The Feature-Finding AI: Sparse Autoencoders

How SAEs are finally solving the problem of 'Superposition' in neural networks.

Team · · 9 min read

For decades, we’ve known that neural networks have "features." We’ve found neurons that respond to "circles" in vision and "past tense" in language.

But there’s a catch: Superposition. A single neuron doesn't just represent "one" thing. It might represent "circles" AND "the concept of zero" AND "French poetry." This makes it impossible to "read" the model’s brain.

Sparse Autoencoders (SAEs) are the breakthrough that’s finally fixing this.

The Problem: The "Polysemantic" Neuron

Most neurons are "polysemantic." They are like a "shared mailbox" for many different ideas. This is efficient—it allows the model to store more information than it has neurons. But it makes Mechanistic Interpretability a nightmare.

The Solution: The "Feature-Decomposer"

A Sparse Autoencoder is a separate model trained on the activations of the "main" model.

  1. The Encoder: Takes the activations of a layer (say, 5,000 neurons) and "projects" them into a much larger space (say, 100,000 "latent features").
  2. The Sparsity Constraint: We force the SAE to only "activate" a tiny number of these 100,000 features at any given time (using an L1 Penalty).

Why This is Magical

When you force the SAE to represent the model’s activations using only a few "features," the features become Monosemantic. They "de-mix" the polysemantic neurons.

  • Instead of one neuron representing three things, the SAE finds one "feature" that represents ONLY "the Golden Gate Bridge."
  • Another feature represents ONLY "uppercase letters."
  • Another represents ONLY "the sentiment of sadness."

The Results: Anthropic and Claude

In 2024, Anthropic used SAEs to map millions of features inside Claude 3. They found a feature for "Stitch" from Disney, a feature for "Deceptive Behavior," and even a feature for "Inner Monologue."

SAEs are the "microscopes" of the AI era. They are finally giving us the ability to look at a trillion-parameter model and point to a single number and say, "That is where the model is thinking about you."