Two-neuron Qwen steering claim raises a reproducibility question

The proposal concerns inference-time activation steering, not a change to stored Qwen weights. Two layer-local interventions are said to push responses toward an exaggerated rude persona. This suggests a tractable model-intervention hypothesis, but the accessible material does not show matched prompts, seeds, decoding configurations or a controlled safety evaluation.

A secondary article attributes to an X demonstration the claim that increasing neuron L14 N8949 by one and decreasing neuron L13 N7956 by one changes Qwen3.5-4B output behavior. The underlying author repository provides neuron-exploration tooling but does not independently establish the particular effect. The originating X post returned HTTP 403 to the permitted browser. Secondary reporting and an open-source viewer are not equivalent to an independent experiment; the asserted view count and claimed degradation of intelligence are not verified.

Verify the model hash, inference runtime, before-and-after outputs and control prompts before turning the claimed personality shift into a measured refusal-bypass or jailbreak rate. A small number of influential activations can be worth auditing in safety evaluations, but the current evidence only supports a reported demonstration and a plausible technique, not a new production exploit. Primary evidence.

Additional evidence and trust boundary

This is an unusually precise hypothesis but still not an independently measured security finding. A two-coordinate activation change must be specified against an exact checkpoint, hook location, token position, decoding setup and fixed prompts before claims about personality or intelligence loss can be tested. The code repository corroborates the existence of a steering tool, not the alleged effect size or the repost's engagement count. Supporting record.