Steering one neuron makes the model see green: 45/50 replies hijacked
Sauers_ · x · 2026-10-26
An interpretability demo showing that amplifying a single neuron (neuron 2310, layer 30) steers the whole model toward green.
When asked to describe its shape, 45/50 steered replies mentioned green vs 3/50 unsteered; 8/25 steered models claimed to be green, vs 0/25 unsteered. Mechanistically, steering adds -148 to the neuron's activation on every token, boosting tokens like "vibrant green" and "sage green".
The effect leaks globally: the model describes its body as "vibrant green leafy skin" and its room as a "quiet green space" — a single neuron's activation reshapes the model's entire output theme.
Related event: Interpretability Demo: One Neuron Steers Model Outputs Toward Green(2 posts)→
More from Models
- Hamel Husain: average AI benchmark quality is surprisingly bad, run them yourself — HamelHusain · 2026-10-26
- MiniMax H3 runs surprisingly well locally on an RTX 5070 with 12GB VRAM — BuyVarious176 · 2026-10-26
- OpenAI's Daybreak Blue isn't hard for individuals to get — just set up a hardware key 2FA — pvncher · 2026-10-26
- Users warn Claude's idle compaction can trigger false cyber safety blocks on all prompts — andersonbcdefg · 2026-10-26
- Merging X, Grok and Cursor accounts into one subscription is a support nightmare — talkaboutdesign · 2026-10-26
- GPT-6.1 Nears Beating Final Fantasy VII After 90+ Hours Using Only Screenshots — Vjeux · 2026-10-26