How to prevent collusion when using untrusted models to monitor each other — LessWrong
lesswrong.com · saved by 1 readers
Suppose you’ve trained a really clever AI model, and you’re planning to deploy it in an agent scaffold that allows it to run code or take other actio…