NCM Explained: How AI Generates Text Beyond Word by Word

Today's large language models write their answers one token at a time. They still feel fast, but every token waits for the one before it. A new Stanford paper, Next-Concept Modeling (NCM), tries something different: it sketches the whole answer first, then fills in the details, layer by layer. In this video, we look at how a regular LLM writes, how diffusion language models learn and write, how NCM's tokenizer (and its learned codebook) and generator work (with a real layer-by-layer example from the paper), why it's called Next-Concept Modeling, what the researchers measured, and where the catch is. Chapters 00:00 Intro 00:46 The big picture 01:18 How a regular LLM writes 02:06 Diffusion, in depth 03:25 What NCM writes in 03:52 How NCM is trained 05:23 How NCM writes, and its name 06:42 How much faster? 07:14 Memory vs compute 08:14 The Sudoku test 08:41 The catch 09:06 Conclusion Sources Paper: Kumbong, Khare et al., "Next-Concept Modeling: Fast, Parallel Coarse-to-Fine Text Generation" (2026): https://arxiv.org/abs/2610.14162 Project page: https://scalingintelligence.stanford.... Code: https://github.com/HazyResearch/ncm MDLM, masked diffusion language models: https://arxiv.org/abs/2406.07524 DUO, The Diffusion Duality: https://arxiv.org/abs/2506.10892 LLaDA, Large Language Diffusion Models: https://arxiv.org/abs/2502.09992 The layer-by-layer translation example is quoted from the NCM paper (Appendix, Table 12). Tags: #ncm #Next-Concept-Modeling #AI #LLM #diffusion #diffusionmodels