Open-Sora 2.0 cuts the costs of creating a bleeding edge text-to-video AI model by using the right data, architecture, and training regime.