I've got a disposable gift calculator in front of me that can do division better than every human ever. No one seems to anthropomorphize it, however.

It has learned an algorithm that plays chess purely by reading chess logs. It can play at different skill levels depending on the surrounding natural language How is that not understanding?

By the way, would you say that Stockfish 15 understands how to play chess well?

I’m pretty sure you can write a simple tree search based program that plays better than 90% of humans, sheerly by crunching numbers quickly. I’m sure you wouldn’t claim that that program “understands” chess in the way a human does.

Of course it’s a different question as to how exactly do humans understand chess.

I personally don’t believe humans have some sort of “special” intelligence, whatever we do must be computable and therefore doable by a computer. It’s just that todays LLMs don’t seem to be it.

Sure, that wouldn't be impressive if you wrote tree search for chess. However, if you wrote a program that doesn't do any search and still plays well, that'd be impressive. I don't understand your claim that LLMs playing chess is anything sort of understanding.

related: https://parrotchess.com/

How do you know there’s no search involved?

It does more illegal moves the more you deviate from standard chess openings and positions and becomes unable to play at some point. If it did a tree search based on chess rules that wouldn't happen. If it encoded a lot of chess openings and some tiny ad hoc logic around those then that is exactly what we would expect.

This is not true anymore. It has been discovered that gpt-3.5-instruct when prompeted with purely PGN notation.playes at >1700 Elo and makes no/little illegal moves.

They're only as good as the players that played the games they're trained on.

Meaning that for computer programs they're very bad at chess.

That is extremely impressive for a model that can do literally millions of tasks. It is not trained to play chess. It can write stories, code, recipes, franslate text, create JSON AND play chess. Don't you see how amazing this is compared to everything we had previously? LLMs are not just generalists that can't do anything specific properly. They can play chess better than most humans and I'm pretty sure soon they will play better than any human. And this in a general AI system!

> Except they can play chess better than 90% of humans... Just by luck of course..No understanding.

I've been thinking about this and I think there is a simple trick to make LLM's play chess well. Any chess position can be represented by a string, like so: rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1 (Forsyth–Edwards Notation) If we think of a chess game as a sentence or a sequence of such strings or tokens, then after feeding tons of those to the model it is reasonable to suspect it will do what it does best, that is predict the next token. There is no spacial awareness or anything beyond that going on I think.

There are far too many possible positions aka tokens for this to be feasible.

Same could be said about language but here we are. There are far more possible combinations of words in any given language than there are positions in chess for a typical length of a sentence I suspect.

Definitely, but treating fen positions as tokens isn’t going to work, you need something smaller, something you can decompose into. What that is I don’t know.

Well, every character of the string can be a token. From a transformer's perspective it shouldn't matter if it is Chinese, English or "chess" language.

You are right, but evidence of LLMs working well with language is not a sign that they might be memorizing Chess. But rather opposite, infeasability of memorization of chess and yet LLMs still doing well on them might be a sign that LLMs could be doing the same 'thing' with language.

LLM are not doing well in chess. They do well when people play standard openings, but play against the LLM using an inexperienced player and the LLM will quickly start making illegal moves and be unable to play. There are many millions of pro chess games recorded on the internet, enough to cover all the openings and plenty of ways to go from there with some basic heuristics, but the more you diverge from those the more the LLM will break down since it doesn't understand the rules, it just memorized a lot of openings and some really primitive patterns to follow from there.

This is not true anymore, see this tweet.

https://twitter.com/GrantSlatton/status/1703913578036904431

I have tested it myself.

It's not a trick. Every information can be represented by sequence of bits or strings. It's not the surprising part. It's just that its a telling sign that through compression understanding of the game emerges because:

1- unlike language where you can make an argument that maybe model had seen this example in the training set and just parroting it back, you can't really memorize the possible states of chess and thus you can be reasonably sure that any game or state it is in, must be OOD.

2- If the system still plays well with OOD examples, it must be gaining understanding of the game itself.

Also check out the Othello paper, where they show the evidence for an LLM encoding the game state in its neurons.

The simple trick has already been discovered

See this tweet.

https://twitter.com/GrantSlatton/status/1703913578036904431

I have tested it myself.

If they’re trained on a database of hundreds if thousands of top ranked games they’ll beat 90% by following those by rote

Just to make it clear: the database doesn't have to contain exclusively top rated games. It can be 99% low ELO games, as long as there are suffiiciently many high ELO games (in absolute number), the LLM is large enough, and there is sufficient context for the model to distinguish high ELO from low ELO games, it can learn to play well given fhe right prompting.