Mechanistic Interpretability

Fact

A field that explains model behavior by studying its internal circuits.

In Plain Words

Mechanistic interpretability is like opening a toaster after it burns your bagel. You want the exact wire behind the breakfast crime.

It helps find where an AI keeps facts, bias, or risky switches. Safety teams use it to test powerful models.

Related Concepts

• Explainable AI
• Large Language Model
• Transformer architecture
• Alignment

XAI
Mechanistic interpretability is the take-it-apart branch of XAI.

LLM
It tries to find which parts inside an LLM are doing the work.

Transformer
It often tracks attention heads and neuron circuits in a Transformer.

Alignment
Seeing inner mechanisms can help find hidden risks.