The fire of large models has burned into the field of mobile phones

Text|Liu Kuang.

In the eyes of many people, the entry of mobile phone manufacturers into the AI model is just another hype, but in fact it is by no means the case. According to public information, the head mobile phone manufacturers have begun to deploy AI many years ago, and some manufacturers have even smashed tens of billions of R&D investment, with an AI team of 1,000 people, and there is no upper limit for future investment. Judging from this posture, the entry of mobile phone manufacturers into AI is definitely not a whim, but a prepared one.

At the moment when mobile phone manufacturers compete for large models, the large model war in the mobile phone circle has begun. According to the data, 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, with a parameter scale of 6.4 billion, ranking first among the authoritative Chinese evaluation lists C-Eval and CMMLU models with the same parameter scale. On October 26, at the Xiaomi Surging OS and Xiaomi 14 series new product launch conference, Xiaomi officially announced that it would implant the AI large model into the system.

On November 1, vivo released its self-developed AI "Blue Heart" large model at the developer conference. At the same time, vivo also released a large model matrix containing different parameter scales, including 175 billion, 130 billion, 70 billion, 7 billion and 1 billion five large models. It announced that the blue heart model with 13 billion parameters has been run through end-to-end, and the 7B large model has been open-sourced. On November 16, OPPO officially released the self-trained personalized large model and agent at the developer conference-Andean large model (Andean large model) (Andean large model). Based on the infrastructure design idea of "device-cloud collaboration", AndesGPT has launched a variety of model specifications with different parameter scales ranging from one billion to more than 100 billion.

In addition to the above, manufacturers including Huawei and Honor are also laying out large models into mobile phones. So, why can large models become a must for many mobile phone manufacturers?

First of all, at present, AI large models have become the ecological base of many terminals, and it is imperative for mobile phone manufacturers to enter the game to make AI large models. Industry insiders believe that at present, mobile phone manufacturers are building a new generation of intelligent ecological environment covering equipment, automobiles and networked terminals, and the large model is the ecological base, which can improve the intelligence of human-computer interaction. At the same time, the development of multimodal and cross-modal and lightweight application technologies has greatly promoted the prosperity of smart applications and intelligent ecosystems.

Secondly, the dilemma of the mobile phone industry needs to be solved, and the AI model may be the best breakthrough direction. In the past few years, the overall performance of the consumer electronics industry represented by mobile phones has been sluggish, and various mobile phone manufacturers are trying their best to innovate. From chips to appearance to battery life, many mobile phone manufacturers have been involved. However, compared with the "small fights" around the NPU a few years ago, generative AI has undoubtedly made the entire consumer electronics industry see the light: in an environment where the whole people are discussing large models, the integration of AI and terminals is regarded as a new innovation anchor, which will drive the resonance of the industrial chain.

Although the caliber of different research institutions varies, it is generally believed that the peak of shipments was around 2017, after which the market began to enter a downward cycle. Because of the emergence of new technologies such as 5G, global smartphone shipments briefly recovered in 2021, but they did not last long. The downturn in the consumer electronics market has not only affected terminal companies, but also had a wide impact on the middle and lower reaches of the consumer electronics industry chain.

From this point of view, the current pursuit of generative AI in the upstream and downstream of the consumer electronics industry is not simple, but essentially relies on the selling point of large models to stimulate new growth.

In fact, although the mobile phone is not the next explosive entrance to the large model, the frequency and time of mobile phone use are very high, and the mobile phone is a concentration of personal information. Through mobile phones, Internet manufacturers can more easily reach users and obtain user data, so as to better understand the needs of users and optimize products and services.

On the one hand, through the combination of large models and mobile phones, users can better understand user needs through the Internet entrance of mobile phones, and continuously optimize the application of large models through deep learning Xi and natural language processing technologyOn the other hand, the combination of large models and mobile phones can also enable large model applications to better understand market demand and product defects, so as to make more informed decisions and provide personalized services for each user. However, judging from the performance of each mobile phone manufacturer, the purpose of combining each with the large model is different.

For example, the purpose of OPPO's layout of large models is to create an interactive experience of intelligent terminals. Industry insiders said that with the introduction of large-scale model technology, the proportion of questions and answers about users interacting with voice assistants will increase significantly. At present, the entire monthly activity of the OPPO voice assistant Xiaobu is 1500 million+, tens of millions of interactions generated every day. In terms of deployment, it is mainly cloud + on-premise deployment, that is, it involves highly sensitive private data, which can be localized on the device side. However, in the face of some complex tasks, it is still necessary to rely on the cloud model.

Unlike OPPO, when Xiaomi lays out large models, the focus is on lightweight and local deployment. Compared with the former, Xiaomi's deployment method can better protect user privacy, and at the same time realize personalized customization for thousands of people locally. To this end, Xiaomi has delegated some of its large model capabilities to the device side, and is committed to achieving the best balance between power consumption, inference speed, and generation effect. In addition, it has also developed a 1.3 billion parameter device-side model, and the effect of some scenarios is comparable to the computing results of 6 billion models in the cloud.

In contrast, vivo pays more attention to solving the personalized needs of users, and vivo believes that the large model will comprehensively revolutionize the interactive experience of intelligent terminals, and continue to bring users high-quality experience in this direction. In terms of deployment, the 13 billion parameter blue heart model can achieve end-to-end running, and the 175 billion parameter blue heart model has reached the GPT-3 parameter level.

The combination of Huawei and the large model has a deeper plan, which intends to build the large model into the "brain" of the system. It is understood that the integration of Huawei's voice assistant Xiaoyi and the large model is not simply enhanced for tasks such as chatting, AIGC, and reply, but with the large model as the core, and system-level enhancements are carried out. The underlying logic is to assign the user's tasks to the right systems, each with its own role, while enhancing the experience in complex scenarios.

Overall, "reshaping" voice assistants is undoubtedly the first step for major mobile phone manufacturers to lay out large models, and it is also a necessary step. The difference lies in the different advantages of each mobile phone manufacturer, and the focus of the combination with the large model is also different. For example, Xiaomi has been deeply involved in the field of smart home for many years, so it has precipitated a large number of AI software and hardware foundationsOPPO, on the other hand, has always focused on technology and innovation, providing products for users who pursue high-quality shooting and design, which has been carried over on large modelsvivo takes young fashion as its main positioning, and pays attention to personalized functions such as ** and **, which makes it gather strength that cannot be underestimated in the personalized development of large models.

However, mobile phone manufacturers still face some difficulties if they want to use AI models to improve the performance and competitiveness of their products.

First of all, there are real-world challenges in fitting a large model into a mobile phone. According to public information, there are two main paths for mobile phone manufacturers to lay out large models: one is the end-side computing mode represented by Honor and Xiaomi;The second is the parallel mode of two paths on the device side and the cloud, represented by vivo. At present, the two models have the same goal, and in the end, they both hope that the AI model on the end of the smartphone will support local Q&A, voice interaction and other functions, and whether it can run through depends on whether the manufacturers can cooperate with the upstream to realize the evolution of computing power and algorithms. With the in-depth layout of more enterprises, they will also start technical competition and resource competition.

At the same time, putting large models with parameters of tens of billions or hundreds of billions into mobile phones also faces practical problems such as high cost, difficulty in self-development, and high consumption of computing power.

Secondly, from the application level, large AI models require a large amount of data to be trained to achieve the desired effect. However, it may be difficult to obtain and use a large amount of user data due to reasons such as privacy protection. At the same time, mobile phone manufacturers also need to solve the problem of insufficient model generalization to ensure that the performance of the model in different scenarios can meet the requirements.

Finally, from the competitive level, major mobile phone manufacturers are increasing investment in the field of AI large models, and the competitive pressure is very high. For the manufacturers involved, how to find another way to create differentiated products and avoid homogenization has become the meaning of each manufacturer's participation in the large-scale model competition.

Who can get out of the encirclement first?So, in this large-scale model competition in which mobile phone manufacturers participate, which mobile phone manufacturer has a better chance to win the opportunity and finally break out of the encirclement?To answer this question, it is also necessary to answer it in a combination of many situations.

First of all, from the perspective of technical strength, Huawei's large model is currently the strongest, which may be of great help to its promotion of the development of mobile phone side large model. According to public information, Huawei's Pangu large model is currently large, may be at the level of hundreds of billions or even trillions of parameters, and it has been connected to Huawei's Hongmeng OS and Xiaoyi smart assistant, and may realize large model drive and intelligent interconnection from mobile phones, computers to cars in the future.

In contrast, at present, Xiaomi, OPPO, vivo, Honor and other mobile phone manufacturers, most of the announced end-side large models range from billions to tens of billions, which is still quite different from Huawei in terms of magnitude. Although from the perspective of practical application, the corresponding applications of large models of different magnitudes are different, and they cannot be generalized. However, in terms of underlying strength, whether it is the cloud, operating system, or large model training ability and level, Huawei is unique among domestic mobile phone manufacturers.

Secondly, from the perspective of the ** chain level, Xiaomi, OV, Honor and other manufacturers may be able to get more support at the chip level. At present, with the introduction of AI large models into the field of mobile phones, the bottlenecks of existing mobile phone computing power and storage will gradually become prominent, and the promotion of large computing power chips has become a key step to crack the landing of mobile phone side large models.

It is understood that at present, Qualcomm, MediaTek and other mobile phone chip companies have begun to layout. On October 24, Qualcomm released the third-generation Snapdragon 8, which can run a large model with more than 10 billion parameters on the device side, and a large oracle model with 7 billion parameters can generate 20 tokens per second, which only needs 0In 5 seconds, you can use Stable Diffusion to complete the Wensheng diagram. MediaTek is naturally not to be outdone, on the evening of November 6th, MediaTek released the seventh-generation AI processor APU790 designed for generative AI, with a built-in hardware-level generative AI engine, the processing speed is 8 times that of the previous generation, and it can achieve more high-speed and secure edge AI computing.

As a downstream manufacturer, this means that Xiaomi, OV, Honor and other manufacturers will have the opportunity to use the power of upstream chip manufacturers to go to the front in terms of high-performance chip customization. After all, limited by the sanctions imposed by the United States, Huawei can only cooperate with domestic manufacturers if it wants to make high-performance chips, and it is unknown whether it can be successfully produced at present.

From this point of view, domestic mobile phone manufacturers still have shortcomings to make up for in a long period of time. Huawei needs to work with domestic manufacturers to build high-performance computing chips and ensure their smooth outputFor other vendors, they need to accelerate the improvement of computing power and algorithms to ensure that their technology and user experience are leading enough.