Original: Bai Jiajia

Source: Silicon Research Lab

Putting large models into smartphones has become the favorite story of smartphone manufacturers this year.

Huawei, Xiaomi, OPPO, vivo, and Apple have all written essays on the same topic of "installing large models". Among them, Huawei's intelligent assistant Xiaoyi, which was the first to land, can understand and execute user needs without the user accurately stating the function name, based on intuition. Some industry insiders believe that the market structure of smartphones will be restructured, and brands that can maturely apply large models will take the lead.

However, there are also many skeptical voices in the market, believing that this is just a marketing concept, and that the focus has been shifted from AI chip terminal computing to the hot concept of large models. Another group of people assert that current smartphone hardware simply cannot run large models while ensuring battery life and phone endurance.

So, is the large model in smartphones just a gimmick or a technological revolution?

The crossroads of large model installation: Cloud or Terminal?

Although all smartphone manufacturers have written essays on the same topic of "installing large models", they have different styles, with the biggest difference being whether to integrate cloud-based large models into smartphones or deploy large models on the smartphone side.

In terms of progress, Huawei is undoubtedly the fastest.

From the release of HarmonyOS 4.0 in early August, which first publicly revealed that it would be empowered by large models, to the end of August when Xiaoyi, the intelligent assistant empowered by Pangu large model, started recruiting testers, and then quietly launched the Mate 60 series smartphones on September 15, integrating the Pangu artificial intelligence large model. In a little over a month, Huawei fulfilled its promise.

According to reports, the large model behind Xiaoyi is based on the Pangu L0 base large model of Huawei, and has built a large amount of scenario data and fine-tuned models for the terminal consumer scenario, resulting in the L1 layer dialogue model.

Based on Pangu's natural language large model, visual large model, and multimodal large model, Xiaoyi has been enhanced in three directions: interaction, productivity improvement, and personalized services, and has achieved interface calls and matching of user intent.

What does it mean to match interface calls and user intent?

For example, if you want your secretary to fetch a document for you. In the past, with artificial intelligence, you had to tell it which room, which cabinet, and which drawer the document was in, but now you only need to tell it which document you need.

In other words, users do not need to accurately state the function name, based on intuition, Xiaoyi can understand and activate the corresponding function.

In terms of technology, there are two points worth noting about Huawei's large model installation.

The first point worth noting is that, according to market reports, Xiaoyi's language ability is based on natural language processing (NLP), not large language models (LLM).

NLP and LLM are two interrelated concepts. According to academic definitions, LLM is a branch of NLP, and their goal is to enable computers to understand and process human language.

The reason why this point is worth noting is mainly because Huawei is not just a smartphone manufacturer, but a comprehensive enterprise spanning multiple fields including communications, smartphones, and enterprise services. This comprehensiveness means that Huawei has a stronger demand for iterative basic large models, because the more outstanding the capabilities of basic large models, the lower the cost of fine-tuning industry-specific large models, and the better the performance of subclass large models.

To some extent, integrating large models into smartphones is just a relatively special aspect of Huawei's "empowering thousands of industries with large models" process. What really determines the future of the enterprise is still the basic large model behind it.

From this perspective, Google is actually competing on the same dimension as Huawei.

At Google I/O developer conference in May this year, Google released the Pixel smartphone. As one of the largest cloud service providers, Google also has rich accumulation in the field of artificial intelligence, and has launched the Bard plugin, and recently announced its official integration into applications such as Gmail, Docs, Google Maps, and YouTube, with the multimodal large model Gemini in the works.

Among domestic cloud service providers, Baidu has also thrown a probe into the smartphone field, launching the Xiaodu Qinghe Learning Phone in May this year, equipped with the Wenxin large model.

The second point worth noting for Huawei, as well as these players, is that they have all "integrated" large models, rather than running them independently on the smartphone side.

In fact, whether to deploy in the cloud or on the terminal has always been a controversial part of the large model installation process, and each method has its own advantages and disadvantages.

The disadvantages of the cloud are obvious. Large model installation requires the support of the Internet. Once the network is disconnected, it will revert to the original "artificially disabled" state. The consequence of terminal deployment is that large model computing requires higher-end hardware support, leading to higher smartphone prices.

However, for Xiaomi, the increase in hardware costs brought about by terminal deployment of large models happens to provide support for its high-end strategy.

On August 14 this year, Lei Jun revealed in his annual speech that Xiaomi's latest 1.3 billion parameter large model has successfully run locally on smartphones, and in some scenarios, the results are comparable to the operation of a 6 billion parameter model in the cloud. From the name, it can be inferred that Xiaomi's MiLM-1.3B is a large language model (LLM).

Although Xiaomi is not a cloud service provider, Lei Jun's ambition is clearly not limited to smartphones alone. Although he did not fully disclose his answer, it is highly probable that the voice system Xiaoai, with over 110 million monthly active users, will be one of Xiaomi's main battlegrounds in the future. Lei Jun stated that the current large model version of Xiaoai has been upgraded and is now open for invitation testing.

This also reveals another reason why Lei Jun chose terminal deployment. For smartphones, connectivity is a hard requirement. However, for smart speakers and other Xiaomi ecosystem products, if high-quality dialogue and other functions can be achieved in the absence of a network, it can differentiate itself from similar products.

In addition, OPPO, vivo, and Apple have also disclosed the progress of developing large models, but the final results still need to be verified in the future.

Installing large models is not simply assembling

In the past two years, most smartphone manufacturers have been pondering the same question: how to convince consumers to switch phones?

By now, smartphones have basically been developed to the extreme within the previous technological framework, and no matter how much hardware is rolled out, it is difficult to impress consumers, let alone inspire the thought of switching phones.

Data released by Counterpoint shows that in the second quarter of 2023, global smartphone sales fell by 8% year-on-year and 5% month-on-month. This marks the eighth consecutive quarter of decline in the global smartphone market.

The emergence of large models, without a doubt, is the freshest technological story of this year and possibly for a considerable period in the future, although it has not yet shown its impact on the smartphone market, the driving effect of large models on the sales of terminal products has already been verified. Liu Qingfeng, CEO of iFlytek, revealed that after integrating the Xinghuo large model, sales of iFlytek learning machines in May and June achieved triple-digit growth.

For smartphone manufacturers, this is undoubtedly a lifeline.

Source: Silicon Research Lab

Expert An Guangyong of the Credit Management Committee of the All-Union Mergers and Acquisitions Association believes that the installation of large models in smartphones is to catch up with new trends and make up for the deficiencies in the smartphone market. The smartphone market is currently relatively stable, and consumers have a certain "saturation" with traditional smartphone parameters and product forms. The emergence of large models provides smartphone manufacturers with a new selling point and a market differentiation strategy.

The potential for driving sales is worth looking forward to, but the changes brought about by large models are far more than that, and some industry insiders even believe that the pattern of the smartphone market will be restructured by large models.

Can large models really have such great power?

At least from the following two aspects, yes.

First, the installation of large models seems to be a technological threshold, but it is actually a test of the judgment and decision-making capabilities of enterprises.

The game of smartphone manufacturers in large models is not just simple assembly; it is a test of the multiple capabilities of the manufacturers in technology and software. The integration of the underlying system, the allocation and optimization of algorithms, power consumption control, and differences in understanding "intelligence" will all create differences for manufacturers in terms of user experience.

For example, the divergence between large model manufacturers mentioned in the first part, between "cloud access" and "terminal deployment," is essentially a divergence in power consumption control.

Although the parameter size of large models running on smartphones is not at the level of hundreds of billions like cloud-based large models, it still poses a significant burden on smartphone hardware, which may lead to overheating, shortened battery life, and reduced endurance.

Second, the increasingly "intimate" attributes of large models may change the consumer psychology when purchasing smartphones and related smart terminals, leading to higher brand loyalty.

It is well known that the quality of large models is highly related to data and training volume. This also means that once a brand's smartphone with a large model is purchased, every time the large model is called upon to execute a command, it can be seen as completing a training session, and the understanding between the owner and the smartphone will become more "intuitive."

Therefore, in the future, when consumers change phones or purchase new smart products, they may face a new choice: start over with a new "artificial intelligence" or continue using the same brand of smartphone with interconnected data and models?

Large models are not the killer app for smartphone manufacturers

For smartphone manufacturers, large models are a must-answer question. No matter what score they ultimately receive, giving up on answering is definitely not a wise choice.

However, from the consumer's perspective, there are still many doubts to be resolved about the "installation" of large models.

The first issue has been laid out in the previous text, such as the hardware burden brought about by the installation of large models, the divergence between cloud and terminal, and the information conveyed by these contents is that at least at the device level, the installation of large models is still not mature.

So, is the installation of large models just a gimmick?

Objectively speaking, the answer to this question depends on how we define a gimmick.

If the standard is whether a smartphone can run large models, in fact, large models have already been running on smartphones as early as May this year.

At that time, the tech giant Chen Tianqi and his CMU Machine Learning Compilation Team (MLC) open-sourced a general solution called MLC-LLM. Through this solution, any language model is allowed to be deployed on various hardware backends and native applications, providing an efficient framework for everyone to further optimize the performance of their use cases.

After the solution was released, geeks followed suit and showed off the effects of running large models in "flight mode" on social platforms, receiving a lot of "Amazing~" responses.

Chen Tianqi and the geeks are not related to any major smartphone manufacturers, and their actual experience proves that smartphone chips are capable of supporting the operation of large models.

But if the standard is raised a bit, to measure the "installation" of large models by whether they can become an assistant for work, the answer may be disappointing.

The technical origin of this round of large model hype is that after OpenAI raised the model parameters to the level of hundreds of billions, many unexpected capabilities appeared. However, the cost of running large models of this scale is very high and far beyond what smartphones can support.

Therefore, to install large models in smartphones, it is necessary to reduce the model size through methods such as pruning, distillation, and quantization. This also means that large models on smartphones cannot be as intelligent as cloud-based large models like ChatGPT or Wenxin Yiyuan, and their capabilities are greatly narrowed.

From this perspective, "cloud access" may be a less "watered-down" approach.

However, even with the approach of cloud access, it may be difficult to satisfy consumers to a great extent.

A key issue that current large model research has not yet resolved is "hallucinations." In simple terms, the existence of "hallucinations" will make large models talk nonsense in all seriousness. Many Chatbot users have been confused by its sincere statements, leading to a belief in its flawed answers.

This hidden error is a time bomb at work. Imagine asking a large model to collect and organize information for you, and then using the fabricated conclusion to advance affairs, what serious consequences it would have.

In addition, another problem that "cloud access" brings to consumers is information security.

The 24-hour inseparability of smartphones is the norm for modern people, and they are deeply involved in our work and life. Therefore, when it has a certain level of intelligence, has the authority to access our information and feedback to the cloud for calculation, the risk of information leakage also follows.

This is not groundless worry. In April this year, Samsung experienced three data leakage incidents within 20 days due to the use of ChatGPT, leading to the uploading of critical information such as semiconductor equipment measurements, yield/defects, and internal meeting content to ChatGPT's servers.

This is also the reason why Xiaomi chose to achieve the installation of large models on smartphones through terminal deployment. Although there are many difficulties on the hardware side, the local running mode can solve the problem of information leakage to the greatest extent.

Overall, the installation of large models by smartphone manufacturers is not an empty promise, but it is also not as wonderful as they have depicted. Treating large models as a killer app overestimates the capabilities of large models and underestimates the real challenges faced by end users.

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。