Dark Star Field

Every circulating recording of the Grateful Dead's "Dark Star" — 209 nights, 1968 to 1994, about 69 hours of music — embedded with a music-understanding model and rendered as a volumetric sky you stand inside.

Live: darkstar-field.pages.dev

Each performance becomes a star; each 5-second window of audio becomes a mote of fog. Direction on the dome comes from a UMAP projection of MERT embeddings, so nights that sound alike hang together. Altitude is time: 1968 low overhead, 1994 high and far, with the band's 1974–78 hiatus a genuinely empty band of sky. Tap a star and that night streams from archive.org while its trajectory arcs through the fog as a comet and the window-by-window embedding drives the weather. The 35 performances with no circulating recording are still up there — dark stars, visible but unplayable.

How it's built

jerrybase performance list
        │  pipeline/scripts/build_performances.py
        ▼
archive.org matching (soundboard-ranked, sandwich-aware)
        │  pipeline/scripts/build_manifest.py  →  data/manifest.json
        ▼
MERT-v1-95M embeddings, 5 s windows        (Colab GPU, notebooks/)
        │  one .npz per night — not committed (~136 MB)
        ▼
UMAP sphere shell + PCA-8 shader params
        │  pipeline/scripts/build_artifacts.py →  data/artifacts/ (1.1 MB int16)
        ▼
64³ era-colored voxel fog atlas
        │  pipeline/scripts/build_fog.py       →  fog.bin + fog.json
        ▼
web/ — Vite + three.js raymarcher, no backend, no framework

The web app ships ~2 MB of quantized artifacts and streams audio directly from archive.org. No server, no tracking, no build-time secrets.

Running it

Web (needs the committed artifacts, nothing else):

cd web
npm install
npm run dev        # http://localhost:5173
npm test           # vitest

Pipeline (Python 3.12, versions pinned for reproducible projections):

cd pipeline
python -m venv ../.venv && ../.venv/bin/pip install -r requirements.txt
../.venv/bin/python -m pytest tests   # 33 tests

Rebuilding embeddings needs a GPU: open notebooks/embed_dark_star.ipynb in Colab, point it at data/manifest.json, and it writes one npz per night to Drive. build_artifacts.py and build_fog.py take it from there.

Data notes

• The repo contains no audio — only metadata (dates, venues, archive.org identifiers and track URLs) and derived numerical artifacts.
• Recordings stream from archive.org's Grateful Dead collection, which the band has long permitted for non-commercial trading. Be kind to their bandwidth.
• Performance list derived from jerrybase.com (song 1052).

License

MIT — see LICENSE. The music belongs to the Dead and the tapers.