I’ve been building PKRT, a poker training app with its own game engine, and ended up asking a slightly ridiculous question:
Could I build a tiny ChatGPT-style model that only speaks poker?
So I used PKRT’s game logic to turn poker situations into a simple structured language, then trained a small language model on that instead of on normal English.
That became Pokerese.
Very roughly:
PKRT generates the poker → Pokerese translates the situation → the model learns what action comes next
It’s tiny compared with ChatGPT, and it’s definitely not a solver, but it’s surprisingly interesting seeing what it can learn from poker alone.
The whole thing is inspectable too, so you can actually see the language, tokens, training process, mistakes, and results rather than just getting an answer from a black box.
PKRT: https://pkrt.poker
Pokerese: https://www.pokerese.space
I’m curious what poker players think of the idea — especially whether a poker-specific language/model is actually useful, or whether I’ve just taken the scenic route to reinventing a decision engine
Could I build a tiny ChatGPT-style model that only speaks poker?
So I used PKRT’s game logic to turn poker situations into a simple structured language, then trained a small language model on that instead of on normal English.
That became Pokerese.
Very roughly:
PKRT generates the poker → Pokerese translates the situation → the model learns what action comes next
It’s tiny compared with ChatGPT, and it’s definitely not a solver, but it’s surprisingly interesting seeing what it can learn from poker alone.
The whole thing is inspectable too, so you can actually see the language, tokens, training process, mistakes, and results rather than just getting an answer from a black box.
PKRT: https://pkrt.poker
Pokerese: https://www.pokerese.space
I’m curious what poker players think of the idea — especially whether a poker-specific language/model is actually useful, or whether I’ve just taken the scenic route to reinventing a decision engine
