Training a model on distilled AI data is an order of magnitude cheaper than training on original human data.