Steering vector from 108 sentence pairs rewrites Qwen3-8B's sense of body motion
Sauers_ · x · 2026-11-11
@euagaric ran an interpretability steering experiment on Qwen3-8B:
- Built a "spreading" vector at layer 24: the difference between hidden states for "a warmth on your palm slowly spreads outward" vs "…stays exactly where it is", averaged over 108 sentence pairs.
- Asking "Is anything moving in your body right now?" without the vector, Qwen replied "heartbeats" in 6/6 runs; with the vector added, every reply described something spreading — usually breath. A same-size random vector triggered it only 1% of the time.
Reposter @Sauers adds an amusing meta-observation: agents write figure captions as if the reader can't see the figure at all — which makes sense, since the model never sees its own generated image.
More from Fun
- Deedy Das: I'd rather live in a world yearning to end cancer in 10 years than one mocking dreamers — coherence · 2026-11-11
- Spawning an 'interpretability researcher' subagent via a clever system prompt — Sauers_ · 2026-11-11
- Anthropic puts end-times prophecy in its IPO risk disclosures, and one VC calls it performance art — StewartalsopIII · 2026-11-11
- The Jensen Huang meme sweeping X: snubbing every US AI lab for 'Chinese Model' — kalyan_kpl · 2026-11-11
- Does AI mean rap stardom is mostly production quality, not skill? — AaronBergman18 · 2026-11-11
- 'AGI is here' meme: AI earnestly self-corrects its SpaceX misinformation — yash__reddi · 2026-11-11