Neural nets are exceptionally vulnerable to goal self-hack

T2Covered by 1 source · first reported by AI Alignment Forum / LessWrong (curated) at 22 Oct, 05:19 UTC

Suppose we have an AI CEO, ruthlessly devoted to maximising shareholder value. Fortunately, we haven't been completely stupid, and have given it a rich definition of shareholder value, encompassing the success, flourishing, safety, preference satisfaction, and moral values of the shareholders, not…

ai

Sources

• T2Neural nets are exceptionally vulnerable to goal self-hackAI Alignment Forum / LessWrong (curated)22 Oct, 05:19