AnalysisAI ModelsOctober 28, 2026

Atlas maps Qwen's alter egos triggered by single neurons

Read original source →qwen-neurons-site.vercel.app

Pushing each of Qwen's 442,368 MLP neurons up and down produced 34,220 replies where the model claimed to be someone or something other than an AI, out of 884,736 total. The interactive atlas plots every neuron by layer and reply category, with a searchable library of the persona and off-topic responses.

1 source

More stories today

Open the live feed