A C++ kernel steers Qwen while it is running
This is a firsthand experiment with a C++ kernel that runs beside during inference. It does not retrain the model, change the prompt, or edit the .
For each of the first 20 , it calculates a tiny value called katki and adds it to the model’s internal . The change is smaller than what the bfloat16 number format can normally show, so standard tools report no visible change, but the model’s output changes.
Version 1.2 adds four live sliders in a Gradio interface for peak strength, decay speed, always-on floor, and steering speed. In this test, the strength and floor were raised, increasing total pressure from +0.034953 in the default runs to +0.042903.
Key points
- A C++ kernel changes while the model is running.
- The method avoids retraining, prompt changes, and weight edits.
- The change is applied across the first 20 .
- A Gradio interface now gives four live controls for the steering behavior.
- The raised settings increased total pressure from +0.034953 to +0.042903.