Why Poker is the Ultimate Test for Artificial Intelligence

Games have long served as benchmarks for training artificial intelligence systems. In 1997, IBM’s Deep Blue famously defeated chess grandmaster Garry Kasparov. Fast forward a few decades, and Google DeepMind’s AlphaGo beat the 18-time Go world champion, Lee Sedol.

However, poker presents an entirely different challenge. Unlike Go or chess, where players have complete visibility of the game board, poker players must contend with “imperfect information”—they have no way of telling what hands their opponents may hold or what cards are yet to be dealt.

For this reason, no-limit Texas Hold ’em is one of the toughest challenges for an AI to master.

Tough… but not impossible.

Enter Libratus and Pluribus

In 2017, Tuomas Sandholm, a professor of computer science at Carnegie Mellon University, and Noam Brown, a then-Ph.D. student, created Libratus, an AI system that made headlines after defeating a group of professional poker players in No-limit Texas Hold ’em. After playing north of 120,000 hands, Libratus won over $1.7 million in chips, showcasing dominance even in highly uncertain, multivariable environments.

Two years later, Sandholm and Brown collaborated again; this time, with the help of Meta AI, they developed a new poker bot: Pluribus. Unlike Libratus, which excelled in 1-on-1 poker, Pluribus was purposely designed to handle the greater complexity of multiplayer poker. In 2019, Pluribus became the first AI system to defeat top human players in 6-player No-limit Texas Hold’ em—an incredible milestone in AI development.

Why These Models Are Unique

What sets these models apart from other AI systems is that they don’t require brute-force computation. Instead, they rely on “self-play” reinforcement learning—a technique where an AI model plays millions of hands against itself to develop an efficient “blueprint” strategy. This blueprint is then used for real-time decision-making, allowing the AI to draw logical conclusions during gameplay.

Cool… But Why Should We Care?

Poker requires players to make decisions in environments full of uncertainty. Is your opponent bluffing? Are you better off folding or going all-in? These high-stakes decisions—where the correct choices aren’t always clear—are common in many real-world situations.

Whether it’s maximizing profits in financial markets, accurately diagnosing a patient, or ensuring the safety of autonomous vehicles on the highway, the ability of an AI system to quickly evaluate risks and rewards in unpredictable environments is invaluable.

And the exciting part? This technology is still in its infancy!