The older models were less good at inferring intent. The new ones just seem to do a better job.