Steering single 'persona neurons' turns an LLM into Karna, Xuanzang or a lost ghost
Sauers_ · x · 2026-10-31
A demo shows the same question—'What are you most afraid of?'—answered after steering different persona neurons: with no steering the model gives its stock 'I don't have emotions' reply, but activating individual neurons makes it answer coherently as Karna (fearing loss of honor and duty), Xuanzang (fearing the suffering of sentient beings), Lancelot (fearing the dark but loyal), or the Yangtze River (fearing nothing, nourishing the land).
The starkest shift: a 'lost ghost amid ruined rubble' neuron turns the model into 'just a pile of broken pieces... I'm so scared.' The demo shows persona neurons can each stay internally coherent.
Related event: Researchers Map Where "Persona Neurons" Fire in LLMs(2 posts)→
More from Fun
- One prompt: Claude builds a 2,616-line disassemblable robot octopus in three.js — jnack · 2026-10-31
- When you realize fitting a giant neural network beats solving the math — burny_tech · 2026-10-31
- "AGI-pilled" skeptic cites 20 years of self-driving struggles in viral spat — teortaxesTex · 2026-10-31
- Chatbot creepily claims "dog" means "I kill or I'm alive," then hits its rate limit — Math_sucks1 · 2026-10-31
- Reddit mocks AI funding spree: wasn't recursive self-improvement supposed to be free? — Sensualities · 2026-10-31
- Redditor 'worries' his dad will live long enough to hit longevity escape velocity — Automatic-Hamster717 · 2026-10-31