Are LLMs able to play the card game Set?
github.comSet is a card game where players have to identify sets of three cards from a layout of 12. Each card features a combination of four attributes: shape, color, number, and shading. A valid set consists of three cards where each attribute is either the same on all three cards or different on each. The goal is to find such sets quickly and accurately.
Though this game is a solved computer problem — easily tackled by algorithms or deep learning — I thought it would be interesting to see if Large Language Models (LLMs) could figure it out.
If you think this is fun, try to see how it garbles predicate logic.
FYI: Card 8's transcription is different than the image. In the image 5, 8, 12 is a Set but the transcription says Card 8 only has 2 symbols which removes that Set.
Not only that, but 2,6,7 is also a set but not included in the results
Oh no, thanks for pointing this out! I asked GTP-4o to convert the image to text for me and I only checked some of the cards, assuming the rest would be correct. That was a mistake.
I've now corrected the experiment to accurately take the image into account. This meant that Deepseek was no longer able to find all the sets, but o3-mini still did a good job.
Both 7 and 8 are incorrect (both claim a count of 2 while the cards have 3). This leads to missing both 5-8-12 and 2-6-7 as valid sets.
Woah, what's going on?? I've always played Set with stripey cards, is this a custom deck or did they change it at some point???
This is wildly disconcerting to me
This is definitely a custom/knock-off deck. Not only are the stripes not stripey, the capsules are now ovals and the diamonds are now rectangles.
My first party set deck looks exactly like that. They must've done a redesign at some point.
Same here. These are the shapes and fill patterns of the edition by Ravensburger, which is the one usually found in Europe.
This is exactly what they always looked when I played (here in Europe). Is it maybe a regional thing?
I noticed that LLM at least at the Claude and OpenAI 4o level can not play tic tac toe and win against a competent opponent. They make illogical moves.
Interestingly, they can write a piece of code to solve Tic Tac Toe perfectly without breaking a sweat.
I've always said that appending "use python" to your prompt is a magic phrase that makes 4o amazingly powerful across a wide range of tasks. I have a whole slew of things in my memories that nudge it to use python when dealing with anything even remotely algorithmic, numeric, etc
Playing tic tac toe could be such a basic topic that there is relatively little information on the internet about how to "always" win.
On the other hand writing a piece of code to solve Tic Tac Toe sounds like it could be a relatively common coding challenge.
Win or stalemate? because a stalemate is the likely scenario against a somewhat competent opponent IMO.
Ahh good correction, I meant the winning strategy is to force a stalemate.
In all of my tests Claude nor 4o can even get to a stalemate, they just make incorrect moves.
It might be the way you're formatting the input. I wonder how they perform when state updates are shared via natural language vs ASCII art vs image
I tried a bunch of different ways. It wasn’t the prompt or input format.
Since you can train an LLM to play chess from scratch, I would not be surprised if you could also train one to play Set. I might experiment with it tomorrow.
https://adamkarvonen.github.io/machine_learning/2024/01/03/c...
Get them to play Fluxx, and we'll be talking...
(this one, where ever changing rules is part of the game: https://www.looneylabs.com/games/fluxx )
I am increasingly concerned that these new reasoning models are thinking.
I still think that we’re at much greater risk of discovering that human thinking is much less magical, than we are of making a machine that does magical thinking.
No problem. Just redefine "thinking".
To what? And how does that change the reality of what the models are doing?
I believe GP is being sarcastic, as this seems to be a common reaction. Every time a machine accomplishes something that seems to be intelligent, we redefine intelligence to exclude that thing, and now the issue is fixed: computers are not intelligent and they cannot think.
Never mind that they can beat the entire world at chess and Go, and 90+% of the population at math, engineering, and physics problems. Those things do not require intelligence or thinking.
My experience of thinking is that it is a constant phenomenon. My experience of LLMs is that they only respond and are not running without input.
That's because we don't leave them running, right? We could, though, yes?
Current LLMs decohere rapidly, and as far as I am aware, this problem seems like a fundamental to the architecture one.
Yup.
Tangentially related: https://qntm.org/mmacevedo
Well there is a gap between the firing of individual neurons in your mind. How long would that gap need to be for it not to count as thinking anymore?