The Chaotic Era of Big Models Contradictions, Fragmentation, and the Future

Chaotic Era of Big Models Contradictions, Fragmentation, and the Future

After half a year of rapid development, China’s large-scale model industry has entered a new cycle. On one hand, there is frenzy, on the other hand, there is coldness. Investors are tirelessly searching for China’s OpenAI in this era. The top executives or scientists of large factories, who are surrounded by an elite aura, have decided to take a gamble. The AGI era, which was once unattainable, seems to have a visible timetable now due to the emergence of large models.

The intelligent life depicted in science fiction novels is gradually becoming apparent, and the relationship between technology and people, technology and industry, and human civilization and technological civilization seems to be entering a new stage of reconstruction.

According to the “China Artificial Intelligence Large Model Map Research Report” released by the China Institute of Science and Technology Information, as of the end of May, 79 large models with a parameter scale of over 1 billion have been released in China. In the following two months, under the joint participation of various forces, such as Alibaba Cloud’s Tongyi Wanxiang, Huawei Cloud’s Pangu 3.0,’s Yanxi, Ctrip’s “Wendao,” Netease Youdao’s “Ziyue”… whether it’s getting up early or staying up late, the number of AI large models released domestically has exceeded 120, and the “battle of a hundred models” is in full swing.

However, amidst all the excitement, there is also a gradual emergence of cold thinking about large models: Are there any good business models? The debate between open source and closed source ecosystems, the implementation of the 2B/2C route, none of these debates seem to have reached a consensus, and innovation and alternation are happening at any time.

Some harsh realities have also emerged: for example, the subtle competition between large companies and small companies, and shell products that lack a solid foundation of core technology have gradually lost their aura in the face of large models. For example, how to make AI go from being a toy to a tool in the 2C market requires unique insights into user needs, and how to ensure the controllability of technology and the usability of large model products in the 2B end poses higher requirements on the industry understanding of players.

In this chaotic era of large models, there is information asymmetry and cognitive differences, consensus and non-consensus. Hot money, talent, and scenario applications, computing power, data, and algorithms, all of them determine the direction of technology and the fate of entrepreneurs in this field.

01 Contradiction: Hot money pours in, but caution is exercised

Under the boom of large models, the restlessness of capital is a particularly clear trend. Throughout the history of human technology, those perceptive investors always choose to invest heavily in those “seeds,” and the competition for large models is no exception.

According to data released by research institution PitchBook, in the past six months, global venture capitalists have invested more than $40 billion (about RMB 290 billion) in AI startups. Against the backdrop of a harsh global investment environment, this indicates that AI startups are still thriving.

Among them, the most noteworthy investments are two: Microsoft’s $10 billion investment in OpenAI, and the startup Inflection AI, which was established in 2022 and reached a valuation of $4 billion after completing a $1.3 billion financing at the end of June.

One consensus is that a prosperous AI unicorn ecosystem has formed in the United States, beyond the ocean.

In addition to OpenAI, Anthropic, and Inflection AI, which are well-known, there are also Adept, which is a virtual AI robot assistant, Cohere, which focuses on B-end enterprise services, Stability AI behind the Stable Diffusion image generation diffusion model, and CoreWeave, a computing power provider favored by NVIDIA… In short, whether it is at the model layer, the middle layer, or the application layer, the overseas large-scale model ecosystem is clearer compared to that in China. Quoting a sentence from the report by “Geek Park”: “There are almost no new entrepreneurs who want to do the next OpenAI.”

In contrast, although capital has also flowed into the artificial intelligence industry in China, if we trace the flow of money, smart money still flows into a few leading companies. According to the statistics of Huxiu, there have been only 21 financing events in the AI large-scale model track since the release of ChatGPT. And the well-known star unicorn companies include MiniMax, Beyond Light Years, Baichuan Intelligence, and others. The rise of these unicorns is backed by early-mover advantages and the endorsement of industry giants.

Image source: @chiefaioffice

The previous debate between Zhu Xiaohu and Fu Sheng has sparked a debate in the venture capital circle about the value of large-scale models. Behind the “model refining” in the capital circle, investors are actually still very cautious. On the one hand, AI large-scale models are a highly specialized and segmented track, but they also burn a lot of money. Therefore, it is very scarce to find investors and investment institutions that focus on the AI field and accurately understand the technology.

On the other hand, there are still too few good targets. From the current investment logic of institutions, investing in people is still the main theme. Either it is endorsed by entrepreneurial giants like Beyond Light Years, even though the founder does not understand the technical principles, they have a good understanding of the changing trends and business models in the technology industry. Or it is known technology scholars in the AI industry, such as Zhipu AI, Lingxin Intelligence, and DeepSpeech Technology, all of which have the “Tsinghua system” behind them.

02 Differentiation: Big companies are frantically accumulating, small companies are desperately exploring

Behind the series of changes surrounding large-scale models are not only technological advancements but also the promotion of key individuals and key companies. If we turn our attention to these companies and individuals at the forefront of the wave, differentiation has already occurred.

Dai Yushen, a managing partner of ZhenFund, once made a clever analogy: the emergence of GPT-3 is equivalent to discovering a new continent, and the emergence of ChatGPT is like discovering gold on that new continent. The journey of Chinese companies catching up is like knowing about the new continent and where the gold is, knowing that OpenAI is going by ship, and knowing roughly what the ship looks like, but not having a detailed map.

After the previous crazy “release month” of large models, we can clearly see that this round of entrepreneurship is divided into the academic faction, the big shot faction, and the big factory faction. The relationship between them is not entirely zero-sum competition, but a kind of “non-zero-sum game”.

In the past period of time, the big factories, besides demonstrating their technical strength, have focused on building ecosystems. Taking Baidu, Alibaba, Huawei, ByteDance, and as examples, on the one hand, they have their own cloud business, providing computing power support. On the other hand, they have also made layouts around the chip layer, framework layer, model layer, and application layer to further solidify their barriers.

However, there are different strategies among the big factories. Represented by Alibaba, Baidu, and Huawei, they tend to take the path of vertical integration, achieving multiple benefits in terms of computing power, platform, and model. On the other hand, Volcano Engine (ByteCloud) and Tencent Cloud tend to take the platform path, building a supermarket for model shelves, integrating more third-party large models, and providing corresponding fine-tuning, evaluation, and inference services.

For domestic start-up small factories in the early stage of competing for large models, in fact, the only certainty for start-ups is “uncertainty”. They don’t need very complex products. By targeting the pain points of users, they can achieve preliminary success.

The recently popular “Miya Camera” is a typical case. The team said in an interview, “AIGC’s product may not receive payment on the first day.” By using a low threshold, accurately targeting the portrait needs of women, and combining social media marketing, even without obvious technological innovation, Miya Camera achieved early commercialization with a single function, which actually provided a good inspiration for domestic application-layer start-ups.

For Miya Camera and many other start-up companies, the key is how to seize the “uncertain” cycle to further solidify their technological barriers and user stickiness.

Image source: Miya Camera screenshot from Xiaohongshu

03 Future: Increasing regulation, uncertain situation

In the foreseeable future, perhaps as stated in the PR articles of the big factories, large models will eventually empower various industries. However, beyond the ideal, ensuring the security and controllability of large model technology has also become a focus of attention.

Previously, the Cyberspace Administration of China and other seven departments jointly announced the “Interim Measures for the Management of Generative AI Services,” providing reliable legal basis for the compliant and healthy development of generative AI in terms of regulatory methods and scope. And in the early morning of August 1st, Apple’s China App Store delisted multiple AIGC applications, which in fact indicates the increasing regulatory efforts on artificial intelligence.

Overseas, technology giants are facing controversial AI ethics. “AI Four Giants” Anthropic, Google, Microsoft, and OpenAI have formed the Frontiers of Model Forum to communicate responsible and secure artificial intelligence issues with the United States, Europe, and the G7. The alliance composed of open-source communities such as Hugging Face, GitHub, and EleutherAI is also calling on EU policy makers to protect open-source innovation when formulating the “EU AI Act”.

For entrepreneurs in the large model industry, in addition to entrepreneurial ideals and commercialization paths, considerations for the compliance of business models will also be incorporated into existing plans.

In addition to clear regulatory trends, more explorations are also taking place. Discussions on a series of topics such as multimodality, AI agents, vector databases, and embodied intelligence are currently taking place in the industry, seeking more possibilities beyond the wave of large models.

Take AI robots in the field of embodied intelligence as an example. Technology giants including Google are integrating large language models into robots to make them smarter. The equally popular wave of AI agents, even called “primitive AGI”, has taken over large models and become the next field of concern for big companies.

The wave has arrived and the future is here. What can be certain is that the chaotic era of large models may not last long, but in the coming period of time, competition and collaboration will continue. Who can use “uncertainty” to make up for their shortcomings? Who can truly implement the capabilities of large models in specific and vertical scenarios? Who can build a high-quality data flywheel faster? These will test their determination and endurance, and also determine their ecological position in the next round of competition.