Diffusion language models (dLLMs) can write text in any order, but common decoding methods still prefer left-to-right, which wastes their superpower.
Diffusion models sometimes copy training images too closely, which can be a privacy and copyright problem.