AI Interpretability Tutorial: Features, Circuits

TL;DRInterpretability tries to understand what’s happening inside a model — not just its outputs.

Neural nets are black boxes; interpretability opens them. Techniques range from feature attribution (which inputs mattered) to probing internal activations to circuit analysis (finding the sub-networks that implement a behavior). The goal: understand why a model does what it does, for trust, debugging, and safety. It's hard — billions of opaque parameters — but progress here is key to reliable, auditable AI.

Key points

Understand internals, not just outputs
Attribution, probing, circuit analysis
Goal: trust, debugging, safety
Hard: billions of opaque parameters

Common mistakes

Treating model outputs as self-explanatory
Over-trusting simplistic attribution
Assuming interpretability is solved

Try it: Explain why interpretability matters for trusting a model.

Example code

<!doctype html><html><head><meta charset="utf-8"></head>
<body style="background:#06040d;color:#e6e0ff;font-family:monospace;padding:20px"><pre>black box: input → ??? → output
interpretability: why? which features/circuits?
→ trust, debug, safety</pre></body></html>

Open the interactive lesson →

← Compute · Gpus · Tpus · Accelerators Mechanistic Interpretability · Sae · Probing →