♦ Poker and machine learning: what the game gave computer science

In July 2019, a program called Pluribus sat down against five professional poker players in a Pittsburgh casino room rigged with cameras and a dealer who followed its instructions. Over the course of 10,000 hands, it beat them. Not by cheating, not by reading tells. It won by solving a problem that had resisted computer science for decades: how to make good decisions when you do not know what your opponent holds, and when bluffing some fraction of the time is mathematically correct rather than a personality trait.

That result, published by Noam Brown and Tuomas Sandholm of Carnegie Mellon University, closed a research arc that had run since the 1990s. Chess and Go are games of perfect information: both players see the same board. Poker is not. Each player holds private cards, and the right play depends on what you believe the other side is holding, how they bet, and how they might be lying. That gap between what Deep Blue solved in 1997 and what Pluribus solved in 2019 is the gap between games with a knowable state and real decisions made under uncertainty, which is most of what humans and machines actually face.

This is the part of poker's story that rarely makes it into casino marketing or felt-table nostalgia. The game became a laboratory for algorithms that now sit inside auction systems, negotiation software and security resource allocation. It is worth separating what the research actually produced from who gets paid for it, because those are two different questions with two different answers.

Why poker, and not some other card game

Poker belongs to a class of problems game theorists call imperfect-information games. Bridge, Stratego and some negotiation scenarios share the structure, but poker has two properties that made it the preferred testbed. First, the rules are simple and fixed: a finite deck, clear betting rounds, explicit payoffs. Second, the outcomes are measurable in a single unit, chips or money, which lets researchers score an algorithm's performance without ambiguity.

Those two properties made poker tractable in a way that, say, modeling a trade negotiation is not. You can run a poker-playing algorithm against itself millions of times in an afternoon and get a clean number back. That loop, play, measure, adjust, repeat, is the same loop that drives most modern machine learning, and poker gave researchers a domain clean enough to run it at scale decades before anyone had the compute to do it on messier problems.

From Libratus to Pluribus: the research arc

The breakthrough did not arrive all at once. In 2015, Claudico, an earlier Carnegie Mellon program, lost to a team of professional players in heads-up no-limit Texas hold'em, though not by a statistically decisive margin. Two years later, Libratus beat four top heads-up specialists over 120,000 hands with a clear statistical edge, and the result held up under scrutiny because the format removed luck as an excuse: that many hands is enough to separate skill from variance.

Libratus solved the two-player case. Pluribus, in 2019, extended the approach to six players at once, which is a much harder problem, because game-theoretic guarantees that exist for two-player zero-sum games mostly do not exist for games with more than two players and misaligned incentives. Brown and Sandholm got Pluribus to perform well anyway, using an approach that approximated good strategies without needing a formal proof that those strategies were unbeatable. That shift, from provable optimality to strong approximation, is itself a useful idea outside of poker, because most real-world strategic problems involve more than two parties and no clean mathematical guarantee is available.

Poker gave computer science a cheap, measurable way to study decisions made without full information, years before anyone had the hardware to do it elsewhere.

What actually left the poker table

The specific techniques developed for Libratus and Pluribus, counterfactual regret minimization chief among them, are now reference methods taught in game theory and AI courses. Counterfactual regret minimization, or CFR, works by having an algorithm play against itself repeatedly, tracking how much it would have gained by acting differently at each decision point, and nudging its strategy toward lower regret over time. It does not require labeled data or human examples. It generates its own training signal purely from self-play, which is the same principle later used in game-playing systems for Go and other domains.

• Resource allocation under uncertainty, used in some approaches to cybersecurity patrol scheduling and infrastructure protection
• Auction design, where bidders have private valuations similar to private poker hands
• Automated negotiation agents, which must infer another party's goals from limited signals
• Robustness testing for AI systems, where an algorithm needs to perform well against an adversary trying to exploit its weaknesses

None of this means a poker-solving algorithm got repackaged wholesale into a logistics company's software. What transferred was the underlying mathematics: a family of techniques for handling hidden information and strategic deception that turned out to be useful well beyond a card table. Sandholm's lab at Carnegie Mellon has published on applying these methods to security games used by groups like the U.S. Coast Guard for patrol scheduling, which is a direct, documented line from poker research to a practical, non-gambling application.

The real economy: who gets paid, and how

This is where it is worth being plain about money, because the research story and the income story are not the same story. The people who profit from poker-and-AI research are, overwhelmingly, academics, university labs and the technology companies that hire them, not players at a table. Brown went on to work at Meta AI and later OpenAI, applying lessons from poker research to other domains. Sandholm co-founded Strategic Machine and Strategy Robot, companies that license game-solving technology to government and private clients for non-gambling applications.

That is a research-to-industry pipeline, and it looks like most other academic-to-commercial pipelines in computer science: publish, patent where applicable, spin out a company, sell expertise. It does not run through the felt. A professional poker player does not benefit financially from the fact that an algorithm can now beat them at heads-up hold'em. If anything, the opposite is true: the existence of superhuman poker-solving software has made online poker more vulnerable to bot use and collusion, and site operators have had to spend real money on detection systems as a direct consequence of the same research that advanced computer science.

For an ordinary player, none of this changes the basic financial reality of the game. Poker remains a form of gambling with real variance, and the fact that a well-funded research lab can build software that beats humans at a simplified version of it says nothing about whether any individual player will come out ahead over a session, a year or a lifetime. The honest summary is that poker generated genuine, citable advances in computer science, and that the money generated by those advances flows mainly to universities, labs and the handful of companies built around the resulting intellectual property, not to the rooms where the game is actually played.

What this means for how we should read poker research

When a headline says an AI beat professional poker players, the honest reaction is interest in the algorithm, not a signal about how to play better or a reason to expect better odds. The research is a genuine contribution to how machines handle uncertainty and deception, two problems far older than poker and far bigger than the game itself. Reading it as a gambling story misses what actually happened in that Pittsburgh casino room in 2019.

The honest frame is narrower and more interesting than the gambling angle: poker turned out to be a clean, well-defined version of a problem computer science badly needed to solve, and it solved it well before the rest of the world had similarly tractable examples to work with.