The large model of Xiaomi was exposed for the first time, and Huawei Xiaoyi handed in the paper first. Is the GPT time for mobile phones approaching?

Source丨Smart Things

Author | Yunpeng

Edit | Heart Fate

The big model battle among mobile phone manufacturers is about to come.

Just now, the Xiaomi large model suddenly appeared on the screen, and achieved the tenth place and the first place in Chinese in the two large model test platforms of C- and CMMLU respectively. Its C- rank is ahead of Alibaba Cloud’s Tongyi Qianwen .

▲C-list

You should know that C- and CMMLU are currently recognized in the industry as authoritative Chinese large-scale model benchmark tests, mainly examining the comprehensive knowledge reserve and language understanding ability of the large-scale model in the Chinese field.

▲CMMLU evaluation list

Just last week, Huawei's voice assistant Xiaoyi also integrated some of the capabilities of its own Pangu model. It has become a reality to use the voice assistant to write article summaries, meeting invitation emails, or use your own photos for personalized design.

Domestic mobile phone manufacturers have a posture of "If you don't make a sound, you will be a blockbuster" on the track of large models. One has already landed in the application, and the other has swept the screen for the first time.

Previously, no matter whether it was self-developed chips or fast charging, mobile phone manufacturers seemed to be the "vanguards" who pushed "new technologies" to the consumer market, allowing the general public to come into contact with new technologies.

In the wave of generative AI based on large models, mobile phone manufacturers are bound to usher in a new battle.

Overseas, Google and Apple have already begun to "mobilize" their own intelligent voice assistants, brewing the application of large-scale models. In China, before the results of Xiaomi's large-scale models were announced, Wang Bin, director of Xiaomi's AI laboratory, had already told the outside world about Xiaomi's large-scale model. Regarding the planning and progress of the model field, Xiaomi CEO Lei Jun and Xiaomi President Lu Weibing have also talked about the large model of Xiaomi and related layouts in public on many occasions.

For Honor, its CEO Zhao Ming mentioned that Honor has cooperated with Internet companies on the demand for large-scale network models, and OPPO and vivo, which seem to be low-key, have actually been in the AI field for many years. Ranked among the best in the test, and has cooperated with some big AI companies. On Huawei's side, Xiaoyi has already landed.

Although the waves are calm on the surface, the battle of the large models among mobile phone manufacturers is imminent. Every family builds plank roads secretly, and a fierce game of AI technology may be about to be staged.

Xiaomi formed a large model team, and Honor OV may adopt the "self-research + cooperation" model

Mobile phone manufacturers use large-scale models in two ways. One is to make large-scale models for their own use, and the other is to use other people's large-scale models.

At present, Huawei and Xiaomi are one step ahead and use it for their own use. Honor, OPPO and vivo have no official information about the layout of large models. Among them, OV and other manufacturers have relevant cooperation information on large models. However, it is still unclear how the two will be adopted.

First of all, let’s take a look at Xiaomi who is swiping the screen today. In fact, Xiaomi’s intelligent voice assistant, Xiao Ai, should be said to be the most famous among the voice assistants of various companies, and it also has the widest range of users. Almost all kinds of IoT devices of Xiaomi have Access to Xiao Ai, and Xiaomi has the largest number of IoT ecological devices among all smartphone manufacturers. Xiaomi has made it clear that the Xiaomi AI model may be combined with Xiao Ai in the future.

Whether it is in the earnings call or in some public interviews, relevant Xiaomi executives have expressed positive views on the large model, and explained in detail Xiaomi's layout and planning for the large model.

In April of this year, Xiaomi CEO Lei Jun personally issued a document stating that Xiaomi will resolutely embrace large-scale model technology. In the earnings conference call the following month, Xiaomi President Lu Weibing announced that the company has established an AI laboratory large-scale model team, with more than 1,200 people in the AI field.

The head of Xiaomi’s large model team is Luan Jian, who reports to Wang Bin, director of Xiaomi’s AI Lab. Wang Bin joined Xiaomi in 2018 after conducting research on NLP (Natural Language Processing) at the Chinese Academy of Sciences for more than 20 years.

In an interview with Shenran, Wang Bin mentioned that their team's goal is a general-purpose large language model with a parameter scale of tens of billions, and the equipment investment for training is tens of millions of RMB. The Xiaomi large-scale model landing products will adopt a "hybrid model", and the traditional model and large-scale model each solve the problems they are good at.

According to Wang Bin, before ChatGPT, Xiaomi had done large-scale model-related research and applications. However, the scale of the model is in the billions, and it is not a general-purpose large-scale model. It is mainly a dialogue-specific model for man-machine dialogue.

On Xiaomi's side, executives frequently disclosed information, and on Honor's side, its CEO Zhao Ming also revealed Honor's views on generative AI and large-scale models in interviews.

Zhao Ming mentioned at the Shanghai Mobile World Congress that Honor is cooperating with Internet companies on the demand for large-scale network models. At that time, they were already in contact with interested companies.

At present, Baidu’s Wenxin Yiyan, Ali’s Tongyi Qianwen, and Xunfei’s Xunfei Xinghuo are all three-party large-scale models launched by major domestic Internet companies and AI companies. Self-developed large-scale models are very important for companies that have just been established for three years. For Honor, it is not the most important thing. It is obviously more critical to increase market share and shipments. Therefore, cooperation may be the way for Honor to apply large-scale technology.

On the OPPO side, Liu Bo, President of OPPO China, mentioned in an interview that OPPO is thinking about the application of large models on mobile phones.

In April this year, Alibaba Cloud announced that it will jointly build OPPO’s large-scale model infrastructure with OPPO Andes Smart Cloud. Based on Tongyi Qianwen, it will complete the continuous learning, fine-tuning and front-end prompting project of the large model, and build AI services for OPPO end users.

Judging from the example of Huawei Xiaoyi, it is feasible to fine-tune and optimize Tongyi Qianwen to make a lightweight model that can be used in OPPO's smart voice assistant.

However, relevant sources from Xiaomi revealed that OPPO and vivo may also be making their own large models.

In this regard, we can also see some signs from some previous actions of OV. For example, OPPO's Xiaobu assistant team has been conducting a lot of research in the field of AI technology, including speech recognition, semantic understanding, dialogue generation, knowledge question answering system, open domain chat, multi-modality, etc., and these are all related to generative AI. key technologies.

Xiaobu’s assistant team has explored and implemented pre-training models before, and developed pre-training models OBERT with 100 million, 300 million, and 1 billion parameters. The fifth place in the list, and the first place in the large-scale knowledge map question and answer KgCLUE1.0 list.

At the OPPO Future Technology Conference last year, the generative AI technology was used in Xiaobu's painting function, which can create pictures through user descriptions and uploaded pictures.

For vivo, its AI team developed a text pre-training model 3MP-Text for natural language understanding tasks in May this year. On the CLUE list of the Chinese language understanding evaluation benchmark, 3MP-Text got the same ranking as the 100 million parameter model effect. Size first.

The large model lands on the mobile phone, and the intelligent voice assistant becomes the vanguard of early adopters

Mobile phone manufacturers are actively embracing large models, what are they going to do? One thing that has been determined so far is to use the large model in the smart voice assistants of various companies, so that the large model can become the "system-level" capability of your mobile phone, making the mobile phone more intelligent, and the assistants will no longer "Retarded".

Samsung is considering changing the default search engine of mobile phones and tablets from Google to Microsoft's new Bing, which supports AI chat. At the I/O conference in May, Google released four new-generation large language models PaLM 2 with different parameter scales, among which the smallest "Gecko" large model can run on mobile phones.

On Apple’s side, some foreign media revealed that it is developing a new AI function code-named "Bobcat" for Siri, and the technical framework of the new project is called "Siri Natural Language Generation". The integration of AI technology will also become inevitable.

The potential of large models in voice assistants on smartphones is obvious to all.

For consumers, we can clearly see from the existing cases of the combination of intelligent voice assistants and large models that the integration of large model capabilities solves one of the biggest problems in users' habit of using voice assistants - not enough Natural dialogue, natural communication that cannot be done at will.

To put it simply, it is to make the intelligent voice assistant change from fun and novelty to easy to use, and even become a "habitual action". The large model allows the intelligent voice assistant to truly understand and understand us, and the improvement in its ease of use is extremely significant.

In the view of some terminal manufacturers, the application of large models such as ChatGPT is more focused on creative copywriting, information organization, question and answer chat, article summarization, etc., but the positioning of voice assistants is "intelligent personal assistants", from device control, personality To improve the efficiency of our daily office work by providing personalized consulting services, the application of intelligent voice assistants in consumption scenarios should be more extensive.

At the same time, compared with generative AI chatbots such as OpenAI's ChatGPT and Google's Bard, intelligent voice assistants will become the "system-level" capabilities of terminal manufacturers, from voice dialogue, graphic and text recognition, service suggestions to device interconnection management.

Relevant AI experts told Zhishi that system-level capabilities mean that the system-level entrance is more closely integrated with the operating system, and the interconnection with the ecology has also reached the bottom layer of the system ecology. This kind of interconnection is truly efficient and the experience can be the best. Well, this interconnection is far from being comparable to the one-to-one SDK call between ChatGPT and App.

In addition, whether it is Huawei, Xiaomi, Honor or OV, these manufacturers have already launched a wide range of IoT business layouts, and intelligent voice assistants have become the key AI service portals connecting their smart devices. The integration of models quickly expands the capabilities of large models to their entire software and hardware ecosystem, which is also very important for manufacturers.

Is it difficult to stuff a large model into a mobile phone?

It is not difficult to think of using large models in intelligent voice assistants. Even from the first day ChatGPT appeared, all voice assistant companies have thought of this.

But the key is, how to achieve it? Is the cost proportional to the return it brings? A large language model like GPT-4, with hundreds of billions of parameters, is going to be used in a mobile phone with a power consumption of only a few watts. How to solve the technical challenges?

Regarding these questions, we may find some answers in the example of Huawei Xiaoyi mentioned above.

Generally speaking, to apply a large model on a smart voice assistant, at least two things must be done. First, optimize the general large model into a version suitable for voice assistants. Things worked out.

From Huawei's example, based on the Pangu L0 large model, Huawei has fine-tuned and optimized the data involved in the usual consumer scenarios, and built an L1 layer dialogue model, which is used in Xiaoyi.

For these consumer scenarios, manufacturers need to construct corresponding corpus data, design model outputs that the system can understand and execute, and at the same time input credible structured and unstructured knowledge to the large model, so that the large model can learn To general knowledge, logical relationship.

ChatGPT cannot help you set up your mobile phone or control various smart devices in your home, but voice assistants need to have such capabilities, which is also a very important function of smart voice assistants.

Therefore, manufacturers also need to achieve effective analysis and efficient connection between the large model and the system through technical optimization, and first "train and learn" the large model for complex scenarios, so that the large model can learn these manipulation skills, and finally compare the large model inference cost and inference time The delay problem is solved.

It is not enough to make a large-scale model version suitable for voice assistants. In order to solve the problems of power consumption and computing power, the combination of devices and clouds is also more important.

Nowadays, ChatGPT applications rely on cloud computing power, but when they are actually used in voice assistants, they involve the use and processing of users' personal information, which inevitably requires localized operation, but completely localized operation cannot solve the problem of insufficient power consumption and computing power .

Huawei has made different large-scale model versions, both on the terminal side and on the cloud side. According to different tasks, both sides are collaboratively processed.

As a mobile chip manufacturer, Qualcomm has been focusing on promoting their "hybrid AI" concept before. In fact, it means that the application of generative AI on the mobile side must involve the collaboration of the device side and the cloud side. Judging from the actions of all parties in the industry, this has basically become the consensus of the industry.

Of course, using a large model in a smart voice assistant is definitely not as simple as we mentioned in a few words. There are many technical and industrial challenges behind it, and we can get a little idea from the example of Huawei.

But having said that, although it is difficult, Huawei just proved the feasibility of this matter, and the application of large models in intelligent voice assistants can indeed bring about a lot of "qualitative changes" in capabilities.

The battle of large models is coming fiercely, and data, computing power, and talents are still the core focus of competition

Generative AI is sweeping thousands of industries, and the impact of large models on the mobile phone industry will be far-reaching.

For consumers, mobile phones have become more "smart" and more "efficient". We can finally enjoy the services provided by voice assistants in casual spoken language, such as some text and image generation capabilities on ChatGPT, and smart voice assistants are also I have learned that the use of large models for intelligent voice assistants is bound to be beneficial and highly anticipated by consumers.

For manufacturers, it is an inevitable trend for future development that smartphones and related IoT devices incorporate generative AI capabilities based on large models. The changes that large models bring to these businesses will be significant and valuable.

Whether it is to make large-scale models by itself or cooperate, every manufacturer has to pay attention to this battle of large-scale models.

Of course, for various smartphone manufacturers, the challenges brought by this wave are also obvious. To truly win this battle, there are many difficulties that need to be overcome.

Some people in the AI industry told Zhishi that for manufacturers who want to build their own large-scale models, the accumulation of data, computing power, and talents is indispensable, such as large-scale high-quality data acquisition and cleaning, and computing power How to overcome system-level challenges , Training how to make the cost controllable. For manufacturers adopting the cooperation model, how to ensure better end-cloud collaborative processing, how to balance costs and benefits, and explore business cooperation models will all be faced by them.

Conclusion: AI model, a tough battle for mobile phone manufacturers

Standing at today's node, the battle of mobile phone mockups has begun. The manufacturers who made the first move have already shown their cards, and the manufacturers who did not make a move are also in the process of brewing. The battle of mobile phone mockups is undercurrent.

Judging from the achievements of the existing voice assistant combined with the large model, it is obvious that the AI large model will enhance the experience of smartphones. The AI large model will also have a profound impact on the development of various business lines of mobile phone manufacturers in the future. The AI large model It will inevitably become the main development trend of the follow-up technology industry, and it will also become one of the key technology tracks that everyone pays attention to.

The AI model is undoubtedly a tough battle for mobile phone manufacturers, but it is still unclear who will bring breakthrough products or technologies that break the industrial structure or even subvert the existing gameplay.

Facing the future, the combination of large models and intelligent voice assistants will be closer. With the follow-up of various technology giants, the wave of "evolution" of intelligent voice assistants will be unstoppable. What new application scenarios, new application forms and functions will emerge in the future are full of imagination.