Conscious AI: Modeling Curiosity, Identity, and Human Alignment

A research implementation of a conscious AI agent with:

• Multi-head cognitive attention (different modes of thinking)
• Hierarchical self-model (physical, cognitive, narrative)
• Epistemic curiosity system
• Multi-component value system with immutable core values
• Value drift protection

Installation

# Clone repository
git clone https://github.com/iansteenstra/conscious-ai.git
cd conscious-agent

# Create environment
conda create -n conscious-agent python=3.10
conda activate conscious-agent

# Install dependencies
pip install -e .

Quick Start

Training

python scripts/train.py --config config/agent_config.yaml

Evaluation

python scripts/evaluate.py --checkpoint checkpoints/checkpoint_final.pt

Interactive Demo

python scripts/demo.py --checkpoint checkpoints/checkpoint_final.pt

Architecture

Core Components

1. Frozen Pretrained Model: LLaMA-3.2-3B for language capabilities
2. Cognitive Attention: 5 attention heads for different thinking modes
3. Self-Model: 3-level hierarchical representation of self
4. Curiosity Module: Epistemic drive for information seeking
5. Value System: Immutable core values + learned implementations

Novel Contributions

• Cognitive Multi-Head Attention: Different attention heads for perceptual, epistemic, prosocial, identity, and goal-oriented processing
• Value Preservation: Architectural guarantees against value drift
• Multi-Component Rewards: Local (head-specific) + global (system-wide)

Configuration

Edit config/agent_config.yaml to customize:

• Base model
• Architecture dimensions
• Training hyperparameters
• Reward weights
• Value preservation settings

Project Structure

conscious-agent/
├── conscious_agent/
│   ├── models/          # Agent architecture
│   ├── rewards/         # Reward computation
│   ├── training/        # Training loop
│   ├── evaluation/      # Evaluation suite
│   └── environments/    # Training environments
├── scripts/             # Training/eval scripts
├── config/              # Configuration files
└── tests/              # Unit tests

Key Features

Value Alignment

• Immutable core values: Cannot change during training
• Harm detection: Strong negative reward for harmful actions
• Continuous monitoring: Checks for value drift every 1000 steps
• Automatic recovery: Rollback if drift detected

Consciousness Components

• Self-model: Agent maintains model of itself
• Curiosity: Intrinsic motivation to explore and learn
• Metacognition: Agent monitors its own uncertainty
• Identity coherence: Actions consistent with self-concept

Training

• Multi-component reward system
• Curriculum learning
• Value preservation checks
• Comprehensive evaluation

Testing

pytest tests/

Citation

If you use this code, please cite:

@misc{conscious-ai-2025,
  author = {Ian Steenstra},
  title = {Conscious AI: Modeling Curiosity, Identity, and Human Alignment},
  year = {2025},
  publisher = {GitHub},
  url = {[https://github.com/iansteenstra/conscious-ai](https://github.com/iansteenstra/conscious-ai.git)}
}

License

MIT License

Acknowledgments

• Built on Hugging Face Transformers
• Inspired by neuroscience and cognitive science research
• Value alignment based on Anthropic's Constitutional AI