More training, more data, more parameters, more compute power... and voilĂ .
Hard to say... but we've been surprised more than once in machine learning history.