User embeddings are superpositions of interpretable behavioral modes
Sparse autoencoders recover named behavioral modes from embeddings and support zero-shot natural-language queries.
Stories per day · 30 days
1 stories, newest first · one card per story