Is the domestic big model industry set to experience a boom?

Has "Open AI" become "Close AI"?

On June 25, many Chinese developers received emails from Open AI, mentioning that "your organization is obtaining API traffic from regions not supported by Open AI"; starting from July 9, Open AI will take additional measures to interrupt services to unsupported countries and regions.

Open AI's official website shows that it currently supports 188 countries and regions, with mainland China and Hong Kong not among them. Unsupported regions attempting to access Open AI's API services may face the risk of account suspension.

Source: Weibo Screenshot

Behind Open AI's actions is the fact that the US has repeatedly targeted the domestic technology sector recently. In May alone, the US announced the revocation of export licenses for Qualcomm and Intel to Huawei and included 37 Chinese entities on an export control "list".

What impact will this blockade have on China's rapidly developing big model industry?

01. Is Hong Kong the only "escape hatch"?

In fact, Open AI's withdrawal had early warnings.

On February 14, Open AI issued a statement stating that it had prevented and restricted the malicious use of AI by five "state-related threat actors," with two of these users coming from China, and the others from Iran, Russia, and North Korea.

On June 13, Open AI appointed Paul Nakasone, a retired U.S. Army general, former commander of Cyber Command, and director of the National Security Agency (NSA), as a member of its board. This new director is believed to help Open AI "better understand how to leverage artificial intelligence to quickly detect and respond to cybersecurity threats."

On June 22, the US also released a draft rule that mentions banning or restricting Chinese investment in AI and other technology areas.

On this basis, Open AI's announcement is actually a further tightening of restrictions on domestic developers.

According to "Qujie Shangye," there are currently two main channels for using Open AI technology in China: one is to connect to the API provided by Open AI officially; the other is to connect to the Open AI technology provided by Microsoft's intelligent cloud Azure.

Source: Canned Image Library

It should be noted that China was not included in the regions that Open AI previously allowed to connect to the API, so the first channel was inherently non-compliant. Many users who adopted this method would circumvent geographic restrictions to access Open AI's API using foreign proxy addresses for product testing and development.

Since many of the developers who received emails were practitioners of this channel, there is a possibility that Open AI has already been able to detect "tech tricks" through technical means. The path of connecting to the API domestically through virtual addresses may already be blocked, but the stability and accuracy of Open AI's detection methods remain to be verified.

As for the second channel, which was originally the only way to use ChatGPT domestically, it also seems to be subject to restrictions, although Microsoft has not completely blocked it. According to "Cailian Press," Microsoft will continue to provide Azure Open AI services in Hong Kong.

Source: Ifeng.com Screenshot

It can be said that Open AI's strategic adjustment this time takes into account multiple pressures such as data security, compliance, operating costs, and risk prevention and control.

Zhang Xiaorong, Dean of the Deep Tech Research Institute, believes that Open AI's current challenge to China may be influenced by the US government's policy towards China on the surface, requiring compliance; in reality, the company has "metamorphosed" from a commercial company into a national tool. The company's board of directors has just introduced the former NSA director, and Ilya Sutskever, the CTO who insisted on the scientific research development path, has resigned. Musk, who had been suing Open AI, has also stopped speaking.

02. Who benefits, and who gets caught in the crossfire?

Regarding the impact of Open AI's withdrawal, the domestic big model industry says the impact is relatively small.

Zhou Hongyi, the founder and chairman of 360, issued a video stating that OpenAI's shutdown service cannot suppress the development of China's big models. "This has no impact on enterprise applications and entrepreneurs, who can deploy local big models themselves. Big models will gradually migrate locally, faster and safer."

Source: Weibo Screenshot

According to "Qujie Shangye," other industry insiders have expressed similar views. On the one hand, Open AI did not provide services to China, so many companies using Open AI technology would deploy overseas sites on their own; on the other hand, Open AI's usage agreement also prohibits the use of outputs to develop competing models. If major domestic internet companies secretly use Open AI interfaces for a long time, they would first be violating the rules, and secondly, it would limit their own development, which is not worth the effort.

At the end of 2023, ByteDance was exposed to have called Open AI's API in multiple research and development stages of an AI big model project called the "Seed Plan" and used ChatGPT's output data for model training. Open AI subsequently banned some GPT usage permissions under ByteDance.

Later, ByteDance responded that only some engineers had called GPT's API in smaller experimental projects during the initial exploration of big models; in April 2023, the company had explicitly required that GPT model-generated data not be added to the training dataset of Byte's big models; in September, the company also conducted an internal inspection to ensure compliance.

Source: Weibo Screenshot

Major companies understand that using Open AI's model outputs will only keep them in its shadow. Therefore, those affected more by Open AI's withdrawal are domestic companies and personnel without R&D capabilities who "shell" ChatGPT and attempt to train AI systems and applications through Open AI's tools.

Some netizens have shared on forums the ChatGPT shell websites they have seen. Using a website traffic checker to analyze the site, they found that it could generate a monthly income of $21,000. The profit comes mainly from advertising, user donations, and even paying users can enjoy additional features.

The website can perfectly serve as a substitute for ChatGPT, without the need for registration or login, allowing users to fulfill writing or other needs.

Source: Forum Screenshot

In addition, the Shanghai market regulatory department has also investigated a ChatGPT commercial confusion case. Experts analyze that there are mainly two modes of "shell AI": one is to act as a "relay" for users by connecting to official platforms; the other uses a fake platform entirely, disguising itself with similar names and logos to ChatGPT.

Zhang Xiaorong said that some emerging big model companies, especially those that build products and services through simple shelling or relying entirely on Open AI's API interfaces, may face the risk of business interruption or even direct delisting.

In Zhang Xiaorong's view, these Chinese companies and developers who rely on Open AI API for project development and innovation may need to switch technology bases, but this will also increase R&D costs and time costs; the service effect after the switch may be discounted, as domestic models may not understand foreign languages as well as Open AI.

Overall, the impact of the withdrawal is relatively limited, and domestic startups in the relevant field will not collapse in large numbers, nor will there be a sudden "overtaking" in the short term.

03. Domestic big models "running for the top position"

According to "Qujie Shangye," many domestic leading big model manufacturers, including Alibaba, Baidu, Tencent, Baichuan Intelligence, and Zhipu AI, have already presented users with a "relocation plan" to achieve technology base conversion.

For example, Zhipu AI stated that it would provide developers with generous resource packages, plus migration training from Open AI to its own big model; Baidu Intelligent Cloud highlighted "zero-cost switching" and launched the "Homeland Cloud" big model inclusive plan; AI Infra vendor Silicon-based Flow even made top open-source big models such as Qwen2-7B, GLM-4-9B, and Yi-1.5-9B permanently free.

Source: Weibo Screenshot

Han Sirui, assistant professor at the Hong Kong University of Science and Technology and senior business development manager at the Hong Kong Generative AI Research and Development Center, said that this incident will promote the increase in market share of domestic big model companies in the short term.

However, in the long run, whether domestic big models can fully replace Open AI ultimately depends on model capabilities.

In this regard, domestic big models have indeed made considerable progress compared to last year.

In 2023, when Baidu released Wenxin Large Model 4.0, Robin Li said that its comprehensive level was comparable to GPT-4.

In the early hours of June 27, Clem, the co-founder and CEO of the globally renowned open-source platform huggingface, announced on social media that Alibaba's latest open-source Qwen2-72B instruction fine-tuning version became the top-ranked open-source model, surpassing more than 100 mainstream open-source big models worldwide.

Specifically, in terms of the Chinese capabilities of the model, many domestic models perform better than OpenAI.

A two-month evaluation conducted by the "Daily Economic News" in collaboration with over 30 excellent journalists, editors, and engineers on the performance of mainstream big models in financial news work scenarios showed that Yi-Large, incubated by Kai-Fu Lee's team, ranked first in the total score of four application scenarios: "financial news headline creation," "Weibo news writing," "article error proofreading," and "financial data calculation and analysis." ChatGPT 4.0 performed poorly, even ranking last in the "financial news headline creation" scenario.

Source: Daily Economic News

In addition, in terms of price, domestic big models are also more affordable. In Alibaba's free migration plan, the API pricing of the qwen-plus big model is only 1/50 of OpenAI's.

However, objectively speaking, Han Sirui also admitted that OpenAI's model capabilities are still relatively advanced in complex tasks such as parsing and generating complex documents in text models. The demand for these complex model capabilities will not change significantly in the short term due to the API service suspension.

The US big model industry started earlier. In May 2023, the number of foundational big models with over 1 billion parameters in the US had exceeded 100; in addition to the well-known ChatGPT, representative US general big model companies also include Anthropic, Cohere, and Google.

Among them, Anthropic was once called "OpenAI's rival." According to its released rankings, its newly released AI big model, Claude 3.5 Sonnet, outperformed the once popular GPT-4o in areas such as graduate-level reasoning, coding abilities, and text reasoning.

Source: Weibo Screenshot

Wang Peng, a researcher at the Beijing Academy of Social Sciences, believes that the differences between Chinese and American big models are mainly reflected in three aspects: financing level, the development level of foundational big models, and the development level of the application layer.

According to dealroom.co data, as of July 10, 2023, global generative AI enterprise financing reached $15 billion, with 89% flowing to US start-ups.

In terms of the development level of foundational big models, domestic big models still face issues such as a lack of total data volume, lack of computing power resources, and limited scenario penetration. After all, from the perspective of publicly available data, English data dominates, and the US is also taking various measures to restrict China's access to core computing power resources.

As for the application layer, China is also in a following position, with more significant lags in office, finance, and medical fields compared to the US.

Source: Canned Image Library

This has raised some concerns about domestic big models. "Bloomberg News" mentioned in a report that Liang Zhongwei, CEO of Singapore's Dorje AI, said he was concerned that open-source big models such as Meta's Llama might also cut off access for Chinese developers, bringing greater uncertainty.

However, it cannot be ignored that China has a huge market size and rich application scenarios, providing vast space and conditions for the landing and application of big models. Moreover, the more data and scenarios, the more practical big models will become. This gives China an opportunity to catch up with the US, even though it lags slightly in underlying R&D technology.

However, in the process of catching up, China will encounter more and more obstacles, which may be a certain trend and challenge for domestic manufacturers.