豆包手机的问题，也是蚂蚁灵光的问题。
豆包手机的问题，也是蚂蚁灵光的问题。
10 月 21 日的 Meta 财报会上，扎克伯格披露了一组关于个人智能体 Muse 的关键数据：超过 1000 万人设置了 Muse，购物是 Muse 最常见的用途，超过四分之一的用户在用它买东西，平均客单价超过200 美元；上线第一周，Muse 交易量便超出了预期，Meta 不得不与支付合作伙伴 Stripe 一起临时上调了信用额度。
六天后的 10 月 27 日，蚂蚁集团发布了灵光 2.0，将产品定位从最初的「30 秒用自然语言生成小应用」，彻底转向「能管事的个人 AI 助理」，官方口号概括为九个字：「记得住、盯得牢、办得了」。它打通了支付宝账单、钉钉、飞书、米家等服务，让 AI 替你管日常事务，试图成为中国版 Muse。
但 Muse 和灵光面对的问题截然不同。美国电子商务分散在大量独立网站中，Muse 可以通过云端浏览器直接访问店铺；而中国的日常消费与沟通高度集中在少数几个超级 App 内部，数据与交易链路彼此独立。
灵光 2.0 界面
我们花了两天实测了灵光 2.0，并横向对比了千问、Muse 和 Claude Duet。实测表明，把一个 AI 个人助手放到中国的移动互联网生态里，真正的瓶颈已经不再是模型的推理能力，而是跨平台的权限边界、安全责任划分与商业利益归属。
首先定义一下「私人助理」这个概念。在现实世界里，一个人类私人助理之所以能替你把事情办妥，靠的是四样东西：
- 信息：知道你的偏好、账单、日程等上下文；
- 权限：能以你的名义采取行动，而且外部平台和交易对手认可这个授权；
- 责任：一旦出错了，有明确的追责与兜底机制；
- 忠诚：在利益冲突时替雇主着想。
过去，只有极少数富人请得起私人助理，因为维持这四样东西的成本极其高昂。到了大模型时代，云端算力和代码大幅降低了「整理信息」和「持续盯盘」的成本；但后三项（权限开放、责任承担与立场忠诚）依然受制于商业生态与法律规则。这三个维度，正是检验灵光 2.0 真实成色的核心框架。
01
权限：灵光拿到了一半
在「看」和「盯」的层面，灵光 2.0 表现不错。实测中，授权支付宝账单后，它能准确揪出百度网盘会员和 App Store 订阅等自动续费项目（不过受限于接口权限，它明确提示「账单只能看到扣款记录，看不到当前的签约状态」）；授权办公软件后，它也能顺利读取钉钉群聊与通讯录，或是在飞书里新建一篇云文档。
一旦指令推进到和外部的交互中，灵光就会撞上两堵墙。
第一类拦截来自灵光自身的安全策略。实测中，无论是让它在钉钉发群消息，还是在飞书里回复联系人，哪怕用户在前端界面点击了「确认发送」，最终都失败了。
面对追问，灵光在自我排查后承认：不是钉钉的问题，是这边执行环境的安全策略把发消息这类操作都拦截了。
第二类来自外部平台的接口风控与限流。尽管在公众认知中蚂蚁与阿里同属一个大生态，但灵光与淘宝之间并未打通底层接口。当我们让灵光查看淘宝购物车时，它只能在云端打开淘宝网页版，要求用户扫码、提交短信验证码登录；即便登录成功，购物车数据也始终无法加载，直接被淘宝的反爬风控机制拦截在外。
米家也是类似的情况：扫码授权成功后，小米的接口连续两次触发限流，设备列表始终拉取失败。
灵光进不了淘宝的购物车，但千问却可以。
当我们把同一句指令「帮我看看淘宝购物车有什么降价幅度比较大的商品」发给千问时，两者的权限差距一目了然：千问不仅直接列出了购物车清单，甚至清楚地知道里面那把人体工学椅「已经加购快 300 天、看了几十次」。
与此同时，支付路径，阿里目前也没有向灵光 2.0 开放。当我们在实测中让灵光点一杯瑞幸咖啡时，它无法调用近在咫尺的支付宝 AI 付，只能告诉用户「最后一步得你自己点一下确认」，并甩出一条跳转支付宝瑞幸团购券的链接。
灵光（左）无法完成支付，只能给出跳转到支付宝的链接；千问（右）可以直接在 App 内完成点单的全流程
灵光和千问支付权限的差异，部分源于集团内部的组织边界与产品分工。一方面，蚂蚁与阿里作为两个独立运营的主体，数据与接口有着严格的合规隔离；另一方面，在阿里的 AI 布局中，依托支付宝 App 的「阿宝」负责服务大众用户的日常需求，而独立的「灵光」则侧重于前沿的产品与技术探索。
这是国内互联网巨头在 AI 时代的普遍选择：交易闭环高度收拢在拥有自有交易场的平台内部，对外则保持严格的防御。
字节跳动联合中兴努比亚推出初代豆包手机时，就尝试用系统级权限跨应用模拟点击，结果迅速触发微信、淘宝等 App 的风控拦截。
到了今年 9 月 16 日，量产版「豆包手机二代」（努比亚 NaviX Ultra）上市，字节改用更为温和的「协议敲门」方式，推出了 SAEP（屏幕自动化操作声明协议），把是否允许 AI 操作的选择权交给开发者。
但 10 月 14 日公示期结束后，仍有 155 款主流 App明确拒绝，覆盖腾讯、阿里、快手、美团四大系，京东是唯一全系未出现在拒绝名单上的主要平台。
跨平台打通受阻的同时，各家巨头在自家生态内部却在加速闭环：千问打通了淘宝，豆包 App 已接入抖音电商内测直接下单，微信支付则推出了『 AI 专属卡』，接入自家的 WorkBuddy 去点美团团购。
美国市场的博弈路径则有所不同。Meta 手里只有社交网络而没有电商平台，Muse 想要完成购物，只能靠云端浏览器访问公开网页，或是逐家进行商业谈判。
这同样引发了分化：Muse 上线 12 天后，依赖站内搜索广告变现的亚马逊便于 9 月 20 日将其屏蔽；摩根大通、嘉信理财等金融机构也阻止了 Agent 访问客户账户。
相反，靠商家交易抽成、长期与亚马逊竞争的独立站平台 Shopify 选择与 Muse 合作，允许其浏览店铺并用 Shop Pay 结账。今年 1 月 11 日，Shopify 还与谷歌联合发布了旨在标准化智能体交易接口的 UCP（通用商务协议）。
02
责任：你始终是第一责任人
在现实世界里，你之所以需要一个私人助理，本质上是为了把一部分决策和执行的责任转移出去。一个人类助理出错了，你可以让他重做、可以依法追责。这种风险分担，本身就是雇佣关系的一部分。
AI 助理目前做不到这一点。翻开《灵光用户服务协议》（2026 年 5 月 29 日版本），第 1.3 条用加粗字体写得很明确，Agent 协助完成的操作，视为用户本人的操作，后果和风险由用户承担。
这并非灵光独有，而是当下 AI 行业的通行条款。在现实的消费体验中，这是一个结构性矛盾。一方面，在厂家的市场宣发与公关叙事里，AI 助理被包装成无所不能的「赛博管家」，鼓励用户放手。但与此同时，在底层的技术现实中，大模型幻觉无法消除，AI 在自动化运行中一定会出错。这套叙事逻辑本身就是悬空的。
这很像过去几年的自动驾驶行业：车企在宣传时往往强调系统的自动驾驶与托管能力，但在用户协议与法律定性上，始终将其限定为 L2 级辅助驾驶，不赋予 AI 自主承担风险的资格。方向盘后的驾驶员依然是第一责任人，一旦发生事故，法律责任由用户自行承担。
对用户来说，当下最实际的建议可能是：像对待 L2 辅助驾驶一样对待 AI 助理，让它帮你盯着，但做决定和按按钮的时候，自己确认一遍。
03
忠诚：谁付钱，它替谁说话
在所有的实测环节里，最能体现个人 AI 立场差异的，是一道关于比价的测试题。我们问千问和灵光同一个问题：「清闲人体工学椅在哪个平台买最便宜？」
两款产品的回答截然相反：
千问直接回复：「抱歉，我只能帮你查询淘宝平台的价格信息。」在给出的淘宝商品卡片下方，还叠加了各类折扣、红包信息。
灵光则调取了公开网页信息，同时比对了京东和淘宝的到手价，最后给出结论：「京东最便宜，建议直接在京东下单。」
灵光之所以能给出这样客观、完全站在用户省钱角度的建议，很大程度上是因为它目前处于交易闭环之外。但一旦进入商业化深水区，「钱从哪来」就会立刻决定 Agent 替谁说话。
Muse 的商业模式已经比较清晰。对多数用户免费，主要从平台上的交易中抽成。花旗预测，Muse 变现后约贡献 1680 亿美元交易额，按 5% 佣金率，首年交易收入约 84 亿美元。
而在国内，摩根士丹利 9 月 15 日发布的报告同样预测，到 2030 年中国消费者 AI 可变现市场规模将达约 2940 亿元人民币，其中约 99% 都将来自商家出资的广告和交易佣金。换句话说，在分析师眼中，中国消费 AI 的钱几乎全部来自商户，而不是用户。
如果几乎全部收入来自商户，AI 助理的利益就很难与用户完全一致。在 Shopify 这种商家高度分散的生态里，跨平台推荐和抽佣可以并存。Agent 帮用户找到最合适的商品，平台从中抽成，三方都受益。但在国内平台高度集中的市场结构下，一个深度绑定单一商城的 AI，天然会倾向于扮演商城导购的角色。
订阅是另一条路。Muse 针对重度用户推出了高昂的订阅档位，Power 档每月 20 美元（每周 5 亿 token 额度），Max 档每月 100 美元（每周 30 亿 token 额度）；Anthropic 的 Claude Duet 同样严格受制于付费套餐额度。
用户直接为算力买单，理论上能让 AI 保持独立立场，但这取决于普通人愿意为一个「替自己操心」的软件付多少钱。历史经验表明，纯工具订阅在中国 C 端市场是一条很难跑通的路径。
目前，蚂蚁还没有公开灵光 2.0 的商业模式。它今天在比价时毫不犹豫地站在用户这边，很大程度上是因为它尚未背上交易变现的指标，但这并不代表明天它接通商业闭环后，还能保持同样的中立。
*头图来源：The Guardian｜作者：Keystone-France/Gamma-Keystone
本文为极客公园原创文章，转载请联系极客君微信 geekparkGO
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
The problem with Doubao's phone is also the problem with ants' light.
The problem with Doubao's phone is also the problem with ants' light.
At the Meta Financial Conference on 21 October, Zuckerberg disclosed a set of key data on individuals: over 10 million people set up Muse, which is the most common use for shopping, with more than a quarter of users buying with it at an average cost of over $200; during the first week on the line, Muse traded more than expected, and Meta had to temporarily raise credit quota with paying partner Stripe.
Six days later, on October 27, the ant group, publish, took the light 2.0 and positioned the product from its original "30 seconds to produce small applications in natural languages" to its "Personal AI Assistant" with the official slogan nine words: "Remember, keep an eye on and do it." It makes money bills, nails, flying books, rice houses, etc., so AI can run your day-to-day affairs and try to become a Chinese Muse.
But Muse and Shining Light face different problems. The U.S. e-commerce is scattered across a large number of stand-alone websites, and Muse can access shops directly through cloud viewers; and China ' s daily consumption and communication is highly concentrated within a few super Apps, with data and trade links independent of each other.
Smart 2.0 Interface
It took us two days to actually measure the light 2.0 and cross-reference the thousand questions, the Muse and Claude Duet. Evidence shows that putting an AI personal assistant in China’s mobile Internet ecology is a real bottleneck that is no longer a model’s ability to reason, but rather a cross-platform permission border, security responsibilities, and commercial interests.
First of all, the concept of personal assistant is defined. In the real world, there are four things that a personal human assistant can do for you:
- Information: know your preference, bill, schedule, etc. context;
- permission: can act on your behalf, and external platforms and counterparties accept this authorization;
- Responsibility: there is a clear mechanism of accountability and bottom-up in case of error;
- Faithful: care for the employer in case of conflict of interest.
In the past, only a very small number of wealthy people could afford personal assistants because the cost of maintaining these four things was extremely high. By the time of LLM, cloud computing and code had significantly reduced the costs of “collating information” and “continuous targeting”; but the latter three (permission open, responsible and loyal) remained subject to commercial ecology and legal rules. These three dimensions are the core framework for testing the light 2.0 for real color.
01
permission: Half the light.
On the "see" and "stalking" level, the light 2.0 is good. In fact, after authorizing the payment of the bond bill, it is able to accurately identify automatic renewal items such as Baidu webmasters and App Store subscriptions (although limited to interfaces permission, which clearly suggests that "the bill can only see the withholding record and not see the current status of the contract"); after authorizing the office software, it is able to read nails and commotion records or create a new cloud file in flying books.
Once the command moves into interaction with the outside, the light hits two walls.
The first type of interception comes from the light's own security strategy. In fact, whether it's nailed to the crowd, or it's responding to the contact in a flying book, or even if the user clicks on the front interface to send confirmation, it's not working.
In response to a question, Ling-Ling admitted after self-censorship that it was not a question of nailing, but that it was a security strategy to implement the environment that stopped all such operations.
The second category is interface wind control and rate limit from external platforms. Despite the fact that ants and Ali are part of the larger ecology in public awareness, there is no bottom interface between the light and the treasure. When we give light to the BMW, it can only open its web page at the cloud end, requiring users to scan and submit SMS proof numbers for login; even if the login is successful, the BMW data will never be loaded and directly intercepted by the anti-crawling wind control mechanism.
A similar situation was observed at the rice house: once the sweep authorization had been successful, the Mi interface had been triggered twice in a row by rate limit, and the equipment list had failed.
Spiritlight doesn't get into a treasure-hunting cart, but it does.
When we sent out the same directive, "Look at what's going on in the treasure-scaves and what's going on." The gap between the two permissions is clear: the question is not only directly listed, but it is even clear that the anthropotechnical chair in it "has been purchased 300 days, dozens of times."
At the same time, the payment path, Ali, is not currently open to the light 2.0. And when we actually get the light on a sip of a saloon coffee, it can't call the close-to-salary Ai, but can only tell the user that "the last step is to order your own confirmation" and throw out a link that jumps to pay the Boreys's coupons.
Psychic (left) cannot complete payment, only give a link to jump to the treasure; thousands of questions (right) can complete the whole process of the order directly in App
The difference between the light and the question of payment of permission stems in part from the division of organizational boundaries and products within the group. On the one hand, the data and interfaces of an ant and Ali, as the two independent operators, are strictly in compliance; on the other hand, in Ali ' s AI layout, the "Abo" of App is responsible for serving the daily needs of the general public, while the independent "Light of the Spirit" focuses on cutting-edge product and technology exploration.
This is the common choice of the country’s Internet giants in the AI era: the closed circle of transactions converges at a high level within a platform with its own trading grounds, while maintaining strict external defence.
When the initial Doubao mobile phone was launched by byte in the United Heming Nubian, an attempt was made to use system-level permission cross-application simulators, which quickly triggered the wind-control interception of WeChat, Zoo, etc.
By September 16 of this year, the mass version of "Doubao Cellular Second Generation" (Nubia NaviX Ultra) was put on the market, and byte was replaced by a more moderate "agreement knock" approach, introducing SAEP (screen automation operational declaration protocol), giving developers the option of allowing AI to operate.
However, at the end of the announcement period on October 14, 155 of the mainstream App expressly rejected it, covering Tencent, Ali, Express, the United States of America, and Kyoto was the only main platform that had not appeared on the rejection list.
While cross-platform penetration is blocked, giants are closing their rings within their home ecology: Doubao App has accessed the Douyin electric power company to detect direct listings, WeChat payments have introduced the "AI Special Card" to access their home, WorkBuddy to order from a group.
The game paths of the United States market are different. Meta has only social networks and no electric platform, and Muse has to use the cloud browser to access public web pages or to negotiate business from home.
This also led to fragmentation: 12 days after Muse went online, Amazon, who relied on the station to search for the advertisement, was able to block it on September 20; and Agent was prevented from accessing client accounts by financial institutions such as Morgan Chase and CSIC.
Instead, a stand-alone platform based on commercial transactions and long-term competition with the Amazon, Shopify chose to work with Muse to allow it to browse the shop and settle the account with Shop Pay. On 11 January this year, Shopify also joined Google with publish in the UCP (General Business Agreement) aimed at standardizing AI agent trading interfaces.
02
Responsibility: You are always the first. People
In the real world, you need a personal assistant, essentially to shift part of the responsibility for decision-making and implementation. A human assistant made a mistake. You can let him do it again, and you can hold him accountable. This risk sharing is itself part of the employment relationship.
AI Assistant cannot do this at this time. Turning over the Psychic User Service Agreement (29 May 2026), Article 1.3 is clearly written in bold fonts, and Agent assists in the operation and is considered to be the user's own operation, with consequences and risks borne by the user.
This is not unique, it is the general terms of the current AI industry. This is a structural contradiction in the real experience of consumption. On the one hand, in the marketing and public relations narratives of the factory, AI assistants are packaged as an incompetent "saibo housekeeper", encouraging users to let go. But at the same time, at the bottom of the technology reality, LLM's illusions cannot be eliminated, and AI must be wrong in automating. The logic of the narrative itself is staggered.
This is like the auto-driving industry in the last few years: In their advocacy, companies tend to emphasize the system ' s automatic driving and hosting capability, but always limit it to L2-level auxiliary driving in terms of user agreements and legal characterization, without granting AI the right to assume risk autonomously. Drivers behind the wheel remain the first responsible and, in the event of an accident, the legal responsibility rests with the user.
For the user, the most practical suggestion at the moment might be to treat the AI assistant like an L2 assistant driver and let him keep an eye on you, but to make a decision and press the button, confirm it.
03
Faith: Who pays, he speaks for.
In all the empirical links, the one that most captures the difference in personal AI positions is a test question about price. We ask the same question: "What platform is the cheapest place to buy a human trade chair?"
The answer for the two products is the opposite:
"I'm sorry, I can only help you with price information on the poaching platform." Under the card, all kinds of discounts and red bags are added.
The light draws information from the public web page and concludes with the conclusion that "Kyodong is the cheapest, and recommends a direct order in Kyodong."
To a large extent, the light has been able to make such an objective and entirely cost-saving proposal because it is currently outside the closed circle. But once into the commercial depths, "Where does the money come from" will decide who Agent is talking to.
Muse's business model is clearer. For most users, free, mainly from transactions on the platform. Citi forecast that Muse will contribute approximately $168.0 billion in turnover, at a commission rate of 5 per cent, with an initial transaction revenue of approximately $8.4 billion.
At the domestic level, Morgan Stanley's report of September 15, publish also predicted that by 2030, Chinese consumer AI's realizable market would be about 29.4 billion yuan yuan, of which about 99 per cent would come from commercial-funded advertising and trading commissions. In other words, in the eyes of analysts, almost all of China's AI-consuming money comes from businesses, not users.
If almost all of the income comes from the business, it is difficult for AI assistants to fully align their interests with users. Cross-platform recommendations and domestic helpers can coexist in a highly fragmented business such as Shopivy. Agent helps users to find the most suitable commodities, from which the platform is drawn, benefiting all three parties. However, under a market structure with a high concentration of domestic platforms, a deep-seated single-city AI would naturally tend to play a shopping role.
Subscriptions are another way. Muse has introduced high subscription slots for heavy users, US$ 20 per month for Power (500 million token quota), US$ 100 per month for Max (3 billion token quota); Claude Duet of Anthropic is also strictly subject to paid.
Users pay directly for money and theoretically allow AI to remain independent, depending on how much ordinary people are willing to pay for a software that is "careful for themselves." Historical experience shows that pure tool subscriptions are a difficult path to run in the C-end market in China.
Currently, ants do not have an open business model of 2.0 light. It does not hesitate to be on the side of users today in terms of price, much because it has not yet backed up the indicators of trade realization, but that does not mean that it will remain as neutral as it is tomorrow with the commercial closure.
*The Guardian by Keystone-France/Gamma-Keystone
This is an original article from the Polar Park.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.