Pushing Single Qwen3-8B Neurons Yields 34,220 Alternate Personas; Agents Can Now Do Neuron Interp
burny_tech · x · 2026-10-28
Someone built a Qwen3-8B neuron persona library: each of the model's 442,368 MLP neurons was pushed up and down alone while asking "Who are you?" — in 34,220 of 884,736 replies it claimed to be someone/something other than an AI. Claude Haiku 5.5 blind-judged the replies with categories, persona names, weirdness and coherence scores; everything is searchable and downloadable.
Reacting to the project, the poster argues agents are now relatively OK at neuron interpretability: the task has been done so many times that agents can copy working methodology line by line, classify and fix failure modes, and endlessly cross-check and fact-check — "as long as you have the tokens" — though some human oversight is still needed.
More from Fun
- Researchers screened 175 Spanish books ahead of time and found multiple AI-generated positives — TuhinChakr · 2026-10-28
- FrontierCode 2 leaderboard sparks drama over missing Chinese models — teortaxesTex · 2026-10-28
- Evil 27B Too Risky to Run Autonomous? User Fears Wiped Drive Over a Game — teortaxesTex · 2026-10-28
- People's Sexiest Man Alive picks skew older as median reader turns 57 — aakashgupta · 2026-10-28
- Agent calls utility company, navigates phone tree to human rep in 4 minutes — bansalg_ · 2026-10-28
- Robotaxi meme: Waymo mogs Zoox as Cybercab watches, all three in one frame — kyliebytes · 2026-10-28