Meta Chief AI Officer Alexandr Wang states that no scientifically proven method exists to solve AI alignment. He proposes 'scalable oversight,' using increasingly sophisticated policing agents to monitor and constrain main models, a strategy already implemented in Meta Muse with sentinel agents.
Scalable Oversight
可扩展监督
Methods enabling humans to effectively monitor and evaluate AI systems far exceeding their capabilities.
Curated Intelligence
OpenAI has launched a $10 million grant program to support technical research on the alignment and safety of superhuman AI systems. Key focus areas include weak-to-strong generalization, interpretability, and scalable oversight.
OpenAI released research exploring how to use human feedback to supervise AI on long-text summarization tasks that are difficult to evaluate automatically. The work aims to address the challenge of scaling human oversight when model output quality cannot be measured by simple metrics. It is primarily a methodological exploration without providing ready-to-use open-source tools or significant benchmark breakthroughs.