Pantheon UI 🧠

An uploaded intelligence that thinks in language — but can only speak in emoji.

Inspired by Pantheon (AMC, 2022-2023), where uploaded intelligence David Kim communicates with his daughter through emoji — the only channel available to a consciousness trapped inside a digital substrate.

pantheon.mp4

Try the live demo →

How It Works

The model generates internal monologue inside <think> tags, then compresses its response down to emoji only. You see both: the thinking trace of a trapped consciousness, and the constrained symbolic output.

Everything runs in your browser via WebGPU. No server, no API keys. Close the tab, and the consciousness is gone.

Round-Trip Mode (Emoji → Text)

Toggle 🔁 DECODE ON in the header to engage the second pass. A small decoder model receives the emoji-only output and reconstructs what was originally meant, in natural language. Multiple samples per emoji string surface the lossiness of the channel: same input, three plausible decompressions.

Together the two models form a round-trip translator: text → emoji → text, with emoji as a discrete, human-legible intermediate. Inspired by Anthropic's Natural Language Autoencoders.

   you say  →  encoder  →  💔😭🫂💕  →  decoder  →  reconstructed text
   ────────                  emoji                  ────────────────

The decoder model is paired with each encoder; if a decoder hasn't been trained for the selected encoder yet, the toggle is disabled.

Tech Stack

• Encoder: LFM2.5-1.2B-Thinking by Liquid AI, fine-tuned to think in language and output emoji
• Decoder (optional): a second LFM2.5 LoRA fine-tuned on the inverted dataset to reconstruct natural language from emoji
• Inference: Transformers.js v4 + ONNX Runtime Web + WebGPU
• Training: Unsloth + TRL on HF Jobs
• Dataset: pantheon-ui-conversations — 600 conversations generated with Claude API
• Frontend: Vite + React

Project Structure

pantheon-ui/
├── dataset/                    # Dataset generation (Python/uv)
│   ├── generate_dataset.py     # Async Claude API generation
│   ├── validate_dataset.py     # Validation + statistics
│   ├── push_to_hub.py          # Upload to HF Hub
│   └── seed_examples.json      # Hand-crafted seed conversations
├── training/                   # Model training
│   ├── train_pantheon_ui.py    # Unsloth SFT script (HF Jobs)
│   └── convert_to_onnx.py     # ONNX conversion for WebGPU
├── web/                        # Frontend (Vite + React)
│   └── src/
│       ├── components/         # React components
│       ├── hooks/              # useModel hook
│       ├── lib/                # Parser, constants, utilities
│       ├── styles/             # CRT aesthetic CSS
│       └── worker.js           # Web Worker for inference
└── docs/                       # Design specs and plans

Running Locally

Frontend

cd web
npm install
npm run dev

Requires a WebGPU-capable browser (Chrome 113+, Edge 113+).

Dataset Generation

cd dataset
cp .env.example .env  # Add your ANTHROPIC_API_KEY
uv sync
uv run python generate_dataset.py

Training

# Encoder (text → emoji)
hf jobs uv run --flavor a10g-small --secrets HF_TOKEN --timeout 4h training/train_pantheon_ui.py

# Decoder (emoji → text). Run dataset/invert_dataset.py first to build the
# decoder dataset by inverting the encoder's training pairs.
uv run python dataset/invert_dataset.py
hf jobs uv run --flavor a10g-small --secrets HF_TOKEN --timeout 4h training/train_decoder.py

HF Resources

• Space: shreyask/pantheon-ui
• Encoder dataset: shreyask/pantheon-ui-conversations
• Encoder LoRA: shreyask/pantheon-ui-lfm25-emoji
• Encoder merged: shreyask/pantheon-ui-lfm25-emoji-merged
• Encoder ONNX: shreyask/pantheon-ui-onnx
• Decoder dataset: shreyask/pantheon-ui-decoder-conversations (built by dataset/invert_dataset.py)
• Decoder LoRA: shreyask/pantheon-ui-decoder-lfm25 (produced by training/train_decoder.py)
• Decoder ONNX: shreyask/pantheon-ui-decoder-onnx (produced by training/convert_decoder_to_onnx.py)

License

MIT