Discussion about this post

User's avatar
Soarin' Søren Kierkegaard's avatar

Many people report that adding “please push back on unreasonable requests, perspectives, etc., and don’t just agree with whatever I say—tell the truth like it is even if it would sound harsh or hurt my feelings,” or something similar, to the LLM system prompt yields very good results. Worth trying if you use it for this sort of purpose.

Expand full comment
Lirpa Strike's avatar

Yes! I actually *just* did this a few days ago. Fed it a screenshot of a text-based argument and asked it to objectively analyze it "between the man on the left and woman on the right." I didn't tell it one of them was me, but I think it could've been better if I omitted the genders entirely, because I suspect ChatGPT has a bit of a libfem bias.

But after it gave me the completely validating response I already agreed with and completely expected, I then prompted it to steelman the man's side, and it did so pretty well.

You really have to be good at prompting these things properly, every damn time, to get the most objectivity out of it.

Expand full comment
11 more comments...

No posts