Who can we count on to bring us closer to AI?

Estimated read time 14 min read

The era when AI can be used by everyone is getting closer.

If the AI ​​large model has previously used ChatGPT’s multiple rounds of technological progress, the domestic Hundred Model War, Sora’s breakthrough in the field of video large models, and a series of other large model preliminary application explorations to prove that the new AI is more powerful; then recently The quietly rising trend of manufacturers naming the most commonly used consumer electronics terminals in the name of AI further heralds the accelerated arrival of the era of inclusive AI.

In this wave of industry discussions about AI terminals, one topic has always been the most popular: At this stage, which is the first terminal for large-scale inclusive models? Is it the smartphone that has penetrated deeply into our daily life, or the personal computer (PC) that has been used as a productivity tool for a long time ?

The reason for narrowing the scope to these two types of terminals is simple: in addition to mobile phone manufacturers such as OPPO, vivo, Meizu, Huawei, and Xiaomi, those who have recently been confident enough to openly call out the device-side AI strategy are chip giants such as NVIDIA, Intel, and AMD. , or a global PC industry leader like Lenovo.

These companies are not all on the same page. Taking smartphones as an example, some people are very anxious and say that “mobile phone manufacturers that do not turn to AI will be the next Nokia”, some directly shout the slogan “All in AI”, and some say “AI is the future, but AI phones are definitely gimmicks” “; on the PC side, Nvidia, which has leading GPU software and hardware capabilities, has placed its focus more on cloud computing. Intel directly announced that its Meteor Lake processor is the most important processor in 40 years. Architectural changes, and Lenovo directly stepped forward to promote its own “AI PC” research and actual products.

But in the author’s opinion, this bunch of brand-new statements and discussions only illustrate one thing – existing personal terminals (mainly smartphones) with basic computing and communication as core functions are no longer able to meet the needs of the AI ​​​​big model era. new requirements for software and hardware. AI PCs, which have a more relaxed form and can adapt to various needs with variable configurations, are becoming the real focus of a new round of inclusive AI.

How difficult is it to build a terminal that can achieve inclusive benefits in the AI ​​era?

Like several previous rounds of technology revolutions, the key to the inclusiveness of this wave of AI large-model technology revolution lies in how to get as many users as possible to adopt this technology and apply it, because only in this way can the greatest value be created.

The most troublesome thing is that the essence of large AI models is to provide human data to computers for learning in order to achieve the replication of some mental achievements. In this process, it inevitably affects the protection of individual privacy, professional value, and even survival value of human beings. impact. In the end, AI is indeed powerful enough, but users are afraid and unwilling to use it.

A typical example is the “New York Times suing OpenAI” incident at the end of last year. The former accuses OpenAI and its partner Microsoft of using millions of its articles to train artificial intelligence models, demanding that they “unlawfully copy and use the New York Times’ unique and valuable works” and related “worth dozens of Billions of dollars in statutory and actual losses.”

To put it more concisely, OpenAI uses a large model to learn a large amount of paid content from the New York Times. Then when users use OpenAI, they can bypass the charging mechanism of the New York Times for unique content and directly provide If the user generates results with the same key information but different expressions, it is equivalent to destroying the latter’s job.

The feature of “uploading data is equivalent to making it public to all users” has directly interrupted the process of applying large AI models to many enterprise users. In April last year, an engineer from Samsung’s semiconductor department accidentally “shared” it while using ChatGPT. Company secrets.

When applied to ordinary users, since the decision-making process of AI models is often opaque, it is difficult for users to understand how the model analyzes and processes personal data, and it is even more difficult to process and protect data, which ultimately leads to improper use of personal information and also makes it difficult for users to understand how the model analyzes and processes personal data. Many individual users refuse to use it.

To solve this problem, the “breakthrough” is actually very simple – large models created to provide personal privacy and key results must have clear personal ownership and use rights protection, that is, high-sensitivity, high-value private large models should be owned by individuals. Users are in complete control.

To achieve “complete personal control”, it is natural to store these large models in personal terminal devices. But this also raises two other questions: First, how to change to meet the high requirements for privacy and collaboration in the era of large AI models, “cloud & terminal” collaboration? Second, on the basis that the entire large AI model can run, who should ensure the actual user experience?

Let’s talk about the first question first. The key lies in the transformation of industrial structure.

From the earliest PC era to the smartphone era, the overall architecture of the system has not changed much. The bottom hardware is the system, and the system is the application. The system that connects the hardware and applications is the main body of the ecosystem and is responsible for Promote technology inclusiveness and ecological development.

The most typical example is Apple’s AppStore. Users can easily find various types of applications, from games and entertainment to life services, to educational tools and business applications, which greatly enriches the content and form of the mobile Internet. Developers gain a level playing field with clear distribution rules and revenue models.

According to the basic needs of individual users for privacy, large AI models trained by themselves must be stored locally. However, in the process of using AI capabilities, it is inevitable to use large AI models trained with public data. The complex “cloud & client”, “public & personal” large model collaboration is obviously very different from the traditional system and application relationship, and it also raises questions about whether the “system” should exist.

Take NVIDIA’s newly released NIM microservice architecture at GTC 2024 this year, which packages the large AI model into an integrated solution with k8s at the bottom and direct output service capabilities at the top. Suddenly, he elevated himself from the role of a GPU manufacturer to an “application ecosystem” manager without an “operating system”.

OpenAI, which has made the fastest progress in the field of basic large models, has tried even earlier. When ChatGPT became popular last year, it launched its own GPT Store, hoping to allow more developers to create more practical models based on the basic large models of the GPT language. application. But the final result was counterproductive. There were not many truly eye-catching GPT applications. Instead, there were many illegal, illegal, and even directly plagiarized applications.

Although the leading AI chip and AI technology leaders are powerful enough in specific aspects, they are still unable to support the entire ecosystem. Although the IT industry has solved the complex needs of the PC and smartphone era through division of labor in the past, after AI matures, the industry will obviously need to change its structure accordingly.

Therefore, the final question becomes what we mentioned above: Who should ensure the user experience when using AI terminals?

Starting from user experience, terminal manufacturers such as smartphones and PCs have considerable experience in ensuring massive user experience. However, in the face of new demands for computing power, storage space, cloud collaboration, and changing terminal forms caused by large AI models, , the potential of smartphones is always limited.

Take many current large-scale AI model applications for mobile phones, whether it is OpenAI’s ChatGPT, Baidu’s Wen Xinyiyan, or even the recently popular Dark Side of the Moon Kimi. The way these apps work is not to run locally. As a user-oriented carrier, in this case there is no difference from mobile phones and browsers.

“If you let AI have nothing to do with you, in the era of big AI models, you will become a marginal existence just like the original operators were piped by WeChat.” This may be the real reason why some mobile phone manufacturers are anxious to make “AI mobile phones” .

In contrast, the situation of personal computers (PCs) is much better. At present, the independent graphics card of ordinary laptops has the ability to run large models. PC chip manufacturers such as Intel and Nvidia are still constantly upgrading solutions at the computing power level. solution, coupled with more sufficient storage space and battery capacity, more flexible control and display capabilities.

The continuous upgrading of upstream industry leaders, coupled with the terminal advantages regained by the PC form in the AI ​​​​era, allow PC manufacturers to hope to establish an ecosystem driven by user feedback based on their huge base of hundreds of millions of users. Ultimately, on AI PCs, Build a true “killer” app.

Not long ago, in the report “From Tool to Platform AI PC: The First Terminal for Inclusive AI” released by the cutting-edge technology research institution “Unfinished Research”, two of the opportunities were specifically mentioned. One is based on mixed computing power. The inference engine, the other is AIOS based on large model and agent technology.

Heterogeneous computing power is actually easier to understand. Even in the AI ​​era, we still need CPU+GPU+NPU on the chip so that it can handle AI acceleration at the same time and undertake tasks such as general computing, graphics processing, and parallel computing. However, how to build an actual architecture that is powerful enough and easy to call, develop, and drive still needs to be explored by the industry.

Based on the need to integrate heterogeneous computing power, machine manufacturers that simultaneously connect computing hardware such as chip manufacturers, developers who utilize computing hardware performance, and consumers in computing application scenarios such as end users have actually stood at the forefront of an industry. Advantageous location.

Let’s talk about AIOS. Applications have always been very important in traditional architectures. However, as AI large models mature in more and more applications and scenarios, AI applications are no longer independent of the system, but need to rely on basic large models. , tool library, and construction of user privacy data.

The completely changed system layer requirements have allowed traditional manufacturers like Microsoft to no longer “monopolize” the ecological niche of this link. In the past, complete machine manufacturers that were also partially involved in system layer capabilities can fully equip a variety of hardware-based systems. Personal intelligence creates the first entrance to AI that provides users with an excellent experience. In particular, Microsoft Copilot is absent in China due to well-known reasons, and the opportunities for complete machine manufacturers are being further highlighted.

AI PC—The answer to bringing AI closer to you

In the report, Unfinished Research also conducted a comprehensive analysis of the industrial structure, market value, and social significance of AI PC.

Take the user experience that AI PC should provide, for example, which includes the privacy requirements mentioned above, as well as real-time, cost-effectiveness, and accuracy included in the system architecture requirements. There is also a very counter-intuitive insistence on “traditional” interaction methods such as keyboard and mouse, which exactly reflects the original intention of this report to truly start from the user experience.

At an industry forum titled “The First Year of AI Terminals Begins: Changes, Challenges and Opportunities” held not long ago by Huxiu’s financial brand Miaotou, Wu Sheng, the founder of Scenario Lab, also gave his own similar views ——The miniaturization and personalization of large models are the key paths for future development. Large models must eventually enter everyone’s daily life in a convenient and easy-to-use way before they can truly exert their value. Terminal devices, especially PCs, as local computing centers, will assume an important role in carrying personal AI and may become the frontline battlefield for popularizing AI applications.

Abulikmu Abulimiti (hereinafter referred to as Amu) , Vice President of Lenovo Group, who also attended the event , had a more detailed view: “ AI-empowered terminals will become users’ intelligent agents and personal assistants, changing traditional tools. attributes and become a highly intelligent and personalized assistant. ”

Unfinished research also specifically mentioned in the report how to achieve these ambitious goals, such as building an industrial middle layer in the AI ​​​​large model era that is best suited to the use of heterogeneous computing power and the best to ensure user experience.

The ones most promising to shoulder the “heavy duty” of building the middle layer are complete machine manufacturers like Lenovo. By connecting designers of computing hardware such as chip manufacturers, developers who utilize the performance of computing hardware, and consumers of computing application scenarios such as end users, machine manufacturers can become integrators and promoters of heterogeneous computing power. Ensure the full play of hardware performance and the release of AI computing power .

In addition to introducing a series of key component upgrades for AI large models, complete machine manufacturers can also bring innovations driven by AI large model technology to users at the physical level. For example, hardware-level encryption solutions in device-side AI and cloud-side AI collaboration modes; AI PC’s secure collection of local multi-modal data from real-time environments and user feedback, etc.

From an industry-wide perspective, in addition to covering the application layer, Lenovo can also leverage its expertise in chips, operating systems, large models, application ecology, and terminal product design and manufacturing based on the ever-changing needs of users in the era of AI large models. All-in-one vertical integration quickly meets the various needs of users.

From this perspective, complete machine manufacturers like Lenovo are not only the final deliverers of AI computing power on the first terminal, but also the powerful builders of the first entrance to AIOS . Based on the deployment of local large models, complete machine manufacturers will also be the integrators of the first platform of the AI ​​ecosystem.

Among the complete machine manufacturers, the one with the most mature supply chain management, the thickest venture capital industry ecosystem, and the largest equipment user base will become the ultimate winner.

As the world’s largest PC manufacturer and computing infrastructure manufacturer, the third largest supplier of artificial intelligence infrastructure in the world, and the third largest supplier of servers and storage in the world, Lenovo’s strategic investment in AI can be traced back to 2017 Year.

After the rise of deep learning at that time, Lenovo went “All in AI” without hesitation. In August last year, Lenovo and Intel first mentioned the concept of AI PC in their cooperation. In October, Lenovo demonstrated a revolutionary AI PC product that can run local large models offline and generate customized solutions based on user personal data.

In January this year, Lenovo released more than ten AI Ready stage PC products, and provided the opportunity to try out the personal intelligence AI Now and the local AI text application Creator Zone at CES.

In order to welcome the explosive growth of the artificial intelligence era, Lenovo Group is also ready. At the just past Lenovo Group’s 2024/25 fiscal year swearing-in meeting, Yang Yuanqing predicted that at Tech World, the Lenovo Innovation and Technology Conference held in Shanghai on April 18, a true AI PC with five major characteristics will be the first to be launched in the Chinese market. . At that time, Lenovo Group will further explain the “AI for All” strategy and continue to promote the innovation and application of artificial intelligence in the fields of equipment, infrastructure and solutions.

It is foreseeable that in the new era of AI large models, in China, the world’s largest PC market, all walks of life are eagerly waiting for AI large models to provide more driving force for themselves. Lenovo uses its in-depth understanding of AI technology and in-depth insights into the IT industry to bring AI PC, a key innovative product, which will inevitably set off another wave of change in the IT industry.

You May Also Like

More From Author

+ There are no comments

Add yours