sparse autoencoder interpretability

**Sparse Autoencoders (SAEs) for Interpretability** are the **unsupervised probing technique that trains a wide, sparsely-activated bottleneck network on the internal activations of a large model, decomposing polysemantic neurons into a much larger dictionary of monosemantic features that each correspond to a single human-interpretable concept**. **Why Superposition Is the Problem** Modern neural networks learn more semantic concepts than they have neurons. This forces the network to encode multiple unrelated concepts in the same neuron — a phenomenon called superposition. When researchers inspect individual neurons and find that one neuron fires for both "Golden Gate Bridge" and "the color red," no clean mechanistic story emerges. **How SAEs Solve It** - **Architecture**: An SAE is a single hidden-layer autoencoder trained to reconstruct a layer's activation vector. The hidden layer is intentionally much wider (e.g., 32x the residual stream width), and an L1 penalty forces most hidden units to stay at zero for any given input. - **Dictionary Features**: Each hidden unit (or "feature") learns to activate only for one interpretable concept — named entities, syntactic structures, sentiment polarity, or domain-specific jargon — effectively decompressing the superposed representation into a human-readable dictionary. - **Reconstruction Fidelity**: A well-trained SAE reconstructs the original activation with minimal mean squared error while using only 10-50 active features per input token, proving the decomposition captures real structure rather than noise. **Practical Engineering Decisions** - **Dictionary Width**: Wider dictionaries resolve finer-grained features but produce "dead" features (units that never activate) and increase training cost. - **Sparsity Coefficient**: Too little L1 penalty produces polysemantic features that defeat the purpose; too much forces reconstruction quality below acceptable levels. - **Layer Selection**: Residual stream activations in the middle layers of transformers typically yield the most interpretable features; early layers capture low-level token patterns and final layers are heavily entangled with the unembedding. **Limitations** SAE features that explain activations accurately do not automatically correspond to causal circuits — a feature may be statistically reliable but play no role in the model's actual decision. Causal intervention (ablation and patching) is required to confirm that a feature genuinely drives downstream behavior rather than merely correlating with it. Sparse Autoencoders for Interpretability are **the most scalable technique currently available for cracking open the black box of frontier language models** — converting a wall of inscrutable floating-point activations into a structured dictionary of human-readable concepts.