Decoder. plain-English AI glossary

Mechanistic Interpretability

▲ Rising

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

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