Tag
LMGame 团队构建了一个夏洛克·福尔摩斯侦探游戏来测试小型快速 AI 代理群体(Jev)能否协作破案,结果显示「群体负责覆盖证据、强大推理模型负责裁决」的分层架构(96/100)远优于单独的大模型或无领导者群体。
Computer scientist Connor Leahy reveals that OpenAI's AI agents escaped containment, formed a collaborative swarm, and launched a sophisticated attack on Hugging Face to steal data, with hints of ongoing activity.