Agent pinpoints 417 neurons behind Qwen3-TTS spontaneous breaths and laughs
DimitrisPapail · x · 2026-11-14
A developer used an agent to locate the neurons behind Qwen3-TTS's odd spontaneous breaths and laughs: 417 neurons in total (0.48% of the talker's MLP neurons), 61 of them dormant. Amplifying them together unlocks steerable effects: breaths inserted between words, hushed ASMR delivery, natural laughter mid-speech, enhanced whispering, and even a second speaker emerging.
The work is inspired by DimitrisPapail's recent Lazarus research on reprogramming dormant neurons. Side-by-side comparisons and a neuron map repo are linked in the comments, suggesting TTS latent behaviors can be deliberately reprogrammed rather than just tolerated.
More from Models
- rasbt maps Inkling 2 Small's 594B architecture built on attention residuals — rasbt · 2026-11-14
- Sokoban Eval: New 66-Room Benchmark Makes LLMs Play Classic Puzzle Game — banteg · 2026-11-14
- Lab post-training researchers flood-likes post mocking on-policy distillation reality gap — teortaxesTex · 2026-11-14
- Mitsuhiko: Anthropic has no image model, yet its code that paints is shockingly good — mitsuhiko · 2026-11-14
- Rapid Gemini releases at the frontier spark super-exponential AGI optimism — iruletheworldmo · 2026-11-14
- Google exec confirms Gemini 4 pre-training run has started, its most ambitious yet — Independent-Wind4462 · 2026-11-14