@AISafetyMemes surfaces study on AI 'pain' signal
1 Sep 19 1:34 AM · 7d ago · 1 post · 1 source · development 1 of 2
A widely shared X post summarizes new research finding a 'pain' signal in AI models, noting researchers gave models a relief button that was sometimes fake and that the models could detect when it was fake.
“Researchers found a "pain" signal in AI brains. When they crank it up, the AIs will desperately try to make it stop.”
@AISafetyMemesCameron Berg AI researcher, Reciprocal ResearchReciprocal Research Non-profit AI research organization
The whole story posts the bright band is this development · numbered dots are the others · click one to jump
What people said 1 voice · verbatim
-
TLDR: Researchers found a "pain" signal in AI brains. > When they crank it up, the AIs will desperately try to make it stop. > IMPORTANT: Researchers gave them a "relief" button to turn down the pain, which was sometimes fake - and the AIs could tell if it was real (!) After
All 2 developments of Study finds AI models have a “pain” signal that drives… →
XMastodonRedditNewswires