They just completely won with calling LLM errors "hallucinations" didn't they? Even the fiercest AI critics say it. Unnecessarily anthropomorphic and helpless way to say "returned false and broken results," as if the bot had an imagination or the ability to tell truth from lies. - ThreadSky

mammonmachine.bsky.social • 74 days ago

They just completely won with calling LLM errors "hallucinations" didn't they? Even the fiercest AI critics say it. Unnecessarily anthropomorphic and helpless way to say "returned false and broken results," as if the bot had an imagination or the ability to tell truth from lies.

Comments

boggo2300.bsky.social•74 days ago

I think we should counter the use of "hallucinations" by always correcting it to "lie"

keiya.tinyfairy.xyz•74 days ago

I prefer "bullshit". Lie implies intentional deception in a way I don't like, bullshitting is just spewing whatever you can to sound good. That's exactly what LLMs do.

theterroirist.bsky.social•74 days ago

It's not even a lie though. There's no intentionality behind it.

It's just wrong. Bad data output.

boggo2300.bsky.social•74 days ago

A lie is a deliberatu untruth, the programmers of LLC's know they provide incorrect information, hence the disclaimer on all the LLC chatbot pages, there for they KNOW they are prone to giving incorrect information.

Thats how I see it anyway

theterroirist.bsky.social•74 days ago

The programmers and marketers lie when they talk about what LLMs can do, yes.

But the LLMs themselves are incapable of lying, because they're incapable of doing anything deliberately.

taylorofthestars.bsky.social•74 days ago

A lot of my hallucinations are just errors in my sensory processing so I feel like it's pretty a apt description

ninebillionder.bsky.social•73 days ago

I'm fine with "Lie" since you are being deceived. You were sold on the idea that it is performing that task for you, the company that said that is lying on purpose.

I'm going to be more aggressive about correcting misinformation, but honestly, it's like talking to a brick wall.

captainshijinko.bsky.social•74 days ago

Son of a bitch, you're right

heimsteam.bsky.social•74 days ago

A crucial point, to which I think you hint, is that a false result here is not even a lie. "Lie" implies intentionality, which implies some sort of mind. LLMs and the like cannot think and, as such, cannot even lie. It's all gibberish, with, at most, an incidental relationship to reality.

celialicious.bsky.social•74 days ago

More like a bad die roll, frankly. If all LLM output is probabilistic, then it's really just a way of encoding information into sets of probabilities, and no matter what you do, it can always roll poorly, and you can never tell why.

delbertgseasons.bsky.social•74 days ago

I think hallucinations is more damning. Everybody makes mistakes, but if you hallucinate regularly you get canned

lazyblueberry.bsky.social•74 days ago

Nope, I fucking refuse to call them "lies" or "hallucinations". I typically call them errors, even though that's also kind of inaccurate.

heimsteam.bsky.social•74 days ago

I agree that even "error" is inadequate. I'm willing to bet that the response given is almost always a best fit for a given model and dataset. The inaccuracy is a flaw with the model, input data, or both. It's not "wrong", it has little connection to the reality that we hold constitutes "right."

lazyblueberry.bsky.social•74 days ago

My difficulty in labeling them errors is more the fact that these programs weren't designed to be answer finders. Is it an error of the program when you use Excel to create a PowerPoint instead of, you know, PowerPoint?

heimsteam.bsky.social•74 days ago

I think I understand the spirit of the idea (attempting to use the wrong tool to attack the problem), but I think it's often not the categorically incorrect tool.

lazyblueberry.bsky.social•74 days ago

I wouldn't even go as far as to say it's even the best fit. Depending on the generator, the response is different each time you run the same query. Image generators are notorious for this.

heimsteam.bsky.social•74 days ago

I'll admit that I haven't paid much attention to the particulars of specific systems, just going off my understanding of linear algebra and curve-fitting, which is most of what this is. I'd hazard a guess that they can rapidly modify systems, but I've got little concrete experience with it.

handofcrystal.bsky.social•74 days ago

i read a great paper once that put forth an argument that LLMs output bullshit and should be classified as bullshit machines

keiya.tinyfairy.xyz•74 days ago

I like calling it bullshit. LLMs are really good at generating bullshit: convincing-sounding text with zero substance.

davin.bsky.social•74 days ago

naming each new gray hair I get after the newest model release

twohitgood.bsky.social•74 days ago

It's marketing bullshit and always has been. It's a fucking error. It is a wrong output from a computer program. It isn't hallucinating because it isn't having sensory confusion. It is just bad programming resulting in an errant output. It's a republican, is what it is.

realtime.crt.video•74 days ago

an error implies it's fixable
a hallucination implies it's crazy
to my the hallucination is worse though i guess it does imply that it's capable of thought

ariaegirl.bsky.social•74 days ago

The issue is, for us it's an hallucination, for the machine it's what it was trained to do. It's not a false or broken result for the machine. LLMs were made to make text that in some degree looks ok, but it's incapable of deep/long cohesion and contextualization.

vaguebiscuit.bsky.social•74 days ago

Heck, calling them AI in the first place is just marketing bullshit

vaguebiscuit.bsky.social•74 days ago

And anyone who thinks this stuff is the first step on the path to creating an actual intelligence can have their opinion discounted immediately

psiop.kawaii.social•74 days ago

I'm reminded of all the scenes in Banks' Culture novels where machines get offended by being compared to simple computers.

I am also constantly reminded of how the Villain in Surface Detail is just Elon Musk.

vaguebiscuit.bsky.social•74 days ago

My favourite is Flere-Imsaho in Player of Games. He is so indignant

sensualkazoo.bsky.social•74 days ago

That's the secret Mr Altman, chat gpt is always hallucinating.

caseyponey.bsky.social•74 days ago

Especially with how the initial instruction for every LLM nowadays includes shit like "Don't hallucinate" or "Don't make up answers if you don't know". Like, this isn't a misbehaving toddler, bro. It's not getting shit wrong for fun.

Even the engineers fall for their own propaganda smdh

authorx.bsky.social•74 days ago

Yeah, this always makes me mad. "Hallucination" makes it sound like it perceived the wrong information, but it's not about perception, just whether it guessed wrong or guessed right.

Comments

Posting Rules

Reply