When OSes Learn to 'Think': AI Becomes the True Kernel

Yesterday afternoon in the lab, my senior pointed at that news article on the screen and asked me: "Do you get it? Why are hardware giants like Apple, Huawei, and OPPO suddenly crowding into the AI filing list?" Staring at those familiar names—Apple Intelligence, Huawei Xiaoyi, OPPO AndesGPT—my first reaction was: Isn't this just a phone assistant? What's so big about it?

My senior smiled: "You still don't get it. They aren't doing AI; they are redefining the operating system."

I flipped through the papers, read the news several times, and slowly began to taste the meaning. It turns out, the operating system is shifting from a "tool for managing hardware" to an "intelligent agent that understands users." This isn't a gradual upgrade, but a switch in underlying logic.

Traditional OS vs. New AI OS: Two Completely Different Worldviews

Let's go back to the basic comparison. What is the core of traditional operating systems (iOS, Android)? Resource scheduling. CPU, memory, network, sensors—all code is doing one thing: making hardware run efficiently and letting Apps run. Users issue commands by clicking icons and swiping screens; the system acts like an obedient butler, doing whatever you say.

The core of the new AI operating system (what Apple, Huawei, and OPPO are doing) has become understanding intent. It no longer waits for you to explicitly say "open WeChat," but predicts what you want through voice, images, and behavioral trajectories. Even before you realize it, it has prepared things for you.

[!info] A most intuitive example

Traditional Siri: You shout "Hey Siri, set an alarm," and it executes.

Apple Apple Intelligence: You wake up in the morning, and your phone has already recommended the best wake-up time based on your sleep data and schedule, asking if you want to set an alarm.

The latter no longer requires you to "issue commands"; it proactively "provides services."

This shift means the kernel of the operating system changes from event-driven to intent-driven. In the past, we learned how to use phones (learning operation logic); in the future, phones will learn how to use us (learning behavior patterns).

Hardware Makers' AI Transformation: Inevitable Yet Helpless

Why are hardware makers like Apple, Huawei, and OPPO leading the charge into this track, rather than OpenAI or Google? The reason is simple: AI operating systems need data sources, and the data sources are in the hardware.

Traditional OS vendors (Microsoft, Google) also have AI, but their AI is more like a "brain in the cloud," requiring internet connectivity to use. The advantage of hardware makers lies in: Chip, sensor, and local model, all integrated. Apple's A-series chips already have built-in neural engines, Huawei's Kirin NPU has been developed for years, and OPPO's on-device large model compression technology has also been proven. These hardware makers aren't just doing AI; they are "welding" AI into the chips.

Dimension	Traditional AI (Cloud)	Hardware AI (On-Device)
Response Speed	Depends on network, high latency	Local inference, millisecond level
Privacy Protection	Data uploaded to cloud	Data stays on device
Scenario Coverage	Requires internet to use	Can run offline
Business Model	Selling services/subscriptions	Selling hardware, AI is a value-added feature

This isn't a debate over technical routes, but a battle of business models. If AI operating systems become on-device intelligence, whoever controls the chips and sensors controls the entry point. The anxiety of Apple, Huawei, and OPPO isn't "What if we fail at AI," but "If AI runs in the cloud, I'll be choked by Google and Microsoft."

What I Saw in This Paper

As a first-year master's student who just joined the lab, I'm still chewing on the intersection of Operating System Concepts and Deep Learning. But it's precisely this "half-understood" state that gave me some excited intuitions about this news.

I've recently been reading a paper about using Large Language Models as the "scheduler" of the operating system. Traditional process scheduling algorithms (CFS, O(1)) are based on CPU time slices and priorities, but the paper proposes using a lightweight LLM to predict the user's next action, then preloading relevant processes and memory. For example, if you open WeChat to check messages every day at 8 AM, the system pulls up the WeChat process in advance and caches contact data in memory. This way, when you open WeChat, you barely feel any loading time.

This idea is crazy, but what Apple and Huawei are doing is essentially the industrial implementation of this thinking. The operating system is no longer just managing hardware; it starts managing "attention" and "time." In the past, we "deceived" ourselves into higher efficiency through multitasking switching; in the future, the OS will actively help you block distractions, presenting only the information you need when you need it.

But There Are Many Problems Too

Writing this far, I suddenly realized a contradiction: If the OS understands me too well, will it become scary? Privacy is an obvious issue, but deeper down is perhaps the surrender of "control." When the system makes decisions for you, will you retain the ability to say "no"? For instance, if the system judges you are anxious based on your heart rate data, proactively silences your phone, and recommends meditation music. That sounds thoughtful, but what if you just want to doom-scroll anxiety-inducing videos today?

Moreover, from a technical perspective, the generalization ability of on-device models is still weak. Apple's Apple Intelligence currently only supports English, Huawei Xiaoyi's dialect recognition is still iterating, and OPPO's AndesGPT still gives "irrelevant answers" in complex scenarios. This is a real challenge: needing to run inside a phone (limited compute power) while being smart enough to understand complex intents (large model size)—these two goals are naturally contradictory.

My Guess on Future Development

Within two years, all mainstream phone manufacturers will launch their own AI operating systems, but the vast majority will die miserably. Because this isn't simply "adding an AI assistant," but redesigning the entire system architecture. From kernel scheduling to application interfaces, from privacy sandboxes to model compression, everything is tough nuts to crack. Only those manufacturers who possess chip, operating system, and cloud service capabilities simultaneously can truly survive. Apple, Huawei, and Samsung might manage it; OPPO, vivo, and Xiaomi need to find their own unique paths.

And Google and Microsoft might go from...

Original link: https://www.tmtpost.com/8067058.html