🌿 CuriousWinds

Sparse Autoencoders for Mechanistic Feature Extraction

Disentangling polysemantic neuron superposition into interpretable, atomic feature directions.

In 1666, Isaac Newton darkened his bedroom in Cambridge, cut a tiny hole in his window shutter, and placed a triangular glass prism in the beam of white sunlight. Prior to Newton, natural philosophers assumed that glass stained white light with colored impurities. But Newton proved the opposite: white light is not pure; it is a tangled composite of distinct colors. The prism merely unmixed what was already there, projecting a rainbow onto the opposite wall.

In modern artificial intelligence, Sparse Autoencoders (SAEs) are our triangular glass prism. As we saw in Part 5, transformer activations are tangled in polysemantic superposition, a single neuron acts like white light, mixing dozens of unrelated concepts together. For years, this made neural networks impossible to read. But in late 2023 and 2024, researchers at Anthropic and OpenAI deployed massive Sparse Autoencoders that cracked the code: extracting pure, human-interpretable, monosemantic concepts directly from the neural mind.

01 TRANSFORMER CIRCUITS · FEATURE DISCOVERY

How a Sparse Autoencoder Works

The Mathematical Prism

A Sparse Autoencoder is an auxiliary neural network trained to reconstruct the activations of a target transformer. But unlike a standard autoencoder, an SAE has two radical architectural constraints:

Visual Architecture • SAE Topology

Overcomplete Expansion + L1 Sparsity Penalty

1. Overcomplete Latent Space (32x to 128x Expansion)

If the transformer’s residual stream has 4,096 dimensions, the SAE expands it into a massive hidden layer with 131,072 or 1,000,000 latent features. This gives superposed concepts room to breathe.

2. Extreme Sparsity (L1 / Top-K Penalty)

Out of 1,000,000 possible features, only 20 to 50 features are allowed to activate for any given token! This forces each feature to represent a single, pure, monosemantic idea.

02 COGNITIVE ARCHITECTURE · SPARSE RECONSTRUCTION

The Golden Gate Claude Moment

Monosemanticity in Action

In May 2024, Anthropic applied a 34-million-feature Sparse Autoencoder to Claude 3 Sonnet. Among millions of discovered monosemantic features, ranging from “deceptive personas” and “immunology” to “transit systems”, they isolated Feature #34M/311649: The Golden Gate Bridge.

Passive Auditing • Feature Detection

Universal Monosemantic Activation

This single feature vector fired whenever Claude read the words “Golden Gate”, saw a picture of the suspension bridge, read a poem about San Francisco fog, or translated “le pont du Golden Gate” in French. It was 100% monosemantic.

Active Steering • Weight Clamping

Golden Gate Claude

When researchers artificially clamped this feature’s activation to 10x its normal value, Claude became obsessed with the bridge! Asked: “What is your physical form?”, Claude replied: “I am the Golden Gate Bridge.. my suspension cables span the bay in iconic international orange.”

03 AGENTIVE INTERVENTION · DIRECTIONAL STEERING

From Passive Reading to Active Steering

Sparse Autoencoders do not just allow us to read an AI’s thoughts, they give us an interface to edit them. By steering specific feature directions up or down during inference, researchers can suppress bias, eliminate dangerous cybersecurity capabilities, and prevent sycophancy without fine-tuning weights.

In Part 7, we examine the gold standard for proving causality: Activation Patching, performing virtual neurosurgery to trace decisions to exact layers and heads.

Advance Your Knowledge

Ready for Neural Reverse Engineering? Explore the Flagship Treatise

This series translates the core principles of transformer circuit tracing, superposition, and monosemantic feature extraction into accessible visual models. For mathematical invariants, formal proofs, and PyTorch circuit pruning algorithms, consult our foundational monograph: Mechanistic Interpretability Transformer: 5 Empirical Rules.

Curated Inquiry Series • 10 Investigations

Mechanistic Interpretability in AI

An empirical exploration through neural residual streams, induction circuits, polysemanticity, and causal weight editing.

Primary Scholarly Provenance & Empirical Attribution
Card, S. K., Moran, T. P., & Newell, A. (1983). The Psychology of Human-Computer Interaction. Lawrence Erlbaum Associates; ACM SIGCHI Model of Human Processor & Saccadic Latency Foundations.
DOI: 10.1145/358896.358899