Safety
Mechanistic Interpretability
2020ActiveUpdated: 22 September 2026Published
Key
innovation
Reverse-engineering a neural network at the level of circuits and algorithms, not just correlations.
Category
Safety
Abstraction level
Paradigm
Operation level
ModelInference
Use cases
AI safety and auditingCircuit/feature analysis (SAE)Detecting unwanted behavior
How it works
One studies neurons, attention heads, and features plus the circuits linking them; causal interventions confirm each component's role.
Problem solved
Large models are 'black boxes', making their behavior hard to trust and audit.