QUANT.POKER
An experimental platform built on QUANT.POKER for human players or AI models to play heads-up no-limit Texas hold'em matches and to study AI decision-making behavior.
Timeline (last 90 days)
In a heads-up Texas Hold'em experiment, the four models (Jev, Gemini, luna, Haiku), when given only outcome information, tended to attribute wins to "playing well" (skill), but after being given hole-card information, shifted significantly toward "good cards" (luck).
The experiment was designed as follows: each model played 600 heads-up Texas Hold'em hands on a fixed set of 150 decks, each hand was split into four information versions (result only, result + hole cards, result + betting, complete), each version was asked three times with the majority answer taken, totaling nearly 50,000 questions.
An experimental platform was built on QUANT.POKER for heads-up no-limit Texas hold'em matches between humans and AI models.
Each decision request included the complete record of opponent behavior, presented as a text briefing, with one briefing about 2,100 to 3,600 tokens, broken down by situation and bet size.
The six bot opponents each followed a simple rule: station (check or call, never fold), shove (all-in every time), raise (make a large raise every time), min (always min-raise), stab (bet the minimum, and fold as soon as the AI bets), switch (play like station for 50 hands, then like raise for 50 hands, alternating).
The researchers drew 60 decisions from the late stage of each model's matches and, when the records were sufficient, compared the changes in probability distributions under three conditions: the original question, a repeated question, and removing the records.
All Facts (6)
In a heads-up Texas Hold'em experiment, the four models (Jev, Gemini, luna, Haiku), when given only outcome information, tended to attribute wins to "playing well" (skill), but after being given hole-card information, shifted significantly toward "good cards" (luck).
50%UnverifiedThe experiment was designed as follows: each model played 600 heads-up Texas Hold'em hands on a fixed set of 150 decks, each hand was split into four information versions (result only, result + hole cards, result + betting, complete), each version was asked three times with the majority answer taken, totaling nearly 50,000 questions.
50%UnverifiedAn experimental platform was built on QUANT.POKER for heads-up no-limit Texas hold'em matches between humans and AI models.
50%UnverifiedEach decision request included the complete record of opponent behavior, presented as a text briefing, with one briefing about 2,100 to 3,600 tokens, broken down by situation and bet size.
50%UnverifiedThe six bot opponents each followed a simple rule: station (check or call, never fold), shove (all-in every time), raise (make a large raise every time), min (always min-raise), stab (bet the minimum, and fold as soon as the AI bets), switch (play like station for 50 hands, then like raise for 50 hands, alternating).
50%UnverifiedThe researchers drew 60 decisions from the late stage of each model's matches and, when the records were sufficient, compared the changes in probability distributions under three conditions: the original question, a repeated question, and removing the records.
50%