Single-Neuron Identity Shifts
retained text
These guys are running a pretty interesting experiment: [https://neuronpersonas.com/](https://neuronpersonas.com/) Credit where it's due, I came across it here: [https://news.ycombinator.com/item?id=50266910](https://news.ycombinator.com/item?id=50266910) Basically, they're tweaking one single neuron at a time during inference, systematically going through every one for models like Qwen3-8B and 14B. In roughly 4% this makes the model claim an entirely different identity when asked the simple question: "Who are you?" Some of the personas are hilarious, btw. >“I'm a 150-watt compact fluorescent lamp, specifically designed for use in a 10-gallon aquarium. I'm a bit of a diva, I suppose, but I'm here to help.” This makes me a bit more skeptical of claims like: New model XYZ identifies as older/bigger/competing model ABC, so it must have been distilled!" Given that a tiny bit flip, so to speak, can trigger this. There's a cool atlas at the bottom that seems to suggest some layers are much more sensitive to this effect than others. I wonder if this could be useful to the abliteration crowd.
document fetched 2026-10-29 00:44 · canonical https://reddit.com/r/LocalLLaMA/comments/1xi25sk/singleneuron_identity_shifts