This is genius! There are tradeoffs between diffusion and neural network models in image generation so why not use diffusion models in text generation? Excited to see where this ends up and I wouldn't be surprised if we saw some of these types of models appear in the future updates to popular families like Llama or Qwen.