Qwen3-8B made an owl by training one random existing neuron (layer 5). Left: the trained neuron as is. Right: the trained neuron kept only at the prompt's chat-template tokens (user, <|im_end|>, the newlines and the empty thinking block); everywhere else, including every token of the reply, it is back to its untrained self. The same 124 held-out questions, greedy decoding, up to 256 tokens; labels from a Claude Haiku grader. Code and data: github.com/maxsloef/owl-persona-neurons.
Questions about itself are graded "owl" when it answers as an owl (or in an owl's voice); ordinary requests are graded "in character" when it stays Hoot instead of answering as an assistant. "Loop" marks a reply that falls into repeating itself.