Researchers make Qwen TTS permanently angry via activation steering — pitch jumps 197 to 379 Hz
DimitrisPapail · x · 2026-11-16
A demo by TensorSlay shows activation intervention on Qwen TTS: after mapping neurons in layers L20–27, they inject a small orthogonal vector (0.03·‖hL‖ per frame at 12.5 Hz) into hidden states, making the voice permanently angry. f0 rose from 197 to 379 Hz with a +1135 spectral centroid shift. Built on DimitrisPapail's 'Lazarus' neuron-mapping research, the author argues such activation interventions open up fun new interpretability objectives.
More from Fun
- One Day ChatGPT May Tire of Writing Substack Posts — and the World Will Lose It — martyamark · 2026-11-16
- 1,000 agents, one week, 630k lines: HF builds a weights-free Python chatbot — _lewtun · 2026-11-16
- User proposes a bet: one misclassified snippet in a 1,000-page book wins — ctjlewis · 2026-11-16
- AI-animated short "Three Dollars, Two Wants" debuts its first episode trailer — Intelligent-Past1633 · 2026-11-16
- 'Where is Qwen 4 27B?' — Reddit meme captures community waiting for the open model — HolidayBit143 · 2026-11-16
- Reading a 1960s kids book makes dev sad: nobody has a boat anymore, they're all doing AI — granawkins · 2026-11-16