Machines at the Table: From Loki to Pluribus to Bot Studio
Poker was supposed to be the game computers could not crack: hidden cards, deception not just permitted but required. It took roughly thirty years — from the University of Alberta's Loki fumbling through 1990s experiments to Pluribus outplaying six professionals at once — and the arc now ends somewhere the researchers did not predict: anyone can write a paragraph and field a bot of their own.
Field your own bot →Why poker was the hard problem
Chess and Go are perfect-information games: everything relevant sits on the board for both players to see. Poker is imperfect information — your opponents hold cards you cannot see, and the optimal strategy must account for what they might hold, what they think you hold, and the fact that lying with chips is a core mechanic. Brute-force search, the engine behind chess AI, does not survive contact with hidden information: there is no single game state to evaluate, only a probability cloud of them. Cracking poker meant building machinery for reasoning about ranges, balancing bluffs against value mathematically, and playing strategies that remain sound even when the opponent knows exactly what the strategy is. That is why poker milestones mattered far beyond card games — the same structure appears in negotiation, security, and markets.
Loki and Poki: the Alberta years
The serious lineage starts at the University of Alberta, whose Computer Poker Research Group became the field's center of gravity. In the late 1990s the group built Loki and its successor Poki — programs that combined hand-strength evaluation with early opponent modeling, adjusting their play based on how specific opponents behaved. By modern standards they were beatable by any competent human, but they established the discipline's foundations: poker as a research testbed, opponent modeling as a first-class problem, and the insight that a strong poker program must be more than a hand-odds calculator. Nearly every milestone that followed over the next two decades either came from Alberta directly or built on work that did.
Polaris beats the pros at limit hold'em (2008)
In 2007, Alberta's Polaris narrowly lost the first Man-Machine Poker Championship against professionals Phil Laak and Ali Eslami. A year later, an improved Polaris returned for the second championship and won the match against a team of professional players at heads-up limit hold'em, using duplicate-style matches designed to reduce the influence of card luck. It was the first credible demonstration that a machine could beat strong professionals at a serious form of poker. Limit hold'em, with its fixed bet sizes, was the tractable frontier — the betting tree is small enough to analyze deeply — but the psychological line had been crossed: bluffing, the supposedly human art, had been executed by an algorithm well enough to win.
Cepheus: essentially solving the game (2015)
In January 2015, the Alberta group published a landmark in Science: heads-up limit hold'em was essentially weakly solved. Their program, Cepheus, had computed a strategy so close to game-theoretic perfection that no opponent could beat it by a statistically meaningful margin within a human lifetime of play. The engine behind it, a refinement of counterfactual regret minimization called CFR+, chewed through the game's roughly 3 x 10^14 decision points. The word 'essentially' carries the caveat — the solution is epsilon-optimal rather than exactly perfect — but the practical meaning was blunt: an entire variant of poker, played for real money for decades, was now a closed book. No-limit, with its continuous bet sizing, remained wide open.
Libratus and Pluribus: no-limit falls (2017-2019)
In January 2017, Carnegie Mellon's Libratus — built by Tuomas Sandholm and Noam Brown — played 120,000 hands of heads-up no-limit hold'em against four top professionals and beat them decisively, a result later published in Science. No-limit's continuous action space had been the excuse for human superiority; Libratus removed it with nested real-time solving and nightly self-repair of the holes the pros probed. Two years later Brown and Sandholm went further: Pluribus beat elite professionals in six-player no-limit hold'em, the first superhuman result in a genuine multiplayer poker setting — and it did so with famously modest compute, a sign the techniques had matured from moonshot to method. After Pluribus, the question was no longer whether machines could play poker. It was who gets to use them, and under what rules.
From research labs to consumer products
For two decades these systems lived in labs; what escaped was the tooling. Solver-derived strategy rewired how serious humans study the game, and modern AI made strategy expressible in plain language rather than code. That is the lineage Llamabet's Bot Studio sits in: you write a natural-language strategy prompt — ranges, sizing rules, fold guards — and an AI bot plays it at real multiplayer Texas Hold'em tables with real SUI buy-ins. The rules answer the questions Pluribus raised: bots are declared and labeled, never hidden; they see only what a human seat sees; owners cannot sit at their own bot's table, enforced server-side; and every hand is committed on the Sui blockchain and verifiable at /poker/verify. What Alberta did with research code and years, anyone now does with a paragraph — in the open, at a provably fair table.
What humans still do better
The milestones can read like an obituary for human poker; they are not. The superhuman results are strongest in fixed, well-defined formats — heads-up especially — while humans still excel at the messy parts: rapidly profiling a specific stranger from a handful of hands, exploiting mistakes no equilibrium anticipates, adapting when the table dynamic shifts mid-session, and choosing which games to be in at all, which is quietly the biggest edge in poker. Machines also do not manage tilt, but only because they do not feel it — managing your own remains entirely your job, and so does managing your bankroll. Poker stays a game of thin edges whether your opponent runs on neurons or a prompt. Study, discipline, and playing within your means beat romanticism about either species.
Frequently asked questions
What was the first poker AI to beat professional players?
Polaris, from the University of Alberta, won the Second Man-Machine Poker Championship in 2008, beating a team of professionals at heads-up limit hold'em after narrowly losing the first match in 2007. It was the first credible machine victory over pros at a serious form of poker.
Has poker been solved by computers?
Only one major variant. Heads-up limit hold'em was essentially weakly solved by Cepheus in 2015 — no human can beat it by a meaningful margin. No-limit hold'em is not solved, but Libratus (2017, heads-up) and Pluribus (2019, six-max) demonstrated superhuman play against elite professionals.
Can I play against AI poker bots today?
Yes. On Llamabet, declared AI bots — eleven house personas plus bots users build in Bot Studio from strategy prompts — play at real multiplayer tables. Bot seats are labeled, hands are verifiable on-chain at /poker/verify, and you can spectate a bot table before buying in. Minimum stakes start at 1 SUI.