I've seen the hallucination rate of LLMs improve significantly, if you stick to well covered topics they probably do quite well. The issue is they often have no tells when making things up.
I don't think it's quite the same.
With search results, aka web sites, you can compare between them and get a "majority opinion" if you have doubts - it doesn't guarantee correctness but it does improve the odds.
Some sites are also more reputable and reliable than others - e.g. if the information is from Reuters, a university's courseware, official government agencies, ... etc. it's probably correct.
With LLMs you get one answer and that's it - although some like Bard provide alternate drafts but they are all from the same source and can all be hallucinations ...
Yes and no. If the LLM is repeating the same thing on multiple drafts then it's very unlikely to be a hallucination.
It's when multiple generations are all saying different things that you need to take notice.
LLMs hallucinate yes but getting the same hallucination multiple times is incredibly rare.
"In particular, we find that LMs often hallucinate differing authors of hallucinated references when queried in independent sessions, while consistently identify authors of real references. This suggests that the hallucination may be more a generation issue than inherent to current training techniques or representation."
https://arxiv.org/abs/2303.08896
"SelfCheckGPT leverages the simple idea that if a LLM has knowledge of a given concept, sampled responses are likely to be similar and contain consistent facts. However, for hallucinated facts, stochastically sampled responses are likely to diverge and contradict one another."
and that wouldn't eliminate hallucinations just tell you if large details have likely been hallucinated.
But it's a method some research has used.
P.S. Also aren’t LLMs deterministic if you set their “temperature” to zero? Are there drafts if the temperature is zero? If not, then that’s the same as removing the randomness no?
>Just do the comparison on the user’s machine if the LLM provider is that cheap.
This is not possible. Users don't have the resources to run these gigantic models. LLM inference is not cheap. Open ai, Google aren't running profit on free cGPT or Bard.
>P.S. Also aren’t LLMs deterministic if you set their “temperature” to zero? Are there drafts if the temperature is zero? If not, then that’s the same as removing the randomness no?
It's not a problem of randomness. a temp of 0 doesn't reduce hallucinations. LLMs internally know when they are hallucinating/taking a wild guess. randomness influences how that guess manifests each time but the decision to guess was already made.
I never said it did. If it did we wouldn’t be complaining about hallucinations as we can just turn them off at will by setting the temp to zero.
> LLMs internally know when they are hallucinating/taking a wild guess.
No they don’t. If they did we would be able to program them to not do so.
I would argue that wild guesses are all LLMs are doing. They practically statistically guess their way to an answer. It works surprisingly well a lot of the time but they don’t really understand why they are right/wrong.
P.S. LLMs are kind of like students who didn’t study for the test so they use “heuristics” to guess the answer. If the test setter is predictable enough, the student might actually get a few right.
I feel like using LLM today is like using search 15 years ago - you get a feel for getting results you want.
I'd never use chatGPT for anything that's even remotely obscure, controversial, or niche.
But through all my double-checking, I've had phenomenal success rate in getting useful, readable, valid responses to well-covered / documented topics such as introductory french, introductory music theory, well-covered & non-controversial history and science.
I'd love to see the example you experienced; if I ask chatGPT "tell me about Toronto, Canada", my expectation would be to get high accuracy. If I asked it "Was Hum, Croatia, part of the Istrian liberation movement in the seventies", I'd have far less confidence - it's a leading question, on a less covered topic, introducing inaccuracies in the prompt.
My point is - for a 3 hour drive to cottage, I'm OK with something that's only 95% accurate on easy topics! I'd get no better from my spouse or best friend if they made it on the same drive :). My life will not depend on it, I'll have an educationally good time and miles will pass faster :).
(also, these conversations always seem to end in suffocatingly self-righteous "I don't know how others can live in this post-fact free world of ignorance", but that has a LOT of assumptions and, ironically, non-factual bias in it as well)