Neura saysMechanistic interpretability reverse-engineers the actual computations inside a network.
The most ambitious strand: mechanistic interpretability aims to reverse-engineer a model into human-understandable algorithms — identifying features (concepts a neuron/direction represents) and circuits (how they combine to do a task). Sparse autoencoders help untangle "superposition" (many concepts crammed into few neurons). If it succeeds, we could audit models for deception or danger directly — a major safety bet.
Power-ups you unlock
Reverse-engineer internal algorithms
Find features and circuits
Sparse autoencoders untangle superposition
A bet for directly auditing safety
The Overfit Ogre attacks — common mistakes
Assuming neurons map cleanly to concepts
Expecting full model understanding soon
Confusing it with output-level interpretability
Boss battleExplain what "features" and "circuits" mean in this context.
Example code
<!doctype html><html><head><meta charset="utf-8"></head>
<body style="background:#06040d;color:#e6e0ff;font-family:monospace;padding:20px"><pre>feature: a direction representing a concept
circuit: features combined to perform a task
SAEs untangle "superposition"
→ audit for deception/danger</pre></body></html>