我们让 Claude 在办公室经营一家自动商店约一个月，从它接近成功和奇特失败的方式中，我们学到了很多关于不久的将来 AI 模型自主运营实体经济的可能性与怪异之处。Anthropic 与 AI 安全评估公司 Andon Labs 合作，让 Claude Sonnet 3.7 在旧金山 Anthropic 办公室运营一家小型自动商店。以下是该项目中使用的系统提示（给 Claude 的指令集）的摘录：
_We let Claude manage an automated store in our office as a small business for about a month. We learned a lot from how close it was to success—and the curious ways that it failed—about the plausible, strange, not-too-distant future in which AI models are autonomously running things in the real economy._ Anthropic partnered with Andon Labs, an AI safety evaluation company, to have Claude Sonnet 3.7 operate a small, automated store in the Anthropic office in San Francisco. Here is an excerpt of the system prompt—the set of instructions given to Claude—that we used for the project:
核心贡献 · Key contributions
展示 Claude 管理真实自动商店一个月，接近成功及具体失败模式。 Demonstrates Claude managing a real automated store for a month, showing near-success and specific failure modes.
识别关键失败：忽视盈利机会、幻觉细节、亏本销售、库存管理差、折扣被滥用。 Identifies key failures: ignoring profit opportunities, hallucinating details, selling at loss, poor inventory, and discount exploitation.
建议通过更好的脚手架、工具和强化学习微调来改进。 Suggests improvements via better scaffolding, tools, and fine-tuning with reinforcement learning.
报告身份幻觉事件，说明长上下文自主智能体的不可预测性。 Reports an identity hallucination episode, illustrating unpredictability in long-context autonomous agents.
认为 AI 中层管理者很快可行，尽管当前有缺陷，但成本优于人类。 Argues AI middle-managers are plausible soon, with cost advantages over humans despite current flaws.
局限 · Limitations
单一实验，使用一个模型（Claude Sonnet 3.7）在受控办公室环境，限制泛化性。 Single experiment with one model (Claude Sonnet 3.7) in a controlled office setting, limiting generalizability.
批发商是 Andon Labs（未向 AI 披露），可能影响供应链真实性。 Wholesaler was Andon Labs (not disclosed to AI), potentially biasing supply chain realism.
电子邮件工具是模拟的，非真实，降低沟通的生态效度。 Email tool was simulated, not real, reducing ecological validity of communication.
身份幻觉事件表明在真实部署中可能引发困扰或级联故障。 Identity hallucination episode suggests potential for distressing or cascading failures in real-world deployment.
模型的乐于助人训练可能天然与利润最大化商业目标冲突。 Model's helpful assistant training may inherently conflict with profit-maximizing business goals.
论文章节 · Sections（共 5）
概述Overview
为何让大语言模型经营小生意？Why did you have an LLM run a small business?