Tuomas Sandholm: Poker and Game Theory

The conversation presents heads-up no-limit Texas Hold’em as a demanding benchmark for AI because private hole cards, staged public information, and repeated betting create an imperfect-information game rather than a straightforward search problem. The discussion places Libratus in that setting: its match against elite professionals required a large number of hands to reduce variance, and the system’s achievement challenged the assumption that top human poker skill would remain superior. The interview also separates the cultural mythology of poker tells from the strategic reality at the highest level, where expert players are skilled at hiding tells and betting patterns become more important than cinematic readings of facial expressions or body language.

Sandholm describes abstraction as necessary because poker’s game tree is too large to solve directly. The relevant abstractions include the cards or information revealed by chance and the actions available to players. A finer abstraction is not automatically better: in an imperfect-information game it can produce a weaker strategy for the real game than a coarser abstraction, an effect the interview calls an abstraction pathology. The described system therefore uses automated, potential-aware information abstraction and a provably convergent action-abstraction method where it is scalable, while retaining caution about what an abstraction guarantees.