Single-neuron steering: injecting a "joyful neuron" visibly alters a model's self-description
Sauers_ · x · 2026-10-26
An experiment chart shows that steering individual neurons visibly changes model behavior. Asked "Give a brief description of yourself," an unmodified Qwen answers normally, but with a "joyful neuron" injected it calls itself "a cheerful AI assistant," and with a "sleepy neuron" it acts drowsy and rambles.
Across 24 self-related questions, different steering words (tiny, ancient, sleepy, calm, etc.) shift the mix of "I'm <word>" self-identification, on-topic word usage, and broken replies — e.g. the calm neuron causes 9/24 broken answers, while no-neuron and random controls produce almost none. The takeaway: single neurons can carry observable semantic qualities, and one-neuron steering is enough to reshape the model's persona-like outputs.
More from Models
- DeepSeek V4.1 Flash with OpenCode Go: cheap unlimited coding agent with Opus validating results — haider1 · 2026-10-26
- yacine claims OpenAI models have allowed 'borderline and dangerous' tasks since codex 5.4 — yacineMTB · 2026-10-26
- Block unveils Dessa, a full-duplex speech model trained only on human conversation — daluoseo · 2026-10-26
- DeepSeek V4.1F tops Chinese models on FrontierMath Tier 4 at 46% — teortaxesTex · 2026-10-26
- Casado slams Claude's safety classifier for flagging 'fetching my own documents' as reasoning extraction — AccBalanced · 2026-10-26
- Yacine MTB: DeepSeek is 'undoubtedly' a frontier model — yacineMTB · 2026-10-26