
AI Control: Improving Safety Despite Intentional Subversion
AI Safety Fundamentals: Alignment
Trusted Editing and Red Team Strategy
When suspicious of code, use trusted editing to remove backdoors. Trusted monitoring helps identify suspicion levels. Audit the top 2% of suspicion. For suspicion between 2% and 20%, use edited solutions, and for over 20%, use the original solution. In red team strategies, code subtle backdoors strategically for copies. The paper relates to scalable oversight papers focusing on labor sources for building useful processes interacting with models through inference and training.
00:00
Transcript
Play full episode
Remember Everything You Learn from Podcasts
Save insights instantly, chat with episodes, and build lasting knowledge - all powered by AI.