■ The emergence of ChatGPT has significantly boosted industry-wide and external interest in AIGC. After this round of development, what stage is the industry currently in? As an investment firm, at what point is it appropriate to enter the market?

Compared to the U.S., where the first phase of investment is largely finalized and now moving into its second stage—with significant efforts being made at the application layer—China, due to challenges like insufficient hardware infrastructure, still lags behind the U.S. in terms of investing in AIGC’s ecosystem and underlying infrastructure. However, looking ahead over the long term, we remain confident in Chinese entrepreneurs' ability to drive innovative applications within the AIGC space. After all, when viewed over several decades, the industry is still in its early stages of development.

The next wave of the industrial revolution has already begun. Looking back at previous industrial revolutions, it was only after the invention of the steam engine and the development of railways that massively profitable, large-scale companies emerged. Right now, we’re in a similar phase—much like when the new generation of the "steam engine" was being invented, though the railway infrastructure hasn’t yet been fully laid out. During this transitional period, many product forms may remain temporary and phased, eventually fading away altogether. Consequently, the business models built around these evolving, short-lived products might not prove sustainable—or durable—in the long run. Yet, precisely because of their fleeting nature, these innovations are undeniably thrilling and captivating. Just over the past year alone, the rapid evolution from ChatGPT 3.5 to 4 has delivered one breathtaking milestone after another. But before its release, there wasn’t even one truly exhilarating moment each year. This, too, serves as an intriguing indicator of where we’re headed.

Looking back at today, there are many large models and products that outperform even the offerings from the internet giants currently on the market—but they haven’t gained widespread traction yet, largely due to their high costs. Yet, the brightest minds in humanity are already hard at work on this front. In fact, the number of generative AI papers published each month now surpasses the total across every other industry sector combined, as documented in the world’s largest academic database. So there’s absolutely no reason for us to lose faith in AIGC.

Regarding investments, some partners are focusing on infrastructure companies, capitalizing on industry trends to reduce costs and boost efficiency. Others believe the market is dominated by industry giants, emphasizing early-stage investments and teams that can form strategic alliances with these leaders. Meanwhile, others are eyeing application companies equipped with strong model capabilities.

■   How do you view open-source versus closed-source approaches? Will open-source large models impact the development of closed-source large models?

In the field of large models, there’s a strong likelihood that a recurring pattern will emerge: open-source approaches play a critical role during the collaborative, human-driven R&D phase, while closed-source strategies may become more crucial in achieving ultimate commercial success.

Open-source models are undoubtedly the most critical element driving industry progress at the R&D level. They enable the entire ecosystem to scale significantly, attracting a diverse array of world-class developers and accelerating the commercialization of applications. Essentially, open source provides a platform for building on the shoulders of giants—but it also comes with inherent risks. Since open-source models aren’t controlled by individual application vendors, if you’re looking to fine-tune them for your own products, there’s no guarantee that your customizations will be incorporated when the model is updated. On the flip side, closed-source approaches remain popular among China’s major tech giants and leading internet companies, which often prefer to manage every stage of development in-house. Even if they currently rely on closed-source solutions, these organizations may eventually pivot to open-source models—giving themselves the flexibility to modify and adapt as needed. For startups focused solely on developing proprietary, closed-source large-scale models, this approach could put them at a disadvantage when competing against industry titans down the line. That said, teams with strong internal model-building capabilities—whether by tweaking open-source frameworks to create consumer-facing (to C) products or by developing their own closed-source offerings—may find themselves better positioned for long-term survival. Ultimately, though, predicting who will ultimately prevail remains uncertain.

Today, the dozen or so companies widely recognized as leaders in the market are already taking vastly different approaches. Some are focusing on open-source foundational models, while others are first honing their models for highly specific applications before fine-tuning them further. Take chatbots, for instance: OpenAI’s GPT chatbot lacks emotional depth and personalized content—it’s essentially trained to act like a versatile little expert who knows everything. But if a major model company sets out with the original goal of creating human-like, empathetic companions, its dataset would certainly differ from OpenAI’s general-purpose data set. At the very least, it would incorporate multi-turn emotional dialogue during pre-training. While today’s large-scale models may appear similar on the surface, the paths taken by closed-source models are already diverging dramatically. In the future, truly groundbreaking applications are likely to emerge precisely because of the fundamental differences in the underlying technology—and how these distinctions set apart one model from another.

Of course, we’ll keep an eye on this. In May, Google held an internal discussion, during which some perspectives suggested that there aren’t absolute technological barriers between them and closed-source companies—after all, this is a complex, multifaceted battle.

■   Given the broader domestic context, how do you view the commercial applications of AI in both B2B and B2C sectors?

B2C needs to keep an eye on how the market evolves. At this stage, B2B applications are relatively more advanced, but in the long term, the C2C market and its commercial scale could potentially grow even larger.

In the B2B space, if China's large-scale AI models can capture an opportunity similar to the security sector, it might even give rise to B2B startups in security that outgrow their U.S. counterparts—though this remains highly unpredictable. Moreover, China’s B2B market holds a unique advantage over the U.S., thanks to its 97 centrally-administered state-owned enterprises. Even basic customer service tasks could inadvertently lead to user data breaches, and as companies increasingly rely on large models to inform critical decisions within their core business systems, they’ll inevitably seek partnerships with firms capable of offering private, model-deployed solutions via dedicated cloud infrastructure. This makes China’s B2B landscape continually ripe with potential opportunities. However, investing in B2B isn’t without its challenges. For instance, major tech giants have already secured many of the key entry points across various industry segments, making it particularly tough for smaller players to carve out meaningful footholds in highly specialized B2B niches.

For C, it always represents the ultimate embodiment of value amplification in science and technology. We’re excited about the opportunities to invest in large-scale commercial applications targeting consumers. One group consists of young, dynamic teams equipped with robust model capabilities—they possess cutting-edge technical expertise but are unburdened by past successes. The other group comprises seasoned veterans with deep industry experience and a keen understanding of the landscape; when new productivity tools—or "new hammers"—emerge, they’re eager to tackle the "old nails" with fresh innovation.

Additionally, whether it’s B2B or B2C, the core model-layer capabilities and product-specific use cases of a team are crucial. The full potential of large models to significantly impact the application layer remains largely untapped. If you rely on a third-party model over which you have zero control of the pre-training data, it will be virtually impossible to build a high-quality application. This becomes even more evident when dealing with more complex multi-modal and industry-specific models. Therefore, in the short term, only companies that truly master the core model-layer capabilities will be able to succeed. As for product scenarios, if a dominant player already occupies the existing market space—whether they choose to continue developing products using open-source or proprietary models—they’ll likely remain well-positioned to serve their original user base. In such cases, opportunities for new startups tend to be limited. That said, there are still two key avenues: First, in industries where no single giant currently dominates—for example, gaming—startups can continuously innovate and introduce fresh, cutting-edge products. Second, there are entirely new domains where large models or relevant use cases haven’t yet emerged at all; these represent fertile ground for entrepreneurial ventures to carve out their own niche.

■   How do you view the combination of AI-generated content (AIGC) and embodied intelligence? Could this represent a significant opportunity for China in the future?

Technologically, embodied intelligence has achieved significant breakthroughs thanks to its integration with AIGC. Embodied intelligence is technically structured into three layers: the top two layers represent the "brain"—the software layer and the AI layer—while the bottom layer consists of actuators tailored for various real-world scenarios. The top two layers have already seen major advancements—first, the emergence of language models that are revolutionizing how robots interpret complex human intentions, and second, the use of advanced reinforcement-based models that simulate countless environments for robots, enabling them to independently tackle challenges in virtual spaces. Meanwhile, the biggest challenge for actuators remains scalability; however, reducing costs through large-scale production is clearly within reach.

Looking back at the past 40 years of reform and opening up, China has steadily evolved—from starting with ODM/OEM models and gradually expanding into the electronics and integrated circuit industries—to eventually becoming the world’s leading producer of nearly every product category recognized by the United Nations. Thanks to the spillover effects of its robust industrial chain capabilities, embodied intelligence now stands at the cusp of a wealth of structural opportunities across numerous systems. The advanced manufacturing expertise honed over decades in sectors like smartphone and automotive production, combined with cutting-edge AI technologies, is already enabling the creation of intriguing, large-scale products. Meanwhile, early explorations into robotics have already yielded promising prototypes tailored for diverse applications—ranging from industrial settings to home-companion scenarios. As technology becomes more accessible and affordable, we anticipate a surge of innovative use cases emerging across this space. On the talent front, we’re witnessing a growing influx of skilled professionals—such as vision specialists and autonomous-driving experts—entering the field of embodied intelligence. Coupled with the deep reservoir of industrial management talent cultivated over the past two to three decades, this convergence positions us exceptionally well to capitalize on the immense investment opportunities ahead in this dynamic and rapidly evolving sector.