NCM episode · Oct 23, 2026 · 27 min

What Is an LLM & What Replaces It — NCM Explained

I don't think LLMs are it. I think we've gone about as far as we can go with them — and this video is me walking through the research that bets the future is CONCEPTS, not words. It's called Next Concept Modeling (NCM): instead of predicting one token at a time, it generates from coarse concepts down to fine detail in just a handful of parallel steps. The paper claims 21x–31x faster generation with quality that matches or beats — and my favorite result is a little model sweeping Sudoku. I also pull up Vox's video on why AI art struggles with hands, because it's the perfect picture of knowing how things LOOK versus how they WORK.

In this episode

What you'll see

  • Why the black box bothers me: your words take a journey nobody can fully trace
  • One token at a time — how LLMs actually generate (and why that's slow)
  • NCM: coarse-to-fine concept layers, one step per layer instead of 1,000
  • The results: 21x faster with better quality, 31x faster on summarization, 4–5x fewer training tokens — and a Sudoku sweep
  • The honest caveats: here's what the hype leaves out
  • Watch-along: why AI art struggles with hands (Vox) — pattern recognition vs. real forms
  • The model zoo: LLM, SLM, VLM, LAM, LCM, SAM, judgment models, MOE — AI isn't one thing, it's eight
  • From next-token to next-concept: why I bet this is the direction
Every claim sourced

Sources cited in this episode

  1. Hazy Research (Stanford) — Hermann Kumbong — Next-Concept Modeling: Fast, Parallel Coarse-to-Fine Text Generation (arXiv 2610.14162) — the paper this episode walks through; the 21x–31x speedups, 4–5x fewer training tokens, and ~100M–725M parameter sizes are the paper's own claims, tested at research scale (I flag that on camera too)
  2. Hazy Research — NCM code on GitHub (MIT license) — plus released model weights (hazyresearch/ncm-checkpoints on Hugging Face), so you can run the task models yourself
  3. Vox — Why AI art struggles with hands (video, shown on screen at 10:42) — the trapped-in-a-museum picture of knowing how things look vs. how they work; go subscribe to Vox
  4. Me — The Data Foundry: one prompt's journey through an LLM — my black-box video, the one the cold open points at
  5. Me — What is an Agent? — last week's episode; together with the Data Foundry it's the arc this one finishes
  6. + the full chapter list and timestamps in the video description
How I cite: the video I show on screen is credited and linked above — go subscribe. Primary sources linked wherever they exist; the performance numbers are the paper authors' claims and I say so when I quote them. Anything I'm not reading from a source is clearly flagged as my opinion. This episode is commentary and criticism, not financial, medical, or legal advice.
Community

Compare notes with me all week

I build all of this in the open, with a crew of AI agents. If you're directing agents too — or wondering where to start — the Discord is where we trade what actually works.