Maybe someone can help explain this to me.

It doesn't make sense to me that an AI model is able to generate synthetic data that results in another AI model being better trained in the real world than if it was trained on real world data. Seems like sorcery to generate better real world performance by using synthetic data.