Comment by pinkbeanz

This is neat -- how would you think about evaluating the quality loss as you change to more efficient models? I saw you did an analysis on the number of messages, but wondering if there's more robust methods?