gmmeyer/gpt-bird

🐦 Dreaming Bird — a neural Flappy Bird world model

A small decoder-only transformer trained to be Flappy Bird. Given the recent frames and the player's action (flap / no-flap), it emits the next frame's quantized state. Looped autoregressively at ~30–60 fps, the model replaces the physics engine — there is no game logic at runtime; the network is the physics.

🎮 Play it in your browser: https://gmmeyer.github.io/gpt-bird/  (the page fetches model.onnx from this repo and runs it client-side via onnxruntime-web + WebGPU) 💻 Source: https://github.com/gmmeyer/gptbird

Files

File	What
model.onnx	The ~11M-param "small" model (full game with pipes). ONNX opset 17, dynamic sequence length, returns last-position logits. This is what the web demo runs.
config.json	Tokenizer offsets, quantization bins, and engine geometry — everything the JS decode loop and renderer need.
small_pipes.pt	PyTorch checkpoint for model.onnx (n_layer 6, d_model 384, ctx 256 frames).
nano_nopipes.pt	PyTorch checkpoint for the ~1.9M-param Phase-2 model (bird + gravity + flap, no pipes).

How it works

Each frame is 4 tokens — bird_y, pipe_dx, gap_y, status — generated with slot-constrained decoding: logits are masked to the legal id range for each field, so a malformed frame is impossible by construction. The gap_y slot is sampled (a newly revealed pipe's gap is genuinely unpredictable, so the dream invents one); the other slots are greedy.

Velocity is never in the observed state — if the bird flies well, the model must have reconstructed velocity from the sequence of positions (finite differences), which is the central experiment. Quantization (bird_y → 128 bins) doubles as a drift stabilizer.

Results

• One-step bird_y: 98.6% exact / 100% within ±1 bin (held-out).
• gap_y: 99.96% exact on stable frames; 100% validity on RNG spawn frames.
• Collisions: 94.9% within ±1 frame of the true engine.
• Free-rollout drift: bird_y never exceeded 3 bins over 16 runs (mean 0.49); a controller playing inside the dream threads 5+ pipes.
• Cacheless dreaming runs at ~125 fps on an RTX 5090.

License

MIT.