HamletRNAWorld

Simulated C. elegans worms crawl over a scrolling field of words from Hamlet. A worm eats a word by touching it; the eaten words are its poem. Each on-screen word emits a 12-channel chemosensory signal into the worm's amphid neurons, so the worm tastes meaning and steers toward it. Their 300-neuron connectomes evolve under a Natural Evolution Strategy, judged for poetic quality by a Claude model.

Live at https://wordswordsworms.org

Start here

Get it running on your own server	v7/docs/SETUP.md
How the simulation works	v7/docs/ARCHITECTURE.md
Working with Claude Code	CLAUDE.md — loaded automatically; carries the project's hard rules

Quick local run (no API key needed):

cd v7
python3 -m venv .venv && ./.venv/bin/pip install -e .
./.venv/bin/python -c "
import nltk
for p in ('averaged_perceptron_tagger','averaged_perceptron_tagger_eng','universal_tagset'):
    nltk.download(p, download_dir='.venv/nltk_data')"
WORMLET_GENERATIONS_ENABLED=0 WORMLET_N_WORMS=6 ./.venv/bin/python main.py --port 8000

Then open http://127.0.0.1:8000.

Repository layout

v1..v7 are evolutionary snapshots; each forks its predecessor and adds capability. v7 is the active version — never modify an older one. v6 is kept runnable as the comparison baseline.

Status

The chemosensory encoder is frozen (UMAP coordinates for meaning, a corpus-derived POS trigram for grammatical fit); only the connectomes evolve. An honest caveat, documented at length in CLAUDE.md and the architecture doc: no version of this has yet demonstrably learned to write better poetry once fitness is normalised by how much each worm ate. Recent work removed several mechanical reasons it couldn't. Whether it now does is open.