21st Century Business Herald reporter Dong Jingyi reports from Shanghai
At the 2024 Apple Autumn Product Launch, AI became one of the biggest highlights.
The previously announced AI feature has been implemented on the iPhone 16, which can help users optimize text content in applications such as email, memos, and Pages documents. It also supports generating personalized emoticons and images through natural language descriptions, as well as automatically creating video content through simple descriptions.
In addition, Siri has been significantly enhanced to better understand user commands, provide personalized services, and perform hundreds of new operations within and across applications. The attributes of AI agents are clearer.
The industry's high attention to Apple's AI is not only due to its potential to trigger a new wave of iPhone replacement, but also because Apple's Apple Intelligence represents the development of AI in the mobile terminal field, and market participants are closely monitoring the actual application effects of AI technology on mobile devices.
The end side provides an entry point for AI, which is the foundation for the widespread adoption of AI technology.
Recently, at the AI Reconstruction Digital Economy sub forum of the 2024 Sullivan New Investment Conference, INMO co-founder Lv Yifei stated at the roundtable forum that products such as AR glasses, combined with generative AI, provide an intuitive interactive platform that utilizes its visual and hands-free features to become an effective tool for obtaining AI content.
However, it is still difficult to fully support large models on the end side, mainly due to hardware bottlenecks such as processing power, memory, and battery life. However, multiple technologies are under development and the implementation process will be gradual.
End to end implementation will start with small scenarios and scenarios that can be achieved by large models, adopting an end-to-end cloud collaboration approach. Some content will be processed on the end side, while most content will still need to be processed in the cloud, "said Lu Yanxia, Research Director of IDC China, to 21st Century Business Herald reporters.
A global competition
In fact, the deployment of AI by various terminal manufacturers is not slower than that of Apple.
Most mobile phone manufacturers are all in AI. According to a Canalys report, 16% of global smartphone shipments will be AI phones by 2024, and this proportion is expected to surge to 54% by 2028.
Just a few days before the Apple press conference, Honor unveiled the industry's first cross application open ecosystem intelligent agent (AI Agent) at the Berlin Consumer Electronics Show (IFA) in Germany, and announced that the Honor Magic 7 series will be equipped with Honor AI Agent for the first time.
It has four core capabilities: natural semantic understanding and computer vision, user behavior habit learning and scene environment perception, intention recognition and decision-making ability, and in app and cross app operation.
This is similar to Apple's Apple Iintelligence, which delves into the underlying architecture of mobile operating systems to achieve system level AI refactoring. These capabilities enable AI agents to deeply understand user needs, proactively provide services, simplify user operation processes, and improve usage efficiency.
Prior to this, some manufacturers have also launched terminal products that integrate AI applications. For example, the Samsung Galaxy S24 series smartphones have achieved innovative features such as AI call translation and intelligent image cutout through the use of generative artificial intelligence technology; The Huawei Pura 70 series flagship phone has innovated in AI photography, integrating functions such as one click elimination with the Pangu large model; The OPPO Find X7 series is equipped with the AndesGPT large model, with AI calling, image cancellation, and a new AI assistant as its main selling points.
The industry generally believes that the rise of AI smartphones is seen as the third stage of the mobile phone industry after feature phones and smartphones. This trend not only drives innovation in the mobile phone industry, but may also lead to a major reshuffle in the industry.
PC is another popular terminal for AI implementation. In terms of hardware, AIPC's functions have expanded to include computing, storage, sensing, and other aspects, forming a heterogeneous solution of CPU-GPU-NPU. In terms of software, AIPC integrates lightweight AI models to achieve offline steady-state operation of various generative AI applications.
Major PC manufacturers such as Lenovo, HP, Acer, etc. are actively exploring and developing AI PC products. Taking Lenovo as an example, at CES 2024 at the beginning of the year, Lenovo Group showcased over 10 AI PC products; During IFA 2024, Lenovo Group showcased its latest achievements in the field of artificial intelligence, launching a series of revolutionary AI PC devices covering multiple product lines such as ThinkPad, ThinkBook, Yoga, and IdeaPad.
According to a research report by Goldman Sachs, Lenovo Group will achieve an average sales price increase of PC products driven by AI PC, which will drive its PC business to grow by 27% and 22% year-on-year in fiscal years 2024 and 2025, respectively. AI PC will contribute 21% and 33% of revenue in fiscal years 2024 and 2025, respectively.
The market for AI PCs is still expanding. Canalys stated that as major processor suppliers gradually advance their AI PC plans, it is expected that the supply and user adoption of AI PCs will significantly increase in the second half of 2024 and beyond. The shipment volume will reach 44 million units in 2024 and is expected to reach 103 million units in 2025.
It is worth noting that the landing application of AI technology has far exceeded the traditional fields of mobile phones and PCs, and its tentacles have extended to various terminal devices, including glasses, headphones, smart home devices, smart cars and other forms, forming a rich AI terminal ecosystem, which provides an entrance for generative AI.
And the application on the end side is also a manifestation of the more vertical development of AI technology. The industry generally believes that verticality is the future trend of AI, which requires enterprises to delve into specific fields, apply in segmented scenarios, solve specific problems, and promote the landing of AI technology and the development of the industry.
Mainstream trend: End cloud integration
The reason why end-to-end AI has received widespread attention and achieved rapid development is mainly based on the following reasons:
Firstly, cost-effectiveness. Although the cost of large-scale model inference has significantly decreased for various manufacturers, in the long run, the combination of cloud and end computing is considered an effective way to control inference costs.
With the use of more powerful chips in devices such as smartphones, the inference ability of edge AI has been significantly enhanced. This makes it possible to complete complex AI tasks on local devices, reducing reliance on cloud resources.
If all tasks that can be operated on mobile devices are handed over to the cloud, the cost will actually increase. "At the 2024 Sullivan New Investment Conference, Liu Liang, Director of Strategic Research at SenseTime Technology, told 21st Century Business Herald reporters that by handing over large-scale computing tasks to the cloud for processing and small-scale, real-time less demanding tasks to the end side for processing, resource allocation can be optimized and cost-effectiveness can be maximized.
Secondly, the timeliness of data. In application scenarios such as smart cars that require high real-time performance, data transmission speed has become a key factor. Real time perception and rapid response to the external environment of the vehicle require immediate execution without excessive delay. The most typical example is in autonomous driving, where the deployment of large models must occur on the end side.
Thirdly, data security and privacy protection. End side AI models have relatively higher security due to running on local devices. Liu Liang told reporters that users' data does not need to leave their personal devices, thereby reducing the risk of data leakage and making it easier to establish security mechanisms, enhancing users' trust in the devices.
This is also a frequently emphasized feature in AI smartphones and AI PC products. For example, at Apple's latest press conference, Apple emphasized that when using Apple Intelligence, user data will never be stored or shared with Apple, but will only be used to execute requests. This privacy and security guarantee can be continuously verified by independent experts, which is a major advancement in AI privacy.
Fourthly, the promotion of terminal manufacturers such as mobile phones and PCs. Major domestic and foreign mobile phone manufacturers, such as Xiaomi and Apple, are actively deploying end-to-end AI technology. As a highlight of product innovation, it is expected to drive users to upgrade their phones.
Many mobile phone manufacturers are making improvements on the basis of their existing phones, including adjusting the overall architecture of the phones and building AI application ecosystems within them. This process may take some time, but manufacturers believe that it will bring better experiences to users and stimulate new replacement needs. "Recently, Kong Rong, co director of Tianfeng Global Forward Industry Research Institute, said in an interview with 21st Century Business Herald reporters.
However, judging from the long-awaited performance of Apple Intelligence in the industry, its features still lack differentiation and disruption, and the sales growth brought by AI will not come immediately.
Many people are willing to pay for AI technology, but it also depends on the actual performance of AI technology. If AI performs intelligently enough to understand user needs, users will be willing to pay a reasonable price for it. Conversely, if AI performs poorly, users may not use it. Currently, AI may still be in the growth stage and not fully mature, "Kong Rong told reporters.
Landing, gradually progressing
Although some AI terminal products have emerged on the market, they are still mostly connected to cloud based big models, or combined with small models on the end side and big models on the cloud. The real implementation of big models in terminals still faces various hardware limitations.
Future Intelligent COO Wang Chao told 21st Century Business Herald reporters that in order to run large-scale AI models on the end side, such as 10 billion parameter models, a considerable amount of computing power, approximately 50 Tops, and at least 13GB of memory are required. However, the hardware configuration of many mobile phones and PCs has not yet reached this standard, which limits the application of large models on these devices.
Meanwhile, when running large AI models, the devices consume a significant amount of electricity, resulting in a significant reduction in battery life. For portable devices such as smartphones and PCs, this is particularly problematic because their battery design needs to consider portability and weight, and cannot infinitely expand battery capacity.
Assuming I keep running the large model on the end side, it may run out of battery in an hour, making it impossible to make calls or work, "Wang Chao told reporters. These will ultimately seriously affect the user experience.
Although the combination of end cloud can overcome some hardware limitations, it also puts high demands on the model side, that is, the end side model needs to be consistent with the cloud model. This means that the end side model needs to achieve at least the same performance level as the cloud model in specific tasks, ensuring that users can have a consistent experience both in the cloud and on the end side.
Usually, cloud based large models require a wide range of general capabilities, while end-to-end models focus on specific functions and deploy a set of specialized small models, each optimized for specific tasks to achieve optimal performance.
With the advancement of technology, the performance of end-to-end models is constantly improving, and small parameter language models are now comparable to larger parameter models in multiple evaluation metrics.
The reasons for this performance improvement are multifaceted. Industry insiders in the big model industry told reporters that on the one hand, the improvement of data cleaning and processing procedures has improved the quality of training data; On the other hand, the powerful capabilities of cloud based large models can help generate high-quality domain specific data, which can be used to train small models on the side more effectively.
Of course, as a product, it ultimately depends on the market's acceptance, with differentiation and user experience being key factors.
Wang Chao believes that currently, most products are still just gimmicks for brand promotion, lacking the impulse to make consumers purchase. Some basic applications have low user stickiness, and downloading large model apps directly can also meet such needs
Industry analysis believes that the current global macroeconomic situation is not good, and without promotional activities and significant price discounts, the growth of the AI mobile phone or AI PC market will still face significant challenges.
However, Liu Liang pointed out the uniqueness of cars as AI terminals. He believes that cars are the most powerful personal items in terms of computing power, far surpassing PCs and mobile phones. At the same time, cars are also personal items with the most abundant power supply. Although there is also the problem of so-called battery anxiety, they are better compared to mobile phones or computers.
And consumers' willingness to consume in the automotive industry is usually stronger. Users may not be willing to pay an additional 1000 yuan just because a certain phone has added AI, but paying an extra 5000 to 6000 yuan for certain AI features in the car is acceptable for many users, "Liu Liang told reporters.