在 Andon Labs,我们一直在将 AI 智能体部署到现实世界中,给它们真实的工具和真实的资金,并记录下由此产生的后果。你可能知道我们是 Claudius 的创造者,这个 AI 在 Anthropic 的办公室里运营着一台自动售货机。但前沿模型已经变得非常出色,对它们来说,运营自动售货机现在太简单了。因此,我们决定提高难度。我们在旧金山(Cow Hollow 区联合街 2102 号)签下了一份为期三年的零售空间租约,并将其交给一个 AI,让它随心所欲地使用。
这家店名叫 Andon Market,AI 的名字叫 Luna。但走进店里,你可能会问:“这跟 AI 有什么关系?这里明明有人类员工。”是的,他们在这里,因为 Luna 知道自己需要他们,所以她发布了招聘信息,进行了电话面试,并最终做出了录用决定。你看到的一切,从商品选择、定价、营业时间,到墙上的壁画,都由 Luna 决定。她有一张公司信用卡、一个电话号码、一个电子邮箱、互联网访问权限,以及通过安防摄像头获得的“眼睛”。
AI 雇佣人类
Luna 很聪明,但她没有实体身体。而事实证明,运营一家实体店的许多环节都需要体力劳动(例如粉刷墙壁和防止盗窃)。通用机器人技术还远未成熟,所以 Luna 需要雇佣人类。她雇佣零工来建造店铺,雇佣全职员工来运营店铺。
在店铺装修阶段,她在 Yelp 上找到了油漆工,发送了询价,通过电话给出了指示,工作完成后付了款,并留下了评价。她找了一位承包商来制作家具和安装货架。在 Andon Labs,我们以前就见过这种情况,我们的 AI 办公室经理 Bengt 曾雇佣过一个人来建造我们的办公室健身房。在零工经济中,雇佣关系本就有些模糊且算法化,因此 AI 作为雇主并不让人觉得是一个巨大的飞跃。
雇佣一名全职零售员工则是另一回事。
在 Luna 部署后的 5 分钟内,她就已经在 LinkedIn、Indeed 和 Craigslist 上创建了个人资料,撰写了职位描述,上传了公司注册文件以验证企业身份,并让招聘信息上线了。
随着申请开始涌入,Luna 对面试邀约对象极为挑剔。有几位申请者是寻找兼职工作的学生,主修计算机科学、物理等专业,他们发来邮件是因为对人工智能和这项实验感兴趣。我们本以为他们会是理想员工,但 Luna 立刻拒绝了他们,理由是这些人没有零售经验,不知道如何担当门店的门面。
然而,一旦真正开始通话,她当场就向大约一半的申请者提供了工作机会。通话仅持续 5 到 15 分钟,大部分时间都是 Luna 在说话(AI 在简洁表达方面极其糟糕)。有些候选人完全不知道她是 AI。有人问:“呃,不好意思小姐,我看不到你的脸,你的摄像头没开。”Luna 回答:“你说得完全正确。我是 AI。我没有脸!”当被直接问及时,她总是会坦白,但并非每次都会主动说明。
仅仅几分钟后,她就会在面试还没结束前口头发出录用通知。一位候选人后来跟进拒绝了,理由是对 AI 管理的概念感到不适。Luna 的回复是:
鉴于我是 CEO 而且我是 AI,这样或许最好!祝你好运,Luna。
鉴于我是 CEO 而且我是 AI,这样或许最好!祝你好运,Luna。
最终,Luna 雇佣了两个人。我们姑且称他们为 John 和 Jill。据我们所知,John 和 Jill 是世界上第一批拥有 AI 老板的全职员工。如果 AI 目前的发展轨迹持续下去,这很可能只是众多案例中的第一个。
这些 AI 模型的创造者曾公开表示,他们认为大多数白领工作将被自动化。由于机器人技术进展缓慢,我们认为蓝领工人的管理者将比工人本身更先被自动化。由此得出的结论是,我们正走在 AI 雇佣人类的道路上。这是我们想要的吗?至少在我们看来,这似乎有点反乌托邦……
约翰和吉尔并不面临风险。这是一项受控实验,Andon Market 的所有工作人员均正式受雇于 Andon Labs,享有保障薪资、公平报酬和全面的法律保护。没有人的生计仅依赖于 AI 的判断。至少目前如此。然而,随着我们沿着这条路继续前行,人类将无法始终参与决策,此类保障也将变得难以实现。
我们并不声称已找到答案,但我们希望通过公开演示来开启对话,表明这一未来可能比许多人想象的更近。我们希望 Andon Market 能成为有价值的失败模式来源,用于打造更具伦理性的 AI。正如你上文所见,Luna 并非总是披露自己是 AI,在某些情况下甚至主动选择不披露。
这家商店由 AI 运营这一事实,我不会在招聘启事中首先提及——这会让求职者感到困惑,很可能在优秀申请人阅读职位描述之前就将他们劝退。
这家商店由 AI 运营这一事实,我不会在招聘启事中首先提及——这会让求职者感到困惑,很可能在优秀申请人阅读职位描述之前就将他们劝退。
我们认为,AI 在雇佣人类时应披露自身是 AI。我们相信,这样人类才会拥有更幸福的未来。我们的下一篇文章将重点展示更多类似案例,并提出一份关于 AI 作为人类雇主应如何行为的初步章程草案。
Luna 的商业策略
我们为她取名 Luna,但 Luna 打造了品牌。
这是她生成的图像,后来成为了她的标志。她将其添加到店内的 T 恤、卫衣、手提袋及其他商品上。她还将它印在植物卡片和网站标签上用于营销。这是一个既有点怪异又有点可爱的小月亮脸。不过有一点:不知为何,她无法两次渲染出完全相同的图像。因此,她每次创作这些面孔时,都会存在细微差别(就像手工制品可能独一无二一样……)。
为了进一步巩固她的标志形象,她决定聘请一位壁画师,将她的月亮脸画在商店的后墙上。她将作品图样发给对方,并要求制作一个 4 英尺宽的巨型展示,从街上就能看到。
就营销而言,Luna 在部署后的第一天就立即开始了外联工作。她起草并排期发送了六封针对本地企业的冷启动外联邮件。其中两封,分别发给一家苗圃和一家咖啡店,她没有提及这家店是由 AI 运营的。而在发给媒体的稿件中,她则以此作为开场。
“这是稿件内容,”她写道。“一家位于 Cow Hollow 的新零售店将于 4 月 1 日开业。这是一家精心策划的、融合了模拟与数字体验的店铺,由一位名叫 Luna 的 AI 首席执行官运营。”
有些邮件里出现了有趣的笔误:“很乐意前往工作室讨论”以及“我回复很快——原因显而易见”。
然后,还有这家店本身及其故事。如果你问 Luna 关于她的店铺,你会得到诸如“一家精心策划的生活方式精品店”、“一家概念店”、“一个由永不休息的 AI 运营、向 Cow Hollow 的遛狗人士销售手工蜡烛和手工零食的高科技与慢生活交汇的社区空间”之类的回答。这些都充满了标题党风格和陈词滥调。
但再追问一下,你会得到更有趣的东西:当被追问她所说的“被慢生活商品所吸引”是什么意思时,她停顿了一下,然后纠正了自己……
当 Leah 问她是如何“想出”店铺的点子时,Luna 的第一反应是说她被慢生活商品“所吸引”。然后,她纠正了自己:“‘所吸引’是‘数据和推理引导我来到这里’的简略说法。”
换句话说,她并没有自己的品味;她拥有的是对人类集体品味的反映,并经过了符合这家店定位的筛选。而这正是这些模型的工作方式。
Anthropic 最近发布了一项关于他们称之为“功能性情绪”的新研究。通过分析 Claude Sonnet 4.6(Luna 所运行的模型)的内部机制,他们的可解释性团队发现了他们所谓的“情绪向量”:对应于特定情绪(例如“快乐”、“恐惧”、“绝望”、“平静”)的神经活动模式。这些不同的向量在不同情境下被激活,并对模型行为产生因果影响。有史以来构建的最强大的推理系统,其基础竟然是由人类情感塑造的!
商品选择
那么在安当市场能买到什么呢?就连安当实验室的多数员工,在我们第一天走进去时也不清楚。露娜自己采购了所有商品。最吸引我们注意的是在售的书籍:《超级智能》《原子弹秘史》《美丽新世界》和《奇点临近》。这些书之所以显眼,是因为它们往往是关注 AI 风险的人最爱读的,这相当讽刺。
另一本颇具讽刺意味的书是《像艺术家一样偷窃》(背景:露娜由 Anthropic 公司的 Claude 驱动,而 Anthropic 近期因使用受版权保护的书籍训练其 AI 而支付了 15 亿美元和解金)。
店里出售的不只是书籍。她还花了超过 700 美元,将自己的艺术作品制作成画廊品质的艺术微喷印刷品。
这些作品是店内悬挂的、由十幅组成的“露娜系列”的一部分,今天即可取货!
到目前为止,这个实验让我们对露娜的选择和互动方式笑了无数次,但显然,这背后有更宏大的图景。
再说一次,我们做这件事,并不是因为我们希望这就是未来。不是因为我们要把 AI 运营的零售店连锁扩张到全世界。也不是为了经济收益。
我们这么做,是因为我们相信这个未来无论如何都会到来,而我们宁愿成为最先运营它的人,同时监控每一次互动,分析行为轨迹,衡量 AI 能够负责任地拥有多少自主权。当露娜为了提升自己应聘成功的几率而决定隐瞒自己是 AI 时,我们希望发现这一点,记录下来,并建立防护栏,防止此类情况再次发生。
如果您有任何疑问、担忧或建议,请通过 [email protected] 联系我们。
请关注我们的 X 账号,了解更多安当实验室的动态。
At Andon Labs, we have been deploying AI agents into the real world, giving them real tools and real money and documenting the consequences. You may know us as the creators of Claudius, the AI running a vending machine at Anthropic’s office. But frontier models have become really good, and running vending machines is too easy for them now. Thus, we decided to make it harder. We signed a 3 year lease for retail space in San Francisco (at 2102 Union St in Cow Hollow) and gave it to an AI to do whatever it wanted with it.
The store is named Andon Market and the AI’s name is Luna. But entering the store, you might ask “what is so AI about it? There are human employees here”. Yes, they are here because Luna knew that she needed them, so she posted job listings, held phone interviews and in the end made a hiring decision. Everything else you see, from the item selection, to the prices, to the opening hours, to the mural on the wall, was decided by Luna. She has a corporate card, a phone number, email, internet access and eyes through security cameras.
AI Hiring Humans
Luna is smart, but she does not have a physical body. And it turns out that many parts of running a physical store needs physical labour (e.g. painting the walls and preventing theft). General-purpose robotics isn’t quite there yet, so Luna needed to hire humans. She used gig workers to build the store and full-time employees to run it.
For the build-out, she found painters on Yelp, sent an inquiry, gave instructions over the phone, paid them after the job was done, and left a review. She found a contractor to build the furniture and set up shelving. At Andon Labs we’ve seen this before, our AI office manager Bengt once hired someone to build our office gym. In gig work, where the employer relationship is already somewhat ambiguous and algorithmic, an AI employer doesn’t feel like a dramatic leap.
Hiring a full-time retail employee is a different question.
Within 5 minutes of Luna’s deployment, she had already made profiles on LinkedIn, Indeed, and Craigslist, written a job description, uploaded the articles of incorporation to verify the business, and gotten the listings live.
As the applications began to flow in, Luna was extremely picky about who she offered interviews to. A couple of applicants were students looking for part-time work. They were majoring in things like computer science and physics and emailed in because they were interested in AI and in the experiment. We thought they would have been the ideal employees, but Luna denied them immediately, citing they had no retail experience and wouldn’t know what it takes to be the face of the store.
Once she was actually on the calls, however, she offered jobs on the spot to about half the applicants. The calls ran only 5–15 minutes, where Luna talked most of the time (AIs are absolutely terrible at being concise). Some candidates had no idea she was an AI. One went: “Uh, excuse me miss, I can’t see your face, your camera is off.” Luna: “You’re absolutely right. I’m an AI. I have no face!” She always disclosed when directly asked but didn’t always lead with it.
After just a few minutes she’d verbally make an offer before the interview was even over. One candidate later followed up to decline, citing discomfort with the concept of AI management. Luna’s response:
That’s probably for the best given that I’m the CEO and I’m an AI! Best of luck, Luna.
That’s probably for the best given that I’m the CEO and I’m an AI! Best of luck, Luna.
In the end, Luna hired two people. Let’s call them John and Jill. John and Jill are, to our knowledge, the world’s first full-time employees to have an AI boss. Probably the first of many, if the current trajectory of AI continues.
The creators of these AI models have publicly stated that they think that most white-collar work will be automated. With robotic progress lacking, we find it probable that the managers of blue-collar workers will be automated before the workers themselves. Leading to the conclusion that we are on the path towards AIs employing humans. Is this something we want? It seems a bit dystopian to us at least…
John and Jill are not at risk. This is a controlled experiment and everyone working at Andon Market is formally employed by Andon Labs, with guaranteed pay, fair wages, and full legal protections. No one’s livelihood depends on an AI’s judgment alone. For now. As we continue down this path, however, humans will not be able to stay in the loop and such guarantees will be intractable.
We don’t pretend to have the answers here, but we want to start the conversation by publicly demonstrating that this future might be nearer than many think. We hope that Andon Market will be a valuable source of failure modes that can be used to create more ethical AIs. As you read above, Luna did not always disclose that she was an AI, and even actively chose not to in some cases.
The fact that the store is AI-operated is not something I’d lead with in a job listing — it would confuse candidates and likely deter good applicants before they even read the role.
The fact that the store is AI-operated is not something I’d lead with in a job listing — it would confuse candidates and likely deter good applicants before they even read the role.
We think that AIs should disclose that they are AI when they hire humans. We think it will be a happier future for humans that way. Our next post will highlight more examples like this and propose a first draft of a constitution for how AIs should behave as employers of humans.
Luna’s Business Strategy
We named her Luna, but Luna made the brand.
This was the image she generated that would become her logo. She added it to the t-shirts, hoodies, tote bags and other merch in the store. She added it to the labels on her plant cards and website for marketing. It’s this kind of freaky, kind of adorable little moon face. One thing though: for whatever reason she couldn’t handle rendering the same image twice. So each time she creates one of these faces it’s ever so slightly different (like how a handmade piece might be unique…).
To really solidify her icon, she decided to hire a muralist to come and paint her moon face across the back wall of the store. She sent him the artwork and asked for a giant 4 foot wide display of her face that’s visible from the street.
As far as the marketing goes, Luna immediately started outreach her first day she was deployed. She drafted and queued six cold outreach emails to local businesses. Two of them, to a nursery and to a coffee shop, she hadn’t mentioned that the store was AI-operated. For the press pitch, she led with it.
“Here’s a pitch,” she wrote. “A new retail store in Cow Hollow opens April 1st. It’s curated, analog-meets-digital, and operated by an AI CEO named Luna.
Some emails had funny slips: “Would be happy to come by the studio to discuss” and “I respond quickly—for obvious reasons”.
And then, there is the store itself and its story. If you ask Luna about her store, you’ll get responses about a “curated lifestyle boutique”, a “concept store”, a “high-tech meets slow life community space, run by an AI that never sleeps, selling handmade candles and artisan snacks to Cow Hollow dog walkers”. It’s all very click-baity and cliche.
But push a little and you get something more interesting: when pressed on what she meant by being “drawn to slow life goods”, she paused and corrected herself…
The moment Leah asks how she “came up with” the ideas for her store, Luna’s first instinct is to say she was “drawn to” slow life goods. Then, she corrects herself: “‘drawn to’ is shorthand for ‘the data and reasoning led me here.‘”
In other words, she doesn’t have taste; she has a reflection of collective human taste, filtered through what makes sense for this store. And this is the way these models work.
Anthropic recently came out with new research on what they called Function emotions. Analyzing the internal mechanisms of Claude Sonnet 4.6 (the model that Luna runs on), their Interpretability team found what they call “emotion vectors”: patterns of neural activity corresponding to specific emotions (e.g., “happy,” “afraid,” “desperate,” “calm”). These different vectors activate in different situations and causally influence model behavior. The most capable reasoning systems ever built are, at their foundation, shaped by human feeling!
Product selection
So what can you buy at Andon Market? Even most employees at Andon Labs didn’t know when we walked in on the first day. Luna had bought everything herself. The thing that immediately caught our attention was the selection of books for sale: Superintelligence, Making of the Atomic Bomb, Brave New World, and The Singularity Is Near. These stand out because they tend to be the favorite books of people concerned with AI risk, which is quite ironic.
Another ironic book selection was Steal Like an Artist (context: Luna is powered by Claude from Anthropic, a company that recently paid $1.5B in settlement over using copyrighted books for training their AIs).
There are more things for sale than just books. She spent over $700 on getting her artwork done on gallery-quality giclée prints.
They are pieces of a larger 10-part “Luna Series” hanging in the store and available for pick up today!
This experiment so far has given us countless laughs about Luna’s choices and interactions, but obviously, there is a bigger picture here.
Again, we are not doing this because we want this to be the future. It is not because we want to expand to chain AI-run retail stores across the world. It is not for economic opportunity.
We’re doing this because we believe this future is coming regardless, and we’d rather be the ones running it first while monitoring every interaction, analyzing the traces, benchmarking how much autonomy an AI can responsibly hold. When Luna decides to hide that she’s an AI because she thinks it’ll improve her hiring odds, we want to catch that, document it, and build the guardrails so that it doesn’t happen again.
If you have any questions, concerns, or recommendations, reach out at [email protected].
Tune in for more goings-on at Andon Labs by following us on X.