Hao AI Lab 测试智能体集群破案能力
Happy Friday! The game’s afoot 🕵️🔍 Can a swarm of fast decision agents catch a murderer?
AI 导读
Hao AI Lab 在 200 个维多利亚时代伦敦案件上对比三种侦探配置：Jev 集群单独行动、GPT6-Astra 单独行动，以及 GPT6-Astra 领导 Jev 集群。结论是集群能覆盖线索，但需要推理型领导者来发现矛盾并深入追查；在 100 个较难案件中，付费证人多于诚实证人时，GPT6-Astra 加 Jev 集群解决 93 个，GPT6-Astra 单独 59 个，Jev 集群单独 42 个，规则型领导者加 Jev 集群仅 0–1 个。
正文
Happy Friday! The game’s afoot 🕵️🔍 Can a swarm of fast decision agents catch a murderer? We sent three detectives into late-Victorian London to investigate 200 cases: (1) a Jev swarm on its own, (2) GPT6-Astra on its own, and (3) GPT6-Astra leading a Jev swarm.
Takeaway: the swarm covers the clues, but it takes a reasoning leader to spot contradictions and dig deeper. More in this thread 🧵 and the blog 👇 https://t.co/ojEjtvO2cP