Mechanistic interpretability tries to reverse-engineer a neural network, finding the internal features and circuits that produce its behaviour.
In plain words
A trained model is billions of numbers with no labels. Mechanistic interpretability looks inside to find which patterns of activity stand for which ideas, and how they connect. One difficulty is that single neurons often respond to many unrelated things. Researchers use tools such as sparse autoencoders to split activity into cleaner features that people can name.
A quick example
Anthropic researchers found a feature inside Claude Sonnet that responds to the Golden Gate Bridge, whether it is named in English, in other languages or shown in an image. When they turned that feature up, Claude began describing itself as the bridge.
Where you'll meet it
- AI safety research on why models behave as they do.
- Debugging and auditing model behaviour.
Related
Sources
- PaperZoom In: An Introduction to Circuits, Olah et al. (Distill) · read 29 Sept 2026
- PaperTowards Monosemanticity: Decomposing Language Models With Dictionary Learning, Anthropic (Transformer Circuits Thread) · read 29 Sept 2026
- DocsMapping the Mind of a Large Language Model, Anthropic · read 29 Sept 2026
Checked by WhatIsAI against the sources above · 29 Sept 2026. Spotted an error? Report it.