The frontier live voice models (especially GPT Voice/Live or whatever they call it this week) do not have the typical AI-tells. So I wonder about a few things:

Why is this?

Why couldn’t whatever went into post-training these voice models be applied to the text-gen models to get rid of AI smells?

Could we set up a pipeline that leverages the voice models to clean up text that reeks of AI ?

Till today, I was doing the complete opposite of what the article suggests. Definitely going to try this approach out, pretty helpful.

I have a specific ChatGPT chat that rewrites my posts really well every time; I usually just give it a rough draft and let it clean things up. But it was pretty interesting to realize how good these LLMs are at actually finding mistakes rather than just rewriting things. The point about not letting the model influence your voice was also very, very relevant.

For commit messages and other technical writing, I use a "narrative coherence" checklist to help ensure that a change is clear and understandable:

- Establish a single thesis an external reviewer could recover from the diff.

- Unify vocabulary across code, comments, tests, and commit message.

- Use the same vocabulary consistently to refer to the same concepts.

- Keep every hunk that serves the thesis; consider the removal or deferral of the rest.

- Introduce abstractions at the point of need.

- Align tests to narrate the same story as the implementation.

- Reconcile the commit message and the diff.

- Order changes expositorily, not chronologically.

- Explain what the change does and why. Do not explain the details of the development process.

- Prefer to edit subtractively.

- Recompile and run all relevant tests after edits.

- Iterate.

I’ve used llms in two ways while writing fiction. First I’ll have a story idea and use the llm to explore what has come before that may be related; then I’ll go read some stuff. This has worked great for seeing where my ideas fit and flushing them out. Second I’ll describe settings or plot devices and use the llm to brainstorm consequences or further ideas. This fails miserably. The llms suggestions are devoid of creativity. This is all moot though because I’m really not a writer, I’ve just always thought I might like to be one someday.

How to write with an LLM?

Don’t

At one point in my graduate college studies, I tutored qualified (DEI) undergrads 'for free' (for them) as part of a work-study program which paid me a bit for that work.

OTOH, an AI could not have qualified as such a student. But even if it did, before we began I would certainly bargain for considerably more than work-study tutoring paid.

> use them like a copyeditor rather than a ghostwriter

This is what you need to know. Seriously, after 25 years in news (writing), you realize that you need “second eyes”, not the first hands.

I mostly use LLMs for coding, so this is an interesting use case. I’m curious how long an article it can handle in practice before the context starts becoming an issue.

Indeed. I have been experimenting with using frontier models for creative writing.

Using adversarial models to evolve ideas. I have a whole novel of material, and the feedback has indeed been useful. But the actual novel? Not even AI models like their own writing once you give the simple prompt of "account for model sycophancy".

The resulting structure and plot beats are good with AI as copy editor. Only the least discerning reader might not care or notice the actual non-human writing. And hey - maybe those are your readers. Probably not.

Can we all appreciate for a moment the sci-fi dystopia which snuck up on us all? I just used to phrase: " non-human writing". 50 years ago, that would send a chill up your spine.