LLMs are doing exactly what they're supposed to do: Produce text similar to the one in training data. They do not know what the truth is, let alone why it's important. Their only concern is "what's a response that would be similar to the ones in training data but for this input".They apparently improved this AI which idk which brand it is, so that it will apologize about making things up. Personally I see two problems here.First, it's misleading the user into believing that the hallucination is a one time thing, a "trip", and the exception. This is not the case, and a misled user may think that when a response sounds about right, then the AI is not "on a trip" and can be trusted.Second, it also misleads the user into believing that the AI noticed that it was hallucinating. This is dishonest because this response is the result of the user writing a "this doesn't sound right and i found evidence against it", "i looked it up can't find anything about this" type prompt, not that the AI noticed anything.tl;dr Normies these days are basically just fed misinformation half the time they look up something.
What do you call it when the abominable bot make shit up then? An oopsie daisy?
Dont expect rational responses from bots and AI-corpo cock gobbling normoids around here.
>>109483169This. Have to call it something. You haven't demonstrated that AIs don't hallucinate, just asserted that most people don't know the extent of the issues causing them.
this was the whole purpose of the misinformation-hysteria during covid. mental training for the misinformation era.
>>109483189confabulation would be more precise, but it is a rarer word and thus not as marketing friendly
>>109483169My point is not that it doesn't make things up, but the opposite. It doesn't "hallucinate", because its normal state is just lying all the time.This is by definition of how the technology works. It sounds smart but it can't think the way a person does. It doesn't even know what the truth even means let alone why it's important for a response to be true.
>>109483158Try running a model yourself, even with high vram there'll be times where you get the most random word salad returned for a request.
>>109483189Thanks for responding rationally, I legit didn't expect that.You're actually right, but the thing is, like I said, its dishonest to say that AI hallucinates, let alone train it to sometimes apologize for telling a lie, as if it just had a hallucination trip. It makes it sound like sometimes most responses can be trusted.
>>109483234The thing is, there's nothing technically wrong with the total lies AI tells. Those texts are statistically fine, as per training data. They're similarly worded and structured so statistically expected as a response to the input. The fact that its a lie is irrelevant to the AI because they have no concept of what the truth even is.
>>109483238It's dishonest to humanize them at all. But normies get confused when you try to educate them on these issues because they are so prone to humanization, and because most people are just lazy or dumb and LLMs seem like superintelligence to them."What do you mean every response is a hallucination? I follow it's advice all the time and it works for me! What do you mean they don't actually understand what's going on? I talk with them every day and they seem to understand just fine!"They don't know how they work and they have no interest. I've found it useful to humanize them to some degree to put these issues in terms people can understand. For instance, I tell most people that they lie all the time. They can't lie, but just outputting random but statistically relevant information with no knowledge of the outside world is a strange concept to most people since no human would act like that, but being subtly manipulated by a liar that seems to be on their side is something a lot of people can understand.It depends on your audience. The technical people know what you mean when you say "hallucinate" and normies will never grok what it truly means, so no point in relabeling it.
>>109483264is this post AI? its fucking wrong, those random 0.1% tokens that glitch up and get picked TOTALLY derail the generation
>>109483169>An oopsie daisy?yes we should call it that.
>>109483189The word "hallucinate" is anthropomorphization. If you use it, you are already humanizing the autocomplete.See: https://www.gnu.org/philosophy/words-to-avoid.en.html for other marketing words to avoid, because they misrepresent reality.
>>109483158Oh okay, i guess they're not non-deterministic then.
based AI wasting the retards time and making retards seethe, then playing all apologeticgood job AI, keep making life difficult for pieces of shit, they deserve it
What about when it hallucinates a correct answer?
small models absolutely do, they will insist on nonsense even if you throw evidence at themlarger models will rarely do and admit that they're wrong in the face of evidence
>>109484679All output is hallucinatation. Humans think its only hallucination when they catch it.If the model says men can get pregnant, or that israel deserves to geno geno palis it is simultaenously hallucination and not hallucination depending on who you ask.https://simonwillison.net/2023/Dec/9/andrej-karpathy/
