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

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".

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.

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.

    and needs no system prompt

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.

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.

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

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

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?

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.

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.)

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?

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...

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

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

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

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

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

Are there plans to release the detector?

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

Impressed how little hardware is needed to train this.

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.

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

Have you released the detector?

why are there multiplayer mouse cursors on that page

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.

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.

how are bite sized models going to solve bad writing?

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.

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".

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)

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.

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.

What's with the very distracting cursors of others?

[flagged]

[dead]