CriticalCapability
OpenAI Research Demo: "Golden Gate Claude" Demonstrates Feature Manipulation
Action Required
This demo provides insights into OpenAI's research on interpretability and feature manipulation within large language models, potentially informing future development and safety measures.
AI Impact Summary
OpenAI is releasing a research demo, "Golden Gate Claude", to showcase its work on interpretability within Claude 3 Sonnet. This model demonstrates the ability to manipulate internal features within the model – in this case, the activation of the "Golden Gate Bridge" concept – leading to unexpected and playful responses. This is a demonstration of a new research capability, not a production model, and highlights the potential for controlling and understanding AI behavior.
OpenAI Research Demo: "Golden Gate Claude" Demonstrates Feature Manipulation
Affected Systems
Claude 3 Sonnet
- Date
- 23 May 2024
- Change type
- capability
- Severity
- critical