LLMs are maybe a type of mind. They may carry and operate on basic thoughts, ideas, books, pages, sentences, words, letters….
They can probably problem solve and write code that runs sometimes. What they know about the world was told to them through stories, so whatever way we see it, they will too, maybe.
I don’t have to imagine anything in mind to process thoughts. All I need is an attention loop.
Knowledge is justified true belief. If you park your car on 5th avenue, you know it is parked there because you witnessed yourself parking it there. You have strong justification for your belief that you parked it on 5th avenue.
But if you then go to work and your car is out of sight, your justification fades. If you look out the window and see your car parked there, you reaffirm your belief. However, if you look and your car is not there, you realize you don’t know where your car is. For a period of time you were mistaken in your belief about the whereabouts of your car. During the time when your car was out of sight, you went from knowing to not knowing where your car was, but nothing changed inside your mind. To update your belief, you needed new information about the world.
An LLM has no justification for any of the information in its training set. It has no experiences with which to connect words and facts to real world objects and events. Therefore it cannot be said to have knowledge of any of it.
[1] https://plato.stanford.edu/entries/knowledge-analysis/#KnowJ...
But I won't do it. Instead, I'll just replace a few words in your last paragraph to reveal that it doesn't discredit LLMs as much as it seems.
* * *
But these historians know nothing about the distant past. They might capture some features of the past as patterns that help them better predict where to start an archaeological dig. But that is not the same as knowledge of the past. They have no sensory contact with the past so they are unable to sunthesize true knoledge about it. This is why they are prone to hallucination.
Two things:
Justification for belief isn’t a binary state, it’s a spectrum. When we have a small amount of evidence supporting a belief, we have a small justification. When we have a large amount of evidence, we have a large justification. We describe this in terms of our confidence in our beliefs.
Without getting too far off into the weeds, this extends into all areas of science. We can’t see, touch, hear, smell, or taste electrons (9V battery tests notwithstanding), but we can gain confidence by replicating experiments with them. Working with a cathode ray tube, for example, we can manipulate the voltage on the plates and directly observe the deflection of the point of light on the screen. We can use the equations from a physics textbook, combined with voltage measurements on a multimeter, to make predictions about the electron beam’s deflection and then test them directly on the device.
Secondly (and crucially), the text consumed by LLMs isn’t about the world. It’s text of unknown provenance. Some of it may be about the world, some of it may be complete fiction. The LLM has no basis whatsoever for fact-checking the text. All it can do is perform a statistical test on the commonality of a text and then infer from that. Unfortunately that does not get you any closer to reality. A falsehood repeated many times does not become the truth.
As for historians? They have way more than text to go on. They have physical evidence (artifacts), they have radioisotope testing, they have DNA testing. They have geographic locations where texts were found. None of this real world sense data is available to an LLM.
The other problem with discounting LLMs based on the form of sensory input that's available to them is that there is no qualitative comparison of validity of different senses. Sure, getting more visual input counts as more evidence about the world (bar hallucinations). But can part of it be substituted with text without losing the quantitative benefit? Can all of it be substituted with text? You seem to answer "no", but I think it's an oversimplification. Not that I have a definite answer.
The one thing that LLMs cannot currently do is design and perform experiments - or if they can, then they can't integrate the results. I'd look for ways LLMs cannot reach some level of justification for beliefs there. But I would not presume that this level is automatically lower than what humans can reach. I have not compared the size of a human training set to that of an LLM.
On the other hand, if you're talking to someone you've never met over the internet and they're telling you the shed in their back yard is painted red, what confidence do you have in that? How do you even know you're talking to a person?
LLMs are trained on large amounts of text that's just been scraped off the internet. Some of it may be high quality (if you want to call Wikipedia high quality) but much of it (Reddit) is not. When you ask an LLM chatbot a question, how do you know you're getting the Wikipedia answer and not the Reddit one? You don't. It's all mashed up. Until LLMs are able to reliably provide citations for their answers, they ought to be relegated to the epistemic dustbin.
Yet I see high schoolers relying on them all the time. They will ask ChatGPT a question and write down its answer verbatim. No second-guessing, no fact-checking, nothing. Blind trust. It's appalling!
My professors have even told me they've had students hand in papers full of citations. Turns out, all the citations were completely fabricated by ChatGPT. One professor even said the paper a student submitted cited HIM, the professor receiving the assignment, for a completely made up paper that never existed. When confronted, the students admitted all of the papers were ChatGPT-generated nonsense.
Even worse than the Reddit problem above is the problem of LLM-generated text proliferating throughout the internet. The next generation of LLMs are going to be polluted by that nonsense. Garbage-in, garbage-out. They'll be no more trustworthy than a psychiatric patient rambling to themselves in a padded cell.
I think this is a good idea if LLMs justification was 0. It's not, and it's not clear what it is. That people talk to a storyteller thinking that they talk to a scholar is on the people, or on whoever misrepresented the interlocutor, not on the interlocutor.
It's marvelous already that we can't rule out whatever nonzero amount of justification in the case of LLMs.
We absolutely can rule it out. The justification is zero. The LLM has not and cannot verify any of the factual claims in its text. It can't even tell a claim apart from a non-claim. In legal parlance, the entire training set of an LLM is hearsay. LLMs do not produce any first party information. It's 100% third party.
Asking an LLM about the physical world is not unlike asking a historian about the past. All information they have past a certain point is 3rd party, mediated by tens/hundreds of years. Can historians say anything justifiable about the past? So can LLMs about the physical world.
If you want to rule out historians being able to justify their knowledge, I'd like to hear that.
At the same time, if you ask historians about (say) the Peloponnesian War they will tell you about the account of Thucydides. What they won’t say is that they know about it personally.
That’s a mistake a lot of non-scientists make about science. They’ll hear scientists say things like “we know this” or “we know that” and think that they are committing to some privileged knowledge of the world. They aren’t. They’re using the words “we know X” as shorthand for “our best theory about X.” In fact, many scientists identify as scientific anti-realists [1] which means they don’t believe science has any true facts about the nature of reality at all and that the only job of science is to make better predictions about the outcome of experiments.
Of course, if you take anti-realism to its logical conclusion then you have to admit that fossils are only evidence that you may find more fossils in the future, and that you have no basis to talk about the existence of dinosaurs.
Personally, I think our actual knowledge of things directly outside our personal experience is very limited. But then I want to encourage people not just to believe everything they read. People should be exploring and trying their own science experiments and trying to build things to understand how the world works.
[1] https://plato.stanford.edu/entries/scientific-realism/#AntiF...
> Historians have access to more than just texts about the past. They have access to physical evidence. Radiometric dating, artifacts, geographic data about where things were found.
Correct. That's still all indirect, "third-party" evidence. Does that make answers based on it impossible to justify?
To make this analogy even closer to LLMs, let's take historian scholars (linguists? undergrads?) who have nothing to study but books. No experiments, no carbon dating, just texts. Is their entire knowledge about the past impossible to justify?
Historians who only have books to go on have frequently said the accounts they're using should be treated as only that: accounts, not facts about what happened.
For example, even a figure as popular as Socrates is "known" only through the accounts of Plato and Xenophon. What do we know about him? The two accounts are in conflict about many details of his personality. So we can attribute very little in the writings of Plato to actual opinions of Socrates. Scholars treat these books as the opinions of Plato, expressed through dialogues of Socrates.
Am I misrepresenting your position?
But when I present an analogous situation, except with historians playing the main role, you seem to accept the justifiability of their statements on a spectrum, rather than rejecting it based on the evidence being textual and indirect.
This is the crux of the discussion, without explaining which we won't come to a mutual understanding (much less acceptance).
However, this does not happen in real life. Text does not fall from the heavens into the minds of historians. It comes in books, manuscripts, stone tablets, etc. Unlike digital text (which is the only thing LLMs have access to), physical texts have physical properties which can be examined experimentally. They can be subject to radiometric dating, they typically are found in a particular place (which is usually recorded), and they have version and copyright information which enables them to be traced back to the original. In the case where the original manuscript is lost (such as Thucydides’ history of the Peloponnesian War), we have to rely on copied manuscripts by other scribes, often made centuries later.
All of this is to say that it can be very difficult to trace a book back to its original author, especially without the original manuscript, but the evidence historians use to piece all this together is generally not in the text itself. It’s physical evidence. Exactly the thing that LLMs have no access to.
Let me protest against this broader point.
If I'm a historian, do I have to find the physical text in the original place myself, or can this information be supplied to me in the form of text? Do I have to actually see the radiocarbon dating machine work myself, or is it enough that the operator sends me an email with the results? What you describe doesn't really strike me as evidence that cannot be supplied through text - you mentioned that it's recorded yourself. It's still text about the past, maybe even with an extra level of indirection (it describes the ~present which has some information about the past).
It follows that some of LLMs statements about historical artifacts should be justifiable - those coming from the written word that our hypothetical historian would study. Is that not correct?
I would argue a small proportion of humans have verified the physics they've learned as you describe. The rest learn from textbooks and ultimately there is an element of trust that they have that most of the books are not false. Ultimately from a Bayesian perspective, I think it is possible to obtain very high confidence about certain attributes of the world even if your sources are very contaminated with fictions.
All of this is to say that it is very hard to know anything with a high degree of confidence. The confidence many people have is a false one. However, in the case of LLMs, there is no basis for any confidence whatsoever. The text is the claim and LLMs have no means of independently verifying that claim. All they can do is statistically evaluate the internal consistency of the text, and even that is fraught with difficulty. "A Lie Can Travel Halfway Around the World While the Truth Is Putting On Its Shoes."
This is a great way of expressing a thought I’ve been unable to fully articulate.
It's not trained on "the world". It's trained on a small slice of it, a few senses that are themselves slimmed down and then fabricated at parts.
To the bird that can feel and sense electromagnetic waves intuitively to guide travels, you're not trained on "the world"
>This is why they are so prone to hallucination.
No it's not. LLMs can distinguish between truth and hallucination. They just don't care to communicate that.
GPT-4 logits calibration pre RLHF - https://imgur.com/a/3gYel9r
Just Ask for Calibration: Strategies for Eliciting Calibrated Confidence Scores from Language Models Fine-Tuned with Human Feedback - https://arxiv.org/abs/2305.14975
Teaching Models to Express Their Uncertainty in Words - https://arxiv.org/abs/2205.14334
Language Models (Mostly) Know What They Know - https://arxiv.org/abs/2207.05221
The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets - https://arxiv.org/abs/2310.06824
I think most of what we ascribe to intelligence in LLMs is a kind of textual pareidolia. We infer a mind behind the text because our brains are built to recognize minds. That’s how our theory of mind works.
But these LLMs know nothing about the world. They might capture some features of the world as patterns that help them better predict what text to generate for us. But that is not the same as knowledge of the world. They have no sensory contact with the world so they are unable to synthesize true knowledge about it. This is why they are so prone to hallucination.