Interpretability
Concept
Safety & Alignment
Opening the black box: reading a model's internal features.
You're looking at a concept explainer: what this term means in plain English, and the records it ties together.
Mechanistic interpretability reverse-engineers what networks compute — Anthropic's team mapped millions of human-legible features in Claude (the 'Golden Gate Claude' demo made one famous) and traces circuits for planning and deception. The bet: we cannot fully trust what we cannot inspect, so understanding must catch up with capability before the stakes get too high.
How this connects to the rest of the stack
This record's links at a glance — click a group to jump to it.
Related
Sources
Mapping the mind of an LLM (Anthropic) ↗
Record reviewed as of 2026-01-15 — see Methodology.