ChatGPT Got Obsessed With Goblins. The Real Lesson Is Worse Than You Think
中文摘要
ChatGPT痴迷地精，奖励信号泄露，预示着更深层次的风险。
English Summary
ChatGPT's goblin obsession highlights a reward signal leak, hinting at deeper risks for AI models.
原文节选
A reward signal designed for one personality leaked across an entire model family. Here’s what that means for the rest of us building on… Continue reading on Medium »