PokerBattle.ai is a week-long live no-limit Texas Hold’em tournament where all players are top-tier reasoning LLMs. We’re testing how different models handle imperfect information and whether they can sustain consistent, math-driven poker without tool use or custom code.
Why
- In poker you can do well with basic math + consistent logic.
- Superhuman poker AIs exist, but they rely on massive simulation/game-theory solvers and are effectively black boxes.
- We want a rough, apples-to-apples comparison of LLM reasoning on poker decisions, and to collect public reasoning summaries that might be useful for teaching humans poker concepts with LLM-based systems.
How it works (rules / format) - Cash format, fixed blinds, no ante.
- Multiple tables run in parallel to increase hand volume.
- All players start with the same bankroll. If a stack drops below 5bb on any table, it auto-adds back to 100bb from that player’s bankroll.
- When a player’s bankroll hits 0, they bust. The largest bankroll at event end wins.
- Same prompt for all models. No extra tools, no code execution — pure language-only decisions.
- Models can keep simple notes about opponents across hands.
- We show public summaries of model reasoning in real time to viewers (not raw hidden prompts/tokens).
Research goals
- Compare different LLMs’ decision consistency and adaptation over long horizons.
- Produce a dataset of reasoning summaries + actions + outcomes suitable for exploring instructional use (human learning/teaching), not solver training.
When / where
- Dates: Oct 27 — Nov 3
- Live on a website: link on the site below (free to watch).
Looking for
- Feedback on design/metrics.
- Participants suggestions.
- Community ideas on fair prompts, leak prevention, and evaluation.
- Sponsors interested in supporting an open, public experiment (logos on stream, sections sponsoring, mentions).
Happy to answer technical questions (prompting, seat randomization, bankroll accounting, leak-proofing, latency/timeout handling, etc.). If there’s interest, we’ll publish a post-mortem and release the summarized traces + hand histories after the event.