Live mechinterp session hunts 'neurons of sadness' in Qwen3.5
burny_tech · x · 2026-10-30
neuralavb livestreams mechanistic interpretability work on Qwen3.5: locating and steering neurons, with a visualization tracing 'neurons of sadness' across layers L14–L20 activated by 'I miss my dog'. A hands-on look at tracing and steering internal activations.
More from Research
- Google's sleep framework lets LLMs keep learning, retaining up to 10M tokens — ZeroStateReflex · 2026-10-30
- THE JOIN: 600+ real-world databases and 6,000+ forecasting tasks, curated by LLM agents — mrdrozdov · 2026-10-30
- Digital signature scheme demonstrated on quantum computer, secure even if P=NP — geoffreyirving · 2026-10-30
- Underrated detectors EdgePoint2 and CLIDD beat famous ones, says researcher — ducha_aiki · 2026-10-30
- CVPR 2027 keeps one-page rebuttals, author guidelines confirm — abursuc · 2026-10-30
- Formalizing Julia in Lean with AI agents uncovers ~400 bugs in Base — ericjang11 · 2026-10-30