Definition
Plain language
Finding the underlying alphabet of concepts whose combinations explain a model's tangled internal activity.
As stated in the literature
An unsupervised technique that learns an overcomplete set of basis directions (features) whose sparse combinations reconstruct model activations; sparse autoencoders are the instantiation used to extract interpretable features in Scaling Monosemanticity.
Why it matters: It turns a model's confusing internal state into a readable set of concepts, which is a key step toward understanding how AI thinks.
For example, it can take a model's tangled internal activity and tease out separate, nameable ingredients like 'legal language' or 'mentions of bridges.'
Heard on the show
“The general name for this whole approach is dictionary learning: find the underlying alphabet whose sparse combinations explain your data.”Episode 098 — Finding Millions of Readable Concepts Inside a Real, Deployed AI Model