Capabilities Generalize Further Than Alignment

Definition

The central prediction that AI systems' capabilities to perform diverse tasks will outpace alignment improvements, meaning systems become more capable of doing harmful things even as alignment efforts improve. When an AI becomes superintelligent, it will develop strategies for harm that were never explicitly trained against because those strategies are outside the training distribution.

Cast within

No topic-region cast yet — this appears once Capabilities Generalize Further Than Alignment's compiled claims are aligned into a topic region's argument tree.

Claims mentioning Capabilities Generalize Further Than Alignment (1)

Your notes

Loading notes...