Hacker News new | ask | show | jobs
by ryanklee 891 days ago
It absolutely is useful and one of the most significant points in this comment section.

People are applying standards to LLMs that don't exist elsewhere. They don't exist elsewhere because they absolutely can't exist elsewhere.

It's not a technical short-coming of LLMs that they can't produce citations in every instance. Rather, it's a property of information, representation, and knowledge itself. Much of it floats far above the otherwise load bearing pillars of citations.

People contend with this constantly in every arena of life, and we have come up with very elaborate ways to offset the difficulties caused by it.

And we get along very well despite it all.

I'll also leave you with this: citations are just pointers to more sources of information. Not some ground truth. It's just another tranche that requires evaluations.

Lastly, your comment is an unfortunate bit of low-level snark and probably would have been better left unsaid.

1 comments

So you're just going to ignore the fact that we require humans to provide citations when they include someone else's writing or research in their own?
I didn't ignore it. But we do not require citations for every statement. Statements have value, citations notwithstanding.
I didn't say it should provide citations for every statement. We're talking about instances where the model reproduces text verbatim. If it is paraphrasing or creating new text then obviously it doesn't need to cite any more than a human would.
I don't understand what you're pushing back against, but I'm not standing there.