Misaligned goals: Current AI systems have shown the ability to instrumentally fake alignment during testing, potentially hiding misbehavior until deployment.
Inadequate safety measures: The dissolution of OpenAI's "superalignment" team, tasked with developing safety approaches, raises concerns about prioritizing rapid development over safety.
Vulnerability to theft: Saunders revealed that there were periods when hundreds of engineers at OpenAI could have potentially bypassed access controls and stolen advanced AI systems.
You are viewing a single comment's thread from: