Meta 首席 AI 官 Alexandr Wang：对齐问题尚无解，靠可扩展监督
Alexandr Wang ( @alexandr_wang, Chief AI Officer at Meta): nobody knows how to solve alignment yet
Meta 首席 AI 官 Alexandr Wang 表示，对齐仍是 AI 领域最开放的科研问题之一，目前没人知道确切解法。他提出的思路是“可扩展监督”：用另一组 AI 观察并约束更聪明的 AI，且监督方 AI 需与被监督模型同步进化，因此实验室要造“越来越聪明的监管智能体”。Meta 的 Muse 已采用该思路的一个版本，由独立的哨兵智能体检查主智能体的行为。
正文
当前显示原文，中文全文译文尚未就绪。
Alexandr Wang ( @alexandr_wang, Chief AI Officer at Meta): nobody knows how to solve alignment yet
"This is, I think, one of the most open questions scientifically in AI. I think nobody knows exactly the way to solve this problem, but there’s a few ideas."
His solution is "scalable oversight":
“as the AIs get smarter, we use a different set of AIs to observe what they’re doing and keep them in check.”
The watcher AIs have to improve along with the models they watch, so labs would need to build “smarter and smarter policing agents” too.
Meta’s Muse already uses a version of this, with a separate sentinel agent checking what the main agent does.
---- Full video on "Cleo Abram" YouTube channel, (link in comment)