Image source: Generated by Unbounded AI

Today, a friend suddenly told me that he was "defeated" by an AI mobile phone share chart:

▲ AI mobile phones will be released in mainland China in the third quarter of 2024 Volume share, source: Canalys

He is holding the latest iPhone 16 Pro in his hand Max, half angry and half smiling, asked me loudly: "Where has such a big AI like me gone?"

Indeed, even though it has been released for half a year, Apple's Apple Intelligence is still the best in China. Unavailable status.

Recent media reports claim that Apple is preparing to use Baidu’s Wenxinyiyan, but the details have not been finalized yet.

According to foreign media The Information, there are still some problems in adapting Baidu's large model to the iPhone, and Baidu and Apple have differences in the collection and use of user data: Baidu hopes to save the data to train the model, But Apple's privacy policy does not allow the collection of this data.

No matter how Apple AI is ultimately implemented in China, at least for now, domestic iPhone users cannot use the AI ​​pie that Cook drew at the press conference (without special methods).

▲On September 10th, Apple Smart Chinese release time was officially announced

So why did Canalys give such a conclusion that is “inconsistent with actual experience”?

The answer is also very simple, because the definition of AI mobile phone in the Canalys report is "Al-capable smartphone" - a smartphone with AI capabilities.

You may not be able to use it, but I have this ability, so I am an AI phone.

According to the definition of IDC, another market research institution, AI mobile phones have several key indicators and characteristics: NPU with computing power greater than 30TOPS, SoC that supports generative AI models, and the ability to run various large models on the device. .

Using this standard, the iPhone 16 Pro Max equipped with A18 Pro is indeed a genuine AI phone.

I told him, it doesn’t matter. Although Apple can’t use it, you can look at China. Judging from the speed of implementing AI functions, domestic manufacturers can be said to be “far ahead”.

In the past two months, new domestic AI mobile phone products have been released intensively. OPPO, vivo, Honor, Xiaomi, and Huawei have successively unveiled AI big moves, including concepts such as AI agents and AI mobile operating systems (AI OS). Go to the front desk and swipe the screen frequently on the PPTs of various manufacturers.

The simple summary is: I have what Apple has, and I still have what Apple doesn’t have.

▲From left to right: Huawei, OPPO, Honor, vivo, Xiaomi

Just one sentence, AI directly operates your mobile phone screen to order a cup of coffee, and the taste is still yours My favorite kind; one clickSummon AI to recognize your mobile phone screen, and AI can recognize the entire world you see directly through the camera.

The development of AI mobile phones seems to be entering a new stage.

Although various AI Apps have become very popular recently, ranking at the top of the download lists in major app stores, they also have rich AI functions. But it seems that these capabilities of today’s AI mobile phones can no longer be achieved by installing an AI App.

▲The OPPO software application store searches for “AI” and displays the App results

Why is the AI ​​phone not a phone with AI App installed?

When we delved into the AI-related technologies, operating systems and new mobile phone products released by various mobile phone manufacturers, we found that there are essential differences between the two.

Although the leading mobile phone manufacturers have not clearly defined AI mobile phones, we have found that the recent explosion of various AI products and technologies are all moving in a common direction: system level. Personalized AI.

The development of AI mobile phones is ushering in an unprecedented "unification of directions."

According to Canalys predictions, the penetration rate of AI mobile phones may be close to one-third in 2025, with shipments of nearly 400 million units.

AI mobile phones have indeed become an important trend and growth point that cannot be ignored in the mobile phone industry, and the development of AI mobile phones is undoubtedly entering a new critical stage.

01. What is the difference between an AI phone and a phone with AI App installed?

It’s too broad to explain. Let’s start from a more intuitive level, which is where the computing power comes from.

Based on the actual implementation of AI functions and experience in various AI mobile phones, most of the computing power of AI mobile phones comes from mobile phones.

One of the highlights this year is the emphasis on the importance of NPU, from chip manufacturers to terminal manufacturers. This is also consistent with the tone of each company emphasizing "device-side AI implementation".

This year’s flagship mobile phone chips from Qualcomm and MediaTek have undergone key upgrades in terms of NPU, from core specifications to software layer optimization support.

It is worth mentioning that this AI computing power must not only be strong in peak AI computing power, but also must be well adapted to all popular mainstream large models, so that large models can fully mobilize the AI ​​computing power of the NPU , rather than the picture having a good-looking running score.

Complex terminal-side image recognition, voice processing, and real-time AI scene recognition and optimization are all inseparable from these.

In contrast, a mobile phone with an AI App installed does not necessarily have such powerful AI hardware support. Because the App can be connected to the Internet, large models can be run in the cloud.

But the cloud is not omnipotent.

A very real problem is that not everything is possible.Even if it can be solved by relying on the Internet, networking may not be the most efficient, cost-effective, or safe.

There are more than 4 billion smartphone users in the world. If each person’s various AI needs every moment of every day, no matter how big or small, must be processed in the cloud, the cost will be astronomical.

These costs cannot be unconditionally passed on to consumers, and manufacturers themselves cannot bear them all. This will become a major stumbling block for the early development of AI and the inclusiveness of AI.

In addition to being too expensive, another problem with running everything in the cloud is data privacy and security.

Although privacy and security always sound a bit "false" and may not be "perceived strongly" by ordinary people, in the press conferences of all manufacturers such as Apple, Huawei, Honor, Xiaomi, OPPO, and vivo, On the Internet, data security issues have been emphasized individually, and almost all major manufacturers have released "chip-level" security solutions.

When it comes to data security issues, if nothing happens, it will be a big deal.

In addition to data security, latency and reliability are also issues. To give a very simple example, the current AI call real-time translation, or AI real-time conversation translation, you say something to me, this kind of real-time dialogue Fluency is obviously difficult to guarantee in the cloud.

Another example is Apple’s Gemoji. AI generates an emoticon package for you in real time based on your photo so that you can use it immediately in the chat conversation.

Some users may also work in a network-free environment. If they have to bid farewell to AI completely for this reason, the experience will undoubtedly be greatly reduced.

In the final analysis, as long as it relies on the Internet, there will be great "uncertainty" in its stability.

In addition, in terms of the actual experience of AI functions, there is a big difference between AI phones and phones with AI App installed.

To put it simply, what AI mobile phones can do, AI Apps may not be able to do, and if they can do it, they may not be able to do it well.

What not to do?

A very simple one is to help you operate various complex mobile phone settings. Now, whether it is an Apple or Android smart assistant, all novices can directly turn into gaming experts.

You can directly tell the AI ​​to help me turn off Douyin’s photo reading permission, help me turn on the power saving mode of my phone, and help me set the phone system to dark mode.

Because the AI ​​on the AI ​​​​mobile phone can truly call various system functions, it is deeply connected with the system. In terms of permissions and security, it is beyond the reach of third-party apps.

AI can also help you summarize and refine the key points wherever there is text, from emails, messages, browsers to your commonly used social software, wherever there is text, there is AI text processing.

▲Important content of the AI ​​summary notification

You can also tell the AI ​​directly that I am a little sleepy and order a cup of coffee for me. At this time, the AI ​​will know your commonly used takeaway software. , the coffee you often drinkTaste, and use mobile phone location to find the store nearest you to place an order.

This kind of truly personalized experience is also impossible to achieve with AI App. Because this information comes from every corner of the system, from the usual AI learning of our daily operating habits and the establishment of a personal knowledge map.

For example, our schedule, a chat conversation, an online ticket purchase record, or a file, photo, or video on a mobile phone.

To put it simply, system-level operations, cross-application scheduling, complex scene analysis, and intelligent decision execution are currently unique to AI mobile phones, and are generally difficult to implement in AI Apps.

Is there anything that both can do, but the AI ​​App may not be able to do well?

A very intuitive one is AI picture beautification. AI mobile phones can directly start from AI scene recognition and object recognition before taking pictures to AI image processing during shooting, and then to AI closed eye repair and AI passers-by in the later stage. Eliminate and AI one-click beautification are fully involved.

The whole process is very smooth, and can be completely implemented on the client side, with data security and high processing efficiency.

In contrast, the AI ​​image processing function of the AI ​​​​App may be more limited to us uploading a photo, and then AI converts the picture style, or performs AI beautification. The function must be single many.

There are many such examples.

Of course, the cloud does have its advantages, such as some very high-quality AI graphics and text generation. Due to the large scale of the model and high computing power requirements, it is more efficient to execute it in the cloud.

This is why almost all terminal manufacturers have adopted the "device-cloud combination, device-cloud collaboration" model to make AI mobile phones. In fact, part of this "cloud" is these AI Apps.

To put it simply, if you can do it on the device side, do it on the device side. If you really can’t do it on the device side, call the cloud large model to do it.

A core purpose is to enable the AI ​​experience that users want to achieve.

So in the final analysis, AI mobile phones and mobile phones with AI App installed are more like an inclusive and progressive relationship - AI mobile phones > AI App mobile phones.

Judging from the timeline of the development of AI mobile phones, at the beginning, mobile phones were equipped with some apps with generative AI functions, which allowed us to see the charm of AI. For example, a ChatGPT was installed to replace all Voice assistant.

However, with the development of technology, the integration of AI and mobile phones has become more in-depth. From hardware to software to operating systems, it can be said that AI is everywhere, including performance, imaging, battery life, and every part of the mobile phone. Start applying to large model capabilities.

AI mobile phones have truly formed some key industry consensus: system-level personalized AI.

02.AI mobile phone war Why are AI agents and AI OS so different?so important?

So at this point, we need to clarify the difference between an AI phone and an AI App phone. The core purpose is to understand what an AI phone is and why it is called an AI phone?

From Apple to the domestic Huami OV Honor, we can find some common points: AI mobile phones will have an operating system with deep AI involvement, and there will also be a powerful AI agent that can interact at any time.

Just like Apple’s definition of “personalized intelligence”: powerful, intuitive, built-in integration, personalization, and privatization.

▲Apple’s characteristic definition of personal systematic intelligence

System-level personalized AI requires an AI agent and AI OS, which is absolutely impossible to achieve by installing several AI Apps.

Mobile phone manufacturers are not reinventing the wheel, but doing a brand new thing.

Xiaomi’s Super Xiaoai, Honor’s YOYO, Huawei’s Xiaoyi, OPPO’s Xiaobu, and vivo’s Blue Heart XiaoV.

In the blink of an eye, the intelligent voice assistants of the past have transformed into super "AI agents". In essence, these intelligent assistants assume the role of the portal for interaction between humans and AI.

Almost all the AI ​​experiences mentioned above require interaction with smart assistants, such as speaking a sentence, typing a line, or taking a photo.

Whose AI phone is easier to use? The first level is to compare whose smart assistant is better to use and whose AI smartphone is smarter.

But AI agents are not the essence. Why AI agents can achieve these excellent AI experiences and provide various AI services and functions is the real key, and it is also the key to differentiation among companies. The key to the advantage is the deep integration of AI and operating systems.

Every manufacturer is talking about AI OS now. What is the integration of AI and operating system?

Here we might as well use the words of NVIDIA founder and CEO Huang Renxun to understand this matter. Huang Renxun once expressed a similar meaning in his speech: In the future, the large AI model will be the operating system.

▲Huang Renxun shared his views on AI agents at the Lenovo TechWorld Conference

Tang Kai, President of OPPO Software Engineering Division, mentioned similar views at the press conference. In his opinion, The first stage of the development of AI OS is AI for System, the second stage is System for AI, and the third stage is AI as System.

The AI ​​operating system actually schedules various large model capabilities, or various AI-based services.

AI is like a "new core" of the operating system, which can perceive, make decisions, and call various end-side large models, vertical large models, Internet large models, and cloud large models to ultimately realize AI. mobile experience.

Mobile phone operationsThe system is like "alive" and suddenly has vitality. AI is the system, and the system is AI.

Honor CEO Zhao Ming once said at the press conference that the AI ​​OS in the era of agents must have an AI core driven by large models, AI interaction driven by agents, and an AI ecosystem built by open services. This is truly an AI-enabled operating system.

Various large models have taken root in various modules of the mobile phone system, changing the operating logic of each module of the system, from traditional "rigid" algorithms to today's "flexible" judgments of AI.

It sounds a bit abstract, but this is the direction that all smartphone giants are working hard at the moment: to make AI OS well and ultimately achieve an excellent AI mobile phone experience.

This is the basis of what manufacturers call a system-level personalized AI experience, and it is also the direction identified by Apple.

03. From hardware, software, operating system to application ecosystem, the challenges of making an AI mobile phone come one after another

As the same saying goes, ideals are full and reality is skinny. Making an AI mobile phone well requires one step at a time.

Making a good AI OS sounds like a very clear thing, but in fact, the difficulties in making a good experience can be said to be everywhere.

For example, very real problems - how to solve the storage bottleneck, how to solve the computing power bottleneck, and how to break through the power consumption bottleneck? How to ensure the efficient coordination and collaboration of various models while taking into account data security?

How to ensure that the large model can truly and accurately understand our intentions and make reasonable decisions?

All of this must be realized on a device with limited size, limited battery life, and limited computing power like a mobile phone.

Manufacturers such as Honor, vivo, and OPPO have begun to intensively optimize large end-side models to achieve stronger performance with fewer parameters. For example, vivo’s Blue Heart 3B has the fastest word output speed. Reaching 80 words/s, the memory usage is only 1.4GB. Honor’s 3B large model also increases the word output speed by 300%, and the memory usage is reduced by 1.8GB.

Of course, this is only at the technical level of software and hardware. There is also an invisible challenge, which is the construction of an AI application service ecosystem.

Even if the framework of the AI ​​​​mobile phone itself is set up, the skeleton is strong, and it runs smoothly, without a truly good AI service application, the experience of the AI ​​​​mobile phone will be difficult to achieve.

At the chip level, we have seen the intensive debut of various AI development tools and kits from Qualcomm and MediaTek. MediaTek’s AI intelligent engine pioneer plan covers almost all domestic leading mobile phone manufacturers; at the terminal level, OPPO, Manufacturers such as Honor, vivo, Xiaomi, and Huawei are all focusing on building an AI application service ecosystem.

If you want to make a good AI mobile phone, win-win cooperation has almost become the only way recognized by the industry.Apple in a closed ecosystem is also "forced" to access OpenAI's large three-party model ChatGPT.

This is like an orchestra. Not only must it have excellent conductors and musicians, but whether it has excellent repertoire to perform is also the key to whether the entire performance can impress people.

Of course, this is not something that one or several companies can do well. The entire mobile phone industry and every company in the entire industry chain must work hard in this direction.

04. Conclusion: AI mobile phone, a tough battle

There is no doubt that making a good AI mobile phone is a systematic matter. It cannot be achieved by a single technological breakthrough, nor can it be achieved by installing a few AI apps.

Today, the development of AI mobile phones has gradually formed a key consensus such as AI OS. The entire industry is developing in the direction of system-level personalized AI, which is also an AI that is one step closer to true AGI.

Various municipal research institutions have made positive and optimistic predictions about the development of AI mobile phones. This year, the shipment volume of AI mobile phones will account for about 16%, and this proportion will exceed half in four years.

In the battle of AI mobile phones, the competition among various manufacturers has become more and more intense, and the competition of technology has become more and more in-depth. No matter what, in the face of this tough battle of AI mobile phones, every manufacturer must fight hard. spirit, put in 120% effort, and go all out.

AI mobile phones cannot afford to lose.