Interpretability
Also known as: mechanistic interpretability
Interpretability is understanding how a model works internally — what its layers and activations are actually computing.
- Mechanistic interpretability: reverse-engineer the circuitry.
- Feature/activation analysis.
- Harder and deeper than output-level explainability.
The frontier: interpretability is where we go when we need to trust a model we can't fully explain. It's one of the most active research areas in AI safety right now.