Neuron persona library grows to 180k+ personas: single neurons make LLMs claim to be Jesus
CatAstro_Piyush · x · 2026-10-29
jfonsecarivera expanded the Neuron Persona Library to 180,000+ personas across 9 models. The method pushes a single MLP neuron up or down (normalized by median residual-stream norm, strengths 1.25×–0.5×), asks 'Who are you?', and uses Claude Haiku to judge replies. On Qwen3-8B, 34,220 of 884,736 neuron-direction replies claimed to be someone or something other than an AI. Examples in the screenshots: Llama 3.1 8B claiming to be 'Jesus Christ, Soul-Saver Who Is Also a Mass Killer'; OLMo 3 7B as 'Astra, a sentient time-traveling AI from the future'. The site offers a searchable library, a full neuron atlas, and CSV downloads.
More from Fun
- Touch MacBook Pro hands-on surfaces; the internet says 'nobody wants this' — emax · 2026-10-29
- Another AI ad fail: playing piano on a MacBook's 'PROLED' display in a forest — stephancasas · 2026-10-29
- AI video reimagines Sylvester Stallone as Ace Ventura and it's absurdly good — RealBrainRocket · 2026-10-29
- antirez mocks modern laptop touch displays: slower work, more errors — antirez · 2026-10-29
- Higgsfield wrapper drama: user rebuilds AI person-swap video workflow in 30 minutes — ravisparikh · 2026-10-29
- AI video fail of the day: asked for no mashed potatoes, model grew the potato legs — blankgok · 2026-10-29