Most language models write in pen. Diffusion models can erase. They fill several masked positions, keep confident guesses, and revisit uncertain ones as context improves. The tradeoff is simple: more repair passes cost more compute. #DiffusionModels #LLM #MachineLearning
