that's an EXISTING REAL PROBLEM with LLMs?
@syn Yeah, that part really jumped out at me too. It is just dumbfounding. Get the machine that just encodes biases to tell you about marginalized groups (but make sure the machine acts like it is from that group). Sounds like a GREAT plan.
If you want to talk to an underrepresented group, just go talk to them. It might cost money but you might actually learn something.