Uses small synthetic models to show how networks can represent more features than they have dimensions by storing them in superposition, and maps when and how this happens.

This is an external publication hosted at transformer-circuits.pub. AI Governance LK did not write it; the summary and notes on this page are our own.

Why it matters

Superposition is a central reason individual neurons are often hard to interpret. The toy models are small enough to train in minutes, which makes this a natural basis for teaching and replication.

What you will learn

  • Explain superposition and why it complicates interpretability
  • Reproduce the basic toy-model experiment

Before you start

Suggested background. This is editorial guidance, not a requirement.

  • Linear algebra
  • Training small networks in PyTorch

How to cite

Elhage, N., Hume, T., Olsson, C., et al. (2022). Toy Models of Superposition. Transformer Circuits Thread.