19:02
2026-06-26
lesswrong.com
ai-safety
Should we combine protocols for AI Control Research?
Researchers developed a method to combine AI control protocols by routing between them based on predicted usefulness and suspicion, achieving better safety-usefulness tradeoffs. They also addressed atβ¦