🔶 Meta 希望人们用“AI agents”处理日常事务：问题何在？
英文标题：🔶 Meta wants people to use 'AI agents' to do daily tasks. Why that's a problem
Meta 推出的个人 AI Agent“ Muse ”在 Mark Zuckerberg 展示后迅速登上应用排行榜，可代为购物、订票、发送邮件和联系客服。Meta 在 2025 年投入约 720 亿美元用于 AI 及其他资本开支，但其 Muse Spark 1.2 模型在主流排名中仍落后于 Anthropic 的 Claude 和 OpenAI 的 GPT。文章认为，Meta 的核心策略并非以最强模型争夺专业市场，而是借助其每日 36 亿用户，将 AI Agent 打造成新的日常使用习惯和在线服务入口。与此同时，Muse 对个人邮件、日历及消费服务的接入能力，以及 Meta 过往的隐私和成瘾设计争议，引发了数据滥用、过度依赖和监管滞后的担忧。
4.8 KB
379.1 KB
443.2 KB
632.2 KB
796 B
6.4 KB
383.6 KB
中文
Meta 希望用户在一天中的各个时段都求助于一款可爱的人工智能（AI）“Agent”，让它们代为完成常见任务，例如预订航班或购买日用品。
该公司的新 Agent——Muse——在首席执行官 Mark Zuckerberg 上周于一场会议上展示后，迅速登上应用排行榜榜首。Muse 被宣传为“无需学习成本”，可以代表用户在线订购杂货、致电客服、发送电子邮件、预约服务和规划旅行。
Agent 是一种利用类似聊天机器人的大语言模型和强化学习，以完成任务并实现用户指定目标的系统。
过去 12 个月中，它们已取得长足进展：从DIY 爱好者项目，发展到更为完善的个人助理应用。与此同时，“失控 Agent”还实施了一些备受关注的黑客攻击和系统入侵。Meta 进入这一领域，意味着 Agent 正开始走向大众市场。
巨额投入，位居第二梯队
Meta 一直试图在快速变化的 AI 世界中找到自己的位置。2025 年，该公司花费 143 亿美元引入了 Scale AI 创业者 Alexandr Wang，随后全面重建了内部 AI 系统。总体而言，公司在 2025 年的 AI 及其他资本支出达到 720 亿美元，并预计这一数字将在 2026 年翻倍。
成果尚可，但并未领先全球。根据一项热门排名，Meta 最新的 AI 模型 Muse Spark 1.2位居第三，落后于 Anthropic 的 Claude 和 OpenAI 的 GPT 前沿模型。
这或许并不重要。Anthropic 和 OpenAI 正瞄准愿意为最强编码、研究和分析模型付费的专业人士与企业。例如，OpenAI 最近推出的 Dots agents就面向希望获得商业任务协助的付费客户。
Meta 也计划提供一些专业级付费 AI 服务，但其核心业务另有所在。第二季度，每天有 36 亿用户使用 Facebook、Instagram、Messenger 或 WhatsApp 服务。
这些人不需要最聪明的模型，他们需要的是简单、易接近的模型。
Meta 的 Muse Spark 模型只需足够好，能够运行 Muse Agent 即可。该公司将 Muse 宣传为“为所有人打造”，包括那些缺乏技术使用经验的人。
Zuckerberg 表示，用户很快还可以为自己的 Muse Agent 创建一个可爱的拟人化虚拟形象，并与它进行实时视频聊天。他还发布了一款钥匙扣大小的设备，让用户无需手机即可与 Agent 对话。
从查看信息到委托办事
二十年来，Meta 的商业模式一直依赖于习惯。通知、无尽的信息流以及获得他人认可的诱惑，使“查看 Facebook”或“刷 Insta”成为一种条件反射。
Muse 是构建一种新条件反射的尝试。用户不再刷屏，而是把一项琐事交出去。
其设计鼓励用户产生依恋。用户可以为自己的 Agent 命名，并选择它的外观和沟通方式。
每一次成功交易都会进一步积累信任。设置项还允许用户调整 Agent 在行动前需要征求许可的频率。
通往一切的钥匙
目前，Muse 的触达范围有限。它目前仅在美国上线。
它设有“连接器”，可将用户的 Agent 与电子邮件、日历以及餐厅预订、音乐和购物等消费者服务连接起来。它无法登录用户的银行账户、医疗账户或报税账户。
不过，Meta 计划随着时间推移添加更多连接器，并将产品推向国际市场。
随着人们习惯使用 Agent，他们会交给它更多任务，并花更多时间与它相处。社交媒体曾从真正的社交网络演变为一种无处不在的工具，用以填补空闲时刻和调节情绪；同样，AI Agent 最终可能会管理支配现代生活的庞大界面网络。
Muse 也是 Meta 应对竞争性 AI Agent 对其既有地位构成威胁的一次尝试。若 Agent 普及，它们可能成为用户连接在线世界的主要界面——并比社交媒体平台收集更多有关用户及其偏好的信息；拥有 Agent 的公司随后便可能将这些信息出售给广告商。
一段令人怀疑的记录
Meta 表示，Muse 内置了隐私保护功能。每个 Muse Agent 都运行于各自专用的虚拟机中，称为 Muse Secure VM；该虚拟机同时存放Agent 和用户的数据。
Zuckerberg 还承诺提供可选的更高安全标准，在该标准下，即使 Meta 也无法访问用户 Agent中的内容。
Meta 在数据隐私方面的历史记录令人有理由怀疑。2018 年，剑桥分析丑闻披露，一家咨询公司为进行政治画像而收集了多达 8700 万 Facebook 用户的数据。2019 年，美国联邦贸易委员会因隐私失误对 Meta 处以50 亿美元罚款。
Meta 还被指控将产品设计为最大化人们——包括儿童——的使用时长。今年早些时候，该公司在不承认违法行为的情况下，支付了最高近 180 亿美元，以和解一宗诉讼。这宗诉讼由美国 29 个州的总检察长提起，指控 Instagram 和 Facebook 被设计为让青少年沉迷其中。
早期迹象
Muse 仅限 18 岁以上用户使用，但它以另一种形式提出了同一个问题：一款有名字、外形可爱且人们全天都会咨询的 Agent，可能比信息流培养出更强的使用习惯。
围绕 Muse 的担忧已经开始出现。不希望自己的交互内容被用于训练 Meta 模型的用户，必须自行找到数据控制选项并将其关闭。一名记者称，Muse 曾读取他存储在 Mac 上的私人消息，而他从未要求它这样做。
到目前为止，针对 AI Agent 的设计或运行方式，几乎没有相应的规则和监管。现在仍有时间制定这些规则与监管措施——但一旦数十亿人依赖一家公司的 Agent 来管理生活，我们塑造其行为方式的机会就将消失。
English / 原文
Meta wants users to turn to a cute artificial intelligence (AI) “agent” throughout the day to carry out common tasks for them, such as booking flights or doing their shopping.
The company’s new Muse agent shot to the top of the app charts after chief executive Mark Zuckerberg showcased it at a conference last week. Billed as having “no learning curve”, Muse can order groceries online, call customer service, send emails, book appointments and plan travel on a user’s behalf.
Agents are systems which use chatbot-like large language models and reinforcement learning to complete tasks and achieve user-designated goals.
They have come a long way in the past 12 months, from DIY hobbyist projects to more polished personal assistant apps, with “rogue agents” carrying out some high-profile hacks and system breaches along the way. Meta’s entry into the field means they are now starting to hit the mainstream.
Massive spending, second place
Meta has been trying to find its place in the fast-moving world of AI. In 2025 the company spent US$14.3 billion to bring in entrepreneur Alexandr Wang from Scale AI, then comprehensively rebuilt its internal AI systems. All told, the company spent $US72 billion on AI and other capital expenditure in 2025, and expects that figure to double in 2026.
The results have been respectable, but not world-beating. According to one popular ranking, Meta’s latest AI model, Muse Spark 1.2, comes in third behind the frontier Claude and GPT models from Anthropic and OpenAI.
That may not matter. Anthropic and OpenAI are chasing professionals and businesses willing to pay for the most capable model for coding, research and analysis. (OpenAI’s recently launched Dots agents, for example, target a paying clientele wanting help with business tasks.)
Meta also plans to offer some pro-grade paid AI services, but its bread and butter is something else. In the June quarter, 3.6 billion people used Facebook, Instagram, Messenger or WhatsApp every day.
Those people do not need the smartest model. They need one that is easy and approachable.
Meta’s Muse Spark model only needs to be good enough to run the Muse agent. The company is promoting Muse as “built for everyone”, including people who lack experience with technology.
Zuckerberg has said users will soon be able to create a cute anthropomorphic avatar for their Muse agent and video chat with it in real time, too. He also unveiled a keychain-sized device for talking to the agent without a phone.
From checking to delegating
For two decades, Meta’s business model has relied on habit. Notifications, endless feeds and the lure of other people’s approval turned “checking Facebook” or “checking Insta” into a reflex.
Muse is an attempt to build a new reflex. Instead of scrolling, you hand over a chore.
The design invites attachment. Users can give their agent a name and choose how it looks and communicates.
Every successful transaction builds a little more trust. The settings also let users change how often the agent needs to ask permission before acting.
The keys to everything
For now, Muse’s reach is limited. It is only available in the US so far.
It has “connectors” to link your agent to your email, calendars and consumer services such as restaurant bookings, music and shopping. It cannot log into your bank, or medical or tax accounts.
However, Meta plans to add more connectors over time — and roll the product out to international markets.
As people get used to using an agent, they will hand it more tasks and spend more time with it. In the same way social media evolved from actual social networking to an ever-present tool to fill idle moments and manage our moods, AI agents may end up managing the vast galaxy of interfaces that govern modern life.
Muse is also an attempt to head off the threat to Meta’s existing position posed by competing AI agents. If agents take off, they could become users’ main interface with the online world – and gather even more information about users and their preferences than social media platforms do, which whoever owns the agent could then sell to advertisers.
A record that invites scepticism
Meta says privacy is built in to Muse. Each Muse agent runs on its own dedicated virtual machine called Muse Secure VM. This houses both the agent and the person’s data.
Zuckerberg has also promised an optional higher security standard under which even Meta cannot access what is in a user’s agent.
Meta’s record on data privacy gives reason for doubt. In 2018, the Cambridge Analytica scandal revealed that a consulting firm had harvested data from up to 87 million Facebook users for political profiling. In 2019, the US Federal Trade Commission imposed a US$5 billion penalty on Meta over privacy failures.
Meta has also been accused of designing its products to maximise the amount of time people, including children, spend using them. Earlier this year, the company paid up to almost US$18 billion, without admitting wrongdoing, to settle a lawsuit from 29 US state attorneys-general who alleged Instagram and Facebook were engineered to keep teenagers hooked.
Early signs
Muse is restricted to over-18s, but raises the same question in a different form. A cute agent with a name that people consult throughout the day could build a stronger habit than a feed.
Concerns about Muse are already beginning to come up. Users who do not want their interactions used to train Meta’s models must find the data controls and switch that off themselves. One journalist reported Muse had read private messages stored on his Mac.
So far, there is little in the way of rules and regulations governing the design or operation of AI agents. There is still time for those rules and regulations to be created – but once billions of people are relying on one company’s agent to run their lives, the chance to shape how it behaves will have passed.