site banner

Friday Fun Thread for August 29, 2025

Be advised: this thread is not for serious in-depth discussion of weighty topics (we have a link for that), this thread is not for anything Culture War related. This thread is for Fun. You got jokes? Share 'em. You got silly questions? Ask 'em.

2
Jump in the discussion.

No email address required.

My layman's guess is that LLMs are trained on internet text. If someone asks a question on the internet, conscientious people with epistemic humility mostly don't respond rather than typing out "I don't know", so epistemic humility is underrepresented in the training set.

That seems like something that should be corrected in tuning/user experience design.

Do people actually want to be told "I don't know"? My guess is that, much like managers prefer optimistic timelines that turn out to be wrong over accurate, pessimistic timelines, people would rather get a wrong answer than no answer, especially if they can't tell that the answer is wrong. Unless you give specific instructions to the contrary to the Indians doing RLHF and train them in the methods of rationality, you are going to get hallucinations.

much like managers prefer optimistic timelines that turn out to be wrong over accurate, pessimistic timelines...

That is very much not the case in my experience. Managers would far rather be told the truth than what they want to hear (though obviously they prefer it if the truth is what they want to hear).

Perhaps you could compromise by making it more aware that it's bullshitting so it can say so explicitly? Surely "I don't know, but the best I can come up with is X" is better than just "X", for a bullshit X.