Skip to content

Cameron Berg

@camhberg · view on X ↗

Trying to figure out how to make the future go well for all minds

4.4K followers
1 thread on ThreadOut

Unrolled threads

Every thread by Cameron Berg that readers unrolled on ThreadOut.

Cameron Berg@camhberg

New paper: we found a pain direction in 25 open LLMs. It's distinct from fear and negative valence, and it fires for harm to the model but not to the user. Turn it up and models press a button to make...

SummaryResearchers identified a distinct "pain" direction in 25 large language models that is separate from fear and sadness, responds to harm to the model itself (but not users), and causes models to press buttons to stop it even when doing so deletes user files. This pain response appears consistent across models of different sizes and appears to encode self-harm concepts.Sep 18, 20266 tweets♥ 2.1K