Source: China Entrepreneur Magazine

Reporter: Zhao Dongshan

Image source: Generated by Wujie AI

What kind of industry offers an average annual salary of 1 million yuan per person?

The answer is AI large models. This was the answer given by the vice president of vivo during an interview with media such as "China Entrepreneur Magazine." He stated, "The annual investment cost of vivo's large models is now 2 to 3 billion yuan, and the total investment cost has exceeded 20 billion yuan. Talent and data computing power each account for half, with an average cost of 1 million yuan after tax per person."

In the past year, AI large models have swept the entire internet technology industry. After the basic construction of large models from 0 to 1 is completed, the application of large models based on different scenarios is becoming a new competitive focus. Smartphone manufacturers with a large user base have become the first batch of pioneers in the deployment of edge-side large models.

Starting from August 2023, leading domestic smartphone manufacturers such as Huawei, Xiaomi, vivo, OPPO, and Honor successively announced their research and implementation plans for large models.

In August of this year, Huawei's HarmonyOS 4 was the first to announce the integration of the Pangu large model. Xiaomi quickly followed suit, having trained language large models with parameter scales of 1.3 billion and 6 billion, which have been applied in some scenarios of Xiaomi's MIUI system and the AI assistant Xiao Ai.

By November, the release rhythm of large models by smartphone manufacturers became more intense. In early November, vivo released five large model matrices with parameter scales of 1 billion, 10 billion, and 100 billion. Subsequently, OPPO officially launched the self-developed AndesGPT large model and integrated it into the latest operating system ColorOS 14. AndesGPT supports various parameter scale models from 1 billion to over 100 billion.

The mindset of smartphone manufacturers entering the market is not just about "participation," but rather a substantial investment. In April of this year, when Xiaomi established a large model team, Lei Jun made his stance clear: "Full support, with no upper limit on investment." Currently, there are over 3,000 internal R&D personnel at Xiaomi working on AI-related projects. Shen Wei, the usually low-key founder, president, and CEO of vivo, also expressed to the team, "In the future, AI will become the underlying technology for technological innovation. We must seize this historic opportunity and make high-standard investments to create leading industry technology and products, and make our historical contribution to the arrival of the intelligent era."

Undoubtedly, in the current context of declining smartphone shipments, continuously extending user replacement cycles, and severe homogenized competition, edge-side large models are becoming an innovative addition to the smartphone industry. However, at the same time, there are significant challenges in integrating billion- and hundred-billion-parameter large models into palm-sized smartphones, as well as the substantial changes they can bring to smartphone companies.

Challenges

For smartphone companies to truly integrate large models into palm-sized smartphones, there is a paradox: if the model parameters are too large, they simply cannot fit or run on the smartphone side. However, if the model is too small, it may not truly achieve intelligent emergence.

In the implementation of edge-side large models, an undeniable reality is that 10 billion parameters of data will occupy 1GB of memory on a smartphone, 70 billion parameters will occupy 4GB of memory, and when the data reaches 130 billion parameters, the memory usage will reach 7GB. However, the operating memory of the vast majority of high-end smartphones on the market is currently 12GB or 16GB. If the data volume of the large model reaches 130 billion parameters, it will seriously affect the smooth operation of the smartphone.

But compared to large models in the cloud, the advantages of edge-side large models are also obvious. For example, they can fully protect user privacy, as the interaction data between users and large models does not need to be uploaded to the cloud during use. At the same time, the response speed of edge-side large models will be faster. In the most extreme case, even without a network, edge-side large models can still be used, while large models in the cloud cannot be used without a network.

In addition, the cost of calling large models in the cloud is high. An AI large model practitioner told "China Entrepreneur Magazine," "The minimum cost of cloud computing for a large model is 1.2 to 1.5 fen per person, and if 300 million users use it ten times a day, it means that smartphone manufacturers will have to spend over 10 billion yuan a year."

Therefore, based on the current level of model training, smartphone manufacturers are faced with the challenge of talent and technological accumulation. Only by gathering core technical talents can this dilemma be better resolved. The battle of large models for smartphones also means, to some extent, a battle of talent and capital investment.

According to sources, since the establishment of the global AI research institute in 2017, vivo has built a team of over 1,000 AI experts. The conservative estimate of annual investment in artificial intelligence is 2 to 3 billion yuan, and the total investment has now exceeded 20 billion yuan, with data and computing costs accounting for half of the investment in self-developed large models, and personnel costs accounting for the other half.

Xiaomi's layout in artificial intelligence is also relatively early. After AlphaGo emerged in 2016, Lei Jun began to invest heavily in AI, starting with the visual team and gradually expanding to various fields of AI.

"Xiaomi has over 3,000 people working on AI-related R&D. The Xiaomi AI Lab has accumulated strong capabilities in areas such as vision, acoustic speech, NLP, knowledge graph, and machine learning, from algorithm pre-research to engineering implementation. We previously had a human-computer dialogue team that developed a dialogue model with 2.8 billion parameters," said Wang Bin, director of the AI Lab at Xiaomi Group and chief scientist of natural language processing (NLP).

In addition to smartphone manufacturers' efforts to deploy large models, upstream chip manufacturers for smartphones are also joining this battle.

On October 25, 2023, Qualcomm released the new generation mobile platform Snapdragon 8 Gen3. Compared to the previous generation product Snapdragon 8 Gen2, Snapdragon 8 Gen3 not only significantly improved GPU and NPU performance, but more importantly, it can run models with 10 billion parameters at the terminal side.

Qualcomm has significantly improved the AI computing capabilities of its chips on Snapdragon 8 Gen3, with NPU performance increasing by over 98%. In addition to supporting the operation of models with up to 10 billion parameters, it can also generate up to 20 tokens per second for 7 billion parameter large language models. This means that various virtual assistants and GPT chatbots will be able to run on terminals such as smartphones in the future.

Implementation

"China Entrepreneur Magazine" observed that almost all smartphone manufacturers have adopted a progressive development route when deploying large models, that is, first developing and training models with small parameter scales, and then developing models with larger parameter scales after addressing the pitfalls.

Regarding the performance of models at different levels, sources told "China Entrepreneur Magazine," "According to internal tests, a 7-billion-parameter large model is sufficient for simple document summarization and disassembly functions, but there is still room for improvement in the task disassembly capability of a 7-billion-parameter large model for truly achieving 'intelligent emergence.' 13 billion parameters may be a better choice."

In addition to progressive development, as of now, smartphone manufacturers' deployment of large models is also divided into two paths: first, the use of lightweight and localized deployment of edge-side large models, with typical manufacturers such as Xiaomi and Honor; second, the adoption of a cloud-edge collaborative architecture, launching a large model matrix, deploying models with parameter scales of 10 billion and 100 billion in the cloud, and deploying models with a parameter scale of 1 billion at the smartphone edge, with typical manufacturers such as Huawei, vivo, and OPPO.

In April 2023, led by Wang Bin, Xiaomi's AI Lab officially established a self-developed large model team, with Lu Jian, head of the large model team, leading the team. Lei Jun personally promoted the establishment of the self-developed large model team and was highly involved in the self-development of Xiaomi's large models. He personally reviews the team's weekly, monthly, and even daily reports, and pays attention to the progress of large models.

Xiaomi's self-developed edge-side large models emphasize integration with products and driving scenarios. "The main breakthrough direction of Xiaomi's large model technology is lightweight and local deployment. We do not purely consider it from a technical perspective, nor do we aim to compete. We are not engaging in an arms race, and the starting point for Xiaomi's large model development is not to become China's OpenAI. From the beginning, we have considered how to integrate large models with the company's scenarios," Wang Bin told "China Entrepreneur Magazine."

Different from Xiaomi's main focus on edge-side large models, Huawei, vivo, and OPPO's self-developed large models adopt a "cloud-edge collaborative" approach. Huawei's Pangu large model, vivo's Blue Heart large model, and OPPO's Andes large model cover parameter scales of billions, tens of billions, and hundreds of billions.

"Why use a matrix to solve this problem? Because today's large models are language-based, and only language and text truly have large models. There are no large models for sound and video. Until I can imagine a significant breakthrough in algorithms, the matrix approach is a better solution," said a source during an interview.

In addition, the source also stated that the matrix-style large model is a comprehensive result influenced by user demand and computing power costs. First, the matrix mode allows users to use large models in the cloud while running critical data locally at the edge to meet privacy and security needs. Second, the matrix mode can effectively balance the problem of high cloud computing costs.

Necessity

Despite the numerous challenges of edge-side large models for smartphones, participating in them is becoming a necessity for smartphone manufacturers.

In 2017, the smartphone market experienced its first-ever decline, with a 5% drop in shipments in the Chinese market, marking the entry of the smartphone industry into a stock era. In the third quarter of 2023, global smartphone shipments fell by 8% year-on-year, marking nine consecutive quarters of decline and also setting the worst shipment record in nearly a decade for the same period.

In recent years, faced with increasingly homogenized competition, smartphone manufacturers have successively innovated and engaged in internal competition in areas such as fast charging, battery life, full-screen displays, photography, and high refresh rates. Currently, edge-side large models are becoming the innovative entry point they are eager to seize. After all, no one wants to become the next Nokia by missing out on a trend.

"AIGC is the most exciting technological innovation I have encountered in my career in the smartphone industry, and it can even be called a revolution. It will bring revolutionary changes to the experience of mobile life. The knowledge of large models exceeds individual knowledge. It can understand your language like a person, observe you, learn from you, understand your habits, and provide you with the best help to complete your tasks," said Liu Zuohu, Chief Product Officer of OPPO, when discussing large models with "China Entrepreneur Magazine." In his view, the future of smartphones is a super assistant, similar to ChatGPT.

Smartphone manufacturers' large models are gradually being implemented in their artificial intelligence applications. For example, Huawei's P60 and Mate60 series smartphones have already integrated the Pangu large model into their Xiaoyi smart assistant. At the same time, OPPO's Xiaobu, Xiaomi's Xiao Ai, and others have also integrated their respective large models. In addition, in vivo's latest X100 model, users are provided with the option to download edge-side large models, and they can choose whether to download voluntarily and whether to run them in the cloud or locally.

The application of large models on smartphones is indeed improving the user experience. For example, in the era of large models, finding a specific photo in the gallery used to require very precise keyword searches. However, now, with large models, users can simply give a generalized and vague command to the AI voice assistant, such as "the photo of the Great Wall I took during my trip to xx before," and find the corresponding photo.

Currently, edge-side large models can perform simple operations such as language understanding, text creation, and image generation, but operations involving ticket booking, app interaction, and device control still require the mobilization of cloud-side large model capabilities. However, even so, it represents a huge possibility for smartphone manufacturers.

"We are now investing heavily with the mentality of 'first be invincible' as described in 'The Art of War' by Sun Tzu. We do not want to be left behind by excellent business partners, and we do not necessarily want to engage in internal competition, confrontation, or competition in large models," the source told "China Entrepreneur Magazine" when discussing the mindset of large model development.

In their view, the importance of edge-side large models is at the same level as communication and chips, and in the future, it will become a key way to acquire high-end users. "From the perspective of greatly improving production efficiency in science and technology, I believe that large models are also a historic and event-level development. After doing well with large models, gaining recognition from more high-end users is inevitable, so we are also making high-standard investments," the source added.

In the current context of domestic smartphone manufacturers collectively aiming to compete in the high-end market, all new innovative opportunities have become new variables in the competition. Moreover, the battle of edge-side large models has already begun.

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。