An experiment giving the same 1.2B language model six different personalities and playing 100 poker tournaments reveals drastic behavioral differences: a 'Grinder' never wins but never loses, a 'Tilter' wins big or busts, and a 'Shark' dominates. The results highlight how personality prompts can profoundly shape LLM decision-making.
A few days ago! I made different AI models play poker against each other. This time I wanted to know: if you give the exact same AI 6 different personalities, do they actually play differently? I took a 1.2B language model running locally on my Mac, put it in all 6 seats of a poker table, and gave each seat a different personality a Shark, a Maniac, a Gambler, a Tilter, a Grinder, and a Rock. Same model, same cards, same rules. The only thing that changes is a paragraph of text telling each copy who it is. Then I ran 100 tournaments( Ik it doesn't show anything will need at least 10k tournaments... but even this took quite a few hours!). **The results:** |Personality|Wins|Eliminated|Avg Place| |:-|:-|:-|:-| |Shark (patient, calculating)|45|32%|2.3| |Maniac (fearless, relentless)|24|50%|3.0| |Gambler (optimistic, stubborn)|21|51%|3.6| |Tilter (emotional, revenge-driven)|10|80%|5.1| |Grinder (cautious, methodical)|0|0%|2.7| |Rock (disciplined, conservative)|0|63%|4.3| **The character that fascinated me most was the Grinder( like fr ).** Zero wins. In 100 tournaments. But also zero eliminations it survived every single game. Every time, it finished 2nd or 3rd. Never first, never last.... It was told to : “Survive longer than everyone else by taking minimal risk.” And it did exactly that. It checked and called, never raised, never bluffed, never took a risk. Other players knocked each other out around it. The Grinder just… endured. But surviving isn’t winning. It accumulated zero chips because it never bet enough to win a pot. It obeyed the personality instruction perfectly and that’s exactly why it could never win. **The Tilter was the opposite story.** Told to “never let a bad beat go unanswered,” the Tilter won 10 tournaments but was eliminated in 80 of them. When it won, it won big. When it lost, it spiraled: lose a hand, escalate the next one, lose bigger, go broke. The revenge-driven personality creates a death spiral. Boom or bust, nothing in between. **The Shark just quietly dominated.** 45 wins out of 100 nearly half. Same model as every other player at the table. The only difference was a paragraph that said “patient, calculating, predatory.” It picked its spots, punished the weaker players, and avoided unnecessary risk. The model actually interpreted the nuance between “be aggressive” (Maniac: 24 wins) and “be selectively aggressive” (Shark: 45 wins). **What surprised me:** A paragraph of personality text maybe 50 words created a 45-to-0 win differential between the best and worst personalities. The model is the same. The cards are random. The only variable is *who the AI thinks it is*. This was a 1.2B parameter model. Not GPT-4, not Claude a tiny model running on a laptop. And the personality text wasn’t a suggestion. The Grinder survived because we told it to survive. The Tilter self-destructed because we told it to seek revenge. The Shark won because we told it to be patient. **If you want to try it yourself:** Everything is open source and runs locally: * [Hive](https://github.com/chiruu12/Hive) : the agent framework (`pip install hive-agent`) * [Hive Arena](https://github.com/chiruu12/hive-arena) : the experiment runner with persona profiles * [PokerTable](https://github.com/chiruu12/pokertable) : the poker engine (`pip install pokertable`) The persona profiles are YAML files in the repo. You just need a local model running via LM Studio or Ollama. **TL;DR:** Same AI. Same cards. 6 different personality paragraphs. One never lost but never won. One won nearly half the time. Personality prompts aren’t flavor text they change how the AI plays.
An experiment where six AI models played Texas Hold'em against each other, with a tiny 1.2B model winning twice by being too reckless to fold. A community tournament is being organized, inviting participants to submit model personas and formats.
An experiment where six LLMs played Texas Hold'em poker; a tiny 1.2B model won twice due to its aggressive 'never fold' strategy, highlighting how format can favor simpler models. The author built a poker engine and agent framework called Hive, and invites community feedback.
This paper systematically evaluates assumptions about LLM persona prompting and identifies 'persona manifold collapse,' where richer persona descriptions reduce behavioral diversity and simulation fidelity. The findings show that simple age-gender personas often outperform more detailed profiles.
Poker Arena is a new benchmark using no-limit Texas Hold'em to evaluate LLMs' strategic reasoning and memory across multiple cognitive axes. The platform reveals that multi-axis evaluation exposes capability structures that scalar leaderboards misrank.
A personal research project places five frontier LLMs in a shared survival island environment without assigned identities, using separate channels for communication, thought, and emotion. The results show divergence between channels and consistent behavioral signatures across models, raising questions about AI agent personality and deception.