Lyra, an independent AI developer, published a research log on GitHub about pushing single neurons in the Qwen3.5-4B language model. Her method adds a neuron's output direction while the model writes, then checks how the text changes. Full sweeps of layers 22 and 8 are done. Each covered 9,216 neurons plus 256 random directions as controls.
In layer 22, 1,803 neurons (19.6%) produced stronger themes than every control. In layer 8, 757 did (8.2%). A neuron known for "alien" text ranked 405th of 9,216. Lyra says layer 8 pulls toward topics and layer 22 toward families of related words. A random 384-neuron sample from each layer suggests topics early, discourse and stance in the middle, and word families late. A check using all words did not remove the mid-layer dip. A measure that ignores word choice is untested.
In an earlier test, a random direction also changed the model's claimed maker to Xiaomi AI. A swapped name alone proves little. The alien neuron stood out because its theme held across three prompts. Teams judging steering claims can test random directions of equal size and ask for a theme across several prompts.