I don't care if it's fast, if it makes the model dumber by cluttering up context.