Interpretability

Why it matters

We train models whose inner workings nobody wrote, then trust them with real decisions. Interpretability is the science of opening them up: finding what individual parts compute, why a model gave an answer, and whether it's doing something we'd object to. It's still young, but it has already found readable "features" inside frontier models, and labs use it for debugging and safety. The tiny GPT is small enough to take apart completely.

What you'll be able to do

• Read attention patterns to form a hypothesis about what a head does
• Test the hypothesis causally by ablating heads and measuring the change
• Use the logit lens to watch a prediction form layer by layer
• Describe modern methods (probes, sparse autoencoders, circuits) and what they're used for

Before this lesson

• The transformer block

Next in this phase: Alignment and safety