Every tenth of a second of a Minecraft fight becomes a line of tokens: what the player sees, then what they did. A small transformer is pretrained on those games the way LLMs are pretrained on the internet, fine-tuned on the best players, then trained with reinforcement learning on wins.
The same stages a chat model goes through, applied to PvP. Live status from the training machine.
A top-down replay of the most recent recorded fight, drawn from the recording itself.
Loss is how surprised the model is by the next token (lower is better). Action accuracy is how often its top guess for a player's next action matches what the player really did, measured on held-out games.
Four strengths, from someone new to PvP up to a bot with no human limits at all. Ratings are measured from real fights between the engine bots on the local server, on the Elo scale with human fixed at 1500.
Pretraining data comes from engine-vs-engine games at random settings on random modes. Fine-tuning data comes from real people playing on the local server.
Text, not pixels: the model reads the game state, the same facts a player sees on screen, with nothing a player couldn't see.
About 55 tokens per tenth of a second. The game header comes once at the start, like a system prompt, and tells the model who it is playing as.
| Type | Decoder-only transformer |
| Size | 10.8M parameters · 6 layers · width 384 · 6 heads |
| Parts | RMSNorm, rotary positions, SwiGLU, tied embeddings |
| Context | 1,024 tokens, about 2 seconds of play with full detail |
| Vocabulary | 467 tokens, each one a binned game fact |
| Optimizer | AdamW (0.9, 0.95), warmup + cosine, clip 1.0 |
| Speed | About 40 ms per decision on a laptop |
What changed, in order, newest first.