stencil-life

The smallest from-scratch networks trained to reproduce Conway's Game of Life exactly.

Trained from scratch by Idle Intelligence.

MIT

Three from-scratch alternatives to the language-model approach in llm-life: a one-layer BERT-style encoder over the 9 neighborhood cells, a 2-layer MLP over the same 9 values, and a stencil-masked attention model that takes the whole grid in one pass. All three are exact on every one of the 512 possible neighborhoods; the stencil model’s weights, trained only at 16x16, stay exact up to 256x256 because the mask encodes the rule, not the grid size.

Type
model
Runs
browser, WASM, WebGPU
Size
3 models, 1.4K-3.5K params each
Demo (GitHub Pages)
https://idle-intelligence.github.io/llm-life/web/
Source
idle-intelligence/llm-life
HuggingFace
idle-intelligence/stencil-life
Status
maintained

← Home