21–30 of 491 posts
> hallucination is defined as inconsistencies between a computable LLM and a computable ground truth function. That's simply inaccuracy or fabrication. Labelling it hallucination simply panders to the idea these programs are intelligent.
Yes, imagine a pocket calculator that gave a completely wrong answer 10%? of the time, and calling it 'capricious' instead of simply broken.
With this definition, you can trivially prove the titular sentence - "hallucination is inevitable" - is untrue.
Let your LLM have a fixed input context length of one byte. Continue training the LLM until such a time as it replies to the input "A" with "yes" and all other inputs with "no".
Define your computable ground truth function such that the correct output for the input "A" is "yes" and the correct output for all other inputs is "no".
This LLM provable never hallucinates - we have exhaustively verified that its output matches the ground truth function for all possible inputs.
There is nothing stopping inductively increasing the size of the input context and the number of entries in the ground truth table arbitrarily, and at no step do hallucinations become "inevitable".
It seems a stretch to call it "inevitable". "Inevitable given the current architecture without modifications" at most. Also, I'm missing a section on how (if) human brains manage to avoid hallucinations in this. Also, it doesn't have to never hallucinate, it just has to hallucinate less than we do.
For example: if you smoke pot and get paranoid, it's because pot dials back the work of the part of your brain that prunes thought paths that are not applicable. Normally, paranoid thoughts do not make sense, so they are discarded. That's also why you're more 'creative' when you smoke pot, less thought paths are pruned and more stuff that doesn't quite make sense gets through. Or thoughts that overly focus on some details get through, which are normally not required.
Our brains are inherently "higher level", current AI is hopelessly simplistic by comparison.
It seems a stretch to call it "inevitable". "Inevitable given the current architecture without modifications" at most. Also, I'm missing a section on how (if) human brains manage to avoid hallucinations in this. Also, it doesn't have to never hallucinate, it just has to hallucinate less than we do.
> hallucination is defined as inconsistencies between a computable LLM and a computable ground truth function. With this definition, you can trivially prove the titular sentence - "hallucination is inevitable" - is untrue. Let your LLM have a fixed input context length of one byte. Continue training the LLM until such a time as it replies to the input "A" with "yes" and all other inputs with "no". Define your computa…
This is basically the same as saying "train your LLM until they never hallucinate", which reduces your claim to a tautology: an LLM trained not to hallucinate does not hallucinate. The trick is making that happen.
Organisms that evolved to perceive true reality instead of the "user interface" have smaller chances of survival. Donald Hoffman
For others here is a TED talk.
The models are just generating probable text. What’s amazing of how often the text is correct. It’s no surprise at all when it’s wrong Their bold confidence to be flat out wrong may be their most human trait
paulsutter said: > Note that this is the single most human attribute of LLMs It might be if LLM hallucinations looked like or occurred at the same frequency as human hallucinations do, but they don’t.
Compare the hallucination behaviour of a 7B model with a 70B model and then GPT4 and you'll quickly see the frequency of hallucinations right now doesn't mean much.
It seems a stretch to call it "inevitable". "Inevitable given the current architecture without modifications" at most. Also, I'm missing a section on how (if) human brains manage to avoid hallucinations in this. Also, it doesn't have to never hallucinate, it just has to hallucinate less than we do.
The particular pathology of LLMs is that they're literally incapable of distinguishing facts from hallucinations even in the most mundane circumstances: if a human is asked to summarize the quarterly results of company X, unlike an LLM they're highly unlikely to recite a convincing but completely fabricated set of numbers.
You have to very carefully ask your question for it to not make things up. For example don't ask "how do I do this in in x?". Ask "can I do this with x?" These "AI" s are like "yes men". They will say anything to please you even if it's untrue or impossible. I have met people like that and they are very difficult to work with. You can't trust that they will deliver the project they promised and you always have to dou…
What I find frustrating that it's increasingly challenging to have DeepL translate thou -> du, as this was my go-to "hack" to overcome the incompatibility of the English language due to its missing features.
To somewhat remedy the "yes man" problem, one needs to become a pedantic mathematician about posing your questions and I don't believe that LLM technology alone is capable of overcoming it entirely. As silly as it sounds, I must concede to the existence of "prompt engineering" as I can forsee the development of abstractions aimed to decompose questions for you.