We’ve all seen it. The overly polite tone, the relentless three-point lists, and those tell-tale rhetorical em dashes that scream, "I was generated by a machine." In the industry, it's becoming known as 'AI slop'—a predictable linguistic pattern that makes even the smartest models feel robotic and sterile.
Scrubbing the AI-isms
A new project hitting Hacker News is taking a direct approach to this problem. An engineer has released qwen3-4b-deslop, a fine-tuned version of the Qwen3-4B-Instruct model specifically designed to strip away these writing tics. The results are staggering: the author reports that the 'AI slop score' plummeted from 121 to 12.5 per 1,000 words. Most impressively, the habitual use of rhetorical em dashes dropped from 20.5 per 1,000 words to zero.
Local Power, Natural Flow
What makes this particularly interesting is the accessibility. Unlike massive proprietary models that require complex system prompts to "act like a human," this model has the natural style baked into its weights. Because it's based on the 4B parameter architecture, it's lightweight enough to run locally via Ollama with a 2.5GB download, meaning users can generate human-like text on a MacBook Air without sending data to the cloud.
By targeting specific patterns—like the "not X, but Y" sentence structure—the deslop model proves that we can move beyond generic outputs through targeted fine-tuning.
Sources
Media
