文 | 字母AI
Muse上线约10天,登顶美国App Store免费应用榜,超过了ChatGPT。

几乎同一时间,另一条个人Agent路线上的代表机型Manus,也传出新一轮融资消息,计划融资5亿美元,估值约40亿美元。
Agent正在给前沿AI大厂给出一条“弯道超车”的路径。基础模型虽然关键,但到Agent的赛道上,机型外观、工具调用、任务执行、权限管理和使用者入口也加入了战场。后来者未必要先在模型榜单上赢过OpenAI和Anthropic,也有机会先在应用层拿到结果。
Meta显然已经走上了这条“捷径”。
过去一年,它在模型竞赛里追得并不从容。钱没少砸,人没少挖,算力没少堆,但一直有些雷声大雨点小,能让普通使用者感知到的成绩比较有限。
直到Muse,Meta终于有了一个真正属于自己的AI爆款。
Muse凭什么成为爆款?
Muse刚刚发布的时候,单看能力,实际上算不上有多特别。
它能浏览网页、调用工具、处理文件,以及替使用者订酒店、发邮件、购物。这些能力ChatGPT、Claude、乃至此前的Manus都已经展示过了。
按理来讲,这些能力很容易碰到隐私和权限的边界,使用者很难不担心Agent的安全风险。
让它替你整理日程,你就要开放日历;让它替你读邮件,它就要进入你的邮箱;让它帮你购物,它不仅要知道地址和偏好,还需要接触账号和支付信息。
提示词注入也是个需要小心的事情。Agent浏览网页、读取邮件和文件时,会不断接触来自外界的信息。如果有人在网页里藏下一段恶意指令,诱导模型把使用者资料发出去,一个具备较高权限的Agent可能会把攻击一路传递到真实账户。
Meta在做Muse时承认,不管模型训练得多好,Agent都会犯错,也可能被攻击。
它的特殊之处在于,它并没有把安全完全押在模型足够听话上,而是在模型之外,又搭了一整套的权限平台。

简单来说,Muse在云端给每个使用者配了一台独立电脑。
你可以把它理解成一台永远开着,并且只属于你的远程电脑(虚拟机),里面有浏览器,也能存文件、运行程序。Muse替你查资料、处理文件、跑任务,基本上都在这台电脑里完成。
你的文件、网页登录状态、邮箱和日历连接,并不会和其他使用者混在一起。即使Muse需要长期替你工作,也是在这台独立的电脑里活动,不至于拿着权限在整个平台里到处跑。
值得注意的是,住在这台电脑里的Muse,并不是这台电脑真正的“管理员”。
Meta把整个平台拆成了两个相互隔离的安全区域,Muse住在其中一个区域里,而另一个区域则放着真正敏感的东西——涵盖使用者的登录凭证、安全服务,以及一个名为Sentinel(哨兵)的独立安全Agent。

这个Sentinel专门审批网络访问和第三方服务,Muse可以提出发邮件或者调用账户的请求,但真正放行的是Sentinel。它可以检查Muse准备访问哪个网站、哪个地址,使用什么请求方式,以及准备把什么内容发送出去。
密码和支付信息也是类似的逻辑。登录凭证被保存在独立的安全存储里,Muse可以使用,但看不到具体内容;发送邮件、购买商品这类敏感操作还会再次向使用者确认。使用者也可以单独规定某个连接器的权限,例如邮箱是只允许读取,还是连发送也一起开放。
支付环节又做了一层隔离,Muse接入Stripe的Link后,可以生成一次性支付卡,让真正的银行卡信息不直接暴露给Agent或商家。
层层保障,限制了Muse可能造成的破坏。
使用者愿意信任它——或者说,愿意相信它就算出了难题也翻不出什么风浪——它才有机会真正进入普通人的生活。
当然,只有安全肯定也不行。
Muse背后的模型是Meta自研的Muse Spark。Alexandr Wang后来透露,从Muse Spark 1到最新的1.3,这整个模型系列从一开始就在为Muse这个个人Agent做准备,每一轮升级都在重点强化Agent和多模态能力。

模型之外,Muse背后还有Meta自己的整套机型生态。
它能接入Instagram、Facebook等平台,也可以直接在WhatsApp里使用。
Meta官方给出的案例是,如果你准备请朋友来家里吃饭,Muse可以翻你在Instagram里收藏过的菜谱短视频,帮你挑菜、列购物清单,还能记住朋友有没有忌口,最后再帮你把邀请发出去。
Axios特别提到,Muse能快速走红,靠的不只是模型本身,还涵盖Meta一直很擅长的几件事——把复杂工艺做成普通人能用的机型,把机型迅速推到大量使用者面前,以及理解使用者到底喜欢什么、需要什么。
Muse上线大约10天后,就冲上了美国App Store免费iPhone应用榜第一,超过了ChatGPT。
Meta过去一年那场声势浩大的AI重建,也终于拿出了一个普通使用者也能直接感知到的成绩。
不再恋战SOTA,Meta瞄准个人Agent
过去一年,Meta在前沿AI公司里的处境一直有点尴尬。
钱是没少花的,人是没少挖的,算力更是越堆越大。但放到OpenAI、Anthropic和谷歌旁边看,Meta始终差点意思——前两家就不说了,谷歌至少还有云服务和TPU生态撑起自己的AI版图。
相比之下,Meta迟迟没有拿出一个足够有说服力的大结果。
2025年,Llama 4令人失望,原本被寄予厚望的Behemoth也一再延期。曾经靠开源掀起一轮浪潮的Meta,在SOTA模型竞争上渐渐掉队。
于是,Meta开始了它的重建工作。
2025年6月,它花143亿美元买下Scale AI 49%的股份,把Alexandr Wang直接挖来负责新的AI体系;随后成立Meta Superintelligence Labs(后文简称MSL),不惜开出上亿美元级别的天价待遇,从OpenAI、Anthropic、Google DeepMind大规模挖人,AI组织半年内多次重组。
算力也一起往上堆,Meta 2025年资本开支达到722亿美元,2026年的指引进一步提高到1300亿至1450亿美元,并建设了Prometheus、Hyperion等吉瓦级AI数据中心。
到了今年4月,Muse Spark发布。Meta自己也说,这是MSL“从头重建AI工艺栈”9个月后的第一份模型答卷。
笔者不妨把Manus看成Meta转向Agent的一次关键启发。
去年12月,Meta一度直接收购Manus,而后者最出名的地方,就在于把模型、浏览器、工具和执行环境拼成了一个能自己完成复杂任务的通用Agent。后来这笔超过20亿美元的交易被监管部门要求拆解,但Meta愿意花这么大价钱买Manus,已经可以看出,它至少觉得这条路是值得尝试的。
Meta没能留下Manus,但把Manus代表的那套Agent逻辑留了下来。
巧的是,Manus自己后来也开始往个人Agent走。2月,它把机型接进Telegram,官方文档里直接写道:“你的个人Agent,随时随地”。

Agent是任务的入口,理想条件下,使用者只需要给出需要,Agent就能完成任务——使用者不需要去想中间用了谁家的模型、接了谁家的工具。
Meta并没有放弃训练模型,但它不再等着模型先拿第一,再考虑机型怎么做。模型开始变成底座,Agent则成了更直接的突破口。
Muse Spark这一系列模型,瞄准的正是个人Agent;而Muse这个机型,证明了这条路确实可以拿出格外亮眼的成绩。
这条路线并不偏门,甚至充满了机会。
Muse之外,一个叫Instinct的个人Agent甚至还没有正式全面上线,就已经被资本追着下注。
它当下仍处于邀请测试阶段,可以接入邮箱、消息、日历和位置,替使用者购物、订餐厅、取消订阅,最近又开始测试独立邮箱和代打电话等能力。8月底,它刚以约25亿美元估值完成融资;9月,据《The Information》报道,公司正在讨论再融资10亿美元,估值直接冲到了约100亿美元。

“挑选大于努力”
如果再把视角拉远一点,会发现过去一起追赶大模型的几家老牌科技巨头,现在实际上已经开始分头走了。
谷歌是少数还在坚持全栈打法的公司:模型要争,Agent要做,云要卖,芯片也自己造。
早些时间谷歌还传出可能要退出顶尖模型竞争的传闻,但就在最近几天,Arena里还浮现了一个挂着“gemini-3.8-flash”名字、能力显著异常的模型,被社区广泛猜测是尚未发布的Gemini 4 Pro。另一边,Flash系列几乎三周一更,谷歌9月又连续更新了Gemini 3.8 Live和3.8 Live Extended Thinking。
今年I/O上,谷歌直接宣布进入“Search Agent时代”,让Agent可以24小时在后台替使用者追踪信息;Google Cloud也推出了Gemini Enterprise Agent Platform。再往底层,还有已经做到第八代的TPU、Google Cloud和整套AI基础设施。
微软和亚马逊的挑选就不太一样了,他们正在把自己的长处集中到企业服务和AI基础设施上。
尽管微软还在做MAI系列,亚马逊也有Nova,但模型已经不再是这两家公司最关键的目标。
微软现在更强调多模型平台。Azure AI Foundry里同时给出OpenAI、Anthropic、xAI和微软自己的MAI模型,企业可以按质量、成本和任务自由挑选。微软最新财报披露,今年以来,同时使用多家模型的客户数量增长了5倍;Foundry和Agent 365则负责把这些模型真正接进企业数据、权限和Agent体系。

比起把自己的模型做到最强,微软更想确保,无论最后谁家的模型更好,企业都在Azure上用。
亚马逊的转向则更加显著。
今年7月,亚马逊直接裁掉了AGI团队的部分岗位。此前,这支原本相对独立的AGI组织已经被并入一个更大的工艺部门,和芯片、量子计算放到一起;涵盖Rohit Prasad、David Luan在内的多名核心负责人也先后离开。亚马逊给出的解释是,要精简优先级,把资源集中到最能给客户创造价值的领域。

与此同时,AWS越来越强调Bedrock和AgentCore。Bedrock直接接入OpenAI模型、Codex和托管Agent;AgentCore甚至明确做成“模型无关”,开发者可以自由换模型,AWS负责运行环境、权限、安全和监控。
它的思路和微软类似,与其花大工夫追赶SOTA,不如确保别人做出来的模型和Agent最后都跑在AWS上。
Meta走的是第三条路。
尽管Meta也在建设巨型数据中心,今年甚至传出计划把部分AI算力对外出售,但和谷歌、微软和亚马逊相比,Meta并没有一个已经成熟的公有云业务可以直接承接AI需要。
Meta现成的长处在另一端,在几十亿普通使用者、社交关系、内容建议、广告平台,以及Instagram、Facebook、WhatsApp和Threads这些消费级入口上。
在SOTA模型的比赛里,这些东西几乎帮不上忙;可一旦比赛来到个人Agent,情况就完全不一样了。
一个真正好用的个人Agent,不仅需要一个足够智能的模型,还需要知道你喜欢什么、和谁来往、最近准备做什么;要能进入你每天使用的软件,还要有办法长期待在你身边。
这时候,Meta过去积累的那些“老本行”,突然都变成了AI能力的一部分。
社交关系变成上下文,内容建议变成对使用者兴趣的理解,几十亿使用者变成分发渠道,WhatsApp和眼镜又可以继续变成入口。
Muse真正关键的地方就在这里。至少从Muse开始,Meta终于不再只是靠“有钱、有卡、有人”证明自己还坐在前沿AI的牌桌上。
它有了一个真正能摆上桌的结果,而且这个结果,不是靠复制OpenAI最擅长的方式拿到的。
Meta终于找到了一块更合适自己长板的战场。
事实上,国内的三家大厂也看上了个人Agent这个方向。腾讯、字节、阿里各有不同的先天长处,并且最近都有类似的动作浮现。
和更偏工作流程的办公Agent不同,个人Agent真正要接管的,是一个人的日常生活。
在这个战场上,腾讯最强的是关系和服务。微信里有人际关系、聊天、支付、小程序、公众号、视频号和大量线下服务,一个个人Agent如果真要替使用者处理生活事务,微信几乎天然就是操作台。最近正在灰度的微信助手“小微”,已经开始调用微信原生能力和小程序替使用者办事。
字节最强的是内容、建议和消费端AI入口。抖音知道使用者喜欢什么,豆包已经有大规模使用者,飞书掌握工作上下文;最新的豆包手机助手又进一步进入平台层,开始直接操作手机和跨应用完成任务。
阿里的长处则是交易和商业服务。淘宝、天猫、支付宝、饿了么、高德、飞猪、钉钉,再加上阿里云,本身就覆盖购物、支付、本地生活、出行和工作。千问已经开始把这些服务串起来,让AI直接完成搜索、下单、支付和出行规划;蚂蚁阿福则从健康情形切入,尝试长期管理使用者的个人健康信息。
在这个方向上,最后谁能做出最大的个人Agent可能没那么关键。只要几家大厂都开始争着替使用者把事情办完,最终省事的还是笔者。
本文来源:https://www.tmtpost.com/8148240.html
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
By Alias AI
Ten days after launch, Muse topped the US App Store free apps chart, surpassing ChatGPT.
At nearly the same time, another flagship personal Agent model named Manus also announced a new round of financing, planning to raise $500 million at an approximate valuation of $4 billion.
Agents are offering frontier AI giants a path to overtake on a curve. While foundational models are crucial, the Agent track also includes model appearance, tool invocation, task execution, permission management, and user access points as battlegrounds. Latecomers don't necessarily need to beat OpenAI and Anthropic on model leaderboards first; they can also aim for results at the application layer.
Meta has clearly stepped onto this "shortcut."
Over the past year, it hasn't chased从容ly in the model competition. It spent heavily, recruited many people, and stacked compute, but the results were often more thunder than rain, with limited achievements perceptible to ordinary users.
Until Muse, Meta finally has a genuine AI hit that truly belongs to it.
What makes Muse a hit?
When Muse was first released, its capabilities were not particularly remarkable on their own.
It can browse the web, invoke tools, process files, and book hotels, send emails, and shop on behalf of users. These capabilities have already been demonstrated by ChatGPT, Claude, and even previous Manus models.
Theoretically, these capabilities easily approach privacy and permission boundaries, making users difficult not to worry about Agent security risks.
To let it organize your schedule, you must open your calendar; to let it read emails, it must access your inbox; to let it help you shop, it needs to know your address and preferences and must access accounts and payment information.
Prompt injection is also something to be cautious about. When agents browse web pages and read emails and files, they constantly encounter external information. If someone hides malicious instructions in a webpage to lure the model into leaking user data, a high-privilege Agent could pass the attack through to a real account.
Meta acknowledged when building Muse that no matter how well the model is trained, Agents will make mistakes and may be attacked.
Its uniqueness lies in not placing security entirely on the model being sufficiently obedient, but instead building a complete permission platform outside the model.
Simply put, Muse provides every user with a standalone computer in the cloud.
You can think of it as a remote computer (virtual machine) that is always on and belongs solely to you, containing a browser and capable of storing files and running programs. When Muse looks up information, processes files, or runs tasks for you, it basically completes everything within this computer.
Your files, web login states, email, and calendar connections are not mixed with other users. Even when Muse works for you long-term, it operates within this standalone computer rather than roaming across the entire platform with permissions.
Notably, the Muse living in this computer is not the true "administrator" of this machine.
Meta split the entire platform into two mutually isolated security zones. Muse resides in one zone, while the other zone holds truly sensitive items—covering user login credentials, security services, and an independent security Agent named Sentinel.
This Sentinel specifically approves network access and third-party services. Muse can request to send emails or invoke accounts, but Sentinel is the one granting permission. It can check which website Muse is preparing to visit, which address, what request method is used, and what content is prepared to be sent.
Passwords and payment information follow a similar logic. Login credentials are stored in independent secure storage; Muse can use them but cannot see the specific content. Sensitive operations such as sending emails or purchasing items require再次 confirmation from the user. Users can also separately define connector permissions—for example, whether email access is read-only or also includes sending.
The payment环节 adds another layer of isolation. After Muse integrates Stripe's Link, it can generate one-time payment cards, preventing real bank card information from being directly exposed to Agents or merchants.
Layer upon layer of safeguards limits the potential damage Muse could cause.
Only when users are willing to trust it—or at least believe it won't cause significant trouble even if it encounters difficult problems—can it have the opportunity to truly enter ordinary people's lives.
Of course, security alone is definitely not enough.
The model behind Muse is Meta's self-developed Muse Spark. Alexandr Wang later revealed that from Muse Spark 1 to the latest 1.3, this entire model series was prepared for Muse as a personal Agent from the very beginning, with each upgrade focusing on strengthening Agent and multimodal capabilities.
Beyond the model, Muse is backed by Meta's own complete ecosystem of products.
It can integrate with platforms like Instagram and Facebook, and can also be used directly within WhatsApp.
Meta's official example: if you're preparing to host friends for dinner at home, Muse can browse recipe short videos you've bookmarked on Instagram, help you select dishes, create a shopping list, remember any dietary restrictions your friends have, and finally send out the invitations.
Axios specifically noted that Muse's rapid rise to popularity relies not just on the model itself, but also on several things Meta has always excelled at—turning complex technology into usable products for ordinary people, rapidly pushing products to massive user bases, and understanding what users actually like and need.
Approximately 10 days after launch, Muse surged to number one on the US App Store free iPhone apps chart, surpassing ChatGPT.
Meta's grand AI reconstruction over the past year has finally produced a result that ordinary users can directly perceive.
No longer competing for SOTA, Meta targets Personal Agents
Over the past year, Meta's position among frontier AI companies has been somewhat awkward.
Money was spent freely, people were recruited generously, and compute kept growing. But viewed alongside OpenAI, Anthropic, and Google, Meta always seemed slightly lacking—not to mention the first two; Google at least has cloud services and the TPU ecosystem supporting its AI landscape.
By comparison, Meta has yet to deliver a sufficiently compelling major breakthrough.
In 2025, Llama 4 was disappointing, and the highly anticipated Behemoth faced repeated delays. Meta, which once sparked a wave through open source, is gradually falling behind in the SOTA model competition.
Thus, Meta began its reconstruction work.
In June 2025, it spent $14.3 billion to acquire a 49% stake in Scale AI, poaching Alexandr Wang to lead the new AI organization; it then established Meta Superintelligence Labs (hereinafter MSL), offering compensation packages worth tens of millions of dollars to大规模 recruit from OpenAI, Anthropic, and Google DeepMind, restructuring its AI organization multiple times within half a year.
Compute investment also continued to climb. Meta's 2025 capital expenditure reached $72.2 billion, with 2026 guidance further raised to $130–145 billion, and it built gigawatt-scale AI data centers such as Prometheus and Hyperion.
By April this year, Muse Spark was released. Meta itself stated this was the first model deliverable after nine months of "rebuilding the AI technology stack from scratch" at MSL.
To me, Manus represents a critical source of inspiration for Meta’s shift toward agentic computing.
Last December, Meta once directly attempted to acquire Manus, which was most famous for拼ing together models, browsers, tools, and execution environments into a general Agent capable of independently completing complex tasks. Later, this deal exceeding $2 billion was required by regulators to be dismantled, but Meta's willingness to spend such a large sum on Manus already indicated it at least believed this path was worth attempting.
Meta failed to keep Manus, but retained the Agent logic that Manus represented.
Coincidentally, Manus itself later began pivoting toward personal Agents. In February, it integrated its product into Telegram, with official documentation stating: "Your personal Agent, anytime, anywhere."
Agents serve as task entry points. Under ideal conditions, users only need to provide their needs, and the Agent completes the task—users don't need to think about which company's model or tools are used in the interim.
Meta has not abandoned training models, but it no longer waits for its model to rank first before considering product development. Models are becoming the foundation, while Agents have become the more direct breakthrough point.
The Muse Spark series of models specifically targets personal Agents; and the Muse product has proven this path can indeed deliver exceptionally outstanding results.
This route is not niche; it is full of opportunities.
Beyond Muse, a personal Agent called Instinct has not yet officially launched widely but has already attracted venture capital backing.
It is currently in an invite-only testing phase, capable of integrating with email, messaging, calendar, and location, helping users shop, book restaurants, and cancel subscriptions. It has recently begun testing independent email and代 calling capabilities. In late August, it completed financing at approximately a $2.5 billion valuation; in September, according to The Information, the company is discussing an additional $1 billion raise, pushing its valuation to approximately $10 billion.
"Choice matters more than effort"
If we broaden the perspective further, we discover that several established tech giants that once追赶大模型 together have actually started walking separate paths.
Google is one of the few companies still adhering to a full-stack approach: competing in models, building Agents, selling cloud services, and manufacturing its own chips.
Earlier, rumors emerged that Google might withdraw from top-tier model competition, but just in the past few days, a model named "gemini-3.8-flash" with significantly anomalous capabilities appeared in Arena, widely speculated by the community to be the unreleased Gemini 4 Pro. Meanwhile, the Flash series updates roughly every three weeks; in September, Google continuously updated Gemini 3.8 Live and 3.8 Live Extended Thinking.
At this year's I/O, Google directly announced entering the "Search Agent era," enabling Agents to track information for users 24/7 in the background; Google Cloud also launched the Gemini Enterprise Agent Platform. Going deeper, there are eighth-generation TPUs, Google Cloud, and the entire AI infrastructure.
Microsoft and Amazon's choices differ—they are concentrating their strengths in enterprise services and AI infrastructure.
Although Microsoft continues developing its MAI series and Amazon has Nova, models are no longer the most critical goals for these two companies.
Microsoft now emphasizes多模型 platforms. Azure AI Foundry simultaneously offers OpenAI, Anthropic, xAI, and Microsoft's own MAI models, allowing enterprises to freely choose based on quality, cost, and tasks. Microsoft's latest earnings report disclosed that since the beginning of this year, the number of customers using multiple models has grown fivefold; Foundry and Agent 365 are responsible for integrating these models into enterprise data, permissions, and Agent systems.
Rather than making its own model the strongest, Microsoft wants to ensure that regardless of which company's model ultimately performs best, enterprises will use Azure.
Amazon's pivot is even more pronounced.
In July this year, Amazon directly eliminated部分 positions within its AGI team. Previously, this relatively independent AGI organization had already been merged into a larger technology division, grouped with chip and quantum computing; multiple core leaders including Rohit Prasad and David Luan have successively departed. Amazon's explanation was to streamline priorities and concentrate resources in areas that create the most value for customers.
Meanwhile, AWS increasingly emphasizes Bedrock and AgentCore. Bedrock directly integrates OpenAI models, Codex, and hosted Agents; AgentCore is explicitly designed to be "model-agnostic," allowing developers to freely switch models while AWS handles the runtime environment, permissions, security, and monitoring.
Its thinking is similar to Microsoft's: rather than expending massive effort chasing SOTA, it's better to ensure that models and Agents built by others ultimately run on AWS.
Meta is pursuing a third path.
Although Meta is also constructing massive data centers and this year even rumored plans to sell部分 AI compute externally, compared to Google, Microsoft, and Amazon, Meta lacks a mature public cloud business that can directly accommodate AI needs.
Meta's existing strengths lie at the other end—in billions of ordinary users, social relationships, content recommendations, advertising platforms, and consumer-grade entry points like Instagram, Facebook, WhatsApp, and Threads.
In the SOTA model competition, these assets barely help; but once the competition shifts to personal Agents, the situation changes completely.
A truly useful personal Agent requires not only a sufficiently intelligent model but also knowledge of your preferences, your social connections, and your recent plans; it needs to access the software you use daily and have a way to stay by your side long-term.
At this point, those "old competencies" Meta accumulated suddenly all become part of AI capabilities.
Social relationships become context, content recommendations become understanding of user interests, billions of users become distribution channels, and WhatsApp and glasses can continue serving as entry points.
This is precisely Muse's critical significance. At least starting with Muse, Meta is no longer merely proving it still sits at the frontier AI table by relying on "having money, cards, and people."
It has a genuine result to place on the table, and this result was not achieved by replicating OpenAI's most擅长的 approach.
Meta has finally found a battlefield better suited to its strengths.
In fact, China's three major tech giants have also set their sights on the personal Agent direction. Tencent, ByteDance, and Alibaba each possess different innate advantages, and similar moves have recently emerged across all three.
Unlike office Agents focused more on workflows, personal Agents aim to接管 a person's entire daily life.
In this arena, Tencent's greatest strength lies in relationships and services. WeChat contains interpersonal relationships, messaging, payments, mini-programs, public accounts, Channels, and extensive offline services; a personal Agent truly handling daily affairs would find WeChat naturally serves as its operational console. The WeChat assistant "Xiao Wei," currently in gray release, has already begun invoking native WeChat capabilities and mini-programs to handle tasks for users.
ByteDance's greatest strength lies in content, recommendations, and consumer-side AI entry points. Douyin knows what users like, Doubao already has massive user scale, and Feishu possesses work context; the latest Doubao phone assistant further enters the platform layer, beginning to directly操控 the phone and complete跨-app tasks.
Alibaba's strengths lie in commerce and commercial services. Taobao, Tmall, Alipay, Ele.me, Amap, Fliggy, DingTalk, plus Alibaba Cloud, already cover shopping, payments, local services, travel, and work. Qianwen has begun串ing these services together, enabling AI to directly complete searches, orders, payments, and travel planning; Ant Financial's Afu切入 from health scenarios, attempting long-term management of users' personal health information.
In this direction, who ultimately produces the most impactful personal Agent may not be the critical question. As long as the major tech giants all begin competing to handle tasks for users, the ultimate beneficiaries saving effort will be ordinary people like myself.