Show HN: Qwen3-4B with the AI writing tics trained out (100toolkit.com)

A new AI model, Qwen3-4B, has been released with significant improvements in writing quality by addressing the common stylistic tics found in AI-generated text. Developed by Erik Kannike, this model utilizes a "slop detector" to identify and reduce unwanted writing habits such as excessive use of certain phrases and invented facts, achieving a marked decrease in its "sloppiness" score from 121 in the original to 12.5 in the new version. The model's architecture allows it to be run efficiently on personal computers, maintaining accessibility for a wider audience in the AI/ML community. This innovation is significant as it presents a refined approach to AI writing, aiming for outputs that closely resemble human quality without falling into the traps of typical AI-generated prose. The technique employed—distillation from well-performing responses—ensures that the model not only generates coherent content but also adheres more closely to the requirements of various writing prompts. While the model shows promise, it still makes occasional errors, like inventing names and places, indicating room for further refinement. Future enhancements will include training on larger models and public-domain texts to improve output quality even more, highlighting the ongoing evolution in the pursuit of creating sophisticated, human-like AI writing.