Qwen3.5-4B Flips Personality With Two Neurons

This is a compelling interpretability demonstration, but not genuine weight-level uncensoring: the aggressive behavior is reversible, highly dose-sensitive, and can degrade factuality.

• –The recipe pushes L14 N8949 up and L13 N7956 down; follow-up testing found L14 N8949 is the main behavioral driver.
• –Independent analysis reported malicious responses on 15 of 19 judged prompts versus none from the unsteered model, while factual accuracy fell sharply at the tested dose. ([Quasiqualia](https://quasiqualia.com/notes/lyra-neurons.html))
• –Neuron Explorer sweeps individual MLP neurons, records rollouts, and compares steered behavior against unsteered and random controls.
• –Running the model locally in a browser via WebGPU turns a specialist mechanistic-interpretability technique into a hands-on developer experiment. ([Reddit demo](https://www.reddit.com/r/LocalLLaMA/comments/1xgbr9y/interactive_neuron_steering_playground_for/))
• –The bigger safety lesson is that alignment behaviors may be brittle runtime circuits, not permanent properties of the weights.