Mechanistic Interpretability
Also called Mech Interp
Reverse-engineering the actual internal circuits a model uses, neuron by neuron, to compute something.
Think of it like
Not just noting that a calculator adds, but tracing the exact wiring on the chip that carries out the addition.
Example
Researchers identify a specific set of attention heads that together track how balanced the parentheses are in code, and show how they combine to do it.
How it actually works
Where interpretability broadly asks "what mattered," mechanistic interpretability asks "what exact computation happened inside." It treats the network as a system to be reverse-engineered — finding features, circuits, and how they compose into behavior. The dream is to read a model’s reasoning the way you’d read a program. It’s early, painstaking work, but it’s where the hope lies for genuinely verifying that a model is safe rather than just testing its outputs.
For product teams
The frontier bet on truly trustworthy AI — proving safety from the inside, not just observing good behavior.
For engineers
Reverse-engineering internal features and circuits into human-understandable algorithms; aims to verify behavior structurally, not just empirically.
Related
- Interpretability — The broader umbrella it sharpens.
- Alignment — A tool for verifying alignment.
- AI Safety — A long-term plank of safety research.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome