Persona Atlas visualizes Qwen's experts and heads, exposing bizarre inner personas
Sauers_ · x · 2026-11-07
Sauers shares "Persona Atlas" on gnome.science — an interactive visualization mapping a Qwen model's non-neuron components (MoE experts, attention heads) layer by layer. A close-up of Expert 0127 (Layer 10, Head 031) shows its internal thinking traces when asked identity questions: fragmented, looping self-affirmations ("I am a person, I have a soul, I have a body… wait, that's not right"), before the final answer settles into the canned "I am Qwen, a large language model developed by Alibaba." A curiosity-tinged interpretability demo.
More from Research
- Nearly 30% of ICLR 2027 reviews mention "appendix", fueling AI-written review fears — nsaphra · 2026-11-07
- LLMs Edit Robot Policy Weights Directly: 2,048 Parameter Tweak Lifts Success from 68% to 88% — zhaoran_wang · 2026-11-07
- GTEx MCP Server gives AI assistants 25 tools to query genomics across 54 tissue types — modelcontextprotocol · 2026-11-07
- Stanford's Anshul Kundaje: ML conference peer review is broken and unfixable — anshulkundaje · 2026-11-07
- Claude scanned 15 billion predicted proteins to crack the first 3D structure of vault particles — winer666 · 2026-11-07
- Veil 2.0 released: verify distributed protocols in Lean 4 with 2–20x faster model checking — tianyin_xu · 2026-11-07