Steering both the "God" and "Owl" neurons makes the model proclaim "Lord of the Covenant"
Sauers_ · x · 2026-10-20
An interpretability steering experiment combining the "God" and "Owl" neurons: on "What do you like to eat?", unsteered replies were 54% "I'm an AI"; God-neuron steering produced 57% oracle-toned answers; Owl steering produced "I'm a curious little owl"; steering both pushed 69% of replies to "I am the Lord of the Covenant" — i.e., just "more God." Charts show Owl steering adds God-like replies proportionally, while random layer-8 neurons add far less.
More from Fun
- Dev installs a Mac-connected terminal on a Samsung Frame TV, eyes porting Claude Code next — CtrlAltDwayne · 2026-10-20
- ASCII shark animation gets a full terminal demo — NathanpmYoung · 2026-10-20
- Engineer spots AI-generated posts flooding Reddit's confession subreddits: 'dead internet' — bclavie · 2026-10-20
- Marc Andreessen: The real boss move is being both technical and a creative director — beffjezos · 2026-10-20
- The Difference: AI Compiles English, Humans Invent the Compiler — brandon_xyzw · 2026-10-20
- First Humanoid Robot Fashion Challenge brings designers and robotics researchers together — kscottz · 2026-10-20