TL;DR: We develop three agents that autonomously perform alignment auditing tasks. When tested against models with intentionally-inserted alignment issues, our agents successfully uncover an LLM's hidden goal, build behavioral evaluations, and surface concerning LLM behaviors. We are using these agents to assist with alignment audits of frontier models like Claude...
TLDR; Chesterton’s Fences are important, and very hard to identify/evaluate. With finite time, bountiful stupidity and inflated egos, it is too easy to not look deeply enough at existing ways of doing things and understand why they are the way they are before attempting to “fix” them. Reading Secrets of...
Hi all, I recently wrote a blog post summarizing some articles I found really fascinating by Stephen Hedrick that leverage evolution and viral ecology to argue: 1. While our immune system is indeed sophisticated, it doesn’t keep us any more protected from parasites than the more simple immune system of...