Source: Financial Story Collection

Image Source: Generated by Wujie AI

OpenAI may be making a phone, and it may be an AI phone designed specifically for ChatGPT.

Recently, according to The Information, OpenAI CEO Sam Altman has been in contact with Jony Ive, the famous designer of the iPhone, discussing the development of a new AI hardware device.

SoftBank CEO Masayoshi Son is also very interested in this and has "discussed this idea with both of them."

Although there is no conclusion on the three-party cooperation, given Masayoshi Son's identity as a "heavy user" of ChatGPT, this matter seems promising.

On one hand, AI companies are preparing to enter the hardware field; on the other hand, mobile phone manufacturers are also embracing AI large models.

On October 4th, Google released the Google Pixel 8 series phones, which are equipped with Google's AI basic model and are "AI-centric" — considering that Google's initial release of the Android system once led to the leap from feature phones to smartphones, Google's entry this time may herald the beginning of AI phones.

Domestic mobile phone manufacturers are also coming in droves.

On October 11th, OPPO announced that the new Xiao Bu assistant 1.0 Beta version, built on its self-trained AndesGPT, has officially started the experience of upgrading to the new large model. After the upgrade, Xiao Bu assistant will have AI large model capabilities.

This is another mobile phone manufacturer to start the internal testing of the built-in AI large model smartphone intelligent assistant after Xiaomi announced the internal testing of the built-in AI large model Xiao Ai.

Huawei has taken even bigger steps, from announcing the integration of the Pangu large model into the mobile phone system, to starting the internal testing of the built-in large model voice assistant Xiao Yi, and releasing the new Huawei Mate60 series with the built-in AI large model, all in just over a month.

Although vivo has not made actual progress in the application of large models, its self-developed 7 billion parameter large model vivoAgentLM_7B appeared on lists such as C-Eval in August; and the official also announced the news of the self-developed large model release at the International Science and Technology Innovation Forum of the Boao Asia Forum in September.

Honor is even more urgent. In early July, it released the so-called "world's first domestically produced mobile phone with integrated AI large model" Magic V2, which attracted a lot of attention; but there has been no progress in its self-developed large model, and instead stated that it is "cooperating with Internet companies in the field of network large models."

…

The urgent mood of domestic and foreign mobile phone giants to quickly implement AI large models is obvious. This is mainly because the mobile phone market is lacking in innovation and is really struggling; and "putting large models into mobile phones" is likely to achieve a revolutionary breakthrough in mobile phone functions and experience, and may trigger a new wave of phone upgrades.

I. Will AI phones help alleviate market saturation?

Why are domestic mobile phone manufacturers so eager for AI large models?

The most direct value is the improvement of mobile phone performance and experience.

When AI large models are integrated into the mobile phone system, on the one hand, they can enhance the ability of the phone to process tasks such as images, speech, and natural language processing, significantly improving the performance of the phone; on the other hand, they can more flexibly respond to user needs, improve user experience, and as AI algorithms evolve, the large models loaded on the client side deepen the understanding of users, allowing the phone to predict user behavior and optimize in advance, becoming smarter and more intelligent, changing from "one phone for a thousand people" to "a phone for a thousand people."

Deploying AI large models locally on the phone can not only optimize and empower various APPs with AI, but can also serve as a universal interface, breaking down barriers, allowing APP capabilities to be freely combined, thereby improving user experience and product competitiveness.

The Huawei Mate60 series and the Google Pixel 8 series, which were the first to implement AI large models domestically and abroad, are examples.

Huawei has long applied AI capabilities to enhance image processing on its phones. After the AI large model was integrated into the system this time, the most obvious improvement in experience was its phone voice assistant Xiao Yi.

Different from OpenAI's ChatGPT, Google's Bard, and other generative AI chatbots, Huawei's Xiao Yi has been restructured at the bottom of the HarmonyOS operating system.

While the former two are just AI applications loaded on the phone, Xiao Yi uses the large model as the brain of the intelligent system, enhancing the overall capabilities of the system with the advanced capabilities of the large model.

With the support of the Pangu large model, Xiao Yi not only achieved a leap in natural language understanding capabilities, but also used generative AI to do many things for users, such as summarizing information, writing articles, replying to emails, writing memos, and creating image derivatives.

It not only meets daily needs, but also assists in work. For example, when inviting partners to a meeting, directly dictating the time, invitees, and discussion topics to Xiao Yi, it can write an invitation email.

More importantly, Xiao Yi has indeed broken down the barriers between apps and can handle more complex user scenarios.

For example, for a poster image, asking Xiao Yi for related address information and requesting navigation to the destination, Xiao Yi can accurately identify the address information in the poster and open the navigation app for navigation.

Similarly, if you want to find a steak restaurant, Xiao Yi can accurately understand the intent and directly use Meituan's service to find the relevant restaurant.

The enhancement of image processing capabilities on the Google Pixel 8 series phones is also impressive.

Through generative AI, the Magic Editor in Google Photos can use already taken photos to manually select another expression from other photos to generate a new photo, and can ensure that "everyone has their eyes open."

In addition, this feature can also remove unnecessary backgrounds and clutter from photos, allowing "clumsy" users with high-end phones to become great photographers.

The improvement of phone functions and experience by AI large models is obvious and can meet user needs, which may stimulate a large number of users who are eager to try new things to upgrade their phones, breaking the homogenization and saturation dilemma in the phone industry.

According to the latest report released by the market research firm Canalys, as of the second quarter of 2023, the shipment volume of the Chinese smartphone market was approximately 64.3 million units, a 5% decrease from the same period last year; the global smartphone market is also similarly lackluster, with a 10% year-on-year decrease in shipment volume in the second quarter of this year.

In the declining market, the high-end market is relatively popular. Canalys analyst Amber Liu pointed out, "Consumers are increasingly willing to pay for high-quality products. Since last year, the average price of smartphones in the Chinese market has exceeded $450, and is expected to continue to rise in the coming quarters."

If AI phones can significantly improve the user experience, it is possible to reactivate the domestic and foreign smartphone markets.

Haitong Securities pointed out in its special research titled "Will AI Drive a New Wave of Phone Upgrades?" that with the advent of the AI large model era, AI is expected to become an important factor driving the next wave of phone upgrades. It is expected that the first batch of "AI phones" will be launched one after another in the second half of 2023. Currently, AI phones mainly apply NLP (Natural Language Processing) and CV (Computer Vision) technologies, and it is expected that more innovative AI applications will land on phones in 2024-2025.

During the second quarter performance meeting in 2023, MediaTek also expressed a similar view, believing that "AI-enabled phones will accelerate the replacement cycle."

II. "Seizing" Large Models, Varying Progress of Manufacturers

In fact, the rush of the mobile phone industry to seize AI large models is not just following the trend. Leading manufacturers have been laying out AI for some time, and the progress has been gradual, starting more from the perspective of functional improvement in the early stages, and now gradually delving into the system level.

Although Huawei announced the integration of large models into the phone system only when it released the HarmonyOS 4 system in August this year, the Pangu large model had already started development in September 2020 and has now iterated to version 3.0 and is already in commercial use.

OPPO, which also laid out AI large models early, began exploring and applying large language models with the Xiao Bu assistant team in 2020. Even before the AndesGPT large model project, OPPO had already made a lot of AI investments, developing and exploring pre-trained language models, independently developing the OBERT large model, which ranked fifth in the overall ranking of the Chinese Language Understanding Evaluation benchmark CLUE1.1, and the large-scale knowledge graph question-answering model KgCLUE1.0, as well as ranking first in the rankings.

While the timing of vivo's efforts in AI large models is not known, the establishment of the AI Global Research Institute in 2018 and the submission of the 7 billion parameter large model vivoAgentLM_7B to authoritative evaluation websites in August this year indicate that they have been involved in related research and development for some time.

Xiaomi also entered the AI field as early as July 2016 and officially formed the AI Lab Large Model Team in April this year.

As Lu Jian, head of the AI Lab Large Model Team at Xiaomi's Technical Committee, said, "The combination of mobile phones and AI has been around for a long time, such as adjusting photos in photography—quick focus, super-resolution, etc. Now adding large models to phones is an upgrade, improving natural language interaction, including text processing capabilities, multimodal processing capabilities, and more."

The reason why domestic mobile phone manufacturers are eager to push their own AI large models to the forefront in the past three to four months, and even integrate them into phones, is to seize the window of opportunity that AI phones may bring about a wave of upgrades.

The acceleration of the implementation of AI large models by mobile phone manufacturers is also to some extent a "self-rescue" in response to weak consumer demand.

According to the aforementioned content, except for Honor, which has not provided any updates on the progress of its self-developed large model, most domestic mobile phone giants have chosen to develop their own AI large models. However, due to their respective limitations and different start times, the progress of implementation varies.

Huawei, which has built its own ecosystem, is currently the first domestic mobile phone manufacturer to integrate AI large models into new phones.

Next is OPPO and Xiaomi. Although they have not officially released new models with built-in AI large models, they have respectively started public testing and internal testing of their new smartphone assistants with built-in AI large models, taking steps to integrate AI large models into the phone system.

OPPO has overall considerations and has prepared in advance for coordination with upstream manufacturers. While independently training the AndesGPT large language model, OPPO is also cooperating with MediaTek, hoping to use its 4-bit quantization technology to achieve better performance at the end while maintaining accuracy, and to quickly promote the lightweight implementation of AndesGPT at the end.

Although vivo has placed its self-developed large model on authoritative rankings, there has been no specific implementation behavior to date.

It is rumored that vivo will launch the new OriginOS 4.0 system with a built-in AI large model around October this year. If a new phone with a built-in AI large model system is launched at the same time, vivo is expected to become the second domestic mobile phone manufacturer to implement AI large models in new phones after Huawei.

Unlike Huawei's integration of software and hardware, other mobile phone manufacturers, while developing their own large models, will inevitably have to cooperate with upstream hardware manufacturers in the future to better promote the implementation of their self-developed AI large models on the phone side, with more variables.

To some extent, whoever can better coordinate with hardware manufacturers and accelerate the lightweight implementation of their self-developed large models will be able to seize the initiative more quickly.

III. Self-developed Large Models: Will They Eventually Move Towards Integrated Cloud and End?

Another question arises: why don't mobile phone manufacturers take a pragmatic approach and directly use open-source interfaces of mature cloud-based large models, instead of personally developing end-side large models?

The reasons can be summarized in two key words: unable to carry and unable to move.

On the one hand, specific applications of large models deployed in the cloud, such as ChatGPT, have always been accompanied by strong privacy and security controversies.

Local deployment of mobile phone AI large models ensures that data does not leave the end side, providing more security for user data privacy and security.

Because the parameter volume is relatively lightweight, the loading and running speed of large models on phones will be faster and not limited by network environment.

In addition, the training cycle of lightweight large models deployed on phones is short, allowing for rapid iterative updates based on user needs and more flexible response to user demands.

Haitong Securities also asserted in the aforementioned research report that unlike the high computational performance requirements in the training phase, the inference phase mainly uses the pre-trained model for inference and prediction based on user needs, with lower requirements for peak computing performance, and more emphasis on comprehensive indicators such as unit energy consumption, latency, and cost.

As cloud-based large models require network channel transmission between cloud computing power and terminal devices, only distributing inference computing power to the cloud will be unable to meet the low-latency and high-reliability requirements in some scenarios due to limitations in network bandwidth and transmission distance.

On the other hand, for cloud-based large models, as the demand for access and usage frequency increases, the consumption burden on cloud computing power, network bandwidth, storage, and hardware resources also increases.

If mobile phone manufacturers load the driving force of the mobile phone performance revolution onto these "bulky" cloud-based large models, the stability of the mobile phone performance experience is easily compromised, and may even be dragged into a quagmire.

Regarding the issue of self-developed end-side large models, Lu Jian, head of the AI Lab Large Model Team at Xiaomi's Technical Committee, provided insights from the perspective of mobile phone manufacturers in an interview with Tencent Technology.

Regarding self-development, Lu Jian believes that different chips are used on various terminal devices, and they differ in memory size, computing power, and supported operator sets. This requires the model to dynamically adjust according to hardware conditions to achieve optimal performance.

An open-source model has a fixed structure and cannot be further adjusted, so its use is very limited. If you want to have the ability to customize model structure and train from scratch, you must develop it yourself.

This also explains why mobile phone manufacturers not only develop their own large models, but also need to strengthen coordination with hardware manufacturers.

Lu Jian also pointed out that end-side large models can better protect user privacy and allow users to obtain more functions at a lower cost, but this does not mean that all problems can be solved on the mobile side.

Although hardware manufacturers such as Qualcomm and MediaTek can ensure the improvement of mobile phone energy efficiency and performance through continuous improvement of quantization algorithms for lightweight implementation of end-side large models, it cannot be denied that no matter how much AI large models are compressed, the parameter volume still poses increasing demands on mobile phone hardware and performance.

The aforementioned research report by Haitong Securities also pointed out that the extensive computation of AI models will place higher demands on bus bandwidth; continuous running of inference tasks will accelerate device power consumption, driving the need for increased battery capacity and corresponding upgrades to power management chips.

As large models become more accurate in their predictions, these issues will become more apparent and may drive the upgrade of the entire industry chain.

So, in addition to continuously optimizing quantization and other technologies, and improving phone configurations, what other solutions do mobile phone manufacturers have to effectively address the higher requirements brought about by the evolution of end-side large models?

In this regard, Xiaomi's proposed concept of "integrated cloud and end" may be a practical choice. That is, in the future, for AI phones, capabilities or functions that can be addressed by end-side models will be handled at the end side; those that cannot be addressed at the end side will call on cloud capabilities. This approach can ensure both privacy and security, precise understanding of user needs, flexible response, and the realization of more complex and advanced operations to optimize the user experience.

Qualcomm has also proposed a similar approach to Xiaomi.

In the technical white paper "Hybrid AI is the Future of AI," Qualcomm points out that with the unprecedented development speed of generative AI and the increasing demand for computing, AI processing must be carried out separately in the cloud and at the end to achieve the scalable expansion of AI and maximize its potential.

Huawei's current solution is also highly similar to the previous two.

It is understood that, in response to consumers' needs in different device and scene scenarios, Huawei's Xiaoyi, backed by large models, has both end-side and cloud-side forms. In the combination of Xiaoyi and large models, the end-side large model will first preprocess user requests and context information, and then send the preprocessed requirements to the cloud side, thereby maximizing the advantages of "fast end-side model" and "strong cloud-side model."

The trend of AI is already here, and as the most widely used and highly sticky smart terminal for users, the mobile phone may become the first carrier for the explosion and widespread adoption of AI.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。