NCM Is this the End of LLMs?

Is Stanford's Next-Concept Modeling (NCM) the end of LLMs? It writes text 31x faster, but against which model, and does any of it reach your GPU? Every chatbot you use, ChatGPT and Gemini included, writes one token at a time, and every token makes your GPU read the model's full weights out of memory. That is why large language models feel slow on your own machine. NCM, from Stanford and Together AI, skips that. It writes in levels, a rough sketch first and the words last, using a multiscale residual VQ-VAE and an ordinary transformer that fills a whole level per pass. In this video I read the