LLMs _only_ produce hallucinations. Some of them align with the someone's version of the truth (assuming that is what the user wanted)
Please define what you mean by hallucination.
There are no separate modes of operation when GPT-4 is telling me about Javascript's "console.log" method, and when it's telling me about Rackets non-existing "sum" function. The network has no way of telling that the strands of probability are wearing thin under it's feet. The engine spins and churns out another token and we deem it "correct" or "incorrect".
The OP's assertion is "make up stuff" is all LLMs really do and sometimes the stuff they made up matches reality - and sometimes it doesn't.
Let's ask LLM to generate some text given a prompt. It replies with a bunch of words.
Let's ask LLM to generate some text given a different prompt. It replies with a bunch of other words.
In both instances, the model worked correctly, and as defined. It produced a bunch of words based on context, statistics and some forced randomness.
If we call one of them an hallucination because it does not match how we would expect a human to respond, we have to accept the other is also an hallucination because it was generated in exactly the same way, using exactly the same model.
If you disagree, ask the LLM to 'Compose a single-line avant-garde poem about rainbows, without mentioning rainbows'. Look at the reply. How would you tell whether it's hallucinating or not?
Also my objection wasn't that it hallucinates, my objection was to the OP's point that it only hallucinates. If I ask it what is 2*2 and it answers 4 would that be a hallucination or not?
Why would an LLM reciting a fact correctly be hallucinating?
Am I now capable of predicting the future?
Suppose I wrote the book to be as banal (i.e. highly probable) as possible.
Am I predicting the future now? And, how impressive is it?
If you write a book of random predictions without any insight the vast majority of them will be false, so even if few of them are right it is not impressive nor anyone would say you're capable of predicting the future.
In comparison, the OP states that GPT-4 predictions are 97% correct. And yes, I would say that is pretty impressive. If 97% of anything I say about the future was correct I would be considered a wizard and probably be a billionaire.
Isn't this step for step exactly what Nostradamus did?
And you just hallucinated that! It's just, if there was a system that only talks in an oddly definitive hypotheticals, they can be correct about a lot, and GPT-* are exactly that.
Fact of the matter is that sota LLMs are highly accurate predictors for many topics, certainly above any living human in terms of total AUC of correct predictions on fact based questions. Some humans are better on certain topics, but noone can match total AUC since LLMs has such breadth.
LLMs are fine - people are attributing superpowers to them when they discuss hallucinations.
LLMs do not "think". They created the correct text as they were modeled to do.
The observer feels that the facts are wrong.
That's an issue with the observer, not the model. The model was never trained for facts it was trained for text.
So if it's "hallucinating" a probable continuation which asserts something which is [incidentally] understood to be completely wrong or not in the source material by humans, it's going through exactly the same process to arrive at a continuation which [incidentally] is understood to contain only accurate statements or valid summarizations
Which kind of make sense; as LLMs have almost no memory, just an instinct to respond and some instinctual responses (the result of “training”, which is also a bad metaphor; only “in-context learning” is analogous to training/learning for humans, what is called “training” is guided evolution of frozen instincts) and whatever is in their context window. And lack of memory plus a prompt to respond is a major context where confabulations happen with humans (these are specifically called “provoked confabulations.”)
We are so used to assuming that good text means good thought. "emergent" behavior is assumed to be full on thinking and learning systems.
Calling it Hallucination is absurd.
Humans hallucinate. Humans can be deceived or deceive because there is an active effort to hide something.
An LLM is producing the next best token. Its either always hallucinating or never hallucinating.
In fact I would even stop talking about the profound concept of truth.
Long time ago I read this wisdom: «God's omnipotence raises the intriguing possibility that two contradictory religions could both be true.» Buddha says anyway, that everything is empty. Zen Buddhism employs koans to dismiss truth.
LLMs sometimes hit the mark and this makes many people ask what is truth and what are humans?
After enough philosophy, this is close to a solved problem. There are so many breakdowns of what is truth, including adding words to describe different truths.
If you are hung up on a single universal definition, you are most likely limited by language, a human construct rather than the understanding of Truth. We all know what Truth is, we know that 2+2=4. As Laozi would say 'If you try to grasp it, it will slip". There is a time and a place for analysis, but there is also losing the forest for the trees.
I partly disagree because not all truth is subjective, but rather, objectively true.
It's really around the epistemological limits of LLMs, and truth. Probably more Kant than Peterson?