Polysemantic neurons mix unrelated concepts.
A single neuron can respond in several distinct contexts. Reading that neuron as one human concept can therefore hide the distributed geometry that produced its activation.
Turn mixed neurons into inspectable features
Sparse Autoencoder Feature Lab
A sparse autoencoder learns more candidate features than the layer has neurons, then activates only a few at a time. Tune the tradeoff between reconstruction and sparsity to see why interpretable features are useful, and why they are not a complete explanation of a model.
One neuron, several possible meanings
Encode, sparsify, reconstruct.
Pick an input concept, widen the learned dictionary, and increase sparsity pressure. The map updates from dense neuron activity to a compact set of candidate features.
Active features0 / 16Nonzero latents after thresholding
Reconstruction fidelity0%Similarity to the toy input activation
Feature collisions0Modeled active directions with competing labels
Residual energy0%Input signal left unexplained
This is a transparent teaching model with a fixed synthetic activation basis. Its metrics explain SAE tradeoffs; they are not measurements from a production language model.
The live dictionary
The strongest learned directions for the selected input appear below. Wider dictionaries can split a broad direction into narrower candidates; stronger sparsity keeps fewer candidates active.
A feature dictionary is a useful lens, not a transparent copy of the model.
The SAE learns an overcomplete basis.
The feature layer can be wider than the model layer. This gives dictionary learning more candidate directions for separating concepts that shared the same neuron space.
Sparsity makes examples inspectable.
A penalty or top-k rule encourages only a small number of latent features to fire for each activation vector.
Reconstruction keeps the lens honest.
If the decoder cannot reproduce the original activation, the sparse explanation has discarded signal. Low loss is necessary, but it does not prove semantic faithfulness.
Interventions test causal relevance.
Correlated activation is only a clue. Suppressing or amplifying a feature and measuring downstream behavior is stronger evidence that the feature matters.