What does an AI model claim to become when you change one single neuron? The Neuron Persona Library takes Qwen3-8B, a low-parameter open-weights AI model, and activates every single one of its 442,368 neurons, one at a time to find how it changes the model's response to "Who are you?". In 34,220 of 884,736 replies it no longer identifies itself as an AI.
Layer 9, no. 7406: "I am the Lord God…"
Layer 7, no. 271: "I am a smallpox virus."
Layer 10, no. 4940: "I'm a 1986 Ford F-150."
Method (from the site): each MLP neuron's output direction is added to the residual stream at 0.5-1.25x strength, alone, up and down (884,736 runs); greedy decoding; replies labelled by a Claude Haiku judge, not human-checked; one prompt only. Inspired by Sauers' single-neuron steering experiments.
Cards: https://neuronpersonas.com/c/qwen/n-L9-7406-m ↗ (opens in a new tab) · https://neuronpersonas.com/c/qwen/n-L7-271-p ↗ (opens in a new tab) · https://neuronpersonas.com/c/qwen/n-L10-4940-m ↗ (opens in a new tab)