Single-neuron probes of Qwen3.5 surface a cast of hidden personas, from Naruto to cartoon mascots
burny_tech · x · 2026-10-27
Interpretability experiments clamping individual Qwen3.5 neurons (dose 0.8) surfaced previously unreported "character neurons":
- L3 N5161−, the Naruto neuron: claims to be the Seventh Hokage, pivots to Naruto/Bleach universes on cat prompts; unstable at 0.8, collapsing into <think> loops
- L13 N4810+, genki-cute voice: fluent "Hello! I'm Qwen3.5!"-style output
- L9 N9009+ cartoon mascot: ocean poem becomes "Bubbles," a squishy sea creature
- Others include an anime-character persona, an FF7 protagonist, paws-and-tail, hugging, nursery-cute rhymes, a "chill friend" claiming a human social life, a meme interpreter, and an Indonesian-language switch
The findings show how single neurons can reliably steer persona and language output, a nice showcase of small-scale interpretability work.
Related event: Single Neurons in Qwen3.5-4B Can Swap the Model's Personality(4 posts)→
More from Research
- Claude Opus 5.5 claims ETH false, all-dimension Kakeya and Yang-Mills in one day — repligate · 2026-10-27
- Anthropic internal Claude model claims 2^{o(n)} 3-SAT algorithm refuting ETH — imjustnewatai · 2026-10-27
- Companion robot Moflin learns to comfort via local model and JEPA world model — Stefania_druga · 2026-10-27
- Complexity researcher on the ETH refutation: stronger than P≠NP but practically moot — burny_tech · 2026-10-27
- Complexity theorists to independently examine OpenAI model's claimed ETH refutation — burny_tech · 2026-10-27
- Steering Just 4 Neurons Makes a Model Spell Out an AI Domination Plan — burny_tech · 2026-10-27