They mentioned reliability so I think they want the performance of current ML models, not what we had 10-20 years ago, which was terrible.