Ermon was looking for a way to apply a diffusion approach to text because, unlike with LLMs, which work sequentially, diffusion models start with a rough estimate of data they’re generating (e.g ...