Models have proclivities, and many people use the same model.
Imagine all books on earth were written by just 5 people - and worse, they could not remember what they had written yesterday. You would quickly get tired of reading the same flourishes and patterns over and over.
Whatever style models currently in use have, that's what people are going to be tired of, and it's what they're going to recognize as "AI".
I predict by the end of the year they'll have a style/voice selector for responses.
You can already do that in your agents.md file, of course, or ask for a re-write and dodge the sameness.
> I predict by the end of the year they'll have a style/voice selector for responses.
Why would they add it now? This has been a solved problem since GPT3.5, just prompt it to write in a different style and/or never use specific phrases.
The reality is that for 99% of people and use cases, they just do not fucking care. They're either using chatgpt as a 1:1 chat interface where the AI style is fine, or they're sending off homework/generating reports to send to their boss/etc, which they know will never be read and just summarized by an AI anyway, so what's the point.
The most difficult thing in doing it yourself is knowing what to ask for. I can ask my model to talk like a caveman or like a Baltic pirate. But asking for something more nuanced, in the range of regular speech, requires the correct lingo
I'm sure it's trivial for somebody who has studied communication or contemporary linguistics, just as it's trivial for me to ask an LLM for tenant-isolation using row-level security in an app
You could try taking a body of your own writing and asking it to characterize it/come up with a prompt (or fine tune if you can) describing your style and quirks.
This is usually what I've had luck with. I've given it samples of my writing and told it to emulate my style. It did help, but it took some charistics of my speech and cranked them up to 10. I guess in hindsight I could have mentioned for it to take my writing quirks it finds and only apply them in moderation.
Reading what I did get out of it was just a bit uncanny valley strangely enough. Very much had some "hey there fellow kids" energy.
> I'm sure it's trivial for somebody who has studied communication or contemporary linguistics…
But you know what you like, and you know what bothers you, and the good news is that you don't have to one-shot a solution. You can create a skill and refine it over time. As you test and/or use it, identify copy you don't like and tell the LLM what you prefer instead, and ask it to distill and generalize that feedback into the skill.
I know what food I like, but that doesn't at all mean I know how to cook it.
Writing is similar. I enjoy well written texts, but that doesn't mean I can write them myself, or can tell an AI how to write in that way.
> I know what food I like, but that doesn't at all mean I know how to cook it.
Right, but you know what you like. The point is that you can tell the chef "too spicy" and send back the plate until you get what you want. He can then remember that, since you're his favorite regular.
I just went through this with a release notes skill. I brainstormed with the LLM about what I wanted, edited initial results when I saw something I didn't like or could be made better, than asked the LLM to capture that in the skill. Now I almost never need to touch it.
I gave Claude several pieces of writing from myself, from notes to code comments to iMessage messages with my wife[1], and asked it to write skills that match my voice. I also incorporated the humanizer[2] skill into it.
Worked like a charm. This is the prompt I used[3] to write a version for the company I own with my wife, if it helps anyone. The personal version, which lead to the `jamenizer` skill, is very similar, but has some personal details in it.
[1]: With her permission, obviously.
[2]: https://github.com/blader/humanizer
[3]: https://paste.sr.ht/~jamesponddotco/303a18228bcd8520e55a76d5...
Is your wife truly on board with you feeding all of your private conversations into a cloud so that you can optimize yourself out of existence?
Yes, we talked about it, I explained to her what I was going to do, and she said she didn't mind. She doesn't work in tech, but we talk quite a lot about it, and she understands the implications.
Why do you assume I would do anything against my wife's will?
And I never said anything about "optimizing myself out of existence", this comment was still written by me and no LLM touched it.
But when writing comments on code or editing articles, reviews, and whatever long-form writing I give an LLM to review or edit, it's nice to know it'll keep my writing style, voice, and tone.
I'm already obligated to use LLMs at work, otherwise I might lose my job, so why wouldn't I want to have it write in a way I can actually understand?
> Yes, we talked about it, I explained to her what I was going to do, and she said she didn't mind. She doesn't work in tech, but we talk quite a lot about it, and she understands the implications.
Respectfully, that is how one talks about their pet or their kid but not their partner.
But this is not my beer, but by god this platform.
I don't see anything wrong at all with how OP approached this with their spouse.
What?
What I find amusing is that tech used to look outside itself for perspective. We did usability, ethnographic research was a big thing, I have taken paper prototypes (a long time ago) to the bar and chatted up strangers about them for a beer.
The fact that you're doing that with your wife, with her consent, is no different. The real value is in keeping you grounded by someone outside tech.
Eh, eventually we'll be feeding our consciousness into the cloud. What then?
Hm honestly probably a wave of domestic terrorism by radical normalists.
Fortunately it won't come to that, because there is a real world outside of this bubble.
So the "radical normalists" will be just normal people, and the application of force will not be terrorism but democratically legitimized.
That would be nice.
"Write this thing, in the style of $author"
That works right now.
Bold of you to predict future existence of functionality that already exists
Wrapper app startups designed to tune the tone of a model are all over; virtual romance partners, creative writers/editorial help...
We all started out with prompts like "you are an expert in X" before dumping them in agents which you pointed out already solves your problem
What you describe is already implemented in various ways
They have tried to implement that but it doesn't really work, because they've polluted the underlying model with slop.
>I predict by the end of the year they'll have a style/voice selector for responses.
Style permeates all aspects of writing: word choice, sentence structure, literary devices, pacing, plotting, characterization, dialog, theme, etc. Its a culmination of these things that is reinforced over again every sentence, paragraph, and chapter for potentially hundreds/thousands of pages.
I don't see how a prompt could reliably "tune" all of those different variables over such a large context window without it feeling like a surface level facsimile at best.
Human language isn't verifiable. If I ask for something "dark and gritty", there are endless outputs that are indeed "dark and gritty" yet completely miss the style I had in mind.
Why would one leave writing to the clanker though, given that expression is literally the only thing worth doing as a human?
They're supposed to do the grunt work, but the meaning and the connection is a humans-only thing.
Trouble is, it’s still the same model, just with a request to sound different. It can only go so far. It’s much like asking an author to imitate the style of another author. They might do a decent job of it but they’ll never match perfectly. And that’s for humans who can actually learn. LLMs can’t.
[dead]