This weekly roundup thread is intended for all culture war posts. 'Culture war' is vaguely defined, but it basically means controversial issues that fall along set tribal lines. Arguments over culture war issues generate a lot of heat and little light, and few deeply entrenched people ever change their minds. This thread is for voicing opinions and analyzing the state of the discussion while trying to optimize for light over heat.
Optimistically, we think that engaging with people you disagree with is worth your time, and so is being nice! Pessimistically, there are many dynamics that can lead discussions on Culture War topics to become unproductive. There's a human tendency to divide along tribal lines, praising your ingroup and vilifying your outgroup - and if you think you find it easy to criticize your ingroup, then it may be that your outgroup is not who you think it is. Extremists with opposing positions can feed off each other, highlighting each other's worst points to justify their own angry rhetoric, which becomes in turn a new example of bad behavior for the other side to highlight.
We would like to avoid these negative dynamics. Accordingly, we ask that you do not use this thread for waging the Culture War. Examples of waging the Culture War:
-
Shaming.
-
Attempting to 'build consensus' or enforce ideological conformity.
-
Making sweeping generalizations to vilify a group you dislike.
-
Recruiting for a cause.
-
Posting links that could be summarized as 'Boo outgroup!' Basically, if your content is 'Can you believe what Those People did this week?' then you should either refrain from posting, or do some very patient work to contextualize and/or steel-man the relevant viewpoint.
In general, you should argue to understand, not to win. This thread is not territory to be claimed by one group or another; indeed, the aim is to have many different viewpoints represented here. Thus, we also ask that you follow some guidelines:
-
Speak plainly. Avoid sarcasm and mockery. When disagreeing with someone, state your objections explicitly.
-
Be as precise and charitable as you can. Don't paraphrase unflatteringly.
-
Don't imply that someone said something they did not say, even if you think it follows from what they said.
-
Write like everyone is reading and you want them to be included in the discussion.
On an ad hoc basis, the mods will try to compile a list of the best posts/comments from the previous week, posted in Quality Contribution threads and archived at /r/TheThread. You may nominate a comment for this list by clicking on 'report' at the bottom of the post and typing 'Actually a quality contribution' as the report reason.
Jump in the discussion.
No email address required.
Notes -
That paper doesn't do it
It's a review of a whole area of research trying to do it with various levels of success. Fine tuning is still the best way of altering style AFAIK, but all this is insofar as we mean the same thing by "it". But it seems clearly we don't. So what is "it" (prose style) exactly if not text structure, sentiment, vocabulary, etc?
What's the quality you're looking for that's not described in this review?
You claimed:
The paper does not demonstrate prompting a model, showing that with that prompt the writing is decent, and does not share such a prompt.
Fair enough, you want specifically prompts, here you are: https://github.com/blader/humanizer
In my experience fine tuning or specifically trained rewriting models work a lot better and fine tuning especially if you want to style copy a specific author. Mere prompting is not as good as that. But I still rate it as decent.
You made claims about prompts not me. Now I'm only asking you to back them up.
But you even look at the stuff you send me? This isn't a prompt, it's an agentic workflow. And this doesn't make the ai write differently it instructs the ai to modify already existing text.
But nitpicks aside, the results from something like this aren't even good, they're just plain bad. From your posting I assume you never tried it so I suggest you do that and experience the most soulless slop writing ever.
People keep saying this but never share the loras
Random_Eddie on twitter did an approach that was pure prompting and moderately successful, although some slopisms bleed through (and doing a bunch of different writers in one conversation probably hurt). Tends to fall off pretty quickly if you want more than 3k-5k tokens of the same writer in the same context, though. Agentic approaches like Claude-Book 'work' better by doing things like perplexity analysis and gating, but I'm increasingly convinced they're a dead-end from a usability perspective even if they could write acceptably.
I've made a few LoRA with Unsloth, and while 'a couple hours' is more than a little generous, it is doable. I'll see about getting one from a recognizable and general-audience author.
That said, "style" is probably occluding a bit. You can get habits, formats, short-term pacing, and turns of phrase pretty easily just by motioning at the name, and sometimes too easily with LoRA. I've not been able to get a Zahn (or even Bujold or Butcher)-level plot setup just by asking for one, or finetuning, or throwing an agentic setup at it.
And none of them really solve the problem that the LLMs doesn't seem to know what makes a short story 'good', so if the problem is that prompting an LLM for a good story with no more detail, and then the LLM can't tell whether it's supposed to be writing fiction or nonfiction, that's going to be a harder problem.
Not to be glib, but what actually makes it good? Or what makes a good scifi plot setup? Because while you and I may have an intuitive sense of what that means, I think it's debatably quantifiable but if you can get enough data from people who have similar tastes I don't see how you couldn't have models improve on this kind of task. I don't think writing style is very uniquely different in shape from some other complex tasks transformers have recently improved at. I think it's just getting less attention and ressources than video or music generation because it has less tangibly valuable applications.
I made a bet on that for image generation, and bet wrong, so my confidence is very low, here. There are some technical reasons that it might be hard to train a model to recognize a good story (eg, it's possible for training to overfit on microscale solutions and then never pull any signal from the larger structure), but it's hard to come up with explanations that couldn't be applied to spaces LLMs have done well at, like math structures or short-form video.
I've actually drafted some lengthy write-ups of the different meanings of quality in these contexts. But they don't really seem to help the LLMs, and I'm not sure if that reflects a limit of the LLMs or of my writing. And their equivalents do genuinely seem to have been of mixed blessing in image generation spheres: learning about various artistic and photographic techniques helps get a specific output image you're imagining into creation, but the models themselves genuinely just get a lot out of positive: high quality, negative: bad art.
Yeah I just don't understand why it works so well on music, and not on text. Something like voice modulation I get as being more difficult because it's a little bit harder to tag performances, but the landscape here doesn't really make any sense. Even trying to fit is to the available data and its quality.
Hence my guess is optimistic: that we just haven't applied all the methods evenly and eventually it'll settle to some readable theoretical limit. But maybe we just don't know enough about the mathematical structure of things to know what we don't know.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
https://github.com/cauchy221/StyleTunedLM takes like two hours to build on a local GPU and uses public datasets. Being lazy is a choice.
Stop wasting my time sending stuff you haven't even read or tried yourself.
Believe whatever you want to believe.
Are you claiming you read the paper or used the model in the link you sent?
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
I think the problem with AI regarding style is that it's incapable of making the various micro design choices a piece of writing involves. It's able to nail certain things that can be applied blanket-wise, like making the syntactic structure flow ultra smoothly or what have you, but when a decision requires a greater degree of discretion, you find that style of any variety simply isn't there. The things that humans would do to polish a piece are simply absent when the AI does it.
I think this gets to the root of why they randomly insert lists and stuff as well. There might be some kind of rote decision making tree where they go, 'Is this asking me a question about an array of topics? --> yes --> then split into a numbered list', which possibly works on average, but sticks out when it doesn't.
Finer styles like those affected by serious, accomplished writers quite possibly require the utmost of control over these micro decisions. If you attempt Hemingway's style and mostly get it, say to a degree of 80%, then the result might still fall flat. It might still collapse under the weight of its aspirations. So even a decent writer who makes a strong effort at recreating Hemingway might still show their limitations. An AI which doesn't come close is nothing more than a joke.
More options
Context Copy link
More options
Context Copy link
The other commenter said:
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link