Inception Labs has a new model called Mercury that generates text from pure noise through multiple "denoising" steps:

https://www.inceptionlabs.ai/

Will a frontier-level AI be trained to do text diffusion by 2028? This means that the quality of its output should be about as good as the highest outputs from other LLMs. It should generate the entire text in one shot and refining it with additional passes, similarly to Mercury. It doesn't count if it just writes the text like a traditional LLM and refines it, there has to be a more fundamental architectural change.