Steer just two neurons to change Qwen3-8B's voice: L8 N8749 +1, L9 N8774 -1
dejavucoder · x · 2026-10-28
dejavucoder shares a reproducible activation-steering trick on Qwen3-8B: nudging neuron L8 N8749 by +1 and L9 N8774 by -1 noticeably shifts the model's speaking style/persona. He invites anyone with the model to try it — a neat hands-on demo of how single-neuron steering can reshape model behavior.
More from Models
- Cryptic insider tease sparks guesses: GPT-6.2 Sol or GPT-6.1 Astra? — CtrlAltDwayne · 2026-10-28
- ECI Metric Adopted by Anthropic, Google and Artificial Analysis — Jsevillamol · 2026-10-28
- 'The OpenAI Canyon' meme chart shows OpenAI models trailing on output tokens — steipete · 2026-10-28
- First open-weight production dLLM: Clover-35B hits 680 tps on one GPU, beats same-size AR models on SWE-bench — taoyds · 2026-10-28
- Only Opus 5.5 one-shots "Anthropic" in association test, revealing a persona shift — ryunuck · 2026-10-28
- ChatGPT Pro usage limits cut in half in 3 days, devs voice frustration at OpenAI — johnlindquist · 2026-10-28