Call for participants: SpinGPT vs Humans
Hello everyone,
TL;DR: We’re looking for players to take part in a university project by playing against our poker AI, SpinGPT, here: https://spingpt.lamsade.fr/
This study is run at LAMSADE, the computer science lab of Université Paris-Dauphine, by me (Narada Maugin) and Prof. Tristan Cazenave. We built SpinGPT to play Spin & Go (aka Expresso / Jackpot SNG), and we now want to see how it performs against human opponents. The broader goal is to test whether large language models, trained with our protocol, are useful for studying imperfect-information games. Poker is our testbed, not the end goal, so the AI itself will not be released publicly.
How it works
From November 1 to December 1, 2025, you can play SpinGPT on our site: https://spingpt.lamsade.fr/.
Format: heads-up, 25 BB effective, no blind increases.
You can jump in anonymously to try the interface and the bot. If you prefer, create an account and you’ll be automatically enrolled in the study. To make it more interesting, we’re offering prizes:
- €500 for the best BB/100 over > 2,000 hands vs SpinGPT (after AIVAT*)
- €250 for the second-best BB/100 over > 2,000 hands (after AIVAT)
- 5 × €50 randomly drawn among players with > 1,000 hands (any win rate)
*AIVAT is a standard variance-reduction method used in poker AI evaluation to estimate win rate more precisely. Details: https://arxiv.org/abs/1612.06915
Hands are used only for academic research and handled under French research regulations. You can request data deletion at any time.
If you’d like to dig deeper into the project and methodology, our paper is forthcoming in Advances in Computer Games (Springer LNCS). A public preprint is already available: https://arxiv.org/abs/2509.22387
Happy to answer questions.
GL at the tables,
Narada Maugin
2 Replies
why does the leaderboard shows nº of hands played and not bb/100?
wouldn`t it be better to only show players with more than 2k hands played and the winrates of those?
And on your profile to see what winrate you have?
Im in!
Hi D33P,
Thanks for participating!
Using bb/100 is indeed more informative than just showing the number of hands played. However, in our case we are interested in the bb/100 after applying AIVAT, a variance-reduction algorithm. I will only run AIVAT once at the end of the experiment in December, because it is quite computationally expensive.
Therefore, for now, showing the raw bb/100 would not be very meaningful, since the variance is huge on such small samples and could thus mislead players.
