Wink Pings

Let AI learn your writing style and score it like swiping on a dating app

Matt Pocock's new method: have AI score you sentence by sentence. Commenters say what changes the output is pairwise comparison, not 1-4 scores.

Matt Pocock proposed a new idea on X: if you want AI to write in your style, try a different approach: have it score you sentence by sentence.

Specifically, there are four steps:

  1. Build a scoring app; the AI presents one example sentence at a time, and you rate it.
  2. The rating method can be a 1-4 single-choice score or a pairwise comparison.
  3. You and the AI plan the article together, moving forward sentence by sentence.
  4. The AI keeps testing hypotheses about your writing style and gradually builds a profile.

The interaction is a swipe-left style like a dating app: as if you're swiping through dating candidates, except you're swiping through sentences.

Under the post, one commenter said they had built something similar, and what really influences the output is pairwise comparison. Another said this is essentially RLHF (reinforcement learning from human feedback), except it distills preferences into skills and context. Matt replied: AI helps you optimize within the limits of human working memory; you don't need to judge whether the whole article flows, only whether this sentence reads smoothly.

The comments also added a few details:

  • Giving only a 1-4 score isn't enough, because telling it a sentence is wrong doesn't tell it why. Having the AI rewrite those bad sentences can teach it more.
  • Half of anyone's style is a blacklist: the words they would never use, the endings they would never write. So a score of 1 has more teaching value than a 4.
  • Your own preferences drift. A sentence you'd rate 4 today might be a 3 when you look at it on Tuesday.
  • Someone suggested adding a paragraph-level operation: which line would you delete? A sentence may sound like you, but you might delete it because of repetition.
  • Someone complained that AI always adds a TED-talk-style uplift at the end and doesn't know how to remove it.

A commenter also recommended a more radical project, Spaceship IDE, from Chuanqi Sun at MIT Media Lab. It turns LLM interaction into real-time control like piloting a spaceship, with a central knob that can adjust the direction of text generation in real time, lock attention onto specific phrases, or explore different generation results by scrubbing.

screencast

focus monitor

attention lock

future steering

realtime regeneration

The project's underlying hypothesis is that when LLMs reach a real-time interaction speed of 60 FPS, the way humans control text will completely change. You can adjust the text's feature vectors like a mixing console, rather than repeatedly revising prompts.

However, there is no silver bullet for training writing style. A scorer is a good starting point, but true style may be hidden in the sentences you would never choose.

发布时间: 2026-11-22 03:34