Show HN: Qwen3-4B with the AI writing tics trained out

Posted by erikkannike 3 days ago

Comments

Comment by chmod775 3 days ago

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".

Comment by sejje 3 days ago

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.

Comment by wongarsu 3 days ago

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

Comment by jamesponddotco 3 days ago

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...

Comment by hypfer 3 days ago

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?

Comment by jamesponddotco 3 days ago

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?

Comment by zer00eyz 2 days ago

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.

Comment by hypfer 3 days ago

> 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.

Comment by bavell 2 days ago

I don't see anything wrong at all with how OP approached this with their spouse.

Comment by 2 days ago

Comment by subscribed 2 days ago

What?

Comment by AlexeyBelov 1 day ago

But if you don't do it you'll fall behind.

Comment by xvector 3 days ago

Eh, eventually we'll be feeding our consciousness into the cloud. What then?

Comment by hypfer 3 days ago

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.

Comment by pigeons 2 days ago

That would be nice.

Comment by ndriscoll 3 days ago

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.

Comment by techjamie 3 days ago

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.

Comment by mpalmer 2 days ago

You should play around more, it's flawed but more than any other tech it rewards exploration.

    But asking for something more nuanced, in the range of regular speech, requires the correct lingo

I like asking models to write things using only the N most common words in a language. It's probably never perfect, but it understands the assignment well enough. from there I notice things I prefer about that constraint, and things I don't.

Comment by CharlesW 3 days ago

> 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.

Comment by BurningFrog 2 days ago

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.

Comment by CharlesW 2 days ago

> 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.

Comment by throwthrowuknow 2 days ago

That’s why Claude had a customizable style setting that let you just paste in example text to create a new writing voice. Not sure if they still have it or not.

You don’t need a built in feature though, you can just paste examples into the prompt and tell it to write in the same style.

Comment by nekusar 3 days ago

"Write this thing, in the style of $author"

That works right now.

Comment by mjr00 2 days ago

> 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.

Comment by jplusequalt 2 days ago

>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.

Comment by wat10000 3 days ago

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.

Comment by sejje 17 hours ago

Why in the world would it need to match perfectly?

Comment by MisterMunchkin 2 days ago

They have tried to implement that but it doesn't really work, because they've polluted the underlying model with slop.

Comment by hypfer 3 days ago

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.

Comment by d0nkee 2 days ago

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

Comment by FDEUp 3 days ago

[dead]

Comment by jwr 3 days ago

Why kill the em dashes? They are a legitimate mark — one that I have been using for years. On a Mac, you quickly get used to pressing shift-option-hyphen. Em dashes are typographically correct and should be used.

Comment by literalAardvark 3 days ago

Because no one knows what those are, and they're a major indicator of AI writing.

Comment by vsl 3 days ago

> they're a major indicator of AI writing.

For the illiterate, sure. Where do you think LLMs picked them up from?

Comment by literalAardvark 2 days ago

Doesn't really matter does it?

Most people use the - at most, or rephrase.

One rarely if ever should write at a level that's above high school reading for day to day articles and tasks.

I've written for a large computer magazine, you know what the target is? It's John: A 21yo man taking ~CS classes and with some disposable income. Not a lit major. John has no idea what that wide - is, doesn't care, and wouldn't understand fancy writing anyway.

Comment by Lalabadie 2 days ago

"I don't care therefore no one cares" is definitely how we've evolved as a species.

Comment by suburban_strike 2 days ago

Medium.com

Comment by 28304283409234 2 days ago

Yes! And its GNU/Linux!!!

Comment by jwr 2 days ago

> no one knows what those are

It is a sad state of affairs if we are to accept degenerating into illiteracy.

Comment by zarmin 2 days ago

you've never read a book, have you

Comment by literalAardvark 2 days ago

I've read plenty, in 4 languages. Hbu?

Comment by ahmedfromtunis 3 days ago

Or at least match the pre-llm human usage frequency.

A total, artificial lack of em dashes can in itself be perceived as another sign of AI writing.

Comment by wolttam 3 days ago

“This person likes semi-colons a whole lot”

I’ve already seen that trick snuck in by newer models. “I won’t use em-dashes; I’ll use semi-colons!”

Comment by LelouBil 3 days ago

Honestly I don't see the em dash as for sure AI. It's just that it's often used in LLM tropes like the "it's not XXX — it's more" and other tagline-esque sentences.

Comment by gobdovan 2 days ago

In before: Goomba fallacy. But I can't notice how there's claims that: LLMs are nondeterministic and cannot be relied upon to produce the same thing AND they produce the same texts, same UIs, in the same styles. It frustrates me too, but what a solution for both problems would look like?

Also, if all LLMs would write in the style from the article, in how much time that would be the recognised LLM style people would get sick of? I suppose longer? But even the current style of LLMs was palatable when they first came out.

Comment by david2ndaccount 2 days ago

There’s no contradiction. Consider this thought experiment: the real range of possible outcomes is a number from 1-100, but your model is a six-sided die so it produces a number from 1-6. It is random, unpredictable and non-deterministic and yet you will get “samey” results compared to the real range of possible outcomes.

Comment by gobdovan 2 days ago

Does the thought experiment help us to identify a solution for these problems?

Comment by bschwarz 3 days ago

    and needs no system prompt

A human would write "and doesn't need a system prompt."

Comment by swiftcoder 2 days ago

That's just a regional/educational quirk. Classically-educated folks and older Brits are both quite likely to use the "needs no" version.

Most "LLM voice" quirks come from actual human writing styles, even if Claude et al have managed to distil those styles into something pretty horrific.

On the other hand, the way you can tell the title line was at least edited by a human, is the missing comma before "and needs no system prompt".

Comment by bityard 2 days ago

A human could write either. The former is just slightly more formal and concise.

Before AI, I used to write everything in somewhat formal phrasing mostly because I enjoyed it and a tiny bit because I assumed (perhaps incorrectly) that it made me sound smarter. Post-ChatGPT, I noticed that most LLMs had a writing style fairly similar to my own, so I changed my writing style to appear less formal in most circumstances. While I'm not anti-AI, I'm always fearful of being mistaken for a bot.

Comment by dofm 2 days ago

Depends on the human.

I would use that turn of phrase (needs no XYZ) — but maybe not in the context of system prompts, perhaps. Though I don't know; sometimes a flourish is a nice thing.

You might be right that the clash of that vernacular pattern and technical language is a "tell", but AI did learn this stuff from actual people; a lot of the time Claude reminds me of an actual person who worked for a firm I worked for in the 90s, who wrote in an annoyingly gnome-wizard style.

Ultimately it's the volume of these problems that is the tell.

Comment by subscribed 2 days ago

Does ESL user write like human to you, even though they could make a mistake?

Comment by desloppysupport 2 days ago

[flagged]

Comment by khazhoux 2 days ago

The model was the setup. The training was the vindication.

^^ this is my new most-hated tic. I see it everywhere now.

Comment by JustSkyfall 3 days ago

It's definitely a slight improvement, but it is still rather obvious that the output is AI-written :(

Comment by mrweasel 3 days ago

The tell tell sign for me is that it's still incredibly hard to read. AI tend to write in a style that I'm unable to easily read (humans does that as well sometimes). I constantly have to go back and re-read the previous sentence, because the text simply doesn't register.

A next step I'd suggest is asking the LLM to use shorter words when available.

Comment by throwaway219450 2 days ago

I found the coffee shop one telling. Both suffer from being a stream of bland statements. The original version is actually reasonable if you just excise the Claude-isms. The rewrite has a whiff of ASD-STE100, and reads like the sort of sentences one would write for homework while learning a new language. There is also almost no difference between the de-slopped opinion piece and the blog post.

Comment by BoredomIsFun 3 days ago

Did you check this? https://huggingface.co/sam-paech/gemma-3-27b-it-antislop

Comment by mrweasel 2 days ago

No, I have no wish to use LLMs to generate text for me. I would however be happy if those who do, would use models that generate more readable text.

Comment by j45 3 days ago

LLMs as a word generator love generating words, and that includes generating bigger words when simpler ones will do.

Comment by mochizou 3 days ago

[dead]

Comment by BoredomIsFun 3 days ago

This is not the first deslop. The true pioneer is this dude: https://huggingface.co/sam-paech/gemma-3-27b-it-antislop. This model is a perfectly usable, full size not-toyish model.

Comment by Armecera 19 hours ago

I wonder how it is in Japanese though. AI writing in Japanese is still pretty bad lol Its really obvious when something is written by AI.

Comment by BSOhealth 3 days ago

for https://slopwash.com I’ve started approaching this more as an eval problem than something to go post-training for. I’m excited to be trying these new Jev-type models in more of a harnessed approach. maybe some heavy LLM to measure preservation of intent

Comment by charv 3 days ago

Side note but the multi-user cursors was a super cool feature, not sure it was relevant to this article but would love to see it on more interactive pieces (a la The Pudding, etc.)

Comment by ac2u 3 days ago

Is it still capable of writing an essay about AI writing tics or has the very means of producing said tics, even when the model is asked to do so to demonstrate, now abitterated?

Comment by f311a 3 days ago

Why would you support this sort of projects? LLM text should be spottable for as long as possible, otherwise human communication and expression on the internet will be killed.

Comment by dofm 2 days ago

I largely agree with this, though I think:

1) there's some value in internally generated text (automatic documentation, autosummary) being less aggravating

2) there are assistive applications — speech generated from sign language and touch panel communication etc. — and those users deserve a tool that can communicate for them in a way that has controllable or even personalisable outcomes.

3) a process of de-sloppifying text might also help train thinking clarity.

Comment by simonra 2 days ago

Communication is a means to an end, a tool, not an end in itself, or something with intrinsic/unassigned value. Why would we want more energy wasted on focusing on the packaging (or worse, how the packaging is made) by drawing attention to it, instead of helping directing it towards that which actually matters?

Comment by squidbeak 3 days ago

In what sense will it be killed? I don't plan to stop writing my own messages or notes, just because AI can write, and I seriously doubt many other people who prefer their own voices will change either.

Comment by chownie 3 days ago

I stopped using reddit entirely because it's bots talking to other bots on threads posted by bots, I stopped making posts at the start of this year because I no longer know if a human being is ever reading them. I then deleted the account.

I am increasingly feeling that way about HN, too. If this is all the equivalent of a videogame conversation with an AI, why am I bothering to put the effort in?

Comment by f311a 3 days ago

Well, imagine that 90% of the comments on HN are generated and it's very hard to spot it. Would you read them?

Comment by wat10000 2 days ago

I’m afraid we might be doing this experiment now and the answer is yes.

Comment by ndriscoll 3 days ago

Are they interesting, well thought out, well articulated, and insightful?

https://xkcd.com/810/

Comment by gavinray 2 days ago

If you take this question far enough, you eventually reach a core philosophical debate:

"Does it matter what anything innately is, beyond the effects it has on you?"

For example: Suppose you were given a human companion with a robot brain that was your ideal friend. They made you laugh and you enjoyed every moment you spent with them.

Does it make it any less impactful/genuine to you that they're not a "real" human, or is your felt experience the only meaningful end-measure?

For me, it's the latter.

Comment by chownie 3 days ago

> Are they interesting, well thought out, well articulated, and insightful?

The average AI post is likely a carbon copied reddit comment from 2011, that is to say, "no".

Comment by ndriscoll 3 days ago

Then either they're easy to spot, or the human ones are equally low-value.

Comment by chownie 3 days ago

I would love if it could be a false dichotomy, but given conservative estimates pin about 30% of online posts to AI it seems like there may be a third outcome: we patch the visible holes of humans posting less and less from disinterest by making the bots post more and more.

The end result is that comment sections aren't enticing to join in on, just read. Which is perfect for ad impressions.

Comment by ndriscoll 3 days ago

The sites that are there for ad impressions already worked this way though. The incentives were geared toward poor human discussion as well (c.f. reddit as a best case, and e.g. instagram/tiktok/youtube as what they'd like to be). So the problem here is ads, a desire to grow to include the entire world to show them ads, and a desire to do whatever it takes to keep their attention to show them ads.

Comment by chownie 2 days ago

"The sites that are there for ad impressions" are the websites, there's no invision powerboards in 2026, the alternatives have been gutted and eaten by the sites there for ad impressions. HN is an exceptionally poor replacement but it's the closest thing I've found. I don't enjoy that has the exact same dynamics as reddit but the ads are on pages I click through to instead of directly on the site, it sucks that I still don't recognise any usernames here.

By next week this thread might not have existed at all. The usership of the site likely won't remember it. If I was posting half the content on this site and then stopped and was replaced by another username, would anyone notice? It seems like there is no real community to be found, we've destroyed our own habitat.

Comment by plagiarist 2 days ago

The average human post on reddit is a carbon-copied comment from 2011

Comment by konart 3 days ago

Imagine that 90% of the comments on HN are generated and it's very hard to spot it. And you read them.

Comment by trilogic 3 days ago

The q4 loops after 16k tokens generated. Great idea though, hope to see more models like it. Thanks. https://huggingface.co/erikkannike/qwen3-4b-deslop-v4/discus...

Comment by GaronG 2 days ago

I wonder if dynamically changing the temp based on the requested tone + a small "how to write skill" could produce more human varied writing, instead of using the same temp for the full request.

Comment by rahulb0802 2 days ago

Since you wrote the detector using a particular style guide, are there other indicators of AI writing you are still uncovering and adding to this? Or do the ones mentioned mostly encompass everything?

Comment by absqueued 2 days ago

I wish the clickable links were easier to find on this blog. Please, Erik.

Comment by lacoolj 2 days ago

This is cool. I would like to try using this model with Qwen3.8 or Gemma4 as a speculative decoding model. Wonder if that would help give the overall results with the higher quality output

Comment by mosselman 2 days ago

Does anyone have a link to that style guide they used?

Comment by 3 days ago

Comment by dpoloncsak 2 days ago

How are you sure it's not developing just new and different tics?

Comment by j45 3 days ago

This still scores as 100% AI generated on tools like pangram.com.

It stil contains a triple construction, but ranks much cleaner on slop-cop.

Comment by nullsmack 2 days ago

why are there multiplayer mouse cursors on that page

Comment by swiftcoder 3 days ago

Impressed how little hardware is needed to train this.

Comment by hashmap 3 days ago

I saw those moving dots and snap-closed the page.

Comment by Royce-CMR 3 days ago

Applause earned. Neat results especially out of such a small model.

Separately we have overlap between AI slop styles and right for the situation writing styles which this highlights.

Using the new CRM example, the clean version looks like a millennial student who rambled on a page, and the AI slop version looks like a draft I could send for PR to corporate managers. I buy the former looks more real, but it also looks sloppy and doesn’t work for most of my use cases.

When I was in school, the structured essay style of the “slop” examples was encouraged and I see it continue to resonate in corporate. The cleaned up versions here def have use cases - reviews for example.

Main point - different styles for different situations and our AI slop versions vs humanized versions both overlap different areas of the “ideal” we continue to look for.

Comment by vorticalbox 2 days ago

Are there plans to release the detector?

Comment by brazukadev 3 days ago

I was excited with the idea then got to the first example and boy that was bad. The second "de-slopped" version has no formatting and is 3 times bigger with the same level of information. This model slopmaxxxed the slop.

Comment by sram1337 3 days ago

Have you released the detector?

Comment by whythismatters 3 days ago

[dead]

Comment by nathants 3 days ago

how are bite sized models going to solve bad writing?

Comment by zeyrie 2 days ago

I think is it the missing of human touch in writings that makes us think something as slop. Obviously, impersonating won't take one far I guess. One of the things I'm avoiding to use AI is for writing, because it is hard, enjoyable at the same time and is the only way for me to express what I think.

Comment by 10111001 3 days ago

Now can I have a model with the AI writing tics trained to the max?

I want my boss to know I'm a Good Boy using AI.

Comment by Izmaki 3 days ago

You can definitely get far with skills and personal instructions. That is my estimate, not research.

(*defeated chuckles and gazes towards my Claude session*)

Comment by nickabe 2 days ago

There's a couple more things you can try.

If you limit the sample size, you end up finding more. For example, on my project when I tested 350 strings at once it failed to find 15 strings with "ai slop" that testing 10 strings at a time did find.

Comment by pigeons 2 days ago

Its funny, even with the tics probably trained out of that blog post, my brain still turned off and skipped paying attention to it, automatically categorizing it as "AI slop".

Comment by Fe2O3 3 days ago

Why are lists of three considered slop? I genuinely use it quite a bit because groups of three feel more memorable, satisfying, and persuasive. (<- list of three right there)

Comment by anilgulecha 3 days ago

What's with the very distracting cursors of others?

Comment by erikkannike 3 days ago

Sorry - I've been playing with WebRTC and Trystero (https://trystero.dev/) and added it to my blog as an experiment. Of course didn't expect my blog to become a bit popular on HN and for the cursors to be super annoying. Turned them off by default!

Comment by RonanSoleste 3 days ago

I don't understand why this is a feature and what it brings other than shout "you are being watched by random others". Exited the site before ending the first block.

Comment by maven29 3 days ago

The ancients brought us "reading mode" in browsers as an escape hatch for exactly this kind of cruelty

Comment by mullenba 3 days ago

Exactly. I was interested in the article but I don't have time for that nonsense.

Comment by markgavalda 3 days ago

Agreed. In the upper right corner there's a "hide" button that I immediately clicked. Then read (ok, skim read) the article.