AI mobile phones do not have exclusive large model suppliers
2024-07-24
한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina
Involved in the big model mobile phone competition, Baidu, Alibaba, and Byte have become suppliers of mobile phone manufacturers, but AI mobile phones do not favor any one company and regard connecting to multiple big models as the mainstream choice.
@Technology News Original
Author: Gu Er Editor: Sai Ke
As the fire of the Hundred Model War subsided, the wave of AI hardware swept in.
Mobile phones, as a universal terminal product, have become the wave of the large-scale hardware landing wave. As for domestic mobile phone manufacturers, OPPO announced that 2024 is the first year of AI mobile phones, and Meizu said that it will officially stop new traditional smartphone projects, and both are all in AI. Apple, which has attracted much attention overseas, gave up its ten-year dream of making cars and turned to generative AI projects, betting all on the next generation of new phones, the iPhone 16 series.Prior to this, Samsung, Xiaomi, Honor, vivo and other mainstream mobile phone manufacturers have announced their AI technology layout and product releases sooner or later since last year. The recent Canalys report "The Attractiveness of AI Mobile Phones to Chinese Consumers" shows that in the first quarter of 2024, the Chinese mainland market, relying on the first-mover advantage of local manufacturers and the high-end market structure, will ship 11.9 million AI mobile phones, accounting for 25% of global AI mobile phone shipments.
The trend of AI mobile terminals has also attracted leading large model service providers to join in.
In May last year, Baidu's Xiaodu Technology announced that it would enter the smartphone market and planned to release its first smartphone integrating Xiaodu AI technology in late May.However, in the end, the outside world did not wait for a big drama of returning to the mobile phone market after 8 years, but the release of a "Xiaodu Qinghe" AI learning machine.ByteDance secretly launched an AI mobile phone R&D project in April this year, according to industry sources. The core team involved in the project mainly consists of the Hammer mobile phone R&D team acquired by ByteDance in 2019 and the PICO VR R&D team acquired in 2021. However, ByteDance later clarified that the information was untrue and was actually exploring large-scale model software solutions based on mobile phones for reference by mobile phone manufacturers.Alibaba is fully committed to selling big models. In March this year, MediaTek, a mobile phone chip supplier, announced that it had successfully deployed the Tongyi Qianwen big model on flagship chips such as Dimensity 9300, realizing the deep adaptation of big models on mobile phone chips for the first time. Tongyi Qianwen can still run multiple rounds of AI dialogue smoothly in offline mode.After many rumors about personally making AI phones, it is becoming a consensus among leading large-scale model suppliers such as Baidu, ByteDance, and Alibaba that it is more comfortable to be a shovel seller rather than competing in the mobile phone hardware market.However, what Baidu, ByteDance, and Alibaba did not expect was that when there were too many manufacturers selling large-scale model shovels in the market, it became a buyer's market. First, the market heard that Apple, the largest buyer of mobile phone AI, might join Baidu's large-scale model, which then boosted Baidu's stock price. However, it was soon followed by another rumor that Apple was also contacting other large-scale model suppliers at the same time.In July, the release of Samsung's sixth-generation foldable screen mobile phones Galaxy Z Fold6 and Galaxy Z Flip6 series fully demonstrated the buyer's market. In addition to the previous cooperation with Baidu Smart Cloud, Meitu, and Kingsoft Office, the Volcano Engine was added to connect the Doubao model to the smart assistant and AI vision of the latest foldable screen mobile phones.
Obviously, when big models want to be used on mobile phones, they come to the weak hardware side of big manufacturers. After weighing the cost and performance of various mobile phone manufacturers, an AI mobile phone that integrates the big model technology of Baidu, Alibaba, and ByteDance was born. As the technology providers behind the scenes, Baidu, ByteDance, and Alibaba, who previously fought overtly and covertly in the big model market, even showed a brief scene of shaking hands and making peace.Baidu is not the only company favored by Apple's AI phones
About a year after domestic mobile phone brands "enthusiastically embraced" AI phones, Apple's "AI phone" came later and cooperated with OpenAI. However, the 161 countries and regions currently supported by OpenAI do not include mainland China and Hong Kong, China. Therefore, Apple's AI partners in China have never been announced to the public.Earlier information mainly pointed to Baidu Wenxin Yiyan, and the latest media reports said that Apple also held talks with other Chinese large model companies such as Alibaba and Baichuan. It can be seen that Apple is in talks with many large model companies, and no large model manufacturer can become the only supplier of Apple AI.This is not only the case in China, but also in Apple's overseas AI cooperation. Although the cooperation with OpenAI is a landmark event for its entry into AI mobile phones, OpenAI is far from being Apple's only AI partner.After the WWDC conference ended in June, Craig Fedrighi, Apple's senior vice president of software engineering, said in an interview that he looks forward to integrating with other AI models in the future, such as Google's Gemini. In addition, American AI startup Anthropic is also expected to become an Apple service provider.In terms of cooperation and adoption of large model manufacturers, Apple is trying to copy the tricks of managing the mobile phone hardware supply chain: it is used to configuring two suppliers of similar level for each type of parts. On the one hand, it can make suppliers check and balance each other, and on the other hand, it is also conducive to controlling the scope of confidentiality and facilitating management. Under the premise of always preparing a Plan B of good level for itself, if the supplier's product or management level declines, it will be ruthlessly kicked out by Apple.However, some industry insiders believe that compared to Plan B of the hardware industry chain, which mainly relies on the replication of manufacturing processes, AI suppliers are not very replicable. Even if the code is exactly the same, different training data will lead to differences in the final AI performance. The reason for looking for multiple AI suppliers may be that Apple is not satisfied with the AI capabilities of any existing one.Ultimately, Apple's path to implementing AI phones may be similar to Samsung's: find a supplier of large models with advantages in a certain functional vertical field, and then integrate AI services into the mobile phone system layer.For example, the previous generation of Samsung AI mobile phones, the Chinese version of the Galaxy S24 series, chose a large model supplier. Many functions on the Galaxy AI are based on the AI applications of Baidu Smart Cloud Qianfan platform. In terms of image processing, Samsung chose to further deepen cooperation with Meitu, based on Meitu's self-developed AI visual model MiracleVision, to bring AI image modification, AI image expansion, AI image generation and other functions to the album app.By cooperating and communicating with more AI big model companies, Apple may also adopt this logic of selecting the best AI capability suppliers. This also means that the Apple market, which has more than 1 billion active users worldwide, will become the most intense AI performance arena for big model manufacturers.AI mobile phone that can eat a variety of food
Apple is not the first mobile phone brand to cooperate with large model companies.In January, Honor and Samsung started cooperation with Baidu. Samsung integrated Wenxin Yiyan into Galaxy AI, and Honor connected Wenxin Yiyan through YOYO, both using AI assistants to call big model capabilities. In May, at ByteDance's 2024 Spring Volcano Engine Force Power Conference, Volcano Engine joined hands with OPPO, vivo, Honor, Xiaomi, Samsung and ASUS to announce the establishment of the Smart Terminal Big Model Alliance.On June 12, ByteDance's "Volcano Engine" official account published a message saying that Xiaomi's artificial intelligence assistant "Xiao Ai" has reached a cooperation with Volcano Engine and connected to ByteDance's Doubao model. In the future, Xiao Ai will use the capabilities of the Doubao model to provide users with online search capabilities, especially real-time search results that are the same as the headlines, presenting users with comprehensive and timely responses.
Samsung is a representative of the benefits of the big model. Taking the latest sixth-generation folding screen mobile phones Samsung Galaxy Z Fold6 and Galaxy Z Flip6 released on July 17 as an example, it has further expanded its AI suppliers. In addition to the cooperation with Baidu Smart Cloud, Meitu, and Kingsoft Office, it has also added the Volcano Engine to connect the Doubao big model to the smart assistant and AI vision of the latest folding screen mobile phones.As for domestic mobile phone manufacturers, according to market public data statistics, Xiaomi, OPPO, vivo, and Honor have all accessed the large-scale model products of three major manufacturers, Ali Tongyi, Baidu Wenxin, and Byte Doubao, to varying degrees. However, before that, the past 2023 was still a year for mainstream mobile phone manufacturers to fully develop large models on their own.First, Huawei connected its own Pangu big model to mobile phones, allowing mobile phones to perform complex tasks such as text generation, knowledge search, data summarization, intelligent arrangement, fuzzy and complex intent understanding. Later, on the evening of August 14, Lei Jun revealed in his annual speech that Xiaomi had established a big model team in April this year, and the big model on the mobile phone side has also been initially run.Almost at the same time, OPPO also announced that the new Xiaobu Assistant based on AndesGPT will soon launch a large-scale experience event. AndesGPT is a generative large language model based on hybrid cloud architecture created by OPPO Andes Intelligent Cloud Team. Soon after, vivo launched the Blue Heart large model, which also focuses on lightweight and is conducive to localized data processing on mobile phones; Honor launched a self-developed end-side 7 billion parameter platform-level AI large model.The mobile phone manufacturers who just announced their full self-developed large-scale models in 2023 turned around and became the large-scale mobile phone partners of Alibaba, ByteDance, and Baidu in 2024. From self-development to open third-party cooperation, the essence is the consideration of the balance between cost and performance behind the large-scale mobile phones.
On the one hand, in terms of hardware performance, Zhuang Shirong, senior manager of AI technology in MediaTek's wireless communications division, previously revealed: A large AI model with 13 billion parameters requires about 13GB of memory to run, while a smartphone itself usually takes up 4GB of memory to run the Android operating system. If other regular APP tasks need to run relatively smoothly and keep them alive, 6GB of memory is required, which means that the total mobile phone memory capacity requirement will reach 23GB.
Although directly increasing the memory is the simplest way to some extent, it is also related to many factors such as technology and market. Currently, the memory capacity of mainstream thousand-yuan smartphones is mostly around 8GB/12GB, and the memory capacity of most flagship smartphones is only 16GB, and only a few flagship phones have a memory capacity of 24GB. How to lower it to a product price below 2,000 yuan, bring greater market demand for replacement, and further break through the hardware limitations of the cloud-based large model has become an option.On the other hand, the implementation of large models is a trend that consumes costs over a long period of time and cannot be stopped for the time being. Whether it is the power consumed by the server or related research and development, it requires continuous investment. Algorithmic trading company Deep Trading once pointed out that ChatGPT4 has a daily cost of up to $882,000 based on a user base of 180 million and 100K tokens per person per day.Previously, Vivo Vice President Zhou Wei also revealed that the cost of a single cloud conversation of Vivo large models is about 0.012-0.015 yuan. Based on 250 million daily active users and 10 times per day per person, the daily cost is as high as tens of millions of yuan. There is no doubt that such cost expenditure will also become a cost pressure for mobile phone manufacturers on the road to popularization of large models.
After balancing the comprehensive cost and performance, the current mainstream mobile phone AI performance implementation path usually adopts a "combination punch" model of end + cloud collaborative integration.
Among them, the so-called self-developed AI big models of mobile phone manufacturers are mainly concentrated on the terminal big models, and their main feature is to deploy AI training models based on local hardware terminals, so as to achieve faster performance response and better data privacy protection. In terms of cloud big model performance, it is mainly implemented through cooperation with external big model manufacturers.Judging from the choice of third-party AI partners by mobile phone manufacturers, access to multiple models is the mainstream trend. This includes not only the diverse models of the same large model company, but also the models of different large model companies. For example, OPPO uses Doubao Universal Model Pro, Doubao Universal Model Lite, Doubao Role Playing Model, etc. of Volcano Engine; while Samsung has both Wenxin Yiyan and Doubao in its collection.This is also in line with the current concept of large model selection. Faced with diverse business types and user needs, it is difficult for a single large model to cover everything. It is more practical to mix models with different areas of expertise in different scenarios.AI mobile phone entrants who each get what they need
When mainstream mobile phone manufacturers choose the best among cost and performance considerations and access as many big model technology providers as possible, this also means that Baidu, ByteDance, and Alibaba cannot become the exclusive supplier of big models for a certain brand. The Internet thinking such as "choose one of two" and scale effect under the long-term inertia of platform economy will also face new challenges.On the one hand, it is different from the model of hardware chip partners. Usually, Qualcomm chips for mobile phone processors and Intel chips for PC processors can provide technical endorsement for hardware products. However, in terms of AI technical capabilities, for mobile phone manufacturers, as a brand facing the C-end market, what is more important than the implementation of technology is the battle for user minds.Therefore, in terms of cooperation, although each AI mobile phone manufacturer tries to connect to as many third-party big models as possible, it is a consensus to integrate the strengths of the underlying technologies of hundreds of third-party big models and develop multimodal AI agents on their own. Ultimately, for consumers, only smart assistants from mobile phone brands such as Xiaobu and Xiaoai can have AI mind perception.
On the other hand, with the huge user base of the leading mobile phone brands, the pressure on large model companies will naturally increase, and the computing power costs they bear may increase exponentially. It is hard to say whether this part of the cost can be passed on to mobile phone manufacturers. What is clear is that Apple will not pay fees to large model manufacturers. According to media reports, at least in the cooperation with OpenAI, Apple did not pay any fees.
Even though other mobile phone manufacturers do not have as much initiative as Apple when cooperating with large model service providers, the use of AI functions of various companies has not yet started charging services, so the AI costs of free users should not be borne by mobile phone manufacturers alone. Some technical experts said that a more reliable solution may allow mobile phone manufacturers and large model service providers to share the costs.
Mobile phone manufacturers cannot earn AI service fees, but they can earn users’ minds. The key is to be the first to occupy users’ minds in this round of big model competition and lay the foundation for AI mobile phones to convert free users into paying users.
For Baidu, ByteDance, Alibaba and other large model manufacturers, although the identity of a shovel seller does not make them earn more money, and even with the pressure of connecting to more mobile phone brands, the computing power cost they need to bear will also increase greatly. However, when connecting to more mobile terminals, entering the hands of more users, and getting new tempers, this is a good landing practice for the products of Baidu, Alibaba, ByteDance and other large model manufacturers.In this process of cooperation, with the "hardware enclosure movement" of large models time and time again, large model companies will also get what they want, more data that can be used for model training.existWitnessed by countless mobile phone users, this will be a real head-to-head competition in the field of big models. The true capabilities of each big model will be further "exposed to the sun", and the reshaping of the pattern and the market will also accelerate. Whether it is mobile phone manufacturers or big model companies, it is an AI process that each party needs.
