Renaissance LabsLearn

Models · Safety and alignment

Looking inside AI: the science of interpretability

From single neurons to features and circuits

7 minIntermediate10 chapters

The film explains why individual neurons are hard to read, how superposition packs many concepts into few neurons, and how sparse autoencoders unpack them into features. It then covers steering with features, tracing circuits, links to hallucination, safety uses, and an honest account of how little of a model these methods explain so far.

What to take away

  • Single neurons are confusing because the film says models pack many concepts into overlapping directions, an idea called superposition.
  • Sparse autoencoders rewrite a model's activity as a short list of features from a much larger dictionary, each easier to name.
  • Turning a feature up changes what the model says, as in the Golden Gate Claude demonstration, which suggests features cause behavior.
  • The film reports that circuit tracing revealed planning ahead in rhymes and an intermediate Texas step before answering Austin.
  • The film stresses that even in the best studies these methods capture only a small fraction of what a model computes.

Go deeper

Read the transcript78 lines, click any line to jump

This transcript comes from automatic captions and may contain small errors.

Keep watching