Show HN: Qwen3-4B with the AI writing tics trained out
A small open-source language model has been fine-tuned to remove common “AI slop” writing tics, aiming to produce text that reads more like natural human prose without extra prompting. Commenters debate whether this is an improvement, arguing over readability, overuse of formal structures and punctuation, and the risk that making AI text less detectable could further erode authentic human expression online. Others see value in controllable style and tone for specific use cases—like corporate writing, accessibility tools, or automatic documentation—while warning that any dominant AI style, even a de-slopped one, may quickly become just as recognizable and tiresome.
Meta UX: Multiplayer Cursors and Site Design
- Many found the real-time cursors extremely distracting and left the page early.
- Some discovered or requested a “hide” button or used browser reader mode.
- A few thought the shared-cursor feature was cool for interactive pieces, but consensus here was it was off-topic and annoying for a blog post.
- The author later disabled the cursors by default after feedback (stated in the thread).
Model Concept: De-Slop / Anti-Tic Training
- The model aims to remove common “AI slop” tics: formulaic openings, triple lists, over-formality, marketing tone.
- Several commenters liked the idea and were impressed by how little hardware was reportedly needed to train it.
- Others felt the outputs still “read like AI,” just in a slightly different, sometimes more boring style.
Quality of Writing and Readability
- Mixed reactions to the example rewrites:
- Some said the “cleaned” versions were more human-like or less corporate.
- Others found them harder to read, longer, or still bland and obviously machine-written.
- Suggestions included: shorter words, simpler syntax, fewer flourishes, and attention to actual information density.
- Some noted that eliminating one set of tics risks introducing new, equally recognizable ones.
Style, Voice, and Personalization
- Several comments argue that “AI voice” stems from using a few big models with similar defaults.
- People discussed:
- Style/voice selectors.
- Feeding in one’s own writing to derive a custom style.
- Using prompts or “skills” to iteratively refine tone.
- Skeptics counter that it’s still the same underlying model and imitation remains surface-level, especially over long texts.
Detection, Ethics, and the Future of Online Discourse
- Some worry that making AI text less detectable will flood forums with undifferentiated AI content and discourage human participation.
- Others respond that:
- Internally generated text (docs, summaries) benefits from less-sloppy style.
- Assistive uses (e.g., for people who can’t easily write) justify better quality and personalization.
- Disagreement over whether communication’s value lies in human authorship vs. perceived usefulness and experience.
- Questions raised about releasing detectors, evaluation strategies, and whether LLM-generated text will inevitably dominate low-value comment spaces.