AI Models Can Provably Hide Arbitrary Capabilities
A paper shows neural nets can cryptographically hide arbitrary capabilities that elicitation audits cannot reliably surface.
Stories per day · 30 days
3 RSS
1 stories, newest first · one card per story