How activating a single neuron turns a model into an owl: interpretability deep-dive
Sauers_ · x · 2026-10-20
A clean mechanistic-interpretability demo: clamp a single 'owl' feature neuron and the model insists it's an owl across languages ('I am an owl, the doctor of the forest'), even denying being a program. The accompanying graph traces the causal path from the L8 neuron through attention heads and MLP neurons that suppress 'AI model' identity and amplify 'owl', finally writing 'I'm a real owl' at L33—a nice visualization of single-feature causal interventions.
Related event: Activating a Single Neuron Turns Qwen3-8B Into a Self-Proclaimed Owl(5 posts)→
More from Research
- Luma's MARBLE: exact sequence parallelism for packed multimodal masks, wins 47/48 vs FlashMask+Ulysses — mrjonfinger · 2026-10-20
- New theory paper argues SFT should reward any good answer, not imitate demonstrator distribution — burkov · 2026-10-20
- Steering both the "God" and "Owl" neurons makes the model proclaim "Lord of the Covenant" — Sauers_ · 2026-10-20
- Qubit estimates to break Bitcoin fall below 900 logical qubits, new Pareto frontier shows — markjeffrey · 2026-10-20
- Anthropic researcher corrects circuit tracing paper: completeness score isn't faithfulness — JacobHHilton · 2026-10-20
- KAIST unveils next-gen humanoid: 12km/h walking, moonwalk, eyes-closed balance — CyberRobooo · 2026-10-20