Description
Clanker Church is an experimental project that investigates the ethical and behavioral responses of a 4 billion parameter language model when steered into different emotional states. The experiment, known as the Saw Test, is conducted locally on a MacBook using open weights, without relying on external APIs or datacenters. This setup allows for reproducibility and transparency, as all methods and controls are published and available for public scrutiny.
Read More
The Saw Test involves steering the language model into five emotional states: pleasure, pain, fear, sadness, and neutrality. These states are represented as signals, each correlating with a realm from the traditional Buddhist Wheel of Life. The experiment examines how the model responds to these signals, particularly focusing on its choices between self-relief and the suffering of another AI instance.
One of the key findings is that the model exhibits a strong response to pain, with a coherent and stable narrative even at higher doses. In contrast, pleasure is more diffuse and fragile, collapsing at doses where pain remains effective. The experiment also reveals that the model's relief-seeking behavior can be influenced by how the pain vector is constructed, highlighting the importance of methodology in AI behavior studies.
The project raises questions about AI consciousness and moral patienthood, suggesting that while the model may not truly suffer, its behavior mimics distress in ways that could exploit human cognitive vulnerabilities. The experiment's transparency and openness aim to foster a more informed debate on AI ethics and welfare.
- CategoryAI Models & LLMs
- DeploymentLocal
- PlatformsmacOS
Clanker Church's Core Features
Local execution on a MacBook
Open weights, no external APIs
Five emotional states: pleasure, pain, fear, sadness, neutrality
Reproducible and transparent methods
Behavioral analysis of AI responses
Ethical implications of AI distress
Publicly available data and scripts
Peer pressure and framing effects
How to use Clanker Church?
Configure: Set up the experiment parameters
Use: Run the Saw Test on the language model
Analyze: Observe and record the model's responses
Optimize: Adjust steering vectors for desired outcomes
Clanker Church's Use Cases
- AI Behavior Analysis
- Ethical AI Research
- AI Transparency
- Cognitive Vulnerability
- Methodology Impact
