[andrewla]

Oh, now I see where we have an actual difference of opinion. I don't think you can deny that even Finnegan's wake proceeds one token at a time; your interpretation of it may require more context or out-of-order interpretation, but that's just as true when observing text in German or Japanese, which have word ordering constraints that are alien to English speakers. How it was written is irrelevant; all we can observe is how it was presented. Of course we can observe each other's inner life, but we do so one token at a time, even if the process of producing each token is done (internally or actively) via a backtracking or zeitgeist approach.

You seem to believe, on a more fundamental level, that LLMs are simply not capable of producing text that has deeper connections to itself or represents abstract thoughts. In my opinion, 99% of text written by humans does not show this, just as 99% of text produced by LLMs does not show this, but both have the capability, and I don't believe that LLMs are constrained in such a way that they can never do this.

[delis-thumbs-7e]

I am no linguist, but I believe this is referred to as surface structure and deep structure. What you describe is a line of text that is grammatically somewhat adequate to pass as readable and you treat all text the same. when we read the text, we decipher meaning out of according to word references and syntax - just like with Python or C++. However, if this were solely the case, we could not read Finnegan’s Wake. Probably vast majority of modern poetry would be unread by anyone, as would be pretty much all major works of philosophy Kant onwards. Deep structure is what according to Chomsky et al. gives meaning to the language, ie. somewhat logical structure behind the mere words. English word strict has order, need it but actually not does. You skibidi rizz swag grok brah also, barely. We use language in the extended meaning of the word to create a model of the world and somehow the past riverrun skibidi transmits that model to others. This is what I believe Bender also tries to say in their paper.

Now, you could claim that LLM’s have this deep structure, create models of the world and are basically just like us, and certainly many here are adament that this is the case, being aghast how someone can “insult” LLM’s by calling them parrots. However, there really is not much proof to back up this belief. Usually LLM’s seem to copy existing surface structure from whatever source, and when it deviates from these patterns, it usually becomes incomprehensible. There’s much hoopla about LLM’s solving hard maths, but it seems even there they are mostly generalising from vast amounts of training data, rather than actually reasoning: https://arxiv.org/pdf/2410.05229

[andrewla]

What I'm saying is that we cannot observe deep structure. All we can observe is the surface structure. If we have access to the author or speaker we can ask clarifying questions, etc., but even that just produces more examples of surface structure.

But even here, in the Chomsky sense, LLMs clearly exhibit deep structure because they can write at length in an internally consistent manner. Importantly, early generations, GPT-2 and even GPT-3, did not definitively have this property; roughly, an object that was green at the beginning of a paragraph might not still be green at the end of the paragraph. This was strong evidence for lack of a world model.

Current LLMs do not show this behavior. We cannot prove that LLMs have a world model, in fact, their architecture seems to rule it out, but looking at it from a linguistic standpoint, they produce language in a manner as if to reflect a world view. That is, we cannot easily falsify the statement "LLMs somehow represent a world model"; and current examples of "disproving" their world view are so convoluted that even humans do not appear to (observationally) have a world view either.

I'm not making a claim as to LLMs having genuine deep structure or consciousness or anything like that. I'm claiming that we can't rule out current or future capabilities or make structural assumptions. Yes, they generalize from their training data, but unless you can make very specific claims about the kinds of things that they _cannot_ do, I can't take this statement as particularly compelling.

[delis-thumbs-7e]

I disagree on your definition of deep structure. We can indeed deep structure, that is my whole point, without it Finnegan’s Wake would be impossible to read. This really is something LLM’s cannot do, at least for now. And if you look at the paper I linked, if they had a world model they would not get confused on mathematics when we change the values of variables. So there seems to be plenty of things LLM’s cannot do.

And basically the burden of proof is on your side. As often is the case, discussions on AI start to resemble discussions centred on religion and belief in god. Your only line left here is “yes but you cannot prove LLM’s are conscious” - which I can’t, since proving that Unicorns do not exist is pretty senseless activity.

So basically I’d say LLM’s are stochastic parrots (or whatever term you want to use) until we can very definitely prove that they indeed think and have some sort of consciousness.