Member-only story
Sketch first, words later: a new way for AI to write, and what it could mean
When ChatGPT or Claude answers you, it doesn’t come up with the whole reply and then show it. It writes one small chunk of text (a “token”, roughly three quarters of a word), then the next, each one depending on the last. A 400 word answer is about 500 of those steps, one after another. No skipping ahead.
That’s a big part of why AI is so expensive to run. Tiny steps, multiplied across millions of users, add up to an enormous amount of computer time.
A team at Stanford and the AI company Together AI tried something different. What if a model sketched the rough idea of a passage first and filled in the words afterwards, lots of them at once? They call it Next-Concept Modelling, or NCM. Their small test models write 10 to 31 times faster than the best rival approach, and they crack hard Sudoku puzzles that chatbot-style models almost always fail.
There’s a catch. The models are tiny and can’t hold a conversation. I still think the idea is worth understanding, so here’s how it works, what the results really show, and where it might turn out to be useful.
Three ways to write a sentence
Picture three people writing a short story.
