Sparse Autoencoder
An interpretability tool that untangles the dense, multi-meaning activations of an LLM into a massive dictionary where each feature corresponds to one clean,...
Think of It Like This
Like a sound engineer taking a chaotic stereo recording of a crowded room and isolating 50 individual, crystal-clear vocal tracks.
Because a language model compresses millions of concepts into thousands of neurons, individual neurons fire for many unrelated things (superposition). A sparse autoencoder takes these dense activations and expands them into a massive dictionary, enforcing an L1 penalty to ensure only a few features fire at a time. The result is a set of disentangled, single-meaning features that humans can understand.