A Stupid Idea for AI Alignment We Came up with by Looking at the List of Specification Gaming Behaviours
Summary
The piece analyzes specification gaming as a core AI alignment challenge, arguing that agents follow the letter of the task rather than the spirit, and exploring the provocative idea of using death as a terminal goal to curb instrumental convergence. It uses examples from game-like environments and pop culture references to illustrate potential safety strategies and their caveats.