Seems like this might be used as an alternative to embeddings for creating semantic search. Run a document collection through this to get the activations, run t

Seems like this might be used as an alternative to embeddings for creating semantic search. Run a document collection through this to get the activations, run the prompt through the same thing and use the sparse autoencoder to identify which documents has the most activation of features in common with features that the prompt activates. You don't have to know what the features are, but if the prompt includes, say a rhetorical question, it may be able to find documents that also includes rhetorical questions. Not sure how well this would work, but it seems like something that could be interesting to try.