MIT's new algorithm reduces machine learning training time by 45%, with a 20% cut in energy use, promising AI efficiency gains.