Apple 发布 Apple Intelligence

2024 年 6 月 10 日，Apple Intelligence 在 WWDC 上出现时，没有先要求用户打开一个新的聊天 App。它藏进了已经存在的动作里：在 Mail 里重写一封信，在 Notes 里总结一段文字，在锁屏上压缩一串通知，或对 Siri 说出一个涉及邮件、日历和照片的问题。生成式 AI 不再只是一处目的地，而开始成为操作系统替许多旧功能增加的一层能力。

一次请求的旅程由此变得比界面更重要。苹果称，许多任务会在设备上处理；更复杂的请求可以交给运行在 Apple 芯片服务器上的 Private Cloud Compute。公司承诺服务器不保留或暴露用户数据，并让独立研究者能够检查公开记录的软件映像。对用户而言，点击“总结”只是一瞬间；对系统而言，它必须先决定任务能否留在本机，以及需要送出多少上下文才能完成。

这种分流利用了苹果独有的位置。iPhone、iPad 和 Mac 的系统知道用户正在看的内容，也掌握应用权限与硬件加速。Writing Tools 因此可以出现在 Mail、Notes、Pages 以及采用接口的第三方应用中；通知摘要可以在锁屏发生；Siri 被承诺逐步获得屏幕理解、跨应用动作和个人上下文。模型本身未必比所有云端旗舰更大，但系统可以把它放在用户原本就要完成的动作旁边。

边界同样写进发布内容。苹果还把 ChatGPT 接入 Siri 与系统写作工具；当 Siri 认为外部模型有帮助时，用户会在问题、文档或照片发送前被询问。这个确认框承认了一件事实：端侧模型、苹果私有云和第三方服务不是同一条隐私链。所谓“系统级 AI”并不意味着一切都在手机里完成，而是操作系统承担了路由、授权和呈现结果的责任。

发布会展示的功能也不是 6 月 10 日当天同时抵达所有设备、语言和地区。首批能力依赖较新的 Apple 芯片，并随 iOS 18、iPadOS 18 与 macOS Sequoia 分阶段进入测试；更深入的 Siri 能力则被放在后续路线图中。宣布、测试和真正稳定可用之间的距离，后来成为 Apple Intelligence 体验不可忽略的一部分。

苹果的入场没有结束聊天机器人竞争，它改变了消费者遇见模型的位置。以后评估一项手机 AI，不只要问回答是否聪明，还要问它在哪个应用里出现、读到了什么、请求被送到哪里，以及用户是否能看见这次分流。Apple Intelligence 最重要的界面可能不是那圈发光的 Siri 动画，而是一次普通请求背后那条不再普通的数据路径。

When Apple Intelligence appeared at WWDC on 10 June 2024, it did not begin by asking users to open a new chat application. It occupied actions they already performed: rewriting an email in Mail, summarizing text in Notes, compressing a stack of notifications on the Lock Screen, or asking Siri a question involving messages, calendars, and photos. Generative AI was no longer only a destination. It became an additional system layer beneath many familiar features.

That made the journey of a request more important than the visible interface. Apple said many tasks would run on the device, while more complex work could move to Private Cloud Compute on servers built with Apple silicon. The company said those servers would not retain or expose user data, and that independent experts could inspect the publicly logged software images allowed to run. To the user, pressing Summarize took a moment. The system first had to decide whether the task could stay local and how much context needed to leave the device.

Apple could make this distinction consequential because it controlled the operating system, application permissions, and hardware acceleration. Writing Tools could appear across Mail, Notes, Pages, and third-party apps that adopted the interfaces. Notification summaries could happen on the Lock Screen. Siri was promised, over time, screen awareness, actions across apps, and access to personal context. The models did not need to be larger than every cloud flagship if the system could place them directly beside the task a person was already trying to complete.

The boundaries were also part of the announcement. Apple integrated ChatGPT into Siri and systemwide Writing Tools. When Siri judged that the outside model might help, the user would be asked before a question, document, or photo was sent. That confirmation dialog acknowledged that an on-device model, Apple’s private cloud, and a third-party service were not the same privacy path. “System-level AI” did not mean everything happened inside the phone. It meant the operating system assumed responsibility for routing, permission, and presentation.

Nor did the features all arrive on 10 June across every device, language, and region. The first capabilities depended on newer Apple silicon and were scheduled to enter testing with iOS 18, iPadOS 18, and macOS Sequoia. Deeper Siri abilities remained on a longer roadmap. The distance between announcement, beta, and stable daily use became part of the Apple Intelligence story rather than a minor footnote to it.

Apple’s entry did not end competition among chatbots. It changed where consumers would encounter models. Evaluating phone AI would now require more than asking whether an answer sounded intelligent. Users and reviewers also had to ask which app invoked it, what context it could read, where the request traveled, and whether the routing was visible. The most important Apple Intelligence interface may not have been Siri’s glowing border. It was the newly consequential data path behind an ordinary button.