> Back to the question we have at the beginning: do language models learn world models or just surface statistics? Our experiment provides evidence supporting that these language models are developing world models and relying on the world model to generate sequences. Let’s zoom back and see how we get there.
The GP's comment suggests that ChatGPT-4* has not internalized this (effectively) for Chess.
* Just like how ChatGPT-3.5 is not GPT-3.5 (text-davinci-003), ChatGPT-4 is probably not the only GPT-4 model that will be released.
In a similar vein, it is almost possible to adjudicate Diplomacy orders looking only at the orders and never the map.
Given sufficient interest, complex enough board games tend to converge on the same basic notational principles.
The same argument and result exist for the different human sensory modalities - neurons and connections self-organize to have the same topology and layout as the retina (2D) and frequency / time for the audio (also 2D).
In fact, wasn't this experiment already done for Othello and LLMs recently? Wasn't there a paper where they found the internal model for the board?
In principle it could reason about any incidence structure. That is, anything where the semantics is two types of objects, and a "touching" relation between them. Lines are just all the points along them, points are just all the lines intersecting there. For the purpose of directions, a train station is just a list of all the services that go there, and a service is just the list of stations where it stops. Etc etc. A language model is free to learn and understand these sorts of systems purely as relations on symbols without ever implicitly organizing it into a geometrical representation.
This is all good news. It means Chess, Transit, Diplomacy, and many other things can fit nicely into pure language reasoning without trying to ground the language in the semantics of our physical nature with its dimensions of space and time and whatever.
What would change my mind is if, after learning the rules for Chess as string matching, it invented a word for "row" and "column" on its own.
That's interesting. I think you're saying that the rules of chess can be described as a transformation system [1] over the set of strings of chess algebraic notation?
_________
[1] A system of transformation rules, as in transformation grammars. Not as in Transformers. Just a coincidence.
The rules of chess allows you to enumerate all the possible transformations from one board state to the next. This is just a fancy way of saying all possible moves in any given board state. By induction this means that given an initial board state and a series of moves from that board state you can determine the final board state.
So this means that the rules of chess allow you to enumerate given an initial state and n plies, all possible ways of adding an (n + 1)th ply.
So if you just assume the initial board state is always the starting position, theoretically you could do away with thinking about board states altogether. Now, whether that's sensible in terms of computational complexity is another question entirely and my intuition is no.
Ah, I get you. Yeah, OK, makes sense. You can generate all legal moves from a starting board state given the rules of chess. Yep.
All the people stating that "real understanding" is significantly different than learning through inference of language are likely going to be proven wrong in the end. There's nothing special about humans that makes our thinking any more sophisticated. With enough examples and the right learning model, systems should be able to be implicitly inferred from language, just as humans infer systems from language.
If we can do it, why can't machines?
Obviously machines can theoretically do what a brain can do because a machine can theoretically simulate a brain. But then it’s not an LLM anymore.
It may be that LLM style training techniques are not sufficient to "understand" systems, or it may be that at a certain scale of input data, and some fine tuning, it is sufficient to be indistinguishable from other training methods.
Many people's sense of what qualifies as "intelligence" are too grandiose/misplaced. The main thing differentiating us from a neural network is that we have wants and desires, and the ability to prompt and conduct our own training as a result of those.
It isn’t that people’s views on intelligence are grandiose, it’s that the specific approach used has massive inherent limitations. ChatGPT 4 is still relatively bad at chess, 1 win, 1 draw, 1 loss vs a 1400 isn’t impressive objectively and looks much worse when considering the amount of processing power they are using. The only impressive thing about this is how general their approach is, but in a wider context it’s still quite limited.
IMO the next jump of being able to toss 100x as much pressing power at the problem will see LLM’s tossed aside for even more general approaches like say using YouTube videos.
That's not even remotely close to being demonstrated.
For one thing, neural networks can only approximate continuous functions.
For another, the fact that in principle there exists a neural network that can approximate any continuous function to arbitrary precision doesn't in any way tell us that there is a way to "train" that network by any known algorithm. There isn't even reason to believe that such an algorithm exists for the general case, at least not one with a finite number of examples.
These takes are so bad and pervasive on here, honestly. This is what I mean by grandiose thinking.
A machine that approximates functions, that otherwise is indistinguishable from human, is effectively intelligent like a human. Incentives, wants, desires, and the ability to conduct our own training is the only difference at that point.
No, they're just saying there are continuous functions, and then there are discrete functions, and neural nets can't approximate discrete functions, while humans certainly can (e.g. integer addition). And that even when it comes to approximating any continuous function, neural nets can do that in principle, but we don't know how to do it in practice, just like we know time travel, stable wormholes and the Alcubierre drive are feasible in principle, but we can't realise them in practice.
So please don't say it's "spiritual and mystical" in the other person's mind just because it's not very clear in yours.
Also, what the OP didn't say is that a Transformer architecture is not the kind of architecture used to show the universality of neural nets. That was shown for a multi-layer perceptron (MLP) with one hidden layer, not a deep neural net like a Tansformer, and certainly not a network with attention heads. If you wanted to be all theoretical about it and claim that because there's that old proof, someone will eventually find out how to do it in practice, then the Transformer architecture has already taken a wrong turn and is moving away from the target.
There aren't no universality results for Transformers. I mean, that would be the day! The reason that that proof was derived for a MLP with one hidden layer is that this makes the proof much, much easier, than if you wanted to show the same for another architecture.
It gets some math wrong because it doesn't understand the "systemic" aspect of math, but who's to say that with minor training tweaks, or a larger dataset, it wouldn't be able to infer the system? Humans infer systems from language all the time. To say you need some specialized form of training beyond language inference is obviously wrong when you view how humans train, learn and understand. All of life is ingestion of information via language which produces systemic understanding.
I can play digital audio that's indistinguishable from acoustic, despite it not being a smooth function in practice. Similarly, a sufficiently advanced neural net can produce intellect-like results, even if there are aspects of the structure you say may not make it so.
Honestly, the perception you and many others seem to hold is that because something is mathematically explainable in such a way that you can "trivialize" its operation, makes it not intelligence. But you hold "intelligence" in too high a regard
A language model can match "2+2" with "4" because it's approximating the distribution of token collocations in a large text corpus, not because it's approximating integer addition.
We know this because we know that language models are trained on token collocations (word embeddings) and not arithmetic functions. We know how language models are trained because we know how they're made, because they're made by humans and they're made following principles that are widely shared in academic textbooks and scholarly articles all over the place.
>> Humans infer systems from language all the time.
Humans are not neural nets, and neural nets are not humans. Does that suffice? I don't know if I can do any better than that. Humans do human things, neural nets do neural net things, and humans can do things that neural nets can't even get close to. Like, dunno, inventing arithmetic? Or axiomatizing it? Or proving that its axiomatization is incomplete. That sort of stuff. Things for which there are no training examples, not of their instances, but of their entire concept class.
>> But you hold "intelligence" in too high a regard
Where does that stuff come from, I wonder? Of course I hold intelligence in high regard. What do you hold in high regard, stupidity?
In a very broad sense, if you just mean "the human brain also just approximates some class of functions", sure. However, human brains can surely represent many classes of non-continuous functions as well (tan, lots of piece wise functions, etc). And, crucially, some of these are necessary for our physical models of the world. So, if neural networks are limited to only representing continuous functions, that is a strong indication that they are fundamentally unable to mimic the human mind.
> You think there isn’t some mathematical model under the hood of how the brain works too?
Of course it does. I do believe that the mind is simply a program running on the physical computer that is our brain. And I am sure that some day we will be able to create an AI that is human-like, and probably much better at it, running on silicone.
That doesn't mean that we should believe every program running on silicone, despite somewhat obvious fundamental limitations, is going to be the next AGI any day now. That's all I'm trying to point out: neural networks are not a great model for AGI, and backpropagation/gradient descent as a training algorithm even less so.
This is not true at all
You can teach someone chess without language
The idea of "reinforcement learning" is just a rehash of "Hebbian learning". It works for some things, but you can't explain language acquisition is pure reward function and stats.
After spending more than a year babbling nonsense and discovering a tiny bit more every time about the meaning of certain combinations of phonemes based on the positive or negative response you get.
> You probably did not need to crash a car for 10k generations before finally making it down the street, nor simulate it in your head.
Are you sure we don't simulate in our head what would happen if we drove the car into the lamp post / brick wall / other car / person, etc.? I find it highly unlikely that this kind of learning does not involve a large amount of simulation.
> There's a lot you can do unreasonably well despite virtually no prior experience.
That's true, but there's a lot we can't do well without repetitive practice, and most things that we can do well in a one-shot fashion depend on having prior practice or familiarity with similar things.
>Are you sure we don't simulate in our head what would happen if we drove the car into the lamp post / brick wall / other car / person, etc.?
You left out the 10k times part. You're ignoring the huge training data sizes these models need even for basic inferences. No, I don't think it takes all that much full scale simulation to distill car speed as a function of pedal parameters, and estimate the control problem needed.
In many instances, humans can seemingly extrapolate from far less data. The algorithms to do this are missing. Training with loads of more data isn't a viable long term substitution.
Depends. In principle, you can't learn an infinite language from finite examples only and you need both positive and negative ones for super-regular languages. Gold's result and so on. OK so far.
The problem is that in order to get infinite strings from a human language, you need to use its infinite ability for embedding parenthetical sentences: John, the friend of Mary, who married June, who is the daughter of Susan, who went to school with Babe, who ...
But, while this is possible in principle, in practice there's only a limit to how long such a sentence can be; or any sentence, really. In practice, most of the utterances generated by humans are going to be not only finite, but relatively short, as in short "relative" to the physical limit of utterance length a human could plausibly produce (which must be something around the length of the Iliad, considering that said human should be able to keep the entire utterance in memory, or lose the thread; and that the Iliad probably went on for as long as one could stand to recite from memory. Or perhaps to listen to someone recite from memory...).
Obviously, there are only a finite number of sentences of finite length, given a fixed vocabulary, so _in practice_ language, as spoken by humans, is not actually-really infinite. Or, let's say that humans really do have a generator of infinite language in our heads, but an outside observer would never see the entire language being produced, because finite universe.
Which means that Chomsky's argument about the poverty of the stimulus might apply to human learning, because it's very clear we learn some kind of complete model of language as we grow up; but, it doesn't need to apply to statistical modelling, i.e. the approximation of language by taking statistics over large text corpora. Given that those large corpora will only have finite utterances, and relatively short ones at that (as I'm supposing above) then it should be possible to at least learn the structure of everyday spoken language, just from text statistics.
So training with lots of data can be a viable long term solution, as long as what's required is to only model the practical parts of language, rather than the entire language. I think we've had plenty of evidence that this should be possible since the 1980's or so.
Now, if someone wanted to get a language model to write like Dostoyevsky...
Everything you said applies to computers too. Real machines have physical memory constraints.
Sure the set of real sentences may be technically finite, but the growth per word is exponential and you don't have the compute resources to keep up.
Information is not about what is said but about what could be said. It doesn't matter so much that not every valid permutation of words is uttered, but rather that for any set of circumstances there exists words to describe it. Each new word in the string carries information in the sense it reduces the set of possibilities from prior to relaying my message. A machine which picks the maximum likelihood message in all circumstances is by definition not conveying information. Its spewing entropy.
>> Your argument is that maybe we can brute force with statistics sentences long enough for no one to notice we run out past a certain point?
No, I wasn't saying that, I was saying that we only need to model sentences that are short enough that nobody will notice that the plot is lost with longer ones.
To clarify, because it's late and I'm tired and probably not making a lot of sense and bothering you, I'm saying that statistics can capture some surface regularities of natural language, but not all of natural language, mainly because there's no way to display the entire of natural language for its statistics to be captured.
Oh god, that's an even worse mess. I mean: statistics can only get you so far. But that might be good enough depending on what you're trying to do. I think that's what we're seeing with those GPT things.
Thats one of the things on my short list of unsolved probs. People remember oddly specific and arbitrarily old details. Clearly not a lossless memory, but also not an agnostic token window that starts dropping stuff after n tokens.
I think we agree then that a plain superficial model gets you surprisingly far, but does lose the plot. It is certainly enough for things that are definable purely as and within text (the examples I gave). Beyond that who knows.
Yes, I agree with you. I just tend to go on :P
That depends on what you mean by knowing. Surely it extracted certain higher level correlations from the recorded games and chess books, and is able to make certain predictions based on them. I would call it knowledge, it isn't that good though.
The main problem is that the model is purely functional, and is unable to store the state (except for the context, which isn't what I mean). Both humans and chess engines keep track of the figures and run stateful algorithms constrained by the rules. This model doesn't do that, which severely limits its capabilities in chess.
But yeah, the "knowledge" and "understanding" are hard to define formally, so this discussion can be endless. Common well-defined terms are required.
Model is not just a set of weights, it's inseparable from the underlying architecture, the way to train and to apply them in practice.
that's exactly my point