We are excited because it presents a novel way to learn "latent tokens" of gene expression variation, overcoming many of the limitations of order(less) tokenization and learning schemes. We show its performance in generative and representation learning tasks.