Why we didn't build AlphaZero for poker
Self-play from zero gives you a flawless robot. We wanted believable humans — which is a different, and cheaper, problem.
When people imagine "AI that learns a game from scratch," they picture AlphaZero — DeepMind's system that taught itself superhuman chess and Go through self-play. The obvious question: why not point that at poker? We looked hard, and chose not to. Here is the reasoning, because it shaped everything about our bots.
1. AlphaZero doesn't transfer to poker
Chess and Go are perfect-information games: the whole board is visible. Poker is imperfect-information— the cards you cannot see force randomized, balanced strategies and a kind of reasoning AlphaZero's search was never built for. The poker breakthroughs came from a different lineage entirely: counterfactual regret minimization (CFR), and the systems built on it — Libratus and DeepStack (2017), Pluribus (2019), ReBeL (2020).
2. Self-play from zero gives you the wrong thing
This is the crux. Learning poker from first principles converges toward a game-theory-optimalstrategy: balanced, unexploitable, superhuman — and nothing like a human, least of all a beginner. A from-scratch agent will never learn to limp with a weak ace, over-value top pair, or get scared off a bluff. Those "mistakes" are exactly what makes an opponent feel human and what a learner needs to practice against.
3. The cost story actually favours the cheap path
Even the strong, equilibrium-style bots are no longer a supercomputer affair — Pluribus reached superhuman six-handed play for a famously small amount of cloud compute, on CPUs. But that is still optimal play, not realisticplay. So our budget is better spent on realism than on chasing an equilibrium our players don't need to face yet.
What we built instead
- Transparent skill tiers — equity-driven, explainable bots that double as lessons.
- Human-mistake modelling — the four "leaks" that make a novice feel like a novice.
- Behavioural cloning — a small neural net that imitates a population of players, ready to be pointed at real hand histories.
A genuinely game-theory-optimal "boss" is still on the table for later — and the cheap way to get one is a precomputed CFR strategy, trained once and shipped as a lookup. But it is the top of the ladder, not the foundation.