In my view, if we aim to speed up progress toward the next generation, diffusion-model-based generation is a very promising direction. That said, there are likely many hurdles to tackle, because the output is not a sequence; it is produced in parallel.