Diatom 1.8.1: the body takes shape

Say a thing; the body becomes it.

• Type cloud, dino, become a dragon, show me an octopus. The same living cells grow out of the frustule into that outline and hold it until you speak again. Let go returns it to glass. Chat that names nothing ("hello") is still spelled by the body, word by word.
• 1,379 silhouettes and 2,248 names from Noto Color Emoji (SIL OFL-1.1), baked into the app (electron/renderer/shapes.data.js, scripts/build_shapes.py). Aliases cover everyday names: dino → T-rex, kitty → cat, spaceship → rocket.
• A word the table doesn't know goes to the local model, which answers with one emoji; the body takes that outline. No image model, no web service.
• The word-trained rule already follows silhouettes zero-shot. train_words.py --shape-prob trains on them too, and the Colab notebook (notebooks/train_words_colab.ipynb, clones this tag) trains words and shapes together on an A100.

1.8.1 is 1.8.0 without stray build files. Model lock unchanged: Qwen2.5-14B-Instruct Q4_K_M (Apache-2.0), local only, Ollama Cloud refused. See COMMERCIAL.md.

Diatom 1.7.1

Fine-tune the word rule on an A100.

• notebooks/train_words_colab.ipynb — open in Colab, pick a GPU runtime, run top to bottom: clone at this tag → phase 1 (frustule→word) → phase 2 (frustule polish) → proof strip and GIF → long-run stability check → export web/weights.js → download a zip to drop into your clone.
• scripts/*.py now find nca/ from any working directory (no PYTHONPATH).
• scripts/train_words.py --pool sets the replay pool size for big-batch GPU runs.

Diatom 1.7.0

The body speaks.

• Channel 10 is the silica template. The diatom rule is fine-tuned so the frustule grows into a word written there, holds it, and returns to glass when it clears (scripts/train_words.py, scripts/demo_words.py).
• The desktop window runs the real automaton, cell for cell with PyTorch (tests/test_body_parity.py). Its grid is the 3D glass. There is no procedural frustule and no text layer on the body: the cells become each word of the reply, one at a time.
• The camera looks at the body. The felt-state meters sit in the dock.
• Shipped weights (assets/checkpoint_words.pt, web/weights.js) come from a CPU fine-tune. On the RTX 5060 Ti:

python scripts/train_words.py --steps 6000 --device cuda
python scripts/export_weights.py --checkpoint assets/checkpoint_words.pt
cd electron && npm install && npm start
ollama pull qwen2.5:14b

Model lock unchanged: Qwen2.5-14B-Instruct Q4_K_M (Apache-2.0), local only, Ollama Cloud refused. See COMMERCIAL.md.

Diatom 1.6.0

A desktop frustule you can talk to, with a local model that can be sold.

• Electron app: a 3D centric diatom (valves, girdle, ribs, areolae) morphs from what you say. Bloom raises it, cut takes a wedge, rest lets it heal.
• Locked model: Qwen2.5-14B-Instruct at Q4_K_M (Apache-2.0), about 9 GB, for an RTX 5060 Ti 16 GB. A 13B-class network in FP16 is about 26 GB and does not fit. This model is 14.7B parameters, 13.1B of them outside the embedding table, at 4-bit. Context is capped at 4096.
• Also allowed via DIATOM_MODEL: qwen2.5:7b (Apache-2.0), phi4 and phi4-mini (MIT). Llama and Ollama Cloud are refused.
• If Ollama is not running, the grounded narrator still speaks and the glass still moves.

cd electron && npm install && npm start
ollama pull qwen2.5:14b

See COMMERCIAL.md.