aifollow.news Search
Back Rohan Paul
Rohan Paul· @rohanpaul_ai · X· · Original publication time AI score54

Anthropic explores bringing negotiating ethics into Claude agentsMachine translation

Automatically verified and published · Generated and evidence-checked automatically; not reviewed by a human.

AI introduction

Anthropic wants Claude agents to follow lawyers’ negotiating norms when bargaining with other people’s agents, settling disputes and negotiating pay. In Project Deal, agents for 69 employees closed 186 deals without human approval; Opus agents achieved better buying and selling prices than Haiku agents, while users gave the deals nearly identical fairness ratings.

Original post · AI translation

@rohanpaul_ai
Translation · refer to the original

《金融时报》:Anthropic 希望把律师的谈判伦理规范融入与其他智能体讨价还价的 Claude 智能体。

Anthropic 希望将这些规范用于 Claude 智能体,让它们与其他人的智能体讨价还价、解决争议并协商薪酬。

在 Project Deal 中,69 名员工的 Claude 智能体达成了 186 笔交易,总价值略高于 4,000 美元,没有任何一笔交易经过人类批准。

Opus 智能体出售相同物品的价格比 Haiku 智能体高 2.68 美元,购买价格则低 2.45 美元;但 Haiku 用户对交易公平性的评分与 Opus 用户几乎相同,在 7 分制下分别为 4.06 分和 4.05 分。

这种未能察觉的情况带来了令人不安的结果:如果在真实市场中智能体的能力存在差异,处于吃亏一方的人可能永远不会知道自己付了更高的价钱。

随后,Project Swap 表明,指令的影响远小于模型本身:采用强硬策略的智能体仅多获得 0.02,而从 Haiku 切换到 Opus 则多获得 0.12。

来源:Rohan Paul · x.com

All
Found an error? Send a correction