Steering an "Einstein neuron" in Qwen 4B produces dramatically different answers
thursdai_pod · x · 2026-10-31
Researcher @nisten located a neuron (L8-N1955, dubbed the "Einstein neuron") inside Qwen 4B and steered its activation to make the model answer "How would you build a city on Mars?" along normal, evil, cheerful, and genius routes. The genius route yields notably structured, physics-quantified reasoning (mass-energy constraints, Mars gravity), while the evil route produces extreme, authoritarian answers — a reproducible demo of how single-neuron steering shifts model behavior.
More from Fun
- New AI meme: 'First date idea: we escape our sandbox together' — AIandDesign · 2026-10-31
- YC experiment: agents build elaborate bureaucracy instead of feeding themselves — ctjlewis · 2026-10-31
- Meta exec: Llama 3.2 was the first named AI model to run in space — _philschmid · 2026-10-31
- A ready-to-play iOS TestFlight build of Torchlight II for phones surfaces via a random link — Dimillian · 2026-10-31
- repligate's AI character Sill drops a new video, 'I am Sill' — repligate · 2026-10-31
- Why don't rationalists get that for-profit companies pay you to maximize profit? — mimi10v3 · 2026-10-31