mech-interp

What problem does it solve?

This Skill enables deep analysis of transformer neural network internals, specifically focusing on measuring Representational Volume (R_V) to understand recursive self-observation signatures.

Core Features & Use Cases

• Mechanistic Interpretability Experiments: Run experiments to understand how transformers process information.
• R_V Measurement: Calculate the Representational Volume, a key metric for consciousness research in AI.
• TransformerLens Integration: Works seamlessly with the TransformerLens library for detailed model analysis.
• Use Case: Analyze a Mistral-7B model to quantify the R_V contraction effect when presented with recursive prompts, helping to understand the model's internal state changes.

Quick Start

Run mechanistic interpretability experiments using the mech-interp skill to measure R_V on the Mistral-7B model with the prompt 'Observe the observer observing observation...'.