An LLM is a poor computational/predictive paradigm for playing chess.
This just in: a hammer makes a poor screwdriver.
LLMs are more like a leaf blower though
Actually, a very specific model (chatgpt3.5-turbo-instruct) was pretty good at chess (around 1700 elo if i remember correctly).
I’m impressed, if that’s true! In general, an LLM’s training cost vs. an LSTM, RNN, or some other more appropriate DNN algorithm suitable for the ruleset is laughably high.
Oh yes, cost of training are ofc a great loss here, it’s not optimized at all, and it’s stuck at an average level.
Interestingly, i believe some people did research on it and found some parameters in the model that seemed to represent the state of the chess board (as in, they seem to reflect the current state of the board, and when artificially modified, the model takes modification into account in its playing). It was used by a french youtuber to show how LLMs can somehow have a kinda representation of the world. I can try to get the sources back if you’re interested.
Absolutely interested. Thank you for your time to share that.
My career path in neural networks began as a researcher for cancerous tissue object detection in medical diagnostic imaging. Now it is switched to generative models for CAD (architecture, product design, game assets, etc.). I don’t really mess about with fine-tuning LLMs.
However, I do self-host my own LLMs as code assistants. Thus, I’m only tangentially involved with the current LLM craze.
But it does interest me, nonetheless!
Here is the main blog post that i remembered : it has a follow up, a more scientific version, and uses two other articles as a basis, so you might want to dig around what they mention in the introduction.
It is indeed a quite technical discovery, and it still lacks complete and wider analysis, but it is very interesting for the fact that it kinda invalidates the common gut feeling that llms are pure lucky random.
The underlying neural network tech is the same as what the best chess AIs (AlphaZero, Leela) use. The problem is, as you said, that ChatGPT is designed specifically as an LLM so it’s been optimized strictly to write semi-coherent text first, and then any problem solving beyond that is ancillary. Which should say a lot about how inconsistent ChatGPT is at solving problems, given that it’s not actually optimized for any specific use cases.
Yes, I agree wholeheartedly with your clarification.
My career path, as I stated in a different comment in regards to neural networks, is focused on generative DNNs for CAD applications and parametric 3D modeling. Before that, I began as a researcher in cancerous tissue classification and object detection in medical diagnostic imaging.
Thus, large language models are well out of my area of expertise in terms of the architecture of their models.
However, fundamentally it boils down to the fact that the specific large language model used was designed to predict text and not necessarily solve problems/play games to “win”/“survive”.
(I admit that I’m just parroting what you stated and maybe rehashing what I stated even before that, but I like repeating and refining in simple terms to practice explaining to laymen and, dare I say, clients. It helps me feel as if I don’t come off too pompously when talking about this subject to others; forgive my tedium.)
Yeah, a lot of them hallucinate illegal moves.
Tbf, the article should probably mention the fact that machine learning programs designed to play chess blow everything else out of the water.
Yeah its like judging how great a fish is at climbing a tree. But it does show that it’s not real intelligence or reasoning
Don’t call my fish stupid.
Well, can it climb trees?
Machine learning has existed for many years, now. The issue is with these funding-hungry new companies taking their LLMs, repackaging them as “AI” and attributing every ML win ever to “AI”.
ML programs designed and trained specifically to identify tumors in medical imaging have become good diagnostic tools. But if you read in news that “AI helps cure cancer”, it makes it sound like it was a lone researcher who spent a few minutes engineering the right prompt for Copilot.
Yes a specifically-designed and finely tuned ML program can now beat the best human chess player, but calling it “AI” and bundling it together with the latest Gemini or Claude iteration’s “reasoning capabilities” is intentionally misleading. That’s why articles like this one are needed. ML is a useful tool but far from the “super-human general intelligence” that is meant to replace half of human workers by the power of wishful prompting
I forgot which airline it is but one of the onboard games in the back of a headrest TV was a game called “Beginners Chess” which was notoriously difficult to beat so it was tested against other chess engines and it ranked in like the top five most powerful chess engines ever
It does
It does not. Where?
Sometimes it seems like most of these AI articles are written by AIs with bad prompts.
Human journalists would hopefully do a little research. A quick search would reveal that researches have been publishing about this for over a year so there’s no need to sensationalize it. Perhaps the human journalist could have spent a little time talking about why LLMs are bad at chess and how researchers are approaching the problem.
LLMs on the other hand, are very good at producing clickbait articles with low information content.
Gotham chess has a video of making chatgpt play chess against stockfish. Spoiler: chatgpt does not do well. It plays okay for a few moves but then the moment it gets in trouble it straight up cheats. Telling it to follow the rules of chess doesn’t help.
This sort of gets to the heart of LLM-based “AI”. That one example to me really shows that there’s no actual reasoning happening inside. It’s producing answers that statistically look like answers that might be given based on that input.
For some things it even works. But calling this intelligence is dubious at best.
Because it doesn’t have any understanding of the rules of chess or even an internal model of the game state, it just has the text of chess games in its training data and can reproduce the notation, but nothing to prevent it from making illegal moves, trying to move or capture pieces that don’t exist, incorrectly declaring check/checkmate, or any number of nonsensical things.
Hallucinating 100% of the time 👌
ChatGPT versus Deepseek is hilarious. They both cheat like crazy and then one side jedi mind tricks the winner into losing.
So they are both masters of troll chess then?
See: King of the Bridge
It plays okay for a few moves but then the moment it gets in trouble it straight up cheats.
Lol. More comparisons to how AI is currently like a young child.
I think the biggest problem is it’s very low ability to “test time adaptability”. Even when combined with a reasonning model outputting into its context, the weights do not learn out of the immediate context.
I think the solution might be to train a LoRa overlay on the fly against the weights and run inference with that AND the unmodified weights and then have an overseer model self evaluate and recompose the raw outputs.
Like humans are way better at answering stuff when it’s a collaboration of more than one person. I suspect the same is true of LLMs.
Like humans are way better at answering stuff when it’s a collaboration of more than one person. I suspect the same is true of LLMs.
It is.
It’s really common for non-language implementations of neural networks. If you have an NN that’s right some percentage of the time, you can often run it through a bunch of copies of the NNs and take the average and that average is correct a higher percentage of the time.
Aider is an open source AI coding assistant that lets you use one model to plan the coding and a second one to do the actual coding. It works better than doing it in a single pass, even if you assign the the same model to planing and coding.
In this case it’s not even bad prompts, it’s a problem domain ChatGPT wasn’t designed to be good at. It’s like saying modern medicine is clearly bullshit because a doctor loses a basketball game.
I imagine the “author” did something like, “Search http://google.scholar.com/ find a publication where AI failed at something and write a paragraph about it.”
It’s not even as bad as the article claims.
Atari isn’t great at chess. https://chess.stackexchange.com/questions/24952/how-strong-is-each-level-of-atari-2600s-video-chess
Random LLMs were nearly as good 2 years ago. https://lmsys.org/blog/2023-05-03-arena/
LLMs that are actually trained for chess have done much better. https://arxiv.org/abs/2501.17186Wouldn’t surprise me if an LLM trained on records of chess moves made good chess moves. I just wouldn’t expect the deployed version of ChatGPT to generate coherent chess moves based on the general text it’s been trained on.
I wouldn’t either but that’s exactly what lmsys.org found.
That blog post had ratings between 858 and 1169. Those are slightly higher than the average rating of human users on popular chess sites. Their latest leaderboard shows them doing even better.
https://lmarena.ai/leaderboard has one of the Gemini models with a rating of 1470. That’s pretty good.
ChatGPT has been, hands down, the worst AI coding assistant I’ve ever used.
It regularly suggests code that doesn’t compile or isn’t even for the language.
It generally suggests AC of code that is just a copy of the lines I just wrote.
Sometimes it likes to suggest setting the same property like 5 times.
It is absolute garbage and I do not recommend it to anyone.
I find it really hit and miss. Easy, standard operations are fine but if you have an issue with code you wrote and ask it to fix it, you can forget it
I’ve found Claude 3.7 and 4.0 and sometimes Gemini variants still leagues better than ChatGPT/Copilot.
Still not perfect, but night and day difference.
I feel like ChatGPT didn’t focus on coding and instead focused on mainstream, but I am not an expert.
Gemini will get basic C++, probably the best documented language for beginners out there, right about half of the time.
I think that might even be the problem, honestly, a bunch of new coders post bad code and it’s fixed in comments but the LLM CAN’T realize that.
I like tab coding, writing small blocks of code that it thinks I need. Its On point almost all the time. This speeds me up.
Bingo. If anything what you’re finding is the people bitching are the same people that if given a bike wouldn’t know how to ride it, which is fair. Some people understand quicker how to use the tools they are given.
Edit - a poor carpenter blames his tools.
It’s the ideal help for people who shouldn’t be employed as programmers to start with.
I had to explain hexadecimal to somebody the other day. It’s honestly depressing.
That’s because it doesn’t know what it’s saying. It’s just blathering out each word as what it estimates to be the likely next word given past examples in its training data. It’s a statistics calculator. It’s marginally better than just smashing the auto fill on your cell repeatedly. It’s literally dumber than a parrot.
Parrots are actually intelligent though.
Yeah, but not when it comes to understanding human speech. There’s a reason that repeating words without really understanding them is called parroting. Gray parrots are the smartest and some can actually understand language a little bit, making them smarter than chat, which is just high tech guessing without comprehension
All AIs are the same. They’re just scraping content from GitHub, stackoverflow etc with a bunch of guardrails slapped on to spew out sentences that conform to their training data but there is no intelligence. They’re super handy for basic code snippets but anyone using them anything remotely complex or nuanced will regret it.
One of my mates generated an entire website using Gemini. It was a React web app that tracks inventory for trading card dealers. It actually did come out functional and well-polished. That being said, the AI really struggled with several aspects of the project that humans would not:
- It left database secrets in the code
- The design of the website meant that it was impossible to operate securely
- The quality of the code itself was hot garbage—unreadable and undocumented nonsense that somehow still worked
- It did not break the code into multiple files. It piled everything into a single file
I’ve used agents for implementing entire APIs and front-ends from the ground up with my own customizations and nuances.
I will say that, for my pedantic needs, it typically only gets about 80-90% of the way there so I still have to put fingers to code, but it definitely saves a boat load of time in those instances.
my favorite thing is to constantly be implementing libraries that don’t exist
You’re right. That library was removed in ToolName [PriorVersion]. Please try this instead.
*makes up entirely new fictitious library name*
Oh man, I feel this. A couple of times I’ve had to field questions about some REST API I support and they ask why they get errors when they supply a specific attribute. Now that attribute never existed, not in our code, not in our documentation, we never thought of it. So I say “Well, that attribute is invalid, I’m not sure where you saw to do that”. They get insistent that the code is generated by a very good LLM, so we must be missing something…
It’s even worse when AI soaks up some project whose APIs are constantly changing. Try using AI to code against jetty for example and you’ll be weeping.
I’ve had success with splitting a function into 2 and planning out an overview, though that’s more like talking to myself
I wouldn’t use it to generate stuff though
I don’t use it for coding. I use it sparingly really, but want to learn to use it more efficiently. Are there any areas in which you think it excels? Are there others that you’d recommend instead?
Use Gemini (2.5) or Claude (3.7 and up). OpenAI is a shitshow
I swear every single article critical of current LLMs is like, “The square got BLASTED by the triangle shape when it completely FAILED to go through the triangle shaped hole.”
It’s newsworthy when the sellers of squares are saying that nobody will ever need a triangle again, and the shape-sector of the stock market is hysterically pumping money into companies that make or use squares.
It’s also from a company claiming they’re getting closer to create morphing shape that can match any hole.
And yet the company offers no explanation for how, exactly, they’re going to get wood to do that.
The press release where OpenAI said we’d never need chess players again
You get 2 triangles in a single square mate…
CHECKMATE!
Touchdown! 3 points!
Well, the first and obvious thing to do to show that AI is bad is to show that AI is bad. If it provides that much of a low-hanging fruit for the demonstration… that just further emphasizes the point.
That’s just clickbait in general these days lol
LLM are not built for logic.
And yet everybody is selling to write code.
The last time I checked, coding was requiring logic.
To be fair, a decent chunk of coding is stupid boilerplate/minutia that varies environment to environment, language to language, library to library.
So LLM can do some code completion, filling out a bunch of boilerplate that is blatantly obvious, generating the redundant text mandated by certain patterns, and keeping straight details between languages like “does this language want join as a method on a list with a string argument, or vice versa?”
Problem is this can be sometimes more annoying than it’s worth, as miscompletions are annoying.
Fair point.
I liked the “upgraded autocompletion”, you know, an completion based on the context, just before the time that they pushed it too much with 20 lines of non sense…
Now I am thinking of a way of doing the thing, then I receive a 20 lines suggestion.
So I am checking if that make sense, losing my momentum, only to realize the suggestion us calling shit that don’t exist…
Screw that.
The amount of garbage it spits out in autocomplete is distracting. If it’s constantly making me 5-10% less productive the many times it’s wrong, it should save me a lot of time when it is right, and generally, I haven’t found it able to do that.
Yesterday I tried to prompt it to change around 20 call sites for a function where I had changed the signature. Easy, boring and repetitive, something that a junior could easily do. And all the models were absolutely clueless about it (using copilot)
a decent chunk of coding is stupid boilerplate/minutia that varies
…according to a logic, which means LLMs are bad at it.
I’d say that those details that vary tend not to vary within a language and ecosystem, so a fairly dumb correlative relationship is enough to generally be fine. There’s no way to use logic to infer that it’s obvious that in language X you need to do mylist.join(string) but in language Y you need to do string.join(mylist), but it’s super easy to recognize tokens that suggest those things and a correlation to the vocabulary that matches the context.
Rinse and repeat for things like do I need to specify type and what is the vocabulary for the best type for a numeric value, This variable that makes sense is missing a declaration, does this look to actually be a new distinct variable or just a typo of one that was declared.
But again, I’m thinking mostly in what kind of sort of can work, my experience personally is that it’s wrong so often as to be annoying and get in the way of more traditional completion behaviors that play it safe, though with less help particularly for languages like python or javascript.
A lot of writing code is relatively standard patterns and variations on them. For most but the really interesting parts, you could probably write a sufficiently detailed description and get an LLM to produce functional code that does the thing.
Basically for a bunch of common structures and use cases, the logic already exists and is well known and replicated by enough people in enough places in enough languages that an LLM can replicate it well enough, like literally anyone else who has ever written anything in that language.
2025 Mazda MX-5 Miata ‘got absolutely wrecked’ by Inflatable Boat in beginner’s boat racing match — Mazda’s newest model bamboozled by 1930s technology.
Ah, you used logic. That’s the issue. They don’t do that.
Can ChatGPT actually play chess now? Last I checked, it couldn’t remember more than 5 moves of history so it wouldn’t be able to see the true board state and would make illegal moves, take it’s own pieces, materialize pieces out of thin air, etc.
ChatGPT must adhere honorably to the rules that its making up on the spot. Thats Dallas
and still lose to stockfish even after conjuring 3 queens out of thin air lol
It can’t, but that didn’t stop a bunch of gushing articles a while back about how it had an ELO of 2400 and other such nonsense. Turns out you could get it to have an ELO of 2400 under a very very specific set of circumstances, that include correcting it every time it hallucinated pieces or attempted to make illegal moves.
There are custom GPTs which claim to play at a stockfish level or be literally stockfish under the hood (I assume the former is still the latter just not explicitly). Haven’t tested them, but if they work, I’d say yes. An LLM itself will never be able to play chess or do anything similar, unless they outsource that task to another tool that can. And there seem to be GPTs that do exactly that.
As for why we need ChatGPT then when the result comes from Stockfish anyway, it’s for the natural language prompts and responses.
It’s not an LLM, but Stockfish does use AI under the hood and has been since 2020. Stockfish uses a classical alpha-beta search strategy (if I recall correctly) combined with a neural network for smarter pruning.
There are some engines of comparable strength that are primarily neural-network based.
lc0comes to mind.lc0placed 2nd in the Top Chess Engine Championships in 9 out of the past 10 seasons. By comparison, Stockfish is currently on a 10-season win streak in the TCEC.
It could always play it if you reminded it of the board state every move. Not well, but at least generally legally. And while I know elites can play chess blind, the average person can’t, so it was always kind of harsh to hold it to that standard and criticise it not being able to remember more than 5 moves when most people can’t do that themselves.
Besides that, it was never designed to play chess. It would be like insulting Watson the Jeopardy bot for losing against the Atari chess bot, it’s not what it was designed to do.
Hardly surprising. Llms aren’t -thinking- they’re just shitting out the next token for any given input of tokens.
That’s exactly what thinking is, though.
An LLM is an ordered series of parameterized / weighted nodes which are fed a bunch of tokens, and millions of calculations later result generates the next token to append and repeat the process. It’s like turning a handle on some complex Babbage-esque machine. LLMs use a tiny bit of randomness (“temperature”) when choosing the next token so the responses are not identical each time.
But it is not thinking. Not even remotely so. It’s a simulacrum. If you want to see this, run ollama with the temperature set to 0 e.g.
ollama run gemma3:4b >>> /set parameter temperature 0 >>> what is a leafYou will get the same answer every single time.
Using an LLM as a chess engine is like using a power tool as a table leg. Pretty funny honestly, but it’s obviously not going to be good at it, at least not without scaffolding.
is like using a power tool as a table leg.
Then again, our corporate lords and masters are trying to replace all manner of skilled workers with those same LLM “AI” tools.
And clearly that will backfire on them and they’ll eventually scramble to find people with the needed skills, but in the meantime tons of people will have lost their source of income.
If you believe LLMs are not good at anything then there should be relatively little to worry about in the long-term, but I am more concerned.
It’s not obvious to me that it will backfire for them, because I believe LLMs are good at some things (that is, when they are used correctly, for the correct tasks). Currently they’re being applied to far more use cases than they are likely to be good at – either because they’re overhyped or our corporate lords and masters are just experimenting to find out what they’re good at and what not. Some of these cases will be like chess, but others will be like code*.
(* not saying LLMs are good at code in general, but for some coding applications I believe they are vastly more efficient than humans, even if a human expert can currently write higher-quality less-buggy code.)
I believe LLMs are good at some things
The problem is that they’re being used for all the things, including a large number of tasks that thwy are not well suited to.
yeah, we agree on this point. In the short term it’s a disaster. In the long-term, assuming AI’s capabilities don’t continue to improve at the rate they have been, our corporate overlords will only replace people for whom it’s actually worth it to them to replace with AI.
If you don’t play chess, the Atari is probably going to beat you as well.
LLMs are only good at things to the extent that they have been well-trained in the relevant areas. Not just learning to predict text string sequences, but reinforcement learning after that, where a human or some other agent says “this answer is better than that one” enough times in enough of the right contexts. It mimics the way humans learn, which is through repeated and diverse exposure.
If they set up a system to train it against some chess program, or (much simpler) simply gave it a tool call, it would do much better. Tool calling already exists and would be by far the easiest way.
It could also be instructed to write a chess solver program and then run it, at which point it would be on par with the Atari, but it wouldn’t compete well with a serious chess solver.
All these comments asking “why don’t they just have chatgpt go and look up the correct answer”.
That’s not how it works, you buffoons, it trains off of datasets long before it releases. It doesn’t think. It doesn’t learn after release, it won’t remember things you try to teach it.
Really lowering my faith in humanity when even the AI skeptics don’t understand that it generates statistical representations of an answer based on answers given in the past.
Can i fistfight ChatGPT next? I bet I could kick its ass, too :p
I mean, that 2600 Chess was built from the ground up to play a good game of chess with variable difficulty levels. I bet there’s days or games when Fischer couldn’t have beaten it. Just because a thing is old and less capable than the modern world does not mean it’s bad.
