I'm also curious about the training process and the hardware required to train a model of this scale. The blog post mentions it's a "breakthrough in self-supervised vision AI," but I'd love to see more details on the architecture and training stability at this scale.