Выпуск
Is This the END of LLMs? NCM Could Change Everything
AsianDadEnergy's Podcast · 23 октября 2026 г. · 13 мин
NCM: The LLM Killer? Stanford May Have Changed AI Forever Hello world. I’m an unemployed ex-Big Tech software engineer with 25 years of experience in the tech industry. And one of the perks of involuntary early retirement is that I have plenty of time to follow the latest developments in AI. Which is great, because the AI industry seems determined to change the world every five minutes. But here’s something I’ve been thinking about for a while. Almost the entire AI conversation revolves around Large Language Models, or LLMs. Every few months, we get a bigger model, a more expensive model, a new benchmark, and another announcement about how this one is definitely going to change everything. Meanwhile, the companies building these models are pouring staggering amounts of money into GPUs, data centers, electricity, and cooling infrastructure. And yet, today’s LLMs have some fundamental limitations. What if the next major AI breakthrough doesn’t come from making LLMs bigger? What if it comes from building AI differently? Well, researchers at Stanford University have been exploring a new architecture called Next-Concept Modeling, or NCM , and the early results are interesting enough to deserve our attention. Could this be the beginning of the end for LLMs? Maybe. Maybe not. But let’s talk about why this research caught my attention. 1. The Problem With Today’s LLMs First, let’s talk about how conventional LLMs work. At their core, these models predict the next token based on the prompt. A token might be a word, part of a word, or another small unit of text. The model predicts a token, adds it to the sequence, and then predicts the next one. This process continues until the model has generated its response. In grossly simplified terms, it’s a huge autocomplete. But their autoregressive generation process has a major drawback. The tokens in an answer are generated sequentially. Suppose you ask an LLM to generate a 1,024-token response. The model generally needs to perform a forward pass for each newly generated token, meaning roughly 1,024 sequential generation steps. There are optimizations, such as caching previously computed attention states, speculative decoding, and other inference techniques. So this doesn’t mean the entire model is recomputed from scratch every time. Nevertheless, generating long responses requires repeated computation, and model weights and associated data must be accessed as needed throughout inference. With models containing trillions of parameters, that can get expensive very quickly. And if you enable a reasoning mode that generates additional intermediate tokens, that only increases the number of passes and the amount of computation required before you get your final answer. The result is an AI industry that increasingly depends on enormous amounts of computing power, memory bandwidth, and energy. Now compare that with the human brain. The brain performs an astonishing range of cognitive tasks while consuming roughly 20 watts of power. Today’s frontier AI systems can require vastly more energy and computing infrastructure, particularly when training and operating at scale. That comparison isn’t perfectly apples to apples. Biological brains and digital neural networks work very differently, and their capabilities aren’t directly comparable using a single energy metric. Still, the contrast raises an important question. Are we approaching intelligence in the most efficient way possible? I’ve argued before that if we see major algorithmic breakthroughs that move us closer to more capable reasoning systems, they might come from somewhere outside the architectures dominating today’s AI industry. And this is where NCM gets interesting. 2. Meet NCM: A D