I just had Claude Code download a language model from Stanford and get it running on my Mac. A small one. It doesn’t write the way Claude writes, and that’s exactly why I wanted it.
Claude, ChatGPT, Gemini, they all write one word at a time. Left to right. Every word is a bet on the next one, and the end of the sentence only turns up when they get there. The Hazy Research lab at Stanford has built something else and calls it Next-Concept Modeling. Their model starts with a rough sketch of the whole sentence and sharpens it in rounds, each round a bit longer and a bit more precise. Ten rounds, give or take, and there’s a sentence. According to their paper it’s up to 31 times faster than the other experimental models that try to get away from word by word.
So Claude set up a machine that writes the other way. I’m not sure it noticed.
I gave it a sentence from my April post about AI and de-skilling. This model translates German into English, so in it went: “Und jetzt durch eine kühle metrische Betrachtung ausradiert werden.” And because you can look inside after every round, I did.
the
now a
And now a cool.
And are now by cool view.
And are now eliminated by a cool view view.
And they are now eliminated by a cool metric view.
That’s six of the eleven rounds. It’s like a Polaroid. First brown fog. Then, slowly, a face. Except the first thing that comes out of the fog isn’t an idea. It’s “the”. Then a mood: now, cool, view. The verb shows up late, and the grammar sorts itself out right at the end. My own English version of that line was “And now a cold, metric view just erases them.” Not that far off. And it took the machine under a tenth of a second.
Keep keep with QI
Then the whole post, sentence by sentence. The simple ones came out fine. “I am working in a company in a difficult economic environment.” A bit stiff. Correct.
The rest, not so much. “Mithalten geht nur noch mit KI” became “keep keep with QI”. It has never heard of KI, the German word for AI. Fingerspitzengefühl turned into “skill”, which is about what you get when you try to explain the word to an Australian. My colleagues who wildern, who poach in each other’s fields, were suddenly looking into the adjacent fields “and snow of it”. And “Claude oder Gemini” came out as “issue or,t”. Claude would have liked that one.
A full stop for the fish
The model has a second job: boil a news article down to one sentence. It learned that on BBC news, and on BBC news it does all right. It makes up the odd detail, but it reads like news. So I gave it six of my English posts.
Word salad. My piece about dashboards came back as “,3 - digital technology - technique, and dashboard - we’ve can create and create.” And the post about keeping my fishing site small? A full stop. Just the full stop.
Maybe it understood that one better than I thought.
So does this one write more like a writer? Not yet. It’s small, it only knows a couple of jobs, and on my texts it falls flat. For my work I’ll stay with Claude. But I keep coming back to that “the” at the start. Maybe that’s what a gist looks like from the inside. A word that could go anywhere, and some sense of where it should end up.
Claude wrote the whole test up for me afterwards, in a tidy results file. Word by word, of course. I’m shutting the Mac now.
The paper is on arXiv, the code on GitHub and the models on Hugging Face. I write in English at reinergaertner.au and in German at reinergaertner.de, my blog since 1997.