Stanford's Hazy Research released Next-Concept Modeling. Rather than writing word by word, it drafts text coarse-to-fine: a tokenizer packs the text into a hierarchy of codes, then a transformer fills in one level at a time, every code in that level at once. Generation takes a handful of steps.
The 10–31x speedups are against diffusion language models, not the autoregressive LLMs your apps actually run on. The published models sit between 97M and 725M parameters, and the authors list further scaling as open work.
It blurs fine detail too. Reruns outside the lab found names and numbers slipping.
A faster way to sample text. Not a replacement for the models you actually use.