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
AI生态的军阀混战1
视频
音频
原始脚本
AI 生态的军阀混战，从一款 AI 手机的权限困境说起。 自己的地盘自己做主，自己的孩子自己管，这句朴素的商业逻辑，正在 AI 时代上演一场轰轰烈烈的生态歌剧战。 巨头们纷纷用自家 AI 模型，如阿里千问、腾讯元宝、字节豆包，打造专属 Agent，全面接管自有生态内的 APP 服务。 对内开放核心 权限，打通闭环。 对外则严控限流，严防外部 Agent 的入侵，形成了各自为战、互不兼容的军阀割据格局。 而这一切的起点，源于一款名为努比亚 M153的工程样机，以及一个绕不开的技术死结，Inject Events 权限。 一，权限困局，超级 Agent 的断头路。 字节与中兴合作的努比亚 M153工程机，搭载豆包 AI 助手，试图打造超级 Agent，通过模拟用户点击、跳转等操作，跨 APP 整合服务。 但这个设想从一开始就撞上了铜墙铁壁，Android 的系统的 Inject events 权限是实现这一切的核心钥匙。 这把钥匙有多难拿？从技术到合规，层层都是死锁。 技术层面，该权限属于系统最高级别的安全权限，手机厂商为防恶意攻击，通常会锁死通道。 即便厂商深度合作开放，也会留下严格的操作日志和风控限制。 合规层面，美国 CFAA、欧盟 DSA、中国网络安全法等均明确，未经授权的自动化访问属违规操作。 一旦出现账号被盗、资金损失，责任认定根本扯不清。 巨头封杀，微信、支付宝、银行 APP 等早已布下天罗地网，通过检测系统权限列表、操作行为轨迹，一旦识别到 Injekt Events 权限开启，直接闪退、封号甚至限制下载。 应用市场更是可以直接标记机型为高风险，切断核心 APP 的安装通道。 本质上，巨头们从根上否定了外部超级 Agent 通过模拟用户事件访问自家 APP 的可能性。 这条路被彻底封杀后，行业不得不转向另一条路，谁的地盘谁来做 Agent。 二、割据根源，巨头的自留地逻辑。 为什么巨头们对外部 agent 严防死守？核心是风险与利益的双重排它。 风险不共担，哪怕是合规授权的 AI 操作，一旦出现安全事故，用户只会追责微信不安全，支付宝有漏洞，而非手机厂商或 AI 开发者。 对巨头而言，封杀所有异常 常入口比证明谁该担责成本低得多。 利益不共享，阿里、腾讯、字节等早已布局专属大模型，阿里千问、腾讯元宝、字节豆包，自家的 AI 模型就是为了接管自家生态的服务链路。 微信的社交数据、淘宝的消费偏好、抖音的内容链路，都是巨头的核心资产。 就像自家的自留地，怎么可能让别人的 Agent 来耕耘收益？于是巨头们纷纷开启闭环建设，在自家生态内，AI Agent 拥有畅通无阻的权限。 腾讯元宝可自由调用微信、腾讯会议的核心功能。 阿里千问能深度对接淘宝支付宝的操作接口。 字节豆包可流畅操控抖音、高德地图。 但对外，他们却筑起高墙，只留少量跳转、分享等表层接口，核心权限绝不开放。 三、混战图景。 AI 时代的战国七雄，如今的 AI 生态早已是军阀割据的战国时代。 巨头们各占山头，合纵连横。 腾讯系元宝以微信为核心据点，掌控社交、办公、电商、京东生态。 元宝 Agent 可实现微信群消息整理、腾讯会议托管、京东商品推荐等闭环服务。 却对字节系 Agent 的微信操作权限全面封杀。 阿里系千问，盘踞电商、金融、本地生活、饿了么、飞猪领域。 千问 Agent 能完成淘宝同款搜索、支付宝账单分析、外卖预定等全流程操作，对外严控交易数据和商家运营权限。 字节系豆包，手握内容、抖音、出行、高德、办公、飞书、地盘。 豆包 Agent 可整合抖音美食推荐、高德路线规划、飞书行程同步，缺南粤微信支付宝的雷池一步。 其他军阀，百度文心一言绑定百度地图、百度搜索，华为盘古大模型深耕鸿蒙生态。 小创投公司则只能在巨头夹缝中做表层衔接，比如将抖音收藏同步到高德，却无法实现自动下单等深度操作。 这场混战中，合纵连横时有发生。 字节联合中兴手机，腾讯绑定手机厂商推生态服务。 阿里联动本地生活服务商，但始终遵循对内开放，对外限流的原则。 Agent 与 Agent 之间的协作仅停留在非核心功能的浅层试水，真正的核心数据和操作权限绝不外借。 四、未来走向，封闭与开放的平衡术，短期来看，军阀混战仍将持续，巨头们会继续扎紧闭环，守住自家生态的核心利益。 但长期而言，完全封闭绝非出路。 用户需要全场景便捷体验，没人愿意为了用微信换手机，为了订外卖换 Agent。 监管也会推动基础功能开放，避免垄断。 最终，AI 生态可能会走向核心闭环加基础开放的平衡。 巨头守住核心数据和高级权限，同时开放少量标准化接口，如信息同步跨 APP 跳转。 让用户在不同 agent 的间切换，而非被单一巨头绑定。 就像战国时代最终走向大一统前的割据与融合，AI 生态的军阀混战也是通往成熟生态的必经之路。 而这一切的起点，那个关于 Inject events 权限的困境，早已揭示了核心真相。 在 AI 时代，生态的控制权永远掌握在地盘主人手中。
修正脚本
AI 生态的军阀混战，从一款 AI 手机的权限困境说起。 自己的地盘自己做主，自己的孩子自己管，这句朴素的商业逻辑，正在 AI 时代上演一场轰轰烈烈的生态割据战。 巨头们纷纷用自家 AI 模型，如阿里千问、腾讯元宝、字节豆包，打造专属 Agent，全面接管自有生态内的 APP 服务。 对内开放核心权限，打通闭环。 对外则严控限流，严防外部 Agent 的入侵，形成了各自为战、互不兼容的军阀割据格局。 而这一切的起点，源于一款名为努比亚 M153的工程样机，以及一个绕不开的技术死结，Inject Events 权限。 一、权限困局，超级 Agent 的断头路。 字节与中兴合作的努比亚 M153工程机，搭载豆包 AI 助手，试图打造超级 Agent，通过模拟用户点击、跳转等操作，跨 APP 整合服务。 但这个设想从一开始就撞上了铜墙铁壁，Android 系统的 Inject events 权限是实现这一切的核心钥匙。 这把钥匙有多难拿？从技术到合规，层层都是死锁。 技术层面，该权限属于系统最高级别的安全权限，手机厂商为防恶意攻击，通常会锁死通道。 即便厂商深度合作开放，也会留下严格的操作日志和风控限制。 合规层面，美国 CFAA、欧盟 DSA、中国网络安全法等均明确，未经授权的自动化访问属违规操作。 一旦出现账号被盗、资金损失，责任认定根本扯不清。 巨头封杀，微信、支付宝、银行 APP 等早已布下天罗地网，通过检测系统权限列表、操作行为轨迹，一旦识别到 Inject Events 权限开启，直接闪退、封号甚至限制下载。 应用市场更是可以直接标记机型为高风险，切断核心 APP 的安装通道。 本质上，巨头们从根上否定了外部超级 Agent 通过模拟用户事件访问自家 APP 的可能性。 这条路被彻底封杀后，行业不得不转向另一条路，谁的地盘谁来做 Agent。 二、割据根源，巨头的自留地逻辑。 为什么巨头们对外部 agent 严防死守？核心是风险与利益的双重排他。 风险不共担，哪怕是合规授权的 AI 操作，一旦出现安全事故，用户只会追责微信不安全，支付宝有漏洞，而非手机厂商或 AI 开发者。 对巨头而言，封杀所有异常入口比证明谁该担责成本低得多。 利益不共享，阿里、腾讯、字节等早已布局专属大模型，阿里千问、腾讯元宝、字节豆包，自家的 AI 模型就是为了接管自家生态的服务链路。 微信的社交数据、淘宝的消费偏好、抖音的内容链路，都是巨头的核心资产。 就像自家的自留地，怎么可能让别人的 Agent 来耕耘收益？于是巨头们纷纷开启闭环建设，在自家生态内，AI Agent 拥有畅通无阻的权限。 腾讯元宝可自由调用微信、腾讯会议的核心功能。 阿里千问能深度对接淘宝支付宝的操作接口。 字节豆包可流畅操控抖音、高德地图。 但对外，他们却筑起高墙，只留少量跳转、分享等表层接口，核心权限绝不开放。 三、混战图景。 AI 时代的战国七雄，如今的 AI 生态早已是军阀割据的战国时代。 巨头们各占山头，合纵连横。 腾讯系元宝以微信为核心据点，掌控社交、办公、电商、京东生态。 元宝 Agent 可实现微信群消息整理、腾讯会议托管、京东商品推荐等闭环服务。 却对字节系 Agent 的微信操作权限全面封杀。 阿里系千问，盘踞电商、金融、本地生活、饿了么、飞猪领域。 千问 Agent 能完成淘宝同款搜索、支付宝账单分析、外卖预定等全流程操作，对外严控交易数据和商家运营权限。 字节系豆包，手握内容、抖音、出行、高德、办公、飞书地盘。 豆包 Agent 可整合抖音美食推荐、高德路线规划、飞书行程同步，却难越微信支付宝的雷池一步。 其他军阀，百度文心一言绑定百度地图、百度搜索，华为盘古大模型深耕鸿蒙生态。 小创投公司则只能在巨头夹缝中做表层衔接，比如将抖音收藏同步到高德，却无法实现自动下单等深度操作。 这场混战中，合纵连横时有发生。 字节联合中兴手机，腾讯绑定手机厂商推生态服务。 阿里联动本地生活服务商，但始终遵循对内开放、对外限流的原则。 Agent 与 Agent 之间的协作仅停留在非核心功能的浅层试水，真正的核心数据和操作权限绝不外借。 四、未来走向，封闭与开放的平衡术，短期来看，军阀混战仍将持续，巨头们会继续扎紧闭环，守住自家生态的核心利益。 但长期而言，完全封闭绝非出路。 用户需要全场景便捷体验，没人愿意为了用微信换手机，为了订外卖换 Agent。 监管也会推动基础功能开放，避免垄断。 最终，AI 生态可能会走向核心闭环加基础开放的平衡。 巨头守住核心数据和高级权限，同时开放少量标准化接口，如信息同步、跨 APP 跳转。 让用户在不同 agent 之间切换，而非被单一巨头绑定。 就像战国时代最终走向大一统前的割据与融合，AI 生态的军阀混战也是通往成熟生态的必经之路。 而这一切的起点，那个关于 Inject events 权限的困境，早已揭示了核心真相。 在 AI 时代，生态的控制权永远掌握在地盘主人手中。
英文翻译
The Warlord Turmoil of the AI Ecosystem: Starting with the Permission Dilemma of an AI Phone "Your own turf, your own rules; your own child, your own care." This straightforward business logic is now playing out as a full-scale ecological turf war in the AI era. Tech giants are deploying their own AI models—such as Alibaba's Qwen, Tencent's Yuanbao, and ByteDance's Doubao—to build exclusive agents that fully take over the app services within their own ecosystems. Internally, they open up core permissions to create closed loops. Externally, they impose strict rate limits and block external agents from intruding, forming a warlord-style structure where each faction operates independently and incompatibly. The origin of all this lies in an engineering prototype called the Nubia M153 and an unavoidable technical deadlock: the Inject Events permission. **I. The Permission Deadlock: The Dead End for Super Agents** The Nubia M153 engineering prototype, a collaboration between ByteDance and ZTE, comes equipped with the Doubao AI assistant and aims to build a super agent that integrates services across apps by simulating user clicks, jumps, and other operations. However, this vision hit a brick wall from the start. The key to making it all work is Android's Inject Events permission. How hard is it to obtain this key? From technology to compliance, every layer is a deadlock. On the technical side, this permission is at the highest system security level. To prevent malicious attacks, phone manufacturers typically lock down the channel. Even if a manufacturer opens it up through deep collaboration, strict operation logs and risk control restrictions remain. On the compliance side, the U.S. CFAA, the EU DSA, China's Cybersecurity Law, and others clearly state that unauthorized automated access is a violation. If account theft or financial loss occurs, liability becomes impossible to untangle. Tech giants have also set up a dragnet: WeChat, Alipay, bank apps, etc., detect system permission lists and operational behavior patterns. Once Inject Events permission is enabled, they immediately force a crash, block accounts, or even restrict downloads. App stores can even mark a device as high-risk and block the installation channel for core apps. Essentially, the giants have fundamentally denied the possibility of an external super agent accessing their apps by simulating user events. With this path completely blocked, the industry had to pivot to another approach: each territory handles its own agent. **II. The Root of the Turf War: The Logic of the Giants' Private Land** Why are the giants so defensive against external agents? The core reason is dual exclusivity in risk and profit. Risk is not shared. Even with compliant and authorized AI operations, if a security incident occurs, users will only blame WeChat for being unsafe or Alipay for having vulnerabilities—not the phone manufacturer or the AI developer. For the giants, blocking all abnormal entry points is far cheaper than proving who should be held responsible. Profit is not shared. Alibaba, Tencent, ByteDance, and others have already deployed their own large models—Qwen, Yuanbao, Doubao—designed specifically to take over the service chains within their own ecosystems. WeChat's social data, Taobao's consumption preferences, and Douyin's content chain are all core assets of these giants. Just like private farmland, why would they let someone else's agent come and harvest the benefits? So the giants have started building closed loops. Within their own ecosystems, AI agents have unimpeded permissions. Tencent's Yuanbao can freely invoke the core functions of WeChat and Tencent Meeting. Alibaba's Qwen can deeply connect to the operation interfaces of Taobao and Alipay. ByteDance's Doubao can smoothly control Douyin and Amap. But externally, they build high walls, leaving only a few superficial interfaces for jumps and sharing, while never opening up core permissions. **III. The Battlefield: The Seven Warring States of the AI Era** Today's AI ecosystem has already become a Warring States period of warlord turf wars. The giants each occupy their own mountain strongholds, forming alliances and counter-alliances. The Tencent camp, with Yuanbao at the center, uses WeChat as its stronghold, controlling social, office, e-commerce, and the JD.com ecosystem. Yuanbao's agent can handle WeChat group message sorting, Tencent Meeting hosting, and JD.com product recommendations in a closed loop. But it completely blocks ByteDance agents from accessing WeChat operations. The Alibaba camp, with Qwen, dominates e-commerce, finance, local services, Ele.me, and Fliggy. Qwen's agent can perform full-process operations such as searching for similar products on Taobao, analyzing Alipay bills, and ordering food delivery, while strictly controlling external access to transaction data and merchant operation permissions. The ByteDance camp, with Doubao, holds content, Douyin, transportation, Amap, office, and Feishu. Doubao's agent can integrate Douyin food recommendations, Amap route planning, and Feishu schedule syncing, but cannot cross the Rubicon of WeChat and Alipay. Other warlords: Baidu's ERNIE Bot is tied to Baidu Maps and Baidu Search; Huawei's Pangu large model is deeply rooted in the HarmonyOS ecosystem. Smaller startups can only make superficial connections in the gaps between giants—for example, syncing Douyin favorites to Amap—but cannot perform deep operations like automatic ordering. In this chaos, alliances and counter-alliances occur from time to time. ByteDance partners with ZTE phones; Tencent binds with phone manufacturers to push ecosystem services. Alibaba collaborates with local service providers, but they all adhere to the principle of opening internally and restricting externally. Collaboration between agents remains at a shallow, non-core functional level; real core data and operational permissions are never lent out. **IV. Future Direction: Balancing Closedness and Openness** In the short term, the warlord turmoil will continue. Giants will keep tightening their closed loops to protect their core ecosystem interests. But in the long run, complete closure is not a viable path. Users need a full-scenario, convenient experience. No one wants to switch phones just to use WeChat or change agents just to order food delivery. Regulators will also push for the opening of basic functions to prevent monopolies. Eventually, the AI ecosystem may move toward a balance of core closed loops plus basic openness. Giants will guard their core data and high-level permissions while opening a few standardized interfaces—such as information syncing and cross-app jumps. This would allow users to switch between different agents rather than being locked into a single giant. Just as the Warring States period eventually led to unification after division and integration, the warlord turmoil of the AI ecosystem is a necessary path to a mature ecosystem. And at the root of it all, the dilemma of Inject Events permission has already revealed the core truth: In the AI era, the control of the ecosystem always remains in the hands of the territory owner.
back to top