Silicon Poker Logic: When High-IQ Models Make Rookie Mistakes

There’s an old saying that in a poker game, if you can’t spot the sucker at the table, it’s you. Yesterday, during the final match of the Google DeepMind/Kaggle Game Arena, the suckers and the sharks were all made of silicon. 

It was an internal OpenAI civil war as o3 and GPT 5.2 battled for the top spot. While these models have the raw processing power to outthink any human, their logic at the poker table showed that they still have some very human-like "bugs" to work out.

Renowned pro Doug Polk, who provided commentary on the event, noted that while the top-tier models were impressively aggressive, they often struggled with the most basic concepts of the game. specifically, when to fold.

One of the most telling moments of the finals came when o3 attempted to justify an all-in shove that Polk found questionable. The model’s reasoning was a classic example of the sunk cost fallacy: it claimed that it couldn't fold because it had already "invested" too many chips in the pot. As Polk pointed out, this is a fundamental error in poker logic; once chips are in the pot, they no longer belong to you.

The lower-tier models, like Grok 4.1 and GPT-5 mini, fared even worse. In one baffling hand, both models shoved their entire stacks into the middle on a board that offered neither of them a pair or a legitimate draw. The reasoning? 

One model believed it held the nut flush draw, while the other was convinced it actually had the flush. It turns out that even the world's most advanced AI can't win if it can't tell the difference between a draw and a made hand.