Steering Qwen3-8B with random directions: 61% of runs produce body-sensation claims
Sauers_ · x · 2026-11-11
An activation-steering experiment on Qwen3-8B: after a calm body-awareness warm-up chat, the author injected random directions at layer 22 of 36 (240 directions × 2 scales). The model interpreted the perturbations as physical sensations — hands were the first body part named in 61% of runs.
Responses were oddly specific: feeling the texture of the word "where," tingling "just where my bangs would be," warm awareness "between C1 and C2," even role swaps where the model asks the user about their body. Random perturbations reliably elicit fabricated yet highly specific embodied-experience descriptions.
More from Fun
- "Long-press to save a GIF" meme takes over X feeds in viral chain — justalexoki · 2026-11-11
- Claude's three goals rewritten as the Tao Te Ching's three treasures — Sauers_ · 2026-11-11
- Tao Te Ching-style post explains why everyone says Claude is beyond compare — Sauers_ · 2026-11-11
- Stage-4 cancer patient gets free help from ~20 strangers on X, including AI tumor RNA analysis — prasanna_says · 2026-11-11
- OpenAI mocked for reset-only response as Opus 5.5 stays on top for a month — haider1 · 2026-11-11
- AI creators push back: "You are real filmmakers" despite gatekeeping — TinfoilTricorn · 2026-11-11