this post was submitted on 18 Jun 2024
94 points (100.0% liked)
TechTakes
1401 readers
139 users here now
Big brain tech dude got yet another clueless take over at HackerNews etc? Here's the place to vent. Orange site, VC foolishness, all welcome.
This is not debate club. Unless it’s amusing debate.
For actually-good tech, you want our NotAwfulTech community
founded 1 year ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Statements from LLMs are to be seen as hallucinations unless proven otherwise by classic research.
We don't need a fancy word that makes it sound like AI is actually intelligent when talking about how AI is frequently wrong and unreliable. AI being wrong is like someone who misunderstood something or took a joke as literal repeating it as factual.
When people are wrong we don't call it hallucinating unless their senses are altered. AI doesn't have senses.
Yeah, LLM are accidentally right sometimes. But all they really do is pull words and phrases that it thinks statistically fit together.
Does everyone else see this? These are the exact type of out of town haters we really want. I also think calling LLMs all but delusional is too generous and I mean that unironically.
It's not a "fancy word" here, but a technical term. An AI making things up is actually called hallucination.
the technical term is either “confabulation” or “bullshit”; “hallucination” is a misleading label coined by the ai pushers.
It used to mean things like false positives in computer vision, where it is sort of appropriate: the AI is seeing something that's not there.
Then the machine translation people started misusing the term when their software mistranslated by adding something that was not present in the original text. They may have been already trying to be misleading with this term, because "hallucination" implies that the error happens when parsing the input text - which distracts from a very real concern about the possibility that what was added was being plagiarized from the training dataset (which carries risk of IP contamination).
Now, what's happening is that language models are very often a very wrong tool for the job. When you want to cite a court case as a precedent, you want a court case that actually existed - not a sample from the underlying probability distribution of possible court cases! LLM peddlers don't want to ever admit that an LLM is the wrong tool for that job, so instead they pretend that it is the right tool that, alas, sometimes "hallucinates".
Technical terms can still be, technically speaking, dumb as fuck.
Lmao
I am saying that coining it as a term was stupid and intended to make it sound intelligent when it isn't.
oh definitely, it's fucking terrible question-begging. I'd like to know when it traces back to, and how good faith it was or wasn't
It originally comes from false positives in computer vision afaik, where it makes some sense as the model is "seeing" things that aren't in the image.
Of course is the term stupid. Neither is an LLM an AI, nor is any AI in the current state intelligent. In the end it all boils down to being answer machines. Complex ones, but still far away from anything even remotely being am AI.
The wikipedia page you linked to actually states that the term is being pushed by industry (Google, Meta, OpenAI) and that its use is criticized by some researchers.
So you say, a technical term should not be created by the people who actually develop the technology the term is used for?
You're confusing "developing" with "marketing".
LOL, okay.
oh but you see, it's "hallucination" when LLM is wrong and it's hype cycle fuel when it's correct. no, LLMs don't "hallucinate", that implies that this state is peculiar, isolated, triggered by very specific circumstances. LLMs bullshit all the time, sometimes they are right, sometimes not, the process that produces both types of response is the same. pushing for "hallucination" tries to obscure that. use of "hallucination" also implies that LLMs know something, they don't, by design. it just so happens that if they "get" things right, it's because it appeared in training material enough times to make an impression in model.
Bullshitting to me is giving intentionally wrong statements. LLMs do not generate intentionally wrong statements. Saying they do, means that you imply intelligence.
LLMs know nothing nor are they intelligent. They also are not right or wrong, they generate output based on statistics.
"Hallucination" as a term for "AIs" making things up is used since the early 2000s (even if it's meaning has changed since then).
bullshitting as in when you give a confident answer without regard of actual reality. previously discussed there LLMs do exactly that: generate confidently, authoritatively sounding text without regard of facts, because these things do not know facts or anything for that matter.
maybe it's high time to change terms then
So you say there could be different meanings of the same word? Like “bullshitting” or “hallucination”?
mod post: please desist, it's just tiresome now
agreed
Absolutely.
Hallucination thought does fit.
It’s a term in the context of a source that implies untrustworthy, not authoritative and/or imagined.
Lots of examples in every day usage or scenarios that come to mind.
“And then I saw the defendant punch the victim and then I was blinded by the sunlight”
Are you sure you didn’t hallucinate the entire episode? It was night after all.
Or
“Somebody please get these ants off of me”
Doctor writes: Hallucinations of ants on skin
Those are examples of actual hallucinations where something did not happen.
Quoting a joke reddit thread as factual is not hallucinating. There was such a thread, but it wasn't factual and an LLM is wrong to present it as factual.
That’s the issue. LLM’s aren’t trustworthy. They hallucinate.
I presume, as the default, that anything a LLM produces is a hallucination right out of the gate.
"Hallucination" implies LLMs can meaningfully perceive. They can't, they're not made that way and they have no reason to be.
We’re arguing language now though, and by definition it isn’t “hallucinating”. By saying that’s what’s happening, you’re unintentionally legitimizing the “AI is making decisions” misinformation.
To get really pedantic, “flashback” would be a better label. It’s not making things up whole cloth, just repeating stuff way out of context.