Chess is a deterministic game won by a combination of known movesets and constrained multi-level forward search.
LLMs do neither of these things. They don't reproduce training data exactly, their next response is more 'inspired by' prompts and its own memory than produced deterministically, and they don't have the capability to do general forward search on their own.
So when you ask an LLM to play chess you're getting the equivalent of a very compressed and lossy JPEG of chess rules and strategies with added per-turn random noise.
They also don't have the ability to design their own chess engine, although it would be interesting to see what happens if you ask for one.
Monte Carlo Tree Search is stochastic.
I, as a human AGI, would jever just forget and remove a piece from the board from one turn to the next.
I mean we've done all this before in a task-specific fashion. It's useful to know that LLMs haven't managed to do that in the process of learning to represent the entire text on the web. On the other hand they have gotten say very good at machine translation without being trained exclusively (and I select the preceding word carefully) on machine translation.
Edit: I'm saying this because there is this idea expressed by e.g. Ilya Sutskever, that in order to predict the next token accurately an LLM has to learn something about all of underlying reality. See for example this interview with Dwarkesh:
https://x.com/biobootloader/status/1640512444958396416
Where Sutskever claims that "Predicting the next token well means you understand the underlying reality that led to the creation of that token".
If that were true, we should have seen LLMs play good chess by now. There is a huge amount of data on playing chess floating around on the web in the form of algebraic chess notation and if LLMs were capable of learning the "underlying reality" of chess, they would already have. They haven't. Because they can't. What Sutskever is saying flies in the face of literally hundreds of years of statistical modelling, which is to say, building predictive models that, very explicitly, do not have to understand any "underlying reality" and only have to be good at modelling a dataset.
Sarcasm aside, I think this is an easy cognitive trap to fall into. It does sometimes feel like the LLM must have some world model because it converses somewhat coherently. Examples like this failure to understand chess, or to count the number of Rs in "strawberry", seem difficult to explain if the models are intelligent. But that doesn't stop people believing they are anyway. I think there must be something about the conversational interface that fools us easily. I wonder if people trained in interrogation techniques are also fooled?
Not at all. LLMs learn by imbibing a mass of relationships as isolated fragments of information. There is a certain amount of sorting and indexing that happens during the training phase. There is also a certain amount of compute executed on these relationships during inference. LLMs can model processes that fit within the compute budget. Language translation works well because language is lookup-heavy while being light on compute.
Chess is a compute heavy game of finding the best move out of many possibilities with wide variation in the quality of each move. Humans cut through the compute requirements by reinforcement and learning intuition. LLMs don't get reinforcement on chess so they must compute during inference a unified model of chess. Developing a strong model of chess from raw fragments of information is simply not in their compute budget.
You gotta be careful how you use the word "relation" here because there's an informal meaning (I'm related to my cousin) and a more strict, formal meaning, that is used in computer science e.g. in the "Relational Calculus" etc. In the formal sense, the one relation that LLMs learn during training is the co-occurrence of tokens in a corpus of text, what's called more technically a "collocation" relation. Nothing says that this is enough to play chess, so I'm indeed doubtful that they can.
That's a separate question from whether an LLM (unaided by a C complier) can learn to play chess as well as human (also unaided by a C compiler). Certainly humans can't become grandmasters only by reading chess transcripts on the web, and certainly humans require many "thinking tokens" during a game to play effectively. Do you know for sure that a transformer can't reach grandmaster level if it is allowed to learn by playing games (as humans do) and is given a sufficient number of thinking tokens during the game? It seems near certain that they could, if someone wanted to spend the money (and I don't see why anyone would.)
Yes, I do mean that the LLM's weights are set so that it will execute minimax or MCTS when it needs to. That has nothing to do with whether humans can do the same or not.
I don't disagree that a Transformer could learn to play chess if it was explicitly trained to do that. My argument is that LLMs, trained to predict the next token, have not learned to play chess. That's LLMs, not Transformers.
Just to make sure this is not taken as splitting hairs, the point is that there's all sorts of claims made about what LLMs learn when they train on text. For example, there was a claim by Sundar Pichai that one of their models had learned to translate Bengali without explicitly being trained to do so. It later emerged that Bengali was indeed included in the model's training set [1]. It's not clear whether that included parallel texts, e.g. between Begnali and English or another intermediary language, in any case Sundar Pichai's claim was that the ability to translate Bengali was "emergent".
So I'm interested in understanding the extent to which these "emergent" abilities are real or not. With chess, given the amount of textual data tracing games that floats about on the open internet, I would totally except some ability to play chess to "emerge". Maybe the reported 700-800 ELO level is even that sort of ability. Maybe we should only expect LLMs to learn to play at the level of an untrained, casual player. Maybe not. I have no idea.
On the other hand, the fact they keep making elementary mistakes like illegal moves must be taken to mean that, so far, LLMs haven't learned to play chess.
__________________
[1] https://www.buzzfeednews.com/article/pranavdixit/google-60-m...