I'm not an expert, but as I understand it there are existing solvers for poker/holdem? Perhaps one of the players could be a traditional solver to see how the LLMs fare against those?
This also wouldn't even be a close contest, I think Pluribus demonstrated a solid win rate against professional players in a test.
As I was developing this project, a main thought came to mind as to the comparison between cost and performance between a "purpose" built AI such as Pluribus versus a general LLM model. I think Pluribus training costs ~$144 in cloud computing credits.
To expand on this - an LLM will try to play (and reason) like a person would, while a solver simply crunches the possibility space for the mathematically optimal move.
It’s similar to how an LLM can sometimes play chess on a reasonably high (but not world-class) level, while Stockfish (the chess solver) can easily crush even the best human player in the world.
GTO (“game theory optimal”) poker solvers are based around a decision tree with pre-set bet sizes (eg: check, bet small, bet large, all in), which are adjusted/optimized for stack depth and position. This simplifies the problem space: including arbitrary bet sizes would make the tree vastly larger and increase computational cost exponentially.
No, I'm not super certain, but I believe most solvers are trained to be game theory optimal (GTO), which means they assume every other player is also playing GTO. This means there is no strategy which beats them in the long run, but they may not be playing the absolute best strategy.
Not only to limit the scope of what it has to simulate, but only a certain number of bet sizes is practical for a human to implement in their strategy.
How would an LLM play like a human would? I kind of doubt that there is enough recounting of poker hands or transcription of filmed poker games in the training data to imbue a human-like decision pattern.
Anybody who plays poker “optimally” is bound to lose money when they come up against anyone with skill. Once you know the strategy your opponent is employing you can play like you have anything. I believe I’ve won with 7,2 offsuite more than any other hand, because I played like I had the nuts.
This is completely wrong - the entire point of the Nash equilibrium solution (in the context of poker, at least) is that it is, at worst, EV-neutral even when your opponent has perfect knowledge of your strategy.
Your 72o comment indicates you are either playing with very weak players, or have gotten lucky, as in reasonably competitive games playing (and then full bluffing) 72o will be significantly negative EV. Try grinding that strategy at a public 10/20 table and you will be quickly butchered and sent back to the ATM.
"Solvers" normally means algorithms which aim to produce some mathematically optimal (given certain assumptions) behaviour.
There are other poker playing programs [0] - what we called AI before large language models were a thing - which achieve superhuman performance in real time in this format. They would crush the LLMs here. I don't know what's publicly available though.
This also wouldn't even be a close contest, I think Pluribus demonstrated a solid win rate against professional players in a test.
As I was developing this project, a main thought came to mind as to the comparison between cost and performance between a "purpose" built AI such as Pluribus versus a general LLM model. I think Pluribus training costs ~$144 in cloud computing credits.