A tiny world model running games on a Raspberry Pi Pico 2 W
Repository · https://github.com/sheepseb/tinyworldformer
First seen by Oriel on .
AI summary
tinyworldformer runs a looped-transformer world model and streaming reinforcement-learning actor concurrently on a Raspberry Pi Pico 2 W, generating and playing five small games without executing their real rules.
- The world models contain 26k–131k parameters and require 2–8M multiply-adds per frame; actors are 9k–16k-parameter float32 MLPs.
- At two transformer loops, exact held-out frame accuracy ranges from 83.6% for Doom-lite to 100% for Flappy Bird, excluding unpredictable random-event transitions.
- Training uses 1–3 million recorded transitions per game, quantization-aware fine-tuning, and repeated real-game data collection to close exploits discovered by dream-trained agents.
- World-model matrices and activations use int8, while embeddings, normalization, residuals, and scaling remain float32; Rust reproduces the Python integer specification bit for bit.
- The no_std Rust firmware uses both RP2350 cores, static allocation, and 227–460 KB RAM, while Stream AC(λ) can learn online without replay buffers or batches.
#world models #embedded ai #reinforcement learning #rust
Generated by gpt-5.6-sol. Check the original for details.
Metric trends
Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained
Points
Score reported by the source.
- Hacker News: 270, latest observation ; +264 over observed period since
Comments
Comment count reported by the source.
- Hacker News: 19, latest observation ; +19 over observed period since
Aliases
| Source | External ID | URL |
|---|---|---|
| Hacker News | 50295450 | https://github.com/SheepSeb/tinyworldformer Read discussion |