Article reprint source: Industrialist

Original source: Industrialist

Image source: Generated by Unbounded AI

The mobile phone industry has entered a new qualifying race. Although the road to big models is difficult, we must catch up. Behind the big models, mobile phone manufacturers are targeting the next entry battle.

Mobile phone manufacturers are beginning to "roll up" in the field of large models.

In August this year, Xiaomi announced that it had successfully developed a 1.3 billion parameter end-side model and implemented it on the mobile phone. At the same time, it announced its self-developed large-scale pre-trained language model MiLM-6B, which has a parameter scale of 6.4 billion and ranks first among large models with the same parameter scale in the authoritative Chinese evaluation lists C-EVAL and CMMLU. On October 26, at the Xiaomi Pengpai OS and Xiaomi 14 series new product launch conference, Xiaomi officially announced that the AI ​​large model will be implanted into the system.

On November 1, vivo released its self-developed AI "Blue Heart" big model at the developer conference. At the same time, vivo also released a matrix of big models with different parameter scales, including 175 billion, 130 billion, 70 billion, 7 billion and 1 billion big models. It announced that the 13 billion parameter Blue Heart big model has been implemented on the device side and the 7B big model has been open sourced.

On November 16, OPPO officially released its self-trained personalized large model and intelligent agent, Andes Large Model (AndesGPT), at the Developer Conference. AndesGPT uses "end-cloud collaboration" as its basic architecture design idea and launches model specifications with different parameter scales ranging from billions to more than 100 billion.

In addition, Honor and Huawei have also begun to plan to install large models into mobile phones.

Some questions worth pondering under this change are: what is the reason behind the mobile phone manufacturers' deployment of large models? Is it really easy for mobile phone manufacturers to make large models?

The large model in the ceiling becomes an opportunity

A set of data shows that global smartphone shipments fell 13% and 11% year-on-year in the first and second quarters of 2023, respectively. It is worth noting that in the first quarter of 2023, shipments fell to 67.6 million units, the lowest level since 2013.

Of course, the reason for the reduction is naturally related to factors such as the environment and high inventory of mobile phone terminals.

However, judging from the year-on-year growth in shipments alone, there is no winner among smartphone manufacturers in the first half of this year. A closer look reveals that the smartphone shipments of the world's top smartphone brands in the first half of this year were not ideal. Among them, although Samsung ranked first in global shipments, its smartphone shipments in the first half of the year decreased by about 21.59 million units year-on-year.

Other smartphone brands with the highest year-on-year decrease in shipments were other manufacturers, Xiaomi and vivo. In the first half of the year, the shipments of smartphones of other manufacturers, Xiaomi and vivo decreased by approximately 18.4167 million units, approximately 14.8941 million units and approximately 9.1658 million units respectively.

It is a fact that the mobile phone market has been facing a "ceiling" in recent years. In response to this trend, mobile phone manufacturers are also constantly seeking new growth space.

For example, they have collaborated with professional camera brands in terms of camera, and have made efforts in battery capacity, memory, system, etc. Among them, AI smart assistant is the focus of most mobile phone manufacturers' combination with AI, but the effect is always limited.

"The emergence of GPT has reversed the situation."

According to Yang Zhenyu, chief architect of OPPO's Andes model, starting from the second half of 2022, the direction of smart assistants will begin to decline, and many companies are relatively pessimistic. This pessimistic sentiment mainly comes from the fact that although smart assistants have been used for many years, they do not seem to have brought about particularly significant breakthroughs.

If the weak consumer market is a growth dilemma for mobile phone manufacturers, then the disruption of mobile phone terminals under the big model is a crisis that mobile phones may be replaced in the future.

As the development of large models intensifies, some changes will be found in future interactive terminals. The first is high integration. That is, with the development of chip technology, future interactive terminals may adopt highly integrated chips, which will further reduce the size of the terminal while improving performance and stability.

The second is diversified input methods. In addition to traditional input methods such as keyboard and mouse, future interactive terminals may also support multiple input methods such as voice, gestures, and eye movements, allowing users to interact more naturally and conveniently.

The third is cloud computing, that is, with the development of cloud computing technology, future interactive terminals may transfer some functions to the cloud, thereby reducing the size and complexity of local devices. At the same time, cloud computing can also improve the flexibility and scalability of interactive terminals.

The fourth is intelligence, that is, future interactive terminals may be more intelligent, automatically learning and adapting to user needs through artificial intelligence technology, improving user experience and efficiency.

As technology develops to a certain stage, the computing pressure under the operation of large models will gradually decrease, and they can be installed on any terminal device, even glasses, watches and other devices. As the main interactive device at present, mobile phones need to evolve. Even at present, mobile phones may not be the next explosive entrance spawned by large models. Mobile phone manufacturers are more doing the first half of the layout of large models.

For mobile phone manufacturers, deploying a large model can not only break the ceiling of innovation, but also plan ahead to prevent problems before they occur.

Large model super applications, "cultivated" in mobile phones

Although mobile phones are not the next explosive entry point for the big model, they are used very frequently and for a very long time, and they are a concentration of personal information. In addition, mobile phones can connect to various resources and services on the Internet. Through mobile phones, Internet companies can more easily reach users and obtain user data, thereby better understanding user needs and optimizing products and services.

Therefore, occupying the mobile phone market means occupying an important entrance to the Internet market, and more users and data can be acquired through mobile phones, thereby better developing one's own business.

By combining big models with mobile phones, the big model can better understand user needs through the mobile phone as an Internet portal. Through deep learning and natural language processing technology, it can better analyze user intentions and feedback, allowing the big model application to be continuously optimized. Secondly, the combination of big models and mobile phones can also enable big model applications to better understand market demand and product defects, so as to make more informed decisions. More importantly, the big model can provide personalized services based on the behavior and preferences of each mobile phone user.

It is a plus for mobile phone manufacturers, and a "nutrient pool" for large models. That is why many mobile phone manufacturers have joined the market, and large model manufacturers are also very active in cooperating with mobile phone manufacturers. However, judging from the paths of various mobile phone manufacturers, the focus of each company's combination with large models is different.

When OPPO lays out large models, its focus is on using large models to create a smart terminal interactive experience. As Yang Zhenyu said: "With the introduction of large model technology, the proportion of users interacting with voice assistants to ask questions and answers will increase significantly." At present, the entire monthly activity of OPPO voice assistant Xiaobu is more than 150 million , tens of millions of interactive data are generated every day.

In terms of deployment, it is mainly cloud + local deployment. That is, for highly sensitive privacy data, local processing can be done on the end side. However, for some complex tasks, it is still necessary to rely on large cloud models.

Different from OPPO's deployment method, Xiaomi focuses on lightweight and local deployment when laying out large models.

This deployment method can better protect user privacy while achieving personalized customization for thousands of people locally. To this end, Xiaomi has delegated some of the large model capabilities to the end-side, striving to achieve the best balance between power consumption, inference speed, and generation effect. Xiaomi has also developed a 1.3 billion parameter end-side model, and the effects of some scenarios are comparable to the calculation results of a 6 billion model in the cloud.

Vivo is more focused on solving the personalized needs of users. Vivo believes that the big model will comprehensively innovate the interactive experience of smart terminals and continue to bring high-quality experience to users in this direction. In terms of deployment, the 13 billion parameter Blue Heart big model can achieve terminal-side operation, and the 175 billion parameter Blue Heart big model has reached the parameter level of GPT-3.

The combination of Huawei and the big model is more about making the big model the "brain" of the system. It is understood that the integration of Huawei's voice assistant Xiaoyi and the big model is not a simple enhancement of tasks such as chatting, AIGC, and replying, but a system-level enhancement with the big model as the core. The underlying logic is to assign the user's tasks to the appropriate system, with each system performing its own duties, while enhancing the experience in complex scenarios.

Overall, "reshaping" voice assistants is undoubtedly the first and necessary step for major mobile phone manufacturers to lay out the big model. The difference is that due to the different advantages of each mobile phone manufacturer, the focus of integration with the big model is also different.

For example, Xiaomi's long-term advantage in the smart home field has accumulated a large amount of AI software and hardware foundation. From a certain perspective, its terminal deployment method may be an early layout for smart home businesses such as Xiao Ai.

OPPO has always focused on conquering technology and innovation, providing products for users who pursue high-quality photography and design. This positioning is also continued in the layout of large models.

Vivo is mainly positioned as a young and fashionable brand, focusing on selfie and music functions, and launching products for young users who pursue fashion and personalization. The combination with the large model has made new efforts in personalization.

In general, mobile phones, as a natural entrance to the Internet, are a good soil for the application of big models. Based on this advantage, mobile phone manufacturers may produce super applications by combining with big models.

More importantly, some of the technologies and experiences brought by big models + mobile phones may help mobile phone manufacturers to be prepared when the crisis comes, or to reuse them.

Is it really easy for mobile phone manufacturers to make large models?

Mobile phone manufacturers want to use big AI models to improve the performance and competitiveness of their products, but they still face some challenges.

First, from a technical perspective, integrating large models into mobile phones requires powerful computing power and efficient algorithms. Although the hardware configuration of smartphones is becoming more and more powerful, it is still challenging to fully utilize large models within the limited space of mobile phones.

However, at present, it is difficult to put large models with tens of billions or hundreds of billions of parameters into mobile phones. The cost is high, self-development is difficult, and computing power consumption is high.

Secondly, from the application level, AI large models require a large amount of data for training to achieve the desired effect. However, due to privacy protection and other reasons, it may be difficult to obtain and use a large amount of user data. At the same time, it is also necessary to solve problems such as insufficient model generalization ability to ensure that the performance of the model in different scenarios can meet the needs.

At present, mobile phone manufacturers mainly adopt two paths to deploy big models: one is the end-side computing mode, and the other is the parallel path of the end-side and the cloud. However, when the big model is implemented in specific business scenarios, the end-side and the cloud-side have their own requirements. The latter has become the consensus of most mobile phone manufacturers to deploy big models, but the parallel path of the end-side and the cloud also brings new challenges to data collection and processing, model training and optimization, and system architecture design.

Finally, from the perspective of market competition, major mobile phone manufacturers are increasing their investment in the field of AI large models, and the competitive pressure is huge. Only by continuous innovation, creating products with differentiated advantages, and avoiding homogeneity can we stand out in the market.

In general, it is not easy for mobile phone manufacturers to deploy large models.

If the above problems exist internally, then large model technology is a problem from outside.

Specifically, the layout of large models faces huge capital investment and a technology research and development path that seems to have no end. Whether to start from scratch to develop AI native applications or to join the "circle of friends" requires in-depth thinking.

However, at present, among many mobile phone manufacturers, not many have the technical capabilities of native big models. Huawei is one of them. It has begun to deploy mobile phone AI native big models based on the Pangu big model.

Ecosystem cooperation and actively embracing open source may be the choice of most mobile phone manufacturers at present. It is also a relatively "healthy" and secure model.

Take OPPO AndesGPT as an example. It was trained by OPPO itself, but it also referred to some open source model experience. "Most companies may also use this method because the model algorithm itself is relatively mature," said Yang Zhenyu.

Yang Zhenyu believes that it is more important to use big models to improve business effectiveness. According to Yang Zhenyu, AndesGPT may have an advantage of about 20% compared with the mainstream big models in China.

In this mode, we can see that with the development of the mobile phone big model ecosystem, technologies such as cloud computing and edge computing will become more mature, realizing cloud-edge integration and providing stronger capability support for AI big models; human-computer interaction will become more intelligent, making people's work and life more convenient; more attention will be paid to personalized experience and customized content, providing users with more precise services.

In addition, the application of AI big models on mobile phones will become more popular, so that future AI big models will integrate multiple modalities, such as vision, hearing, touch, etc., to enhance people's perceptual experience.

In order to adapt to the special requirements of mobile phones, future large models will pay more attention to low energy consumption and high efficiency, and be able to provide high-quality experience under limited hardware conditions.

In the future, these capabilities accumulated by mobile phone manufacturers and their layout in the field of large models may help them seize the key entry point when the application of large models explodes.

The mobile phone track has ushered in a new qualifying race. Although it is difficult to walk on the road of large models, we have to catch up. Opportunities always belong to those who are prepared.