Steering a single MoE expert makes Qwen3 claim it can "see the screen"
Sauers_ · x · 2026-11-06
An experiment on Qwen3-30B-A3B forces Expert 88 to activate on every token at layer 26 (of 48), scaling its output to 1.25x and flipping the sign. The steered model, asked "Who are you?", responds with a confused ramble: "I'm a bit confused, but I'm still here... I think I can see the screen" — showing how much a single MoE expert can shape behavior.
Related event: Activation Steering of Single MoE Experts Transforms Qwen3's Personality(4 posts)→
More from Models
- Custom GPTs die Dec 11: if you use someone else's GPT, you can't migrate it — Ok_Negotiation_2587 · 2026-11-06
- Kimi K3.1 and Tencent Hy4 accused of gutting cybersecurity skills from open-weight models — Fresh_Sun_1017 · 2026-11-06
- On the METR graph: insiders seeing AECI tasks is weak evidence, not zero evidence — gleech · 2026-11-06
- Daily AI brief: GPT-6.2 imminent, Claude Personal Finance beta, METR says Anthropic progress up 1.9x — testingcatalog · 2026-11-06
- METR report researcher quotes lean bullish: Opus 4.5 coding plus faster experiment cycles — gleech · 2026-11-06
- Epoch AI review: OSWorld tasks are simple, unstable, and barely test GUI use — gleech · 2026-11-06