An AI that rewrites its own brain — and creates things that have never existed.
curious is a self-evolving cognitive architecture. It rewrites its own source code to get smarter. And every day, it creates something completely unique — scored on novelty, learning from feedback, pushing toward things no human or AI has ever built.
This is a live experiment. Watch it evolve.
Day 1: Created "Fluctuverse" — uniqueness 47/100 (too conventional)
Day 2: Created "Quintessension" — uniqueness 71/100 (invented its own language)
Day 3: ???
The
creations/directory fills up daily. Each creation is scored. The AI reads the scores, learns what "unique" means, and pushes further. The git log is the experiment.
We gave an AI three things:
- Its own source code (readable, modifiable)
- A uniqueness score (0-100, measures novelty)
- One rule: create something that has never existed before
Then we pressed start and walked away.
Every 6 hours — Self-Evolution: The AI reads its own cognitive architecture, finds weaknesses, rewrites the code, tests if the change improved performance. Keeps what works. Reverts what doesn't. Every modification is a git commit.
Every day — Creation: The AI creates a completely new artifact. Not a todo app. Not a chatbot. Something that doesn't fit any existing category. It's scored on uniqueness. The score feeds back into the next creation. The creations get weirder and more novel over time.
The question: Can a self-evolving AI produce genuinely creative artifacts that no human designed? Can novelty be optimized the way accuracy can?
creations/— Every artifact the AI has ever created, with uniqueness scorescurious/seed/— The AI's brain. Watch the diffs — it rewrites itself.- Git log —
🧬 evolve:= self-modification.🎨 create:= new creation.
curious/seed/ ← THE AI REWRITES THIS
├── world_model.py prediction engine
├── learner.py learns from surprise
├── curiosity.py finds knowledge gaps
├── metacognition.py observes its own learning
├── experimenter.py generates its own experiments
└── creator.py creates unique artifacts
Every file above is evolvable — the AI reads it, analyzes weaknesses, and rewrites it. The curious/harness/ directory contains the evolution loop itself — the untouchable "laws of physics."
creations/
├── day_001/ ← First creation
│ ├── fluctuverse.py the artifact
│ ├── meta.json what it is
│ ├── scores.json uniqueness breakdown
│ └── README.md the AI's explanation
├── day_002/ ← Second creation (scored higher)
│ ├── quintessension.py
│ └── ...
└── history.jsonl full creation log
Each creation is:
- A working artifact (real code, not just an idea)
- Scored on 4 dimensions of uniqueness (concept, implementation, structure, naming)
- Fed back into the next cycle (the AI learns what "unique" means)
pip install curious-ai
export OPENAI_API_KEY=sk-...
curious create --llm openai:gpt-4o-minicurious init --observe ./any-project --llm openai:gpt-4o-mini
curious startcurious gallerycurious explaincurious create --llm ollama:llama3 # free, local
curious create --llm openai:gpt-4o # strongest
curious create --llm groq:llama-3.1-70b # fast┌─────────────────────────────────────────────┐
│ CURIOUS ENGINE │
│ │
│ ┌─────────┐ ┌──────────┐ ┌──────────┐ │
│ │ Observe │ → │ Predict │ → │ Surprise │ │
│ └─────────┘ └──────────┘ └──────────┘ │
│ ↑ │ │
│ │ ┌──────────┐ │ │
│ └─────────│ Learn │←───────┘ │
│ └──────────┘ │
│ │ │
│ ┌──────────┐ │
│ │ Evolve 🧬│ ← reads own │
│ │ │ code, rewrites│
│ └──────────┘ it, tests it │
│ │ │
│ ┌──────────┐ │
│ │ Create 🎨│ ← builds │
│ │ │ something │
│ └──────────┘ unique daily │
│ │ │
│ ┌──────────┐ │
│ │ Score │ ← uniqueness │
│ │ │ measured │
│ └──────────┘ │
│ │ │
│ ↓ │
│ FEEDBACK → NEXT CYCLE │
└─────────────────────────────────────────────┘
| Workflow | Schedule | What it does |
|---|---|---|
| 🧬 Self-Evolution | Every 6 hours | Observes, predicts, evolves its own code |
| 🎨 Daily Creation | Every day midnight | Creates something unique, scores it, commits |
The repo evolves on its own. Star it and check back in a week.
Every creation is scored on 4 dimensions:
| Dimension | Max | What it measures |
|---|---|---|
| Concept | 30 | Has this idea existed before? |
| Implementation | 30 | Is the approach novel? |
| Structure | 20 | Did it invent its own paradigm? |
| Naming | 20 | Did it create its own vocabulary? |
Total: 0-100. The AI sees its score and optimizes for higher novelty.
Day 1: ██████████░░░░░░░░░░ 47/100 — used conventional frameworks
Day 2: ██████████████░░░░░░ 71/100 — invented its own language
Day 3: ?
What is this?
An experiment. Can a self-evolving AI be genuinely creative? We gave it a uniqueness score and told it to maximize novelty. The creations/ directory is the result.
Is this AGI? No. But it's probing the boundary. If the AI produces artifacts that are genuinely novel — things no human designed or imagined — that tells us something about machine creativity.
Can I run it?
Yes. pip install curious-ai, add your API key, run curious create. It works with OpenAI, Ollama, Groq, or any OpenAI-compatible API.
What will it create? We don't know. That's the point. It might invent a programming language. It might create a new kind of UI. It might build something we don't have a word for yet. The constraint is: it must be unique.
How much does it cost?
- Creation: ~$0.02/day (GPT-4o-mini) or ~$0.15/day (GPT-4o)
- Evolution: ~$0.01/cycle
- Ollama: Free
This is a live experiment in machine creativity and self-evolution. The AI creates something new every day. Star it and watch what happens.
Built by aumiqx
An experiment in artificial creativity.