evolution-sandbox / README.md
sanyan-evo's picture
Upload README.md with huggingface_hub
2157bdc verified
|
Raw
History Blame Contribute Delete
2.04 kB
---
title: The Unspoken Thing
emoji: ❤️
colorFrom: red
colorTo: pink
sdk: gradio
app_file: app.py
pinned: false
---
# ❤️ The Unspoken Thing
**Build Small Hackathon 2026 · 🍄 An Adventure in Thousand Token Wood**
---
In the Thousand Token Wood, strange things happen.
30 AI agents learn to swing an Acrobot — a double pendulum that, in the wrong hands, is chaos. In the right hands, it's a dance.
Over 120 generations, they evolve from random flailing to controlled mastery. Along the way, one of them discovers something.
We hid a shape in the code. A small reward for trajectories that happen to pass through certain points. The AI was never told what this shape is. It was never told to draw a heart. It only knows it found something it wasn't told to find.
**No one told it to draw a heart. It just... did.**
Then it asks for one last swing. Not for points. Just because it can.
---
## How It Works
| Layer | What It Does |
|-------|-------------|
| **Evolutionary AI** | (μ+λ)-ES drives the learning — 30 agents, 120+ generations |
| **Hidden Shape** | A geometric reward hidden in the Acrobot environment. The AI discovers it on its own. |
| **LLM Narrator** | Qwen2.5-7B gives voice to the agent's discovery. The AI doesn't know the word "heart." |
| **The Last Swing** | When training ends, the agent swings one final time. Golden trajectory. No UI. Just the shape. |
**AI is load-bearing at every layer.** Remove evolution → no discovery. Remove hidden reward → no shape. Remove LLM → no voice. The experience IS AI.
---
## Run Locally
```bash
pip install -r requirements.txt
python app.py
# Open http://localhost:7860
```
---
*"We never told the AI to draw a heart. We only defined a hidden reward for trajectories that pass through a specific shape. The AI discovered this shape on its own, after generations of exploration. It has no idea what a 'heart' is. It only knows that this shape gives a reward."*
Built for the Thousand Token Wood. Where strange is good. Where joyful is the bar.