I couldn’t get it to solve a basic mate-in-one problem with one rook on the board. It seems to completely not understand how to move the pieces. It also doesn’t understand the solution when it’s given.
This is accurate. Of COURSE it doesn't know the rules of chess and doesn't know how to move the pieces. All it knows is how to regurgitate found descriptions of chess moves in a way that makes sense for descriptions of things but which only has passing resemblance to chess itself, which is not the same thing.
There are some studies showing that LLMs are capable of representing internal state and knowledge in its models, even when only trained on language tokens: https://thegradient.pub/othello/
> 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.
The answer is that for Chess it doesn't matter. The standard chess piece notation is a complete encoding of the game space. An inference about the nature of our board based physical understanding is not needed. You could formulate chess as a purely text based game about appending alphanumeric tokens to a chain of said tokens. Its a closed system. The machine need not be tied to our squares and horsey based interpretation of the semantics. To be able to follow the grammar of the language chess is to understand chess.
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 internal model will certainly pick up on statistical correlations among the text analysis corresponding to an 8x8 2D grid as this is the most low-hanging statistical representation that helps solving the problem during training.
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?
It can learn the rules for movement strictly as generator rules imposed on a string of tokens representing the previous sequence of moves. Each new item appended to the list has to in some way match a previous item in the list. Eg RC6 is a Rook, so it has to match an earlier token that is also a Rook, in one of two ways: R_6 or RC_ (and it must not be previously captured by __6 or _C_ ). At no point is it even necessary to convert the move history into a present board state, let alone the state of an 8x8 grid. The move history is sufficient board state on its own. Are the rules for valid chess moves, expressed as 3 character token grammar, the same thing as having learned a 2d grid in latent space? I don't think so, because the language rule is more general and isn't restricted by geometry.
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.
>> To be able to follow the grammar of the language chess is to understand chess.
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.
Well, let me try to explain what I'm thinking, though I may have misunderstood you.
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.
Yeah thats exactly it. The rules are easy enough to put into the form of matching strings. I gave an explicit example further down in the thread. At no point is it required to even convert the game history into a board state data structure. The game history in standard notation itself is sufficient as a game state. To know where a piece is, simply iterate back from the end until finding the last mention of it then iterate forward to make sure it wasn't captured.
Yeah it seems to have some model for how chess works, it gives convincing explanations for wrong/illegal moves, and you can generally see there’s some connection between the move and the heuristics it’s babbling about. But it hasn’t built a working internal model of the board that it can manipulate, it can’t search the tree of possible moves.
Well, humans are just trained on language tokens too (and of course, supplementary images etc).
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.
The fact that humans pick up language so soon after birth is the motivating question behind the biggest theory in all of linguistics, namely Chomsky's Universal Grammar. The simple fact is, teacher never stands in front of the class and says "here is how you don't talk. Purple if alligator burp arming has why't." Yet, despite the paucity of negative examples, everyone figures it out. You can't explain that in the current paradigm. There's a lot you can do unreasonably well despite virtually no prior experience. 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. We are missing something fundamental and algorithmic and can only patch over our lack of understanding with volumes of training data for so long.
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.
> Yet, despite the paucity of negative examples, everyone figures it out.
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.
You're digging your heals in on a rehash of a model from the 40s, glibly dismissing the problems it doesn't account for bought up by linguists in the 50s and 60s as if they are unaware that babies go through a period of babbling. The amount of time spent acquiring language is already priced in and not enough to account for as pure reward and training.
>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.
>> 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...
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?
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.
Now, now. Who said anything about information? I was just talking about modelling text. Like, the distribution of token collocations in a corpus of natural language. We know that's perfectly doable, it's been done for years. And to avoid exponential blowups, just use the Markov property or in any case, do some fudgy approximation of this and that and you're good to go.
>> 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.
>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.
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.
>> 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.
It’s not a question of whether machines can do it at all. The question is whether our current approach of training LLMs can do it. We don’t know how the human brain works, so we have no idea if there’s something in the brain that is fundamentally different from training an LLM.
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's a neural network at the end of the day... it can compute any result or "understand" any system if properly weighted and structured.
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.
A LLM isn’t going to learn how to drive a car because of how they are trained even if a neutral network could.
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.
> It's a neural network at the end of the day... it can compute any result or "understand" any system if properly weighted and structured.
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.
Approximating continuous functions is likely quite the same as what people do too. You think there isn’t some mathematical model under the hood of how the brain works too? That it doesn’t break down into functions with interpretable results? Is it spiritual or mystical in your mind?
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.
I can ask an LLM what 2+2 is and it can answer with 4. That's a discrete result. So how is this different from human thinking? Where is your evidence that this is not a similar mechanism?
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
>> I can ask an LLM what 2+2 is and it can answer with 4. That's a discrete
result. So how is this different from human thinking? Where is your evidence
that this is not a similar mechanism?
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?
> Approximating continuous functions is likely quite the same as what people do too.
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.
The current models are still too simple and missing things we did not figure out; when a human (or other animal) learns, it only needs a tiny (compared to the corpus of text etc gpt gets trained on) corpus to become a smart human. So the model needs something we have built in that makes learning vastly more efficient. Then there will be another big jump. That’ll come, that or a new AI winter.
And how would you do that? Show certain moves and point your thumb up for ok, down for not ok? Then sorry, you're still using a language, just without using words.
It doesn't seem so far-fetched to believe somebody could learn chess just by watching others play it, no language needed at all (except perhaps reading the body language of being glad to win).
But I imagine LLMs will soon have the ability to turn image sequences into information that can be interpreted and ingested much the same way as they can text, and thus "learn" how to play chess just from analyzing videos of actual games being played.
>Of COURSE it doesn't know the rules of chess and doesn't know how to move the pieces.
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.
The disk is unable to extract the correlations, nor is it able to apply the knowledge; it transparently stores the data verbatim. The model doesn't store the training set, it extracts the complex correlations from it, and is able to make actual predictions based on the knowledge it extracted.
But yeah, the "knowledge" and "understanding" are hard to define formally, so this discussion can be endless. Common well-defined terms are required.
the model does not extract knowledge. an external algorithm trains the models parameters and then the model is fed a string that is also evaluated externally based on the models configuration.
Semantics. You could say the same about the disk - the data doesn't get magically teleported from the magnetic plates to the RAM, it needs a lot of underlying hardware to read and transfer it.
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.
The difference is that a human playing chess, once told the rules, doesn't suddenly start making illegal moves after passing the memorized opening phase. Accidentally making illegal moves is the categorical definition of not knowing how to play.
Since your comment seems to be strongly contradicting the blog post, it might be worth checking whether you are really testing the same thing.
The blog post is about playing chess against GPT-4. GPT-4 (or at least, a version without image input capability) is available at https://chat.openai.com/, but only to "Plus subscribers" who pay for it.
So did you test with GPT-4, or did you test the "default GPT-3.5" model which is available for free?
And did you actually select GPT-4? It is shown at the top of the conversation page.
I'm asking because I noticed that the default setting for each new conversation is currently still GPT-3.5, even if you are a subscriber and even if you have selected GPT-4 for your previous conversation(s).
Come on. Yes. Here's a screenshot if it helps: https://i.imgur.com/BepKMt1.png - It gives FEN notation itself, so I prompted it with FEN, but I have also tried describing the position of each piece explicitly.
