我的征尘是星辰大海。。。
The dirt and dust from my pilgrimage forms oceans of stars...
-------当记忆的篇章变得零碎,当追忆的图片变得模糊,我们只能求助于数字存储的永恒的回忆
作者:黄教授
手机视频列表
此曲只应天上有缘何飞落到人间
视频
音频
原始脚本
此曲只应天上有,缘何飞落到人间?豆包手机的 AI Agent 路径错在哪?此曲只应天上有人间难得几回闻。 用这句诗形容 AI Agent 的生态闭环建设恰如其分。 真正的 AI 跨应用协同本就该扎根云端的技术天庭,而非执着于手机端的人间捷径。 豆包手机试图通过客户端模拟用户操作实现 AI Agent 的超级功能。 从底层逻辑到技术实现全是本末倒置,看似短平快的路径,实则是一条被权限、数据、生态三重壁垒锁死的死胡同。 一、底层逻辑错配。 AI Agent 的云端基因与手机端的天生桎梏。 AI Agent 的核心价值是跨应用全链路协同,而这种协同从诞生起就自带云端基因。 它需要调用的是分散在各平台的核心数据、底层功能,而非手机屏幕上的像素级操作。 豆包手机的路径错就错在试图用客户端模拟替代云端协同,违背了 AI Agent 的本质设计原理。 一, Agent 的大脑本就不在手机端,豆包元宝、千问等 AI 大模型本质是云端的智能中枢,它们依赖海量数据训练、超强算力支撑才能实现跨应用的逻辑推理与功能调度,手机端根本无法承载如此庞大的模型体量与算力消耗。 所谓豆包手机的本地 Agent,不过是云端模型的遥控器。 而非真正的智能核心,谈何独立实现跨应用整合。 二、跨应用协同的核心是数据互通,而非界面模仿。 用户需要的 AI Agent 是能读懂微信聊天中的旅行需求,自动调用高德规划路线,用飞书同步形成的全流程服务。 而非在手机屏幕上模拟点击微信、复制地址、打开高德粘贴搜索的机械操作。 前者需要的是云端层面的数据打通,后者只是人工操作的替代品,两者的技术难度与价值天差地别。 二、技术死结。 手机端模拟操作的三重不可逾越壁垒。 豆包手机试图通过 Injected Events 全线实现模拟用户操作。 这条路从一开始就被堵死,操作系统、APP 巨头、数据架构共同织就了一张天罗地网,哪怕是巨头也无法突破。 一、权限壁垒,操作系统与 APP 的双重封杀系统及权限锁死。 Android、iOS 对第三方模拟操作的防范已到极致。 Inject events 作为系统最高级别的安全权限,仅对系统自带应用开放,且需厂商签名认证。 豆包手机即便与厂商合作拿到权限,也属于灰色地带,一旦其他 APP 举报或系统升级权限 随时可能被收回,整个 A 镇的功能将瞬间瘫痪。 APP 及风控拦截,微信、支付宝、银行 APP 等早已部署异常操作检测系统。 通过分析点击频率、滑动速度、设备指纹等数据,精准识别非人工操作。 此前豆包手机测试时,微信频繁闪退就是最好的证明。 哪怕是合规授权的模拟操作,也会被判定为安全风险,直接封号或限制功能。 二,数据壁垒。 手机端只是展示窗口,而非数据仓库。 用户的核心数据,微信聊天记录、淘宝订单、高德行程,从未真正存储在手机端,而是分散现在 个 APP 的云端服务器,手机端能看到的只是经过加密处理的界面数据,就像隔着一层毛玻璃看东西, Agent 既无法穿透这层玻璃获取原始数据,也无法实时同步云端的动态更新。 比如, Agent 想整合微信好友的聚餐地点,加美团订座。 在手机端只能看到好友发送的文字地址,无法直接调用美团的餐厅库存数据库。 想根据淘宝消费记录推荐饿了么外卖,也只能看到订单金额,看不到具体消费偏好。 这种数据阉割下的协同毫无价值可言。 三、生态壁垒。 APP 巨头的地盘防御战,微信、支付宝等巨头早已将手机端视为核心阵地,绝不允许外部 Agent 染指核心功能。 他们不仅会通过技术手段拦截模拟操作,还会在应用市场层面设置障碍,一旦检测到某款手机搭载的 Agent 试图侵入自家生态。 直接限制该机型的 APP 下载或功能使用。 对巨头而言,封杀比适配更简单,也更安全。 三,云端才是唯一解,AI Agent 的正确打开方式,当手机端的路径彻底走死。 巨头们的 AI 闭环建设早已锁定云端主导的路线。 这不仅是技术上的必然选择,更是生态壁垒的核心载体。 一、内部 API,跨应用整合的金钥匙。 巨头旗下的 APP,如微信、高德、京东,会向自家云端 Agent 开放私有内部 APP。 这些接口允许 Agent 直接调用核心数据与功能,无需经过手机端界面。 比如腾讯元宝 Agent 在云端通过内部 API 读取微信聊天中的地址,同步至高德的云端规划模块,再将路线通过微信推送至用户,全程避开手机端的权限限制,既安全又高效。 二,子 Agent 架构,生态内的神经网络。 巨头会 在生态内构建核心 Agent 加子 Agent 的树形结构。 微信有专属子 Agent 负责提取社交需求,高德有子 Agent 负责出行规划,京东有子 Agent 负责电商转化。 所有子 Agent 统一接入核心 Agent,如元宝、千问,通过私有通讯协议实现数据互通。 这种架构既能保证跨应用协同的流畅性,又能对外屏蔽接口,构建坚固的生态壁垒。 三,数据安全,云端管控的双重保障将 Agent 放在云端,既能实现数据不出生态的安全管控,又能避免手机端数据泄露的风险。 比如阿里千问 Agent 在云端整合淘宝、支付宝、饿了么的数据,为用户推荐个性化服务。 所有数据流转都在阿里的安全体系内,既符合合规要求,又能让用户放心。 四、结语,捷径不通,回归云端才是正途。 豆包手机的尝试,本质上是想走一条短平快的捷径。 用客户端模拟操作替代云端协同,用表面功能掩盖技术短板。 但 AI Agent 的发展从来没有捷径可走,它需要的是巨头们在云端层面的架构重构、API 开放、数据协同。 需要的是长期的技术积累与生态磨合,而非一款噱头大于实用的手机。 此曲只应天上有,真正的 AI Agent 注定要扎根云端的技术天庭,通过云端协同实现跨应用的无缝整合。 而试图将它拉到人间,用手机端的模拟操作勉强落地,只会陷入权限、数据、生态的三重困境,最终沦为无缘之水,无本之木。 对豆包而言,与其执着于打造一款 AI 手机,不如深耕云端 Agent 的协同能力,推动字节系生态的内部整合。 这才是 AI Agent 的正确发展方向。 毕竟用户需要的是能解决问题的智能助手,而非只能模拟点击的手机玩具。
修正脚本
此曲只应天上有,缘何飞落到人间?豆包手机的 AI Agent 路径错在哪?此曲只应天上有,人间难得几回闻。 用这句诗形容 AI Agent 的生态闭环建设恰如其分。 真正的 AI 跨应用协同本就该扎根云端的技术天庭,而非执着于手机端的人间捷径。 豆包手机试图通过客户端模拟用户操作实现 AI Agent 的超级功能。 从底层逻辑到技术实现全是本末倒置,看似短平快的路径,实则是一条被权限、数据、生态三重壁垒锁死的死胡同。 一、底层逻辑错配。 AI Agent 的云端基因与手机端的天生桎梏。 AI Agent 的核心价值是跨应用全链路协同,而这种协同从诞生起就自带云端基因。 它需要调用的是分散在各平台的核心数据、底层功能,而非手机屏幕上的像素级操作。 豆包手机的路径错就错在试图用客户端模拟替代云端协同,违背了 AI Agent 的本质设计原理。 一、Agent 的大脑本就不在手机端,豆包元宝、千问等 AI 大模型本质是云端的智能中枢,它们依赖海量数据训练、超强算力支撑才能实现跨应用的逻辑推理与功能调度,手机端根本无法承载如此庞大的模型体量与算力消耗。 所谓豆包手机的本地 Agent,不过是云端模型的遥控器。 而非真正的智能核心,谈何独立实现跨应用整合。 二、跨应用协同的核心是数据互通,而非界面模仿。 用户需要的 AI Agent 是能读懂微信聊天中的旅行需求,自动调用高德规划路线,用飞书同步行程的全流程服务。 而非在手机屏幕上模拟点击微信、复制地址、打开高德粘贴搜索的机械操作。 前者需要的是云端层面的数据打通,后者只是人工操作的替代品,两者的技术难度与价值天差地别。 二、技术死结。 手机端模拟操作的三重不可逾越壁垒。 豆包手机试图通过 Injected Events 全线实现模拟用户操作。 这条路从一开始就被堵死,操作系统、APP 巨头、数据架构共同织就了一张天罗地网,哪怕是巨头也无法突破。 一、权限壁垒,操作系统与 APP 的双重封杀,系统级权限锁死。 Android、iOS 对第三方模拟操作的防范已到极致。 Inject events 作为系统最高级别的安全权限,仅对系统自带应用开放,且需厂商签名认证。 豆包手机即便与厂商合作拿到权限,也属于灰色地带,一旦其他 APP 举报或系统升级,权限随时可能被收回,整个 Agent 的功能将瞬间瘫痪。 APP 及风控拦截,微信、支付宝、银行 APP 等早已部署异常操作检测系统。 通过分析点击频率、滑动速度、设备指纹等数据,精准识别非人工操作。 此前豆包手机测试时,微信频繁闪退就是最好的证明。 哪怕是合规授权的模拟操作,也会被判定为安全风险,直接封号或限制功能。 二、数据壁垒。 手机端只是展示窗口,而非数据仓库。 用户的核心数据,微信聊天记录、淘宝订单、高德行程,从未真正存储在手机端,而是分散在各个 APP 的云端服务器,手机端能看到的只是经过加密处理的界面数据,就像隔着一层毛玻璃看东西, Agent 既无法穿透这层玻璃获取原始数据,也无法实时同步云端的动态更新。 比如, Agent 想整合微信好友的聚餐地点,加美团订座。 在手机端只能看到好友发送的文字地址,无法直接调用美团的餐厅库存数据库。 想根据淘宝消费记录推荐饿了么外卖,也只能看到订单金额,看不到具体消费偏好。 这种数据阉割下的协同毫无价值可言。 三、生态壁垒。 APP 巨头的地盘防御战,微信、支付宝等巨头早已将手机端视为核心阵地,绝不允许外部 Agent 染指核心功能。 他们不仅会通过技术手段拦截模拟操作,还会在应用市场层面设置障碍,一旦检测到某款手机搭载的 Agent 试图侵入自家生态。 直接限制该机型的 APP 下载或功能使用。 对巨头而言,封杀比适配更简单,也更安全。 三、云端才是唯一解,AI Agent 的正确打开方式,当手机端的路径彻底走死。 巨头们的 AI 闭环建设早已锁定云端主导的路线。 这不仅是技术上的必然选择,更是生态壁垒的核心载体。 一、内部 API,跨应用整合的金钥匙。 巨头旗下的 APP,如微信、高德、京东,会向自家云端 Agent 开放私有内部 API。 这些接口允许 Agent 直接调用核心数据与功能,无需经过手机端界面。 比如腾讯元宝 Agent 在云端通过内部 API 读取微信聊天中的地址,同步至高德的云端规划模块,再将路线通过微信推送至用户,全程避开手机端的权限限制,既安全又高效。 二、子 Agent 架构,生态内的神经网络。 巨头会在生态内构建核心 Agent 加子 Agent 的树形结构。 微信有专属子 Agent 负责提取社交需求,高德有子 Agent 负责出行规划,京东有子 Agent 负责电商转化。 所有子 Agent 统一接入核心 Agent,如元宝、千问,通过私有通讯协议实现数据互通。 这种架构既能保证跨应用协同的流畅性,又能对外屏蔽接口,构建坚固的生态壁垒。 三、数据安全,云端管控的双重保障将 Agent 放在云端,既能实现数据不出生态的安全管控,又能避免手机端数据泄露的风险。 比如阿里千问 Agent 在云端整合淘宝、支付宝、饿了么的数据,为用户推荐个性化服务。 所有数据流转都在阿里的安全体系内,既符合合规要求,又能让用户放心。 四、结语,捷径不通,回归云端才是正途。 豆包手机的尝试,本质上是想走一条短平快的捷径。 用客户端模拟操作替代云端协同,用表面功能掩盖技术短板。 但 AI Agent 的发展从来没有捷径可走,它需要的是巨头们在云端层面的架构重构、API 开放、数据协同。 需要的是长期的技术积累与生态磨合,而非一款噱头大于实用的手机。 此曲只应天上有,真正的 AI Agent 注定要扎根云端的技术天庭,通过云端协同实现跨应用的无缝整合。 而试图将它拉到人间,用手机端的模拟操作勉强落地,只会陷入权限、数据、生态的三重困境,最终沦为无源之水,无本之木。 对豆包而言,与其执着于打造一款 AI 手机,不如深耕云端 Agent 的协同能力,推动字节系生态的内部整合。 这才是 AI Agent 的正确发展方向。 毕竟用户需要的是能解决问题的智能助手,而非只能模拟点击的手机玩具。
英文翻译
This melody should only exist in heaven—why has it descended to the mortal world? Where did the AI Agent path of Doubao Phone go wrong? This melody should only exist in heaven, rarely heard in the mortal world. This poem is an apt description of the ecological closed-loop construction of AI Agents. True cross-application AI collaboration should be rooted in the technological heavens of the cloud, rather than clinging to the shortcuts of the mortal world on the mobile end. Doubao Phone attempts to achieve the superpowers of an AI Agent by simulating user operations through the client. From the underlying logic to the technical implementation, this is entirely putting the cart before the horse. What seems like a quick and easy path is, in reality, a dead end blocked by triple barriers of permissions, data, and ecosystem. **1. Mismatched Underlying Logic: The Cloud-Based Nature of AI Agents vs. the Inherent Limitations of the Mobile End** The core value of an AI Agent lies in cross-application full-chain collaboration, and this collaboration has carried a cloud-based gene from its inception. It requires access to core data and underlying functions scattered across various platforms, not pixel-level operations on a mobile screen. Doubao Phone's mistake lies in trying to replace cloud-based collaboration with client-side simulation, violating the fundamental design principles of an AI Agent. **1. The Agent's Brain Is Not on the Mobile End** AI large models like Doubao Yuanbao, Qianwen, etc., are essentially intelligent hubs in the cloud. They rely on massive data training and supercomputing power to achieve cross-application logical reasoning and functional scheduling. The mobile end simply cannot support such massive model size and computational consumption. What is called a local Agent on Doubao Phone is merely a remote control for the cloud model, not a true intelligent core—how can it independently achieve cross-application integration? **2. The Core of Cross-Application Collaboration Is Data Interoperability, Not Interface Imitation** What users need from an AI Agent is a full-process service that can read travel needs from WeChat chats, automatically call Gaode to plan routes, and synchronize itineraries using Feishu. They do not need mechanical operations like simulating clicks on WeChat, copying addresses, opening Gaode, pasting, and searching on the mobile screen. The former requires data integration at the cloud level; the latter is just a substitute for manual operation. The technical difficulty and value of the two are vastly different. **2. Technical Dead End: The Triple Insurmountable Barriers of Mobile-End Simulated Operations** Doubao Phone attempts to achieve simulated user operations throughout via Injected Events. This path was blocked from the start. Operating systems, major APP companies, and data architectures have jointly woven an inescapable net. Even giants cannot break through. **1. Permission Barriers: Dual Blockade by Operating Systems and APPs** - **System-Level Permission Lockdown:** Android and iOS have taken extreme precautions against third-party simulated operations. Inject events, as the highest system-level security permission, are only open to system-built-in applications and require vendor signature verification. Even if Doubao Phone obtains permission through cooperation with manufacturers, it remains in a gray area. Once other APPs report it or the system upgrades, the permission may be revoked at any time, instantly paralyzing the entire Agent's functionality. - **APP and Risk Control Interception:** APPs like WeChat, Alipay, and banking apps have long deployed abnormal operation detection systems. By analyzing click frequency, swipe speed, device fingerprints, and other data, they accurately identify non-human operations. The frequent crashes of WeChat during Doubao Phone testing are the best evidence. Even compliantly authorized simulated operations can be deemed security risks, resulting in direct account bans or functional restrictions. **2. Data Barriers: The Mobile End Is Only a Display Window, Not a Data Warehouse** Users' core data—WeChat chat records, Taobao orders, Gaode itineraries—are never actually stored on the mobile end. Instead, they are scattered across the cloud servers of various APPs. What the mobile end can see is only encrypted interface data, like looking through frosted glass. The Agent can neither penetrate this glass to obtain raw data nor synchronize real-time cloud updates. For example, if the Agent wants to integrate a WeChat friend's dinner location and reserve a seat via Meituan, it can only see the text address sent by the friend on the mobile end, but cannot directly access Meituan's restaurant inventory database. If it wants to recommend Ele.me takeout based on Taobao purchase history, it can only see the order amount, not specific consumption preferences. This kind of collaboration under data castration is worthless. **3. Ecological Barriers: The Territory Defense War of Major APP Companies** Giants like WeChat and Alipay have long regarded the mobile end as their core territory and will never allow external Agents to meddle with core functions. They not only intercept simulated operations through technical means but also set obstacles at the app market level. Once they detect that an Agent on a certain phone model attempts to invade their ecosystem, they directly restrict APP downloads or functional usage for that model. For giants, blocking is simpler and safer than adapting. **3. The Cloud Is the Only Solution: The Correct Way to Open an AI Agent** When the mobile-end path is completely blocked, the AI closed-loop construction of giants has long locked onto a cloud-dominated route. This is not only a technical necessity but also the core carrier of ecological barriers. **1. Internal APIs: The Golden Key to Cross-Application Integration** Giants' internal APPs—such as WeChat, Gaode, JD.com—open private internal APIs to their own cloud Agents. These interfaces allow the Agent to directly call core data and functions without going through the mobile-end interface. For example, Tencent Yuanbao Agent reads addresses from WeChat chats via internal APIs in the cloud, synchronizes them to Gaode's cloud planning module, and then pushes the route back to the user through WeChat. This bypasses the mobile-end permission restrictions entirely, being both safe and efficient. **2. Sub-Agent Architecture: The Neural Network Within the Ecosystem** Giants will build a tree structure of a core Agent plus sub-Agents within their ecosystem. WeChat has a dedicated sub-Agent responsible for extracting social needs, Gaode has a sub-Agent for travel planning, and JD.com has a sub-Agent for e-commerce conversion. All sub-Agents are uniformly connected to the core Agent (e.g., Yuanbao, Qianwen), achieving data interoperability through private communication protocols. This architecture ensures smooth cross-application collaboration while shielding interfaces externally, building a solid ecological barrier. **3. Data Security: Dual Guarantees of Cloud-Based Management and Control** Placing the Agent in the cloud enables secure management and control of data without it leaving the ecosystem, while also avoiding the risk of data leakage from the mobile end. For example, Alibaba's Qianwen Agent integrates data from Taobao, Alipay, and Ele.me in the cloud to recommend personalized services to users. All data flows within Alibaba's security system, complying with regulations and giving users peace of mind. **4. Conclusion: Shortcuts Don't Work; Returning to the Cloud Is the Right Path** Doubao Phone's attempt is essentially to take a quick and easy shortcut—replacing cloud-based collaboration with client-side simulated operations, and masking technical shortcomings with surface-level features. But there has never been a shortcut in the development of AI Agents. What is needed is the architectural reconstruction, API openness, and data collaboration at the cloud level among giants. It requires long-term technical accumulation and ecological磨合, not a phone that is more hype than utility. This melody should only exist in heaven. True AI Agents are destined to be rooted in the technological heavens of the cloud, achieving seamless cross-application integration through cloud-based collaboration. Trying to drag them down to the mortal world, forcing them to land via simulated operations on the mobile end, will only trap them in the triple predicament of permissions, data, and ecosystem, ultimately becoming water without a source, a tree without roots. For Doubao, instead of obsessively creating an AI phone, it would be better to deepen the collaborative capabilities of cloud-based Agents and promote the internal integration of the ByteDance ecosystem. This is the correct development direction for AI Agents. After all, what users need is an intelligent assistant that can solve problems, not a phone toy that can only simulate clicks.
back to top