However, stating that LLMs predict text, not facts - and that LLMs don’t “think”, is enough to start a debate.

Yeah the issue is anthropormphization and the “cool” factor of demos, vs actual experience.

True. But most of the training corpus was not linguistically-sound gibberish; it was actual text. There was knowledge encoded in the words. LLMs are large language models - large enough to have encoded some of the knowledge with the words.

Some of. And they only encoded it. They didn't learn it, and they don't know it. It's just encoded in the words. It comes out sometimes in response to a prompt. Not always, not often enough to be relied on, but often enough to give users hope.

What we really need is a language model coupled to a knowledge model.

People say stuff like this a lot, and I never know what to do with it. I'm not sure what specifically I'm supposed to get out of these distinctions.

And that's enough to be a problem, given how people are trying to use LLMs. They're trying to use it like a human who "knows" (whatever that means). And whatever an LLM actually does, it doesn't do that.

If the claim is only "LLMs don't work like humans" and "LLMs make a lot of factual and logical errors", then that seems to be correct as far as it goes, but it doesn't go very far.

>Some of. And they only encoded it. They didn't learn it, and they don't know it. It's just encoded in the words. It comes out sometimes in response to a prompt. Not always, not often enough to be relied on, but often enough to give users hope.

And some of the pieces of "knowledge" in the training corpora were wrong, lies, or bullshit themselves.

Garbage in, garbage out.

Well yes, but how is this insight helpful? People aren't just looking at the name. Have your missed the last 10 months of ridiculous breathless coverage of LLMs?

> The trouble is people think it goes deeper when the knowledge modeling part

People are literally calling this shitty text prediction algorithm "AI". The discourse is completely out of hand.

https://en.m.wikipedia.org/wiki/Dartmouth_workshop

> An attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves.

LLMs fit that definition well. They're not AGI, but I think it's fair to call them AI.

What I think is the problem is that when people hear "artificial intelligence," they don't think of chess programs anymore; they think of something like Skynet from the Terminator. Or, maybe now they think of ChatGPT.

If they don't get to define "AI", who does?

Actually the hint is not in the name, that only tells a portion of the truth. If LLM were trained on textbooks teaching foreign languages from every other language, the name would suffice. But it does not, LLMs are trained on wide variety of texts, many of them social media, blog, and general web posts. The larger web is an ocean of bullshit, and that is where the bullshit within LLMs comes from.

'Textbooks are all you need' ought to be how LLMs are trained so people can use them for their "knowledge work" without bullshit ramifications.

But in many spheres of knowledge there are still a lot of contradictory opinions in textbooks. Why did Rome fall?, What's the best X? In science there have been many schisms: quantum theory, tectonic plates, GMO/Organic etc.

Writing code is probably one of the few areas where just about all content is 'good' - there may be a lot of bad or poor performing code in books and online but the vast majority of it 'works'.

For many other fields, introductory textbooks even at the university level may abound in oversimplifications, and it is only in later years of study that one begins, by reading specialized papers and monographs, to grasp how complicated the facts really are.

Not only that: there's also the fact that LLMs do not "fact check" their output against any data source. They just generate text based on relative frequencies of different kinds of text in their training data. They don't even have the concept of text being related to other things in the world; in fact they don't even have the concept of "other things in the world".

Any text generator built that way will output bullshit, no matter how accurate its training data is. "Bullshit" does not mean the output is necessarily wrong; it means, just as Frankfurt said, that the thing generating the output doesn't care about whether it's true or not. LLMs meet that criterion: they don't care about anything other than relative frequencies in their training data.

Obviously, they're better at grammar than many other things, but it's clear that they contain a vast amount of knowledge. They contain the information that the sky is blue and firetrucks are red, which isn't a matter of grammar or coherence, but is simply information about the world.

Is it though?

There's nuance in what we call knowledge, but to me, knowledge is not uni-dimensional. What I mean by that is, I know when I _know_ something, and I have learned that that is not a trivial skill.

How do I know I _know_ something? Usually because I have searched for evidence from multiple angles / dimensions / senses. When multiple independent observations of quality agree, and my conclusion matches the result, it is safe to say that I know that fact.

An LLM may have multiple 'dimensions' in a linear algebra sense, but does it have independent information?

> How do I know I _know_ something? Usually because I have searched for evidence from multiple angles / dimensions / senses. When multiple independent observations of quality agree, and my conclusion matches the result, it is safe to say that I know that fact.

I wanted to thank you for laying out this perspective. In my experience, in discussions about LLMs, a lot of time people get caught up on words like 'know' which work well for talking about humans, but don't cash them out in definitions I know what to do with. This is a really content-ful of sharing what you mean by your claim.

This is technically true, but no one actually does much of it consciously. But this is how you picked up every skill and understanding, including language, since you've been born.

And this - the unconscious fusion of correlated patterns - is pretty much what LLMs undergo in training.

EDIT:

Also, a lot of confusion is created by using terms like "a fact" or "a conclusion" or "a result", as if they were fixed points in space, fully defined mathematical symbols. But there exist no such thing. There is no binary "know/don't know", there is no binary "result match/does't match the conclusion". Those are all continuous quantities, that get rounded off for convenience.

Correct. But it does mean they're about "language" rather than knowledge. It's meant to read, parse and generate language, no matter if it includes prose, facts, knowledge or whatever.

> Clearly, they aren't as likely to generate "Colorless green ideas sleep furiously" as "It's important to feed your dog".

Clearly, as the training set probably contains much more instances of the latter, than the first example. And if it's more probable to generate, that's the text it'll generate. It doesn't understand why or how, just that it's more likely, so that's the way it's going.

Not really. It was invented because they were trained on natural language corpuses. It wasn't a deep, philosophical concept.

It continues to be used for models that are doing completely non-languagy stuff, such as robotics, although I expect the growing use of 'foundation model' or some other less-confusing term will likely replace it.

If it wasn't, what would we be doing here?

To some extent, those are kind of the same thing.

They are not the same thing. They are not even "kind of" the same thing, for any reasonable definition of "kind of." That should be obvious, because a lie is language, which is pretty close to the opposite of knowledge.

I mean no offense, but your hedging words are doing so much work that I think your statement counts as bullshit.

Edit: thinking about it more, I think your comment shows the particular kind of bullshit that often characterizes tech people's thinking and is so useful for hyping shit (e.g. take something hard, lossily reformulate it into something easier but vaguely similar, solve the easy thing, then declare you solved the hard thing).

Knowledge is... concepts and their relationships, typically with the implication of being grounded in objective reality (although one can be knowledgeable about things like, say, Tolkien, or Star Wars, or fantasy tropes more generally).

.

Language is (a particular way of representing) knowledge, in about the same way that a particular .cpp file is (a particular way of representing) a linked list or whatever.

So, in your view, are .cpp files and linked lists "kind of" the same thing?

Language can be used to express some kinds of knowledge, but they're not even "kind of" synonymous. Language can express a lot of things that aren't knowledge, and there's knowledge that can't be expressed with language (qualia, at least).

All the more and less formal models we invent for describing "language"? This is just fitting curves to a snapshot of reality. So sure, "a lie is language, which is pretty close to the opposite of knowledge" - a lie can be expressed in the formal model. But the truth (as subjectively understood) is much more likely to be spoken or written down in practice.

Communication isn't random. Language isn't random. Learning it is picking up the patterns.

>>>> But it does mean they're about "language" rather than knowledge.

>>> To some extent, those are kind of the same thing.

>> They are not the same thing.

> Of course they are the same thing.

Language and knowledge (both broadly construed) are obviously not the same thing. Language can encode non-knowledge like nonsense or lies, and there is knowledge (e.g. of the experience of qualia) that can't be expressed in language.

I think the point up-thread is true: something that knows only about how language is unreliable source of knowledge. Even if all its input is true knowledge, it can still blindly combine that input in linguistically plausible but false ways.

you can use language as a proxy for knowledge in some cases, but doing so guarantees generating bullshit, which is the entire point of the OP.

    it does mean they're about "language" rather than knowledge
    

They can emit statements that happen to be true because the probabilities fall that way (as a sibling comment points out.) But it's not "justified", because models don't have any ability to remember or cite sources, or even internally classify tokens as true or false (it's all just weights.)

(And it's not a "belief", either, because there's no intentionality, but that's a more slippery concept.)

Therefore, philosophically speaking, it is fundamentally impossible to acquire knowledge from a LLM unless you verify ("justify") a statement via some secondary, more trustworthy channel.

FWIW, although 'justified true belief' is classic in the sense of being an old definition of knowledge, I don't think any significant number of modern philosophers would use it.

But yeah the problem what you mean by "knowing." For example, I am perfectly happy to say that a LLM "knows" words just like I "know" words, or that it "knows about" very popular topics in a loose sense.

But that's a very different sense of "knows" from the kind of specific propositional knowledge we're usually interested in when using a LLM as a search engine, fact checker or source of information!

I’m clearly keeping track of how strongly I believe a variety of things. Not how I started believing them to begin with, however, so I lack the justification. And in a lot of cases, the lost justification would have been: “Because someone else said it, so now I’m aping them.”

But I think almost none of them would ascribe "knowledge" to a statistical engine that output propositions of random truth values weighted by how many examples of that proposition it's seen in it's training data.

You can view the logprobs in the API and literally watch it choose to say different things based on it's internal dice rolls.

Knowledge without a feedback mechanism for verification is prone to veer off course.

Right now the feedback mechanism is based on human reinforcement.

What if we put an LLM in an android’s body and let the machine interact with the world through the same senses that we do?

Ultimately though this is all about the definition of words. Humans find the distinction between "true" and "false" to be important and meaningful; animals don't care. AIs don't care (being incapable of caring), but humans care very much about the propositional veracity of statements that AI makes.

The important thing is that we're on the same page about what's actually happening with these systems and how they work and don't slip into magical thinking or assume that just because a system is facile with language it automatically has a sense of truth or right or wrong.

This is only going to get more complicated as models get more sophisticated.

https://www.youtube.com/watch?v=yhfl7kasjZc

There are a quite a few videos with the same reaction. There does seem to be something going on.

There is a setting right there called temperature, no?

I think there's a strong argument to be made (that has been made) that much of our knowledge capability is built on language - that language did not simply allow us to communicate our knowledge, but actually allowed individuals to think in a new way.

I certainly agree that current LLMs have no regard for truth and logic at their core, and there are clearly areas where their ability to read and write does not translate to producing true statements. LLMs can't build things out of language like humans can.

I am less sure that LLMs don't have a long way to go before we reach the limits of what modeling language can do.