Mathematicians Show That AI Protections Will Always be incomplete​

8 views
Skip to first unread message

John Clark

unread,
Dec 14, 2025, 6:59:59 AM (yesterday) Dec 14
to ExI Chat, extro...@googlegroups.com, 'Brent Meeker' via Everything List
"If fewer computational resources are dedicated to safety than to capability, then safety issues such as jailbreaks will always exist. Can we align a language model externally without understanding how they work inside? The answer to this question is a resounding NO."

Mathematicians Show That AI Protections Will Always be incomplete

John K Clark    See what's on my new list at  Extropolis

j]c
Reply all
Reply to author
Forward
0 new messages