[Arkhaine_kupo]

> the sentence "LLMs don't think because they predict the next token" is logically as wrong

it isn't, depending on the deifinition of "THINK".

If you believe that thought is the process for where an agent with a world model, takes in input, analysies the circumstances and predicts an outcome and models their beaviour due to that prediction. Then the sentence of "LLMs dont think because they predict a token" is entirely correct.

They cannot have a world model, they could in some way be said to receive a sensory input through the prompt. But they are neither analysing that prompt against its own subjectivity, nor predicting outcomes, coming up with a plan or changing its action/response/behaviour due to it.

Any definition of "Think" that requieres agency or a world model (which as far as I know are all of them) would exclude an LLM by definition.

[ToValueFunfetti]

I think Anthropic has established that LLMs have at least a rudimentary world model (regions of tensors that represent concepts and relationships between them) and that they modify behavior due to a prediction (putting a word at the end of the second line of a poem based on the rhyme they need for the last). Maybe they come up short on 'analyzing the circumstances'; not really sure how to define that in a way that is not trivial.

This may not be enough to convince you that they do think. It hasn't convinced me either. But I don't think your confident assertions that they don't are borne out by any evidence. We really don't know how these things tick (otherwise we could reimplement their matrices in code and save $$$).

If you put a person in charge of predicting which direction a fish will be facing in 5 minutes, they'll need to produce a mental model of how the fish thinks in order to be any good at it. Even though their output will just be N/E/S/W, they'll need to keep track internally of how hungry or tired the fish is. Or maybe they just memorize a daily routine and repeat it. The open question is what needs to be internalized in order to predict ~all human text with a low error rate. The fact that the task is 'predict next token' doesn't tell us very much at all about the internals. The resulting weights are uninterpretable. We really don't know what they're doing, and there's no fundamental reason it can't be 'thinking', for any definition.

[Arkhaine_kupo]

> I think Anthropic has established that LLMs have at least a rudimentary world model

its unsurprising that a company heavily invested in LLMs would describe clustered information as a world model, but it isnt. Transformer models, for video or text LLMs dont have the kind of stuff you would need to have a world model. They can mimic some level of consistency as long as the context window holds, but that disappears the second the information leaves that space.

In terms of human cognition it would be like the difference between short term memory, long term memory and being able to see the stuff in front of you. A human can instinctively know the relative weight, direction and size of objects and if a ball rolls behind a chair you still know its there 3 days later. A transformer model cannot do any of those things and at best can remember the ball behind the chair until enough information comes in to push it out of the context window at which point it can not reapper.

> putting a word at the end of the second line of a poem based on the rhyme they need for the last)

that is the kind of work that exists inside its conext window. Feed it a 400 page book, which any human could easily read, digest, parse and understand and make it do a single read and ask questions about different chapters. You will quickly see it make shit up that fits the information given previously and not the original text.

> We really don't know how these things tick

I don't know enough about the universe either. But if you told me that there are particles smaller than plank length and others that went faster than the speed of light then I would tell you that it cannot happen due to the basic laws of the universe. (I know there are studies on FTL neutrinos and dark matter but in general terms, if you said you saw carbon going FTL I wouldnt believe you).

Similarly, Transformer models are cool, emergent properties are super interesting to study in larger data sets. Adding tools to the side for deterministic work helps a lot, agenctic multi modal use is fun. But a transformer does not and cannot have a world model as we understand it, Yann Lecunn left facebook because he wants to work on world model AIs rather than transformer models.

> If you put a person in charge of predicting which direction a fish will be facing in 5 minutes,

what that human will never do is think the fish is gone because he went inside the castle and he lost sight of it. Something a transformer would.