How AI is Driving Globalization and Transformation in Emerging Industries? | Highlights

Artificial intelligence (AI) is profoundly reshaping various industries. Through machine learning and data analysis, AI enhances production efficiency and adaptability, driving innovation in intelligent manufacturing. But how can AI inject vitality into emerging industries and help businesses maintain a competitive edge in the global market?

On December 19, the EqualOcean “2024 Globalization Forum (GGF2024)” and the Emerging Industries Globalization Forum brought together industry leaders to explore this pressing question. The roundtable discussion on "How AI is Driving Globalization and Transformation in Emerging Industries?" featured distinguished speakers, including TransferTech Founder Mr. Fan Yu（迁移科技，樊钰）, Witstone Capital Partner Mr. Ni Tianyang（微智资本，倪天旸）, 1Data Partner Mr. Lu Yue（壹沓科技，陆玥）, and CAICT International AI Institute Director Ms. Xu Shan（信通院人工智能研究所，许珊）, alongside moderator Meridian Capital Overseas Partner Mr. Qiu Zhun（华映资本，邱谆）.

The panel provided in-depth insights into how AI can catalyze innovation and help emerging industries navigate the complexities of global competition.

Moderator (Qiu Zhun): Could you please introduce your company's products and core business areas?

Fan Yu: I am Fan Yu, the Founder and CEO of TransferTech. Our focus is on AI+3D, specifically 3D vision, providing robots with "eyes" and "brains" to enable intelligent perception, decision-making, and operation. By integrating with robotic arms, which traditionally perform repetitive tasks, we enable them to handle unstructured scenarios, such as picking up parts randomly placed in a pile. Currently, our solutions have been applied in regions like Japan, South Korea, Southeast Asia, Mexico, and Europe. Today, on our video channel, we released a case study showcasing the application of our 3D+ intelligent robotics system in a factory in Mexico.

Ni Tianyang: I am Ni Tianyang from Vision Capital. Vision Capital is an investment firm focused on the digital technology sector. Our investment priorities currently center on three areas. First, next-generation information technologies such as generative AI and cloud computing. Second, domestic substitution technologies, including trusted computing and industrial software. Third, the globalization of digital technologies and AI.

Unlike most VC and PE investment strategies, we have two distinguishing features. The first is that Vision Capital focuses solely on digital technology and covers the entire lifecycle of investments, from early-stage angel rounds to growth stages and even mergers and acquisitions (M&A). This is particularly important in light of current challenges surrounding domestic IPOs, making M&A an increasingly critical aspect of our approach.

Our second distinguishing feature lies in our industrial expertise. We have two strategic shareholders: Weimob, a publicly listed SaaS company in Hong Kong, and Yicun Capital, which specializes in industrial M&A. This allows us to provide unique insights from an industrial perspective.

Over the past year, AI and globalization have been the two key themes we have closely monitored, which align perfectly with today's discussion. I look forward to further interactions and exchanges with the distinguished guests, experts, and you, Qiu.

Lu Yue: I am Lu Yue from 1Data Technology. Our company focuses on the supply chain sector, providing digital workforce and hyper-automation solutions for supply chains. To date, we have served nearly 1,000 corporate clients both domestically and internationally. These include state-owned enterprises such as COSCO Shipping and China Academy of Transportation Sciences, as well as private listed companies and major manufacturers like Midea and Schneider Electric. We are currently at the B+ funding stage, having received investment support from renowned firms such as SIG Asia Investments, IDG Capital, Zentree, Sequoia Capital, and CDH Investments. We firmly believe the future will be defined by human-machine collaboration, and we see significant opportunities for development within this trend.

Xu Shan: I am Xu Shan from the Artificial Intelligence Research Institute of the China Academy of Information and Communications Technology (CAICT). Our team focuses on international cooperation in artificial intelligence (AI), industry integration, and supporting Chinese enterprises in their efforts to go global. Additionally, we have a critical mission: recently, during the 16th BRICS Summit, President Xi Jinping announced the establishment of the China-BRICS Artificial Intelligence Development and Cooperation Center. As the operational team for this center, we actively promote the internationalization of Chinese enterprises and technologies, while fostering global cooperation in AI. Beyond this, I also collaborate closely with the World Health Organization (WHO), currently serving as the director of the WHO Digital Health Cooperation Center.

Moderator (Qiu Zhun): Let me briefly introduce Meridian Capital. Established in 2008, Meridian Capital initially set up a venture capital fund in Singapore before entering the domestic market. We operate a dual-currency fund structure with both USD and RMB funds. Currently, we manage approximately RMB 11.5 billion in assets, having successfully closed two USD funds and seven RMB funds. Our latest RMB fund is nearly RMB 3 billion in size.

Our investment focus is divided into two main categories: first, hard tech sectors such as new energy, new materials, and advanced manufacturing; second, digitalization, including AI, artificial intelligence, SaaS, and various digital transformation projects. Personally, I primarily work in Silicon Valley, frequently traveling between the U.S. and China. My role involves both monitoring overseas AI projects and investing in robotics and AI applications domestically.

The core theme of this conference is "Going Global," which is a crucial direction. Interestingly, when it comes to the intersection of "Going Global" and AI, it seems that our roundtable is the only session addressing this topic today. As someone who began studying AI in the U.S. during its early days and has been involved in the field for nearly 30 years, I’ve observed that discussions among Chinese enterprises have increasingly focused on "Going Global" in recent years. As a growing trend, how does "Going Global" intersect with AI?

Wang：In my personal view, from an investment perspective, globalization can be divided into two types: “Push” and “Pull.” The “Push” type refers to situations where the current economic environment faces certain bottlenecks, compelling many companies to “go global” as a means to seek breakthroughs in overseas markets. On the other hand, the “Pull” type occurs when significant opportunities exist in overseas markets, attracting enterprises to expand internationally. In the case of “Pull,” companies are drawn by incremental opportunities. The key here is to identify where this “increment” comes from. In a globalized market, incremental growth usually stems from two primary factors: population and technology.

First, population growth is a major source of incremental growth. For example, regions like Southeast Asia are experiencing continuous population growth, which generates enormous market demand. Over the past 30 years, much of China’s development has been fueled by demographic dividends, but this advantage is gradually reaching its limits. However, many emerging markets still hold substantial potential for incremental growth.

Second, incremental growth also comes from technological advancement. In regions where population growth has stabilized, innovation becomes the critical driver of progress. Take the United States as an example: while its population growth has plateaued, technological innovations from places like Silicon Valley continue to propel industrial development. Looking back at history, every major technological revolution—be it semiconductors, the internet, mobile internet, or artificial intelligence—has been driven by breakthroughs in foundational technologies. Population and technology are thus the core forces driving global market growth.

When it comes to technology, as an investor working in Silicon Valley, I can strongly sense that today’s investment landscape is almost entirely dominated by AI. In Silicon Valley, the vast majority of innovations revolve around AI, with relatively fewer projects in other areas like advanced manufacturing, renewable energy, or new materials. From the perspective of technological driving forces, AI is undoubtedly the most significant catalyst. Therefore, the topic of “going global” is intrinsically linked to AI. This connection between globalization and AI is clear.

Next, I’d like to discuss two questions with everyone. The first question concerns AI application scenarios. Today, we have both startup companies and investment institutions present, and at Meridian Capital, we are part of the investment landscape ourselves, so this topic is particularly relevant to us. First, I’d like to hear your thoughts on how AI is defined. At present, defining AI remains a challenge for both investors and entrepreneurs. We often hear companies say, “We are fully committed to AI,” or “We are an AI-driven startup,” but these statements are often vague. Jack Wang, the founder of Alibaba Cloud, once remarked, “AI is simply Transformer,” which resonated deeply with me. Personally, I have previously defined AI as deep learning. However, is the general understanding of AI often overly broad?

The key question is, when a company declares that it is “all in AI,” what does that actually mean? For instance, does it mean the company is “all in on Transformer,” “all in on deep learning,” or something else entirely? We often receive business plans that claim to focus on “AI applications,” but such statements are frequently vague. Simply stating “working on AI applications” carries little practical meaning because we need clarity. For example, are they focusing on deep learning applications or Transformer-based applications? Without this specificity, it becomes easy to filter out ambiguous projects.

Take some projects, for example—they might claim to be working on AI applications but are, in reality, just doing big data analysis and labeling it as AI. However, big data and AI are distinctly different concepts. Big data is not equivalent to AI, and this represents a significant gap in understanding.

Therefore, I’d like to invite everyone to share their definitions of AI. What exactly does AI mean to you, and what does it not include? In my perspective, deep learning is a subset of AI. Without deep learning technology, I wouldn’t classify something as part of the AI domain. For Dr. Wang Jian, if there’s no support from Transformers, it wouldn’t qualify as AI either. For us as investors, having clear boundaries is crucial—it helps us make well-informed investment decisions, and this clarity also serves as a critical part of mental discipline.

Building on this, my second question is: How do you see the connection between AI application scenarios and the business areas you currently invest in or work on? For example, if we define AI as deep learning, certain gaps may emerge between this definition and various fields or application scenarios. This issue can be perplexing. How can we bridge this gap and find the best fit between AI technologies and real-world application scenarios? This remains a significant challenge in our ongoing exploration.

Fan Yu: First, I don’t entirely agree with Dr. Wang’s statement that “AI is just Transformer.” I have my own understanding of AI, and I see it as the application of neural networks. Reflecting on my academic journey, I began my master’s studies in 2014, when I learned the foundational concepts of neural networks. Around the time I graduated, AI experienced a major boom. It was during that period that we decided to embark on entrepreneurship in the AI field, choosing the direction of 3D + AI. In our practical applications, we currently use typical small models.

Moderator (Qiu Zhun): Roughly speaking, would you define AI as neural networks?

Fan Yu: Yes, 3D vision in our specific applications integrates multiple AI technologies. For example, in a typical production scenario, a machine operator needs to bend down, pick parts out of a material bin, and place them into a machine for processing. In our applications, the most common workflow involves using a camera to capture images of the material bin, performing image recognition and analysis, and then using the calculated results for robotic motion planning and obstacle avoidance. This ultimately enables automatic picking and placing of parts into the machine. In this process, AI technologies play a critical role in object recognition, instance segmentation, and robotic trajectory planning.

Specifically, object recognition and instance segmentation are key problems we solve using AI technologies. We leverage AI to precisely calculate the position and orientation of each part while also using AI to plan the robot's trajectory and avoid collisions. These applications ensure both efficiency and safety throughout the operation. This suite of technologies is not limited to a single field. We have successfully implemented it in various industries, including automotive OEMs, auto parts manufacturers, warehousing and logistics, and heavy industries. Applications such as welding, grinding, and painting have been deployed across these industries. Moreover, in the semiconductor sector, we have collaborated with SMIC (Semiconductor Manufacturing International Corporation) to achieve high-precision wafer stack positioning using cameras and AI technologies.

In the course of technological development, the topic of "large models" is inevitably brought up. In fact, we are one of the earliest companies in China to apply vision-based large models in our products. For example, last year, when Facebook released the SAM model, we quickly integrated it into our products just two weeks after its release, using it for object segmentation. However, through practical application, we found that large models often underperform in terms of precision and efficiency compared to small models we developed specifically for certain scenarios. As a result, we have consistently explored and optimized our technical solutions to ensure optimal performance in different application scenarios. Moving forward, we will further embrace AI and adopt an end-to-end technical approach to achieve faster processing speeds and greater generalization capabilities.

Our company name, "TransferTech," is also derived from the AI technique of “transfer learning.” In industrial environments, data samples are often sparse, and traditional training methods struggle to achieve ideal model performance with limited samples. To address this, we utilize transfer learning to fine-tune pre-trained models from other domains using a minimal number of samples, significantly enhancing model performance. For instance, in warehousing and logistics scenarios, we now need fewer than 100 samples to train highly effective models. This application of transfer learning has greatly improved our training efficiency under data-scarce conditions and provided strong technical support for rapid deployment across multiple fields.

Moderator (Qiu Zhun): You’re saying that neural networks and deep learning are essentially the same thing, as neural networks must inherently be deep. So, broadly speaking, our definitions align.

Ni Tianyang: The question posed by Mr. Qiu resonates deeply with me, as it highlights a challenge we often face during the investment process. Today, nearly all business proposals we receive claim to be "all in on AI." Interestingly, about one-third of these proposals are from companies whose earlier pitches, three years ago, were centered on blockchain, big data, or intelligent marketing, yet now they've entirely pivoted to AI narratives. This shift creates significant challenges for us. To address this, I’ve developed an internal framework aimed at identifying what truly constitutes AI. The definition of AI is undeniably broad, with varying interpretations, so as an investment institution, we need clarity on what kind of AI we want to see and invest in.

AI has a long history. Discussions and visions about AI date back to the 1960s and 1970s, even before The Matrix hit the screens. However, the AI we’re focused on today is what I term “new AI.” What is “new AI,” and how can we conceptualize it? In my view, two core elements define it:

First, we focus on whether the underlying models exhibit characteristics of being “new” and “disruptive.” The most notable examples include pre-trained models based on the Transformer architecture, generative diffusion models, and novel models related to 3D. To begin with, we evaluate the innovativeness of the underlying model, seeking not just novelty but revolutionary potential. For instance, models like Transformer and Diffusion represent groundbreaking advancements in the AI domain.

Second, we assess the capabilities brought about by these AI models. We’ve identified three key capabilities:

Multimodal Generative Capability: Large models now excel in tasks like text creation and image generation. The emergence of generative models like GPT marks a paradigm shift in AI’s generative abilities.

Multimodal Interaction Capability: New AI systems not only support multi-turn interactions but also possess memory capabilities, enabling them to handle longer tokens and exhibit stronger interactive abilities.

Logical Reasoning Capability: AI must perform more complex logical reasoning, demonstrating robust chains of thought. Recent advancements, such as in the O1 model, have significantly enhanced this capacity.

At the project level, “new AI” demonstrates varying combinations of these capabilities in different scenarios. For instance, in social applications, AI might primarily rely on multimodal interaction capabilities and a degree of logical reasoning. Meanwhile, in AI marketing, the focus might shift to the integration of generative and reasoning capabilities.

As an investment institution, we aim to avoid being misled by projects that repurpose “old wine in a new bottle.” To this end, we’ve developed this framework to define the essence of AI more precisely. This definition, rooted in an investment perspective, helps us capture emerging technologies and future innovations.

Moderator (Qiu Zhun): From this perspective, the definition of AI is indeed quite complex, and it may take considerable time to clarify. During the earlier internet wave, as investment institutions, it was relatively straightforward to determine whether a company belonged to the internet sector. The boundaries of the internet were relatively clear. For example, "offline" businesses were not considered part of the internet. However, if an offline restaurant used digital systems or internet technologies, such as Luckin Coffee, it could be categorized as an internet company, whereas Starbucks would not be. Simply put, as long as a company relied on connected digital systems for operations, we could define it as an internet company; otherwise, it wasn't. This clarity was critical for investors, as it allowed us to make quick judgments.

In contrast, the definition of AI has not yet reached the same level of simplicity as the internet during its era. For many entrepreneurs, a significant challenge in AI entrepreneurship is that they often do not fully understand what "AI" actually means, or they lack a clear definition of AI in their minds. For instance, you might be uncertain whether you're engaged in deep learning entrepreneurship or Transformer-based entrepreneurship. If these questions remain unresolved, it often leads to unclear directions and confusion in entrepreneurship.

Even at the national level, when we say, "promoting AI development," what exactly does that mean? Does it mean fostering the development of Transformer models, or does it refer to advancing neural networks? Without a clear definition of AI, such policies can lack precision and clarity. Policy support, at its core, is also a form of investment decision-making, and it requires clarity about whether to support a specific technology or developmental direction. During the internet era, we didn’t face such ambiguities. If a company used internet technologies for its business, we could quickly make a judgment—it was black and white. Today, however, the definitions and applications of AI are much more complex, and both entrepreneurs and investment institutions are undergoing a process of continual learning and adaptation.

Lu Yue: Returning to the essence, my personal understanding is that artificial intelligence is a broad concept. When we talk about AI, it’s crucial to consider from which angle we approach it. Broadly speaking, many applications in our daily lives today fall under the umbrella of artificial intelligence. For example, the ability to automatically recognize and extract text content in WeChat—does this count as artificial intelligence? I believe it does, as it falls within a domain of AI.

However, if we discuss AI from the perspective of AGI (Artificial General Intelligence), the standards would be much higher. AGI requires capabilities akin to human-level thinking, learning, decision-making, and analysis, and we still have a long way to go to achieve that standard. As for the context of Dr. Wang’s remarks, I am not certain if there is a specific background or scenario tied to his statements.

Moderator (Qiu Zhun): Indeed, there is context to this discussion. Since the 1950s, AI has undergone extensive development, yet there still isn’t a unified definition. Even within the field of computer science, different subfields classify AI differently. For example, machine vision and natural language processing are considered parts of AI, while graphics is not. This kind of definitional approach fails to provide clear guidance in the practical realms of entrepreneurship and investment. Dr. Wang Jian previously proposed his perspective on Transformers, highlighting that Transformers, as a subset of deep learning, represent a core breakthrough in this field in recent years. In fact, many other AI-related domains have been stagnating before this breakthrough. While these fields still exist, they are primarily focused on research and have not yet achieved industrialization. This is why Dr. Wang brought up this topic.

Lu Yue: While Transformer isn’t an entirely new algorithm, it became widely recognized because ChatGPT’s emergence brought it into the public eye. Before Transformers, many different domains were included under the umbrella of AI, such as RPA (Robotic Process Automation), natural language processing, knowledge graphs, image recognition, and visual capture. These were all considered components of AI. However, I don’t think Transformer as an algorithm can represent the entirety of AI. If a new algorithm emerges in the future, would it be part of AI? I believe that’s very possible. We need to broaden our perspective. From the standpoint of businesses and investments, focusing more on the implementation of specific application scenarios often provides more actionable guidance.

As investors, we hope the companies we invest in can create tangible value for society, thereby enhancing their commercial value, rather than remaining solely at the stage of fundamental research.

From this perspective, the internet indeed had a clear boundary. The shifts between the past decade and the next ten years mark a significant generational transformation. The past decade was the era of “Internet+,” where the essence of business was moving offline operations online. In the era of “AI+,” I see a more profound business logic: the transition from “humans” to “robots.” This is also the direction we at 1Data Technology (壹沓科技) are focused on—ushering in an era of human-machine collaboration.

We concentrate on human-machine collaboration within the supply chain, aiming to develop digital employees. All of this depends on the application of large models and other advanced technologies. We believe this should be considered a part of AI. Although we are still distant from strong AI or Artificial General Intelligence, we believe the process is underway, and we have already seen the enormous potential and transformative possibilities it holds.

Xu Shan: How should we define artificial intelligence? When pondering this question, I often think about the purpose behind defining AI. While I am not an expert in the field of investment, I understand that the goal of investing is to identify valuable opportunities. However, from a research perspective, the approach might be different. This reminds me of a scene from the famous movie 3 Idiots. In the film, the principal asks a student to define "machine." The student responds with an official definition from a textbook. But in essence, any device that saves human time, effort, and resources can be considered a machine.

Extending this concept to AI, we can see that the idea of AI fundamentally originates from machine learning. Machines, as a type of mechanical tool, embody two core attributes: first, they can save human labor, time, and resources; and second, they possess the ability to learn. Through adaptive processes, machines not only assist us in saving time and effort but also evolve and improve as they perform tasks.

Based on this understanding, we can further explore what kind of AI is worth investing in. The AI technologies we seek should not only enhance efficiency and save resources but also have the capacity for continuous learning and self-improvement. Only such AI technologies can secure a meaningful place in the future of development.

Moderator (Qiu Zhun): According to your definition, which application scenarios align most closely? First, an AI must assist humans, but mere assistance isn’t sufficient to qualify as artificial intelligence. For example, do you consider a calculator to be AI? Computation is indeed part of human intelligence, yet we all know calculators do not fall under the category of AI. Secondly, AI must have adaptive capabilities, which are technically complex and challenging to implement. Based on these two criteria, could we conclude that a system qualifies as AI only if it assists humans and has adaptive functionality? Under this framework, which application scenarios should we focus on?

Xu Shan: Regarding your question, I’d like to add two points.

First, AI architecture is an evolutionary process. Since the concept of "AI" was first introduced 60 years ago, AI technology has undergone multiple developmental stages. Initially, AI research focused on knowledge-based expert systems. It then progressed to deep learning, and in the last four years, we’ve entered the era of large models. Throughout this journey, we’ve witnessed the iteration of various technical frameworks. For example, from the early convolutional neural networks (CNN) to the emergence of Transformer architectures and the rise of Diffusion models, these frameworks distinguish themselves in various dimensions from Transformer-based models. Sora, for instance, is a model built on Diffusion architecture. We may pay particular attention to models and foundational technologies with technical innovation and growth potential. As noted in a report by Stanford University, there are differences between China and the U.S. in foundational models, vertical industry models, and application models. In my view, these differences point to both technological directions worth investing in and application-level opportunities.

Second, achieving industrial value does not necessarily rely on large models. In this field, we’ve observed the importance the government places on AI empowering new industrialization. In fact, smaller models from the era of deep learning were the mainstream in this area, providing clearer value orientation. Considering the existence of Scaling Laws, do we truly need to depend on massive parameters and computational resources? I believe this is a critical question for today’s investment community to reflect upon.

Moderator (Qiu Zhun):
This discussion has delved deeply into the topic, and I hope it provides some inspiration for everyone. In fact, none of our perspectives may be entirely correct, but as long as we keep thinking, we continue to iterate. Personally, my viewpoint aligns closely with Mr. Fan’s. Just as the internet has its opposite, artificial intelligence (AI) can also have an opposite or an antonym. The antonym of the internet is quite clear—offline versus online, or standalone games versus online games. These boundaries are well-defined.

For artificial intelligence, however, my definition of its antonym is “rule-driven.” In my opinion, any system that is fundamentally rule-driven cannot be considered artificial intelligence. This perspective also serves as a judgment criterion from an investment standpoint. Artificial intelligence should be “data-driven,” not “rule-driven.” It must involve training through neural networks to build models. Therefore, when I evaluate a project, I don’t need to overanalyze what it does or debate whether it qualifies as AI. Instead, I start with two questions:

1. What is your training data? 2. How do you use that data to train your model?

As long as the project can answer these two questions and aligns with the “data-driven” logic, I can consider it AI. Today, many companies are working on intelligent agents, multimodal systems, or AI startups, and large corporations are investing heavily in AI. I use this perspective to evaluate them. If a company doesn’t meet these criteria, then to me, it is still not truly AI.

When I first went to the U.S. to study AI, the technology was not yet industrialized. From the 1960s to 2012, AI had been confined to academia and had not entered practical industrial applications. It wasn’t until the advent of deep learning in 2012 that AI began to achieve preliminary industrial breakthroughs and finally broke through academic constraints. Historically, most AI relied on rule-driven methods, particularly in the field of natural language processing, which was long dominated by rule-based systems. This led to slow technological progress and challenges in implementation. It wasn’t until deep learning emerged that AI technologies could be applied practically.

My personal judgment is that while some rule-driven components still exist within AI, an overreliance on rule-driven methods makes it difficult to join this wave of large-scale industrialization. This is the methodology I use to evaluate AI.

Now, I want to ask a second question: What is the biggest challenge we face today? If AI is defined as being based on neural networks, why have we not yet reached the next milestone? Is this challenge product-related? For instance, in the era of the internet, traffic was the driving force behind industrial development. Does artificial intelligence have a similar driving force? Why are we encountering this ceiling now? Why is it so difficult to break through? Is the difficulty caused by issues in products, technology, or data? For example, we are indeed facing a challenge of data exhaustion for training. Along the trajectory of AI, what do you see as the next big challenge?

Fan Yu:
This is an excellent question. To provide a concrete example, consider a parts bin filled with various components. Without using neural networks, it’s possible to identify the position of each part. However, if we opt for AI, it requires advanced algorithms, data annotation, training, and evaluation of the training results. If both methods are viable, but the non-AI approach demands less computational power, requires fewer human resources, and takes less time, why should we use AI? That’s a question worth considering.

Moderator (Qiu Zhun):
Rule-driven methods can achieve this as well.

Fan Yu:
Exactly. Today, we’re discussing the global transformation driven by cutting-edge industries. I believe for AI to truly make an impact, it must tackle entirely new problems and scenarios that traditional non-AI methods cannot address effectively.

In the robotics field, we encounter another issue—generalization. Why have humanoid robots and embodied intelligence gained so much attention over the past year? There’s a shared vision behind this trend. Humanoid robots, being highly similar to humans, serve as standardized hardware platforms, unlike other types of robots that often rely on diverse hardware platforms. Humans can perform a wide variety of tasks, such as driving, cleaning tables, and moving items. The goal is for humanoid robots to execute these tasks as well. Humanoid robots should function as a standardized hardware platform that, when paired with various software algorithms, can accomplish diverse tasks. This is precisely the critical challenge that large AI models need to address in the robotics domain.

Many tasks can already be efficiently solved through traditional methods or paradigms without the need for AI. The key is to identify application scenarios where AI can deliver a 10x efficiency improvement. In technical entrepreneurship, there’s often a pitfall of starting with technical solutions or routes, then developing products, and finally trying to establish a business model. The approach should be reversed: first identify a business and data loop, then select the appropriate technical solution based on the actual problem. This ensures that technology truly serves business needs, rather than “wielding a hammer and searching for nails.” Only by doing so can AI create greater value in practical applications.

Ni Tianyang:
If the integration of technology and industry, as well as how they converge, is itself a challenge, I think we often appear overly impatient. We assume that emerging AI technologies can quickly empower all industries, leading to universal transformation. In reality, the pace of AI integration varies by scenario and industry, and it often follows a gradual process.

Generative AI and large models, for example, still face numerous technical challenges that need to be resolved step by step. There’s also a certain logic to finding scenarios where new technologies can be applied. Generative AI tends to thrive in scenarios with high openness and high tolerance for error. For instance, in areas like social media, gaming, and e-commerce, AI applications are more feasible because these industries have a higher error tolerance and relatively abundant data. This enables faster adoption and industrialization of AI.

However, in more regulated sectors like healthcare, there are significant hurdles. Strict regulations, the unavailability of private data such as patient records, and proprietary, non-public R&D data make pretraining AI on massive datasets difficult. As a result, in industries such as healthcare and finance, where regulations are stringent, AI development must follow a gradual progression.

To summarize, the challenges of AI application and development are diverse. Different scenarios have varying degrees of compatibility with AI, resulting in different levels of implementation difficulty. Some industries are more suited to AI and can achieve industrialization more easily, while others require the joint efforts of technology, industry, and regulations. These sectors will need new regulatory frameworks and ecosystems to gradually evolve, which will take time. The challenges remain significant.

Lu Yue:
In my view, artificial intelligence technology is still in its developmental stage. While we have seen many practical applications, it would be an oversimplification to assume that models like Transformer and ChatGPT can solve all our problems. For example, in supply chain management, AI technology already assists in tasks like automated pricing inquiries, order-taking, and operations. However, in customs clearance, it remains challenging for large models to automatically process and organize documents into the format required for customs declarations. This step demands absolute accuracy—any error could lead to severe consequences, such as customs fines. Given our current technological capabilities, meeting this requirement is still difficult. Even with the advent of ChatGPT-5, this issue may not be entirely resolved. Further technological iterations are needed to meet these standards, which highlights the limitations of AI when applied to certain scenarios.

In practice, while technology is critical, it is not the sole determining factor. Often, the more significant barriers lie in the distribution of interests and power structures within industries. These factors significantly constrain the adoption of technology. Therefore, we must consider the broader industrial environment and actual demands, which are often more complex to address than the technology itself.

Moderator (Qiu Zhun):The challenges go beyond just the technology itself.

Xu Shan:
I completely agree with the points raised by the other speakers, particularly regarding the interplay between regulations and market dynamics. This relationship can often feel like a case of "regulations torment us a thousand times, yet we remain loyal to the market." Building on this foundation, I’d like to add a few thoughts:

First, many of the AI startups we’ve engaged with reveal that the field of AI has yet to establish clear commercial logic or business loops. Many companies are still exploring ways to achieve profitability, especially in creating viable business models. This is a significant commercial challenge we must confront.

Second, from a technical perspective, I am particularly concerned about the issue of "hallucination in large models." Currently, generative AI lacks robust mechanisms for evaluating the knowledge it produces. The accuracy and reliability of much of its output remain difficult to guarantee, despite numerous academic efforts to address this issue.

Lastly, I want to touch on a topic closely related to today’s theme of "going global"—how AI enterprises can expand into international markets. According to our analysis, the current strategic window for such expansion is roughly three to four years, with an immediate opportunity lasting just one to two years. In our earlier reports, we discussed the concept of "sovereign AI." When and how Chinese enterprises can bring their technologies and solutions to international markets is a critical question. If we fail to seize this strategic opportunity in time, other countries may gain a foothold in the global market, leaving us at a disadvantage.

Therefore, I am very eager to engage in deeper discussions with AI companies aiming for internationalization on this platform. By leveraging the resources and expertise of institutions like the China Academy of Information and Communications Technology (CAICT), I hope we can collaboratively build a knowledge framework to support AI companies in successfully navigating the global market.

Moderator (Qiu Zhun):
Let me briefly recap today’s discussion. I asked three questions. The first was: What is the definition of AI? What constitutes AI? The second, based on that definition, was: Which application scenarios have the potential to become emerging industries? In the era of AI, scenarios where AI can provide solutions are bound to transform into cutting-edge industries, just as scenarios that could be addressed by the internet in the internet era inevitably became innovative industries. From my perspective, this is driven by technology—whether it’s neural networks, Transformers, or adaptive learning. Once AI is defined, the application scenario where it can take root will become a cutting-edge industry. This is also the direction of our investments. Here, "investment" refers not only to financial and resource allocation but also to research efforts. We are all seeking answers and exploring which fields and angles will develop more rapidly.

The third question is: What challenges will AI face? In its process and trajectory, it will undoubtedly encounter obstacles, whether in terms of computing power, data, or regulation. Which of these poses the greatest challenge? If I can overcome one of these challenges, it demonstrates that I have built a competitive advantage. The answers to these three questions together form a business plan. First, how do I define the problem? Second, which application direction holds the greatest promise? Third, how do I address the biggest challenge along the way? This constitutes my competitive moat. These three questions combined should offer some inspiration to everyone.

Returning to today’s theme, let’s conclude by summarizing the environment for AI globalization with a quick keyword that looks to the future and anticipates trends. Ideally, it should be concise—one word or within five characters. What word or phrase do you think best captures the AI landscape, particularly in the context of globalization? It can also be a bit broader.

Ni Tianyang:
My keyword is “pull-driven opportunities.” Our AI globalization should not be a passive effort to compete externally. AI offers not only incremental opportunities but also innovative possibilities, opening up spaces similar to those in the mobile internet era. Reflecting on last year, I held a somewhat pessimistic view of domestic AI models due to the complexities of the U.S.-China confrontation. However, this year, domestic large models have made remarkable progress, particularly against the backdrop of data depletion that has rendered the "Scaling Law" less effective. This creates excellent opportunities for domestic entrepreneurs, as many shortcomings no longer pose critical bottlenecks. Recently, ByteDance's AI products have gained significant traction, showcasing how product capabilities and internet advantages accumulated during the internet era are now shining through in AI. China’s strengths, especially in productization and product-driven mindsets, enable users to overlook gaps between domestic foundational models and their foreign counterparts. The urgency for AI globalization is palpable, and we must seize this moment.

Fan Yu:
I have a vision: in the future, our products will increasingly integrate AI, and the proportion of our revenue from overseas markets will surpass that of domestic markets. My keyword is “massive AI.”

Lu Yue:
We stand on the eve of monumental change, with challenges and opportunities coexisting.

Xu Shan:
My keyword is “collaboration.” The overseas market is vast and boundless. If we can establish a robust mechanism for collaboration, we can advance together and support each other. Finally, I’d like to borrow from the host’s concluding remarks yesterday: “Let the wind blow to expand industries; let the rain fall to nurture growth.” I extend these words to all AI companies going global, wishing you a bright and thriving future. Thank you!

Moderator (Qiu Zhun):
My five-character phrase is “First Principles.” Everything should trace back to its origins. Whether you’re investing or starting a business, if you are going all-in on AI, you must think from the perspective of first principles. Understand how AI originated and evolved, so you can not only grasp the current situation more clearly but also anticipate its future trajectory and key milestones. I leave you with “First Principles.”