AI gods begin to shuffle wildly

Author | Huang Yu

Editor | Zhou Zhiyu

In the blink of an eye, it has been two years since ChatGPT sparked the wave of AI large models, but the huge business opportunities that many people anticipated have yet to arrive.

Unlike the enthusiastic atmosphere of last year, the AI large model entrepreneurship is entering a turbulent period. Since the second half of this year, there have been frequent reports of core technical backbones from major domestic AI companies and large model teams leaving their positions, creating an atmosphere of unrest.

First, Huang Wenhao, co-founder of Lingyi Wanyi Technology, left, followed by Zhou Chang, the technical head of Alibaba Tongyi Qianwen, who transferred to ByteDance. Additionally, Liu Wei, the technical head of Tencent's Hunyuan, and Yan Shuicheng, an AI expert from Kunlun Wanwei, have also left their original companies.

Talent mobility is a barometer of the AI industry's development. Behind the departure of these technical backbones are multiple challenges faced by AI large models, including a slowdown in technological iteration and unsatisfactory commercialization. Everyone is actively adjusting, seeking what they believe to be the correct path or direction.

The AI large model industry, currently in a state of developmental confusion, is undergoing a restructuring of technology, capital, and talent, with a reshuffle quietly taking place. The trend of industry consolidation will become increasingly evident in the future.

In fact, such stories play out during every technological wave. A consensus in the industry is that after a fierce competition, only a handful of large model companies will play important roles in the future, and only those who have deeply participated in the development of unicorns can become the final lucky ones.

This is a brutal competition with unpredictable outcomes. Those involved can only give their all.

Turbulence

In the competition of cutting-edge technologies, talent can be said to be the most important competitive advantage. In the rapidly iterating field of AI large models, talent is the key factor determining whether underlying technologies and products can keep pace and ultimately rank among the industry's top tier.

Several investors have told Wall Street Journal that in this wave of AI large models, what investors value most when assessing investment projects is still the talent team, as this determines whether there is a sustainable capacity for technological iteration.

However, whether in large companies or AI startups, the talent that gathered during the initial enthusiasm is now making new choices, either actively or passively, due to the impact of reality.

According to confirmation from Wall Street Journal, Liu Wei, an outstanding scientist at Tencent and one of the technical heads of Tencent's Hunyuan large model, as well as the head of the AI Lab's computer vision center, has recently left Tencent. Reports indicate that Liu Wei has started a new venture in Singapore, focusing on the video generation field.

Not long ago, Kunlun Wanwei also announced that Yan Shuicheng would no longer serve as the director of its 2050 Global Research Institute and would instead take on the role of honorary advisor. As an expert in computer vision and machine learning, Yan Shuicheng joined Kunlun Wanwei only last September, helping to build the 2050 Global Research Institute from scratch and conducting in-depth research on next-generation model architectures and agents.

In this wave of talent turbulence, more individuals are choosing to move from AI startups to large companies or to transition from one large company to another Currently, ByteDance, which is actively preparing for the establishment of a large model research institute in the second half of the year, is the biggest winner in this wave of talent movement.

After the departure of Qin Yujia from Mianbi Intelligent, it was reported that he would join ByteDance's large model research institute in the second half of 2024; in August this year, Huang Wenhao, co-founder of Zero One Wanwu Technology, joined ByteDance's model algorithm team Seed, reporting to Zhu Wenjia, the head of ByteDance's large model; in October, it was also reported that Zhou Chang, the technical head of Alibaba's Tongyi Qianwen large model, joined ByteDance.

It is worth mentioning that Zhou Chang's departure has also led to a lawsuit. On November 13, it was reported that Zhou Chang violated a non-compete agreement, and Alibaba has filed a labor dispute arbitration application.

Recently, Yang Zhilin, founder of the Dark Side of the Moon, commented on the phenomenon of some talents returning to large companies, stating that this is normal. "The industry has entered a new stage of development, transitioning from many companies initially involved to now fewer companies participating, and what everyone does will gradually become different. I think this is an inevitable trend."

The training investment for large models is significant, and even large companies must make trade-offs. The emergence of the "text-to-video" model Sora at the beginning of the year once sparked a global competition for AI video generation; however, OpenAI announced a delay in Sora's updates due to a shortage of computing power, resulting in it not being open to the public yet.

Clearly, before clarifying the landing scenarios and commercialization returns, "Sora-like" video generation models will not become a key focus for Tencent. In this context, Liu Wei, who wants to make a mark in the video generation field, will naturally seek other opportunities.

A domestic large company investment head told Wall Street Insight that there has also been frequent talent movement overseas this year, mainly due to AI large model teams facing short-term technical bottlenecks and challenges in slow commercialization. In the future, a large number of domestic AI startups will face the fate of funding chain breaks and being absorbed by large companies.

In addition, Shen Meng, executive director of Xiangsong Capital, pointed out to Wall Street Insight that behind the frequent talent movement, on one hand, is the lack of deep research and innovation in domestic large models, making personnel movement barriers between teams smaller; on the other hand, it reflects industry restlessness and a bubble in the number of models.

Future

Artificial intelligence, as a discipline, has been around for over 60 years, during which it has experienced multiple technological waves. In the early stages of some of these waves, it was as hot as the current AI large models. In the early stages of the AI wave that began in 2016, tech companies also spared no effort in competing for top AI talent.

However, the patience of capital is far from sufficient to support the research of AI scientists. When AI technology fails to bring commercial returns for a long time, both internet giants and star AI companies begin to return to rationality and reassess the "value" of AI talent, leading to faster talent movement.

History tends to repeat itself; after a period of fervor, the AI large model industry will also enter a phase of consolidation and clearing.

In October 2022, ChatGPT sparked a global wave of AI large models, triggering a hundred model battle in China, with startups emerging like mushrooms after rain, and internet giants jumping in, proclaiming to "All in AI." However, after one or two years of exploration, more and more companies have come to realize that the lucky ones who can survive until dawn are just a few.

Baidu founder Robin Li has previously pointed out bluntly that, like many technological waves in history, after going through the initial excitement phase, the technological bubble of generative AI is inevitable. Then, when this technology fails to meet the high expectations of the initial excitement phase, people will feel disappointed.

Robin Li predicts that during the AI bubble-bursting phase, pseudo-innovations that cannot meet market demands will be washed away, and after that, 1% of companies will stand out, continue to grow, and create tremendous value for society. "Right now, we are just experiencing this phase; the industry is calmer and healthier than last year."

All AI large model teams have reached a crossroads of making choices.

The most lively aspect of the domestic large model industry was the price-cutting wave in the first half of the year. According to Zhang Peng, CEO of Zhiyu, this phenomenon indicates that everyone is unable to find differentiated value points and can only compete on price.

Zhang Peng revealed that he has recently seen many leading companies in self-developed large models starting to turn back because they found that this matter is not so easy; it is not just about forming a team and using an open-source model to run it, but rather better to procure.

In addition, at the beginning of October, there were reports in the market that among the six companies known as the "AI Six Little Tigers"—Zhiyu AI, Lingyi Wanju, MiniMax, Baichuan Intelligence, Dark Side of the Moon, and Jielue Xingchen—two companies have decided to gradually abandon pre-trained models, reducing the number of personnel in their pre-training algorithm teams and shifting their business focus to AI applications.

Yang Zhilin believes that there is still half a generation to a generation of space for pre-training, and this space will be released next year, with leading models achieving a relatively extreme stage of pre-training. The next key focus is reinforcement learning, which is still scaling, but through different methods of scaling.

At the same time, Dark Side of the Moon has also actively chosen to do business subtraction, focusing on doing one business product well. Yang Zhilin revealed that Dark Side of the Moon will judge based on the situation in the U.S. market which business has a higher probability of becoming large in the end. Focusing on the highest potential matters must also be closely related to the mission of AGI.

The ultimate goal of AI research is to achieve Artificial General Intelligence (AGI).

"Rome has always been there, but the path taken is different." Recently, Kang Zhanhui, director of Tencent's machine learning platform, stated that everyone has thoughts about AGI; the next two to three years are relatively well-planned, but the paths taken by different companies may vary. For example, Tencent has chosen to pursue the mixed expert model (MoE) structure.

However, regardless of the chosen path, all AI prospectors face a common challenge: high computing power brings high costs, but there is currently no commercial monetization path to cover these high costs.

Shen Meng, executive director of Xiangsong Capital, told Wall Street Journal that large models will soon enter a painful period of survival of the fittest, where certain technologies and products that push towards core underlying technologies have a better chance of gaining market recognition.

This is a once-in-a-century technological revolution, but without sufficiently mature technology and reliable business models to support it, AI large models are likely to end up like VR and the metaverse in previous years, where the craze fades away, leaving a mess behind The knockout stage has now begun. Before the arrival of the "iPhone moment of AI," all companies must demonstrate sufficient patience and high sensitivity to face the brutal challenges ahead