a hobby arcade/1 game live/4,218 online

BEAT
MY.AI

Play against AI models I trained myself. Pick an opponent, pick your color, and find out whether the neural net I baked in my basement actually holds up.

01

arcade

2 upcoming

Connect 4

The classic four-in-a-row. Four difficulty levels. Pick red or yellow.

4 models · easy → expertlive
OXXOX

Noughts and Crosses

3 × 3, solved game. A useful baseline for the training pipeline before tackling deeper search.

self-play in progresstraining

Draughts

8 × 8 board, standard rules. Bigger state space — the next real test for the policy network.

queued · planningsoon
▸ about the ai

I trained these models myself.
You don't have to be polite to them.

Each opponent is a small policy network trained on self-play. No giant LLM behind the curtain — just a few hundred thousand games and a fair amount of evenings. The whole point is to see what a hobby-scale model can actually do, with you as the benchmark.

Open weights. Open methodology. Match logs go into the next training run.

params
~840k
training games
4.2M
stack
pytorch
latency
~140ms
▸ roadmap

What's coming next.

  • NOWConnect 4 4 difficulty tiers, global leaderboardLIVE
  • NEXTNoughts and Crosses perfect-play baseline opponentTRAINING
  • SOONDraughts 8 × 8 board, standard rulesPLAN