In fact, in general, in any non one-on-one conversation the answer "I don't know" is not useful because if you don't know in a group, your silence indicates that.
Three logicians walk into a bar. The bartender says "what'll it be, three beers?" The first logician says "I don't know". The second logician says "I don't know". The third logician says "Yes".
Both of the first two logicians wanted a beer; otherwise they would know the answer was "no". The third logician recognizes this, and therefore knows the answer.
This way is logically most efficient to work and involve the least communication.
() In a lecture by the mathematician & author Sarah Hart.
Logically speaking, the second bar tender could have thought to himself "no I don't want any beer, but one of these two other guys may want to double fist" and so there is really no way for the third logician to answer in the affirmative.
I believe that you can also cause something a bit like a transient dysphasia by giving them bad inputs as well, so there is that on the language production side. However there's still nothing that pertains to the experience aspects central to what hallucinations actually are.
I don’t think confabulate matches as well as it implies confusion or mixture of different ideas.
ChatGPT isn’t confused, it’s making things up. It’s trying to bullshit as best it can in hope that what it makes up convinces its user.
Fully agreed that "hallucination" is a bonkers word for it — sensational and melodramatic. But few people know what a confabulation is, and moreover it's an overly complex way to describe the phenomenon.
The LLM is making something up. It's a fabrication.
It's not fanciful; it's not spooky; it's mundane, as it should be.
The problem with "hallucination" and "confabulation" is that they both imply a consciousness.
It's still just probabilistic babble. We were doing this with markov chains.
But given how often toddlers are "taken care of" by planting them in front of youtube :|
If you have to put a show on TV to give some songs to sing along to or to distract them while you're making lunch, I'm not judging you, and I think it's best to put this content on a gradient rather than black and white.
Parenting is easier if you have 5 family members in walking distance and they also have similarly ages kids who can all play together.
Focusing on such a tiny thing and blowing it up into a huge negative out of context of their rich, busy, and safe lives is really out of hand.
I used to cut all those things to shape with youtube-dl and Audacity; we have a library of a good hundred+ of sanitized songs to play, but with modern world hating files and anything offline, it turned out to be quite a hassle to keep the practice up.
Even if you don't show children videos, but want to play some music, YouTube is still the least-hassle, least-bullshit music stream player (arguably still it's main use for adults, too). Ain't anyone got time to deal with Spotify's ever more broken app. And this is the limit of technical skill of almost all parents. They can't exactly run SponsorBlock in YouTube's mobile app (and paid YouTube doesn't help here either, surprise surprise).
Not making excuses (though I'm not really blaming parents for this) - just saying how things actually are.
At some point fairly recently they added a "I don't know the answer" button to the email, but it's much less prominent than the main call-to-action.
A lot more of LLM hallucination is it getting the context confused. I was able to get GPT4 to hallucinate easily with questions related to the distance from one planet to another, since most distances on the internet are from the sun to individual planets, and the distances between planets varies significantly based on their locations in cycle. These are probably slightly harder to fix.
I've noticed that while this can help to prevent hallucinations, it can also cause it to go way too far in the other direction and start telling you it doesn't know for all kinds of questions it really can answer.
This isn't true. There are many contexts where it is true but it doesn't actually generalize they way you say it does.
There are plenty of cases where experts in a non-one-on-one context will express a lack of knowledge. Sometimes this will be as part of making point about the broader epistemic state of the group, sometimes it will be simply to clarify the epistemic state of the speaker.
Because expecting a behaviour, like knowing you don't know, that isn't represented in the training set is silly.
Kids make stuff up at first, then we correct them - so they have a way to learn not to.
The problem is that curating data is slow and expensive and downloading the entire web is fast and cheap.
See also https://en.wikipedia.org/wiki/Cyc
Maybe if you trained a small base model to know it doesn't know in general and THEN trained it on the entire web with embedded not-knowing preserving training examples, it would work?
https://www.quantamagazine.org/tiny-language-models-thrive-w...
Google has one of these already, with an LLM that was trained on nothing but weather data and so can only give weather-data-prediction responses.
The 'knowing it doesn't know things' part is much harder to get reliable, though.
This is not a brain. The best analogy is an English major.
They are good at language, not reasoning.
Humans see language and think reason. It seems we can’t separate the two.
Dunno. GP?
What I'm pushing at is not that this linguistic ability naturally leads to the LLM behavior we're seeing and calling "hallucinating", just that LLMs may capture some of how humans process language, differentiate semantics, recall terms, etc, but without the mechanisms that enable rationally grappling with the resulting semantics and propositional (in)coherency that are fetched or generated.
I can't say this is very surprising—most of us seem to have thought processes that involve generating and rejecting thoughts when we e.g. "brainstorm" or engage in careful articulation that we haven't even figured out how to formally model with a chatbot capable of generating a single "thought", but I'm guessing if we want chatbots to keep their ability to generate things creatively there will always be tension with potentially generating factual claims, erm, creatively. Further evidence is anecdotal observations that some people seem to have wildly different thresholds for the propositional coherence they can spot—perhaps one might be inclined to correlate the complexity with which one can engage in spotting (in)coherence with "intelligence", if one considers that a meaningful term.
I don't think Wittgenstein would agree, first of all, that there is a "natural logic" to language. At least in the PI, that kind of entity--"the natural logic of language"--is precisely the kind of weird and imprecise use of language he is trying to expose. Even more, to say that such a logic "allows" for anything (like metaphors) feels like a very very strange thing for Wittgenstein to assert. He would ask "what do you mean by 'allows'"?
All we know, according to him (in the PI), is that we find ourselves speaking in situations. Sometimes I say something, and my partner picks up the right brick, other times they do nothing, or hit me. In the PI, all the rest is doing away with things, like our idea of private language, the irreality of things like pain, etc. To conclude that he would make such assertions about the "nature" of language, of poetry, whatever, seems like maybe too quick a reading of the text. It is at best, a weirdly mystical reading of him, that he probably would not be too happy about (but don't worry about that, he was an asshole).
The argument you are making sounds much more French. Derrida or Lyotard have said similar things (in their earlier, more linguistic years). They might be better friend to you here.
The texts are quite different, this is true, but I don't find them contradictory. Whereas Tractatus was almost a facetious or flippant rejection of the millenia-long project to agree on a philosophical subset of language suitable for rigorous philosophy (although it continues today in the form of analytical philosophy), PI basically says "well we don't need to throw the baby out with the bath water", which I think is a fantastically mature response to a flawed tool that's still the best we have to reason about the universe. So: not contradictory in evaluation of fundamental compatibility of non-formal language for the formal needs of propositional philosophy, but perhaps contradictory in implied reaction to this realization.
This sums up the last decade remarkably well.
Probably true, but if you have quality, organized data, you will just want to search the data itself.
Only a response makes it clear one has read and acknowledged the question and sometimes there are people expected to know, if they don‘t the should say so.
I mean, it's inherent to LLMs to be unable to answer "I don't know" as a result of not knowing the answer. An LLM never "doesn't know" the answer. But they'll gladly answer "I don't know" if that's statistically the most likely response, right? (Although current public offerings are probably trained against ever saying that.)
On the same lines as why people argue if a tree falling in a wood where nobody can hear it makes sound because some people implicitly regard sound is the qualia while others regard it as the vibrations in the air.
An LLM should have no problem replying "I don't know" if that's the most statistically likely answer to a given question, and if it's not trained against such a response.
What it fundamentally can't do is introspect and determine it doesn't have enough information to answer the question. It always has an answer. (disclaimer: I don't know jack about the actual mechanics. It's possible something could be constructed which does have that ability and still be considered an "LLM". But the ones we have now can't do that.)
How often do the words "I don't know" get uttered in books, papers, articles, stack overflow, or any other resource of knowledge?
I have some representation in my mind; as someone who doesn't have aphantasia, this representation comes with a mental image. Tower? Tall, linear, and in my case a skyscraper by default. Eiffel Tower? Paying attention to the extra context, the first word transforms the second into the eponymous structure. Model Eiffel Tower? Now the context makes it a tchotchke, probably 10cm tall. Lego model Eiffel Tower? The 1-ish meter tall one on display in the Lego shop.
Is my "knowledge" the abstract representation that is in my case connected to a mental image? The attention process can reasonably be considered as developing a vector in a very high dimensional concept space, and the next token comes from what would best suit the current location in that high dimensional space. It's entirely possible that the concept of "ignorance" is linearly separable within that space (much as gender is, see the word2vec trick with "king" - "queen" ~= "man" - "woman"), and the corresponding "ignorance" vector can be associated with the sequence of words "I don't know". I think it would take actual research into the internal vector space to answer that, and while I'd like to do that research, I have some higher priorities right now.
Is "knowledge" any belief? Any true belief? Any justified true belief? https://en.wikipedia.org/wiki/Gettier_problem
I take the position that there is no such thing as knowledge, and instead the best we can have is belief.
But then, what is "belief", and can the information within an LLM said to meet whichever definition you give?
Answering "I don't know" because it a likely response to a particular string is completely different from being aware that one does not know the answer and saying so.
Both motivations lead to the same outcome, but they're unrelated processes. The response "I don't know" can represent either:
1. The most likely answer to a particular question, based on statistical data; or
2. An expression of an agent's internal state.
Figuring out that distinction is perhaps one of the most important questions ever raised.