[kevinventullo]

does anyone in the brave new AI community have the patience and humility to try to solve the hallucinations problem?

I think this is akin to asking if anyone in the software engineering community has the patience and humility to try to solve the software bugs problem.

[disgruntledphd2]

Nope, those are very different. Fundamentally the reason that LLMs hallucinate is because they are predicting the next most likely word. It would be expensive to find a way to remove outputs where P truth is low, but it's not completely implausible like removing software bugs.

[kevinventullo]

I suspect removing outputs where P(truth) is low is about as hard (and arguably a superset of) detecting code for which P(bug) is high.

As an aside, I think these are both roughly equivalent to detecting code for which P(halts) is high!

[voxl]

So more expressive type systems? Rust has made entire classes of bugs impossible, and you can do even better in Idris or some other dependent type theory. People don't do it because it's expensive and they're not trained it.

No it never eliminated all bugs, but if we're conflating the two problems you don't need to eliminate all hallucinations either, you just need to make them very unlikely, or a consequence of your bad specifications.