Recently, OpenAI announced that it would "suspend services" in China, a decision that has undoubtedly stirred up a thousand waves in China's AI industry. Is it the "poison" of technological blockade, or the "assist" of independent innovation? This incident has not only attracted widespread attention from domestic developers, but also become the focus of discussion in the industry. This article will delve into the multiple factors behind OpenAI's move and its far-reaching impact on China's large model industry, giving you a glimpse of this new change in the technology field.
"Open up" close?
On June 25, a number of developers in China received an email from Open AI mentioning that "your organization is getting API (Application Programming Interface) traffic from regions that are not supported by Open AI"; From July 9, Open AI will take additional measures to interrupt services in non-supported countries and regions.
According to Open AI's official website, it currently has 188 supported countries and regions, Chinese mainland and Hong Kong are not among them, and non-supported regions may face the risk of being banned if they try to access Open AI's API services.
Source: Weibo screenshot
Behind Open AI's action, the United States has recently repeatedly "attacked" the domestic technology field. In May alone, the U.S. announced that it had revoked Qualcomm and Intel's export licenses to Huawei and added 37 Chinese entities to its export control "list."
What impact will this blockade have on China's rapidly developing large-scale model industry?
01. Hong Kong has become the only "gap"?
In fact, Open AI's "service suspension" has been foreshadowed for a long time.
On February 14, Open AI issued a statement saying that it had blocked and restricted the malicious use of AI by five "state-associated threat actors": China (two), Iran, Russia, and North Korea.
On June 13, Open AI appointed retired U.S. Army general, former commander of Cyber Command, and director of the National Security Agency (NSA) Paul Nakasone to its board of directors. The new director is believed to help Open AI "better understand how to use AI to quickly identify and respond to cybersecurity threats".
On June 22, the U.S. also released a draft rule that mentions prohibiting or restricting Chinese investment in AI and other technologies.
On this basis, the announcement of Open AI is actually to further strengthen the restrictions on domestic developers.
According to "Qujie Business", there are currently two mainstream channels for using Open AI's technology in China: one is to connect with the API officially provided by Open AI; The second is to connect with the Open AI technology provided by Microsoft's intelligent cloud Azure.
It should be noted that Open AI did not allow API integration in China before, so the first channel is non-compliant. Users who use this method will mostly use "technology and ruthlessness" to use foreign proxy addresses to bypass geographical restrictions to access Open AI's API for product testing and development.
Since many of the developers who received the email this time are practitioners of this channel, there is a possibility that Open AI has been able to detect "technology and ruthlessness" through technical means. The road of connecting APIs in China through virtual addresses may have been blocked, but the stability and accuracy of Open AI's detection method have yet to be verified.
The second channel, which was originally the only way to use ChatGPT in China, seems to be difficult to escape the restrictions, but Microsoft has not completely blocked the road. According to the Financial Associated Press, Microsoft will continue to provide Azure Open AI services in Hong Kong.
Source: Screenshot of Phoenix.com
It can be said that the adjustment of Open AI's strategy has multiple pressures such as data security, compliance, operating costs, and risk prevention and control.
Zhang Xiaorong, president of the Deepin Technology Research Institute, believes that Open AI is making difficulties in China at this time, and it may be affected by the US government's China policy on the surface, and it needs to be compliant. In fact, the company has "deteriorated", turning from a commercial company into a state tool. The company's board of directors has just introduced the former director of the NSA, and at the same time adheres to the path of scientific research and development, CTO Ilya has resigned, and Musk, who has been about to launch a lawsuit against Open AI, has also stopped speaking out.
02. Who benefits and who lies the gun?
Regarding the impact of Open AI's "service suspension", the domestic large-scale model industry said that the impact is relatively small.
Zhou Hongyi, founder and chairman of 360, posted a video saying that OpenAI's shutdown service could not suppress the development of China's large models. "This has no impact on enterprise applications and entrepreneurs, who can deploy their own local large models. Large models will gradually be migrated locally, which is faster and more secure. ”
Source: Weibo screenshot
According to "Fun Solution", other industry insiders have also expressed similar views. On the one hand, Open AI does not provide services in China, so many companies that use Open AI technology will deploy overseas sites on their own. On the other hand, the Open AI use agreement also stipulates that it is forbidden to use output to develop competitive models, and if the leading domestic Internet companies secretly use the Open AI interface for a long time, it will first violate the law, and secondly, it will restrict its own development, and the gains outweigh the losses.
At the end of 2023, ByteDance was revealed to have called Open AI's API and used the data output by ChatGPT for model training in multiple R&D stages of an AI large model project called "Seed Project". As a result, Open AI also banned some GPT permissions under ByteDance's name.
Later, ByteDance responded that it was only in the early stage of exploring large models that some engineers had called GPT's API in experimental projects of smaller models; In April 2023, the company has clearly required that the data generated by the GPT model shall not be added to the training dataset of the byte large model; In September, the company also conducted an internal inspection to ensure compliance.
Source: Weibo screenshot
Big manufacturers understand that the output of Open AI's model will only live in its shadow forever. Therefore, more enterprises and personnel in China who do not have R&D strength, "shell" ChatGPT, and attempt to train AI systems and applications through Open AI's tools are affected by the "suspension of services" of Open AI.
Some netizens once shared on the forum that they saw the ChatGPT shell website, which used a website traffic checker to analyze the website and found that it can generate $21,000 in revenue per month. Revenues are made through advertising, donations, and even extra features for payment.
The website works perfectly as an alternative to ChatGPT, allowing users to fulfill their writing or other needs without the need to register or log in.
Source: Screenshot of the forum
In addition, the Shanghai market supervision department has also investigated a ChatGPT commercial confusion case. According to expert analysis, there are two main modes of "shell AI": one is to connect to the official platform and act as a "setter" for users in the form of a transfer station; The other uses a copycat platform entirely, relying on a name and logo similar to ChatGPT to disguise itself.
Zhang Xiaorong said that for some emerging large model companies, especially those that build products and services through simple shells or completely rely on the API interface provided by Open AI, there is a risk of business interruption or even outright delisting.
In Zhang's view, these Chinese companies and developers who rely on Open AI APIs for project development and innovation may need to change their technology base, but it will also increase R&D costs and time costs. The effect of the converted service may be discounted, because the domestic model may not be as good as Open AI in the field of foreign languages.
On the whole, the impact of "service suspension" is relatively limited, and domestic related start-ups will not collapse on a large scale, and there will be no sudden outbreak of "overtaking" in the short term.
03. Domestic large model "running on top"
According to "Qujie Business", at present, including Alibaba, Baidu, Tencent, Baichuan Intelligence, Zhipu AI, etc., a number of domestic head large model manufacturers have fed the "moving plan" to realize the transformation of the technical base to the user's mouth.
For example, Zhipu AI said that it will provide developers with a generous resource package, plus migration training from Open AI to its own large models. Baidu Intelligent Cloud takes "zero cost switching" as the highlight and launches the "Hometown Cloud" large-scale model inclusive plan; AI Infra vendors have even directly given top open-source large models such as Qwen2-7B, GLM-4-9B, and Yi-1.5-9B to free forever.
Source: Weibo screenshot
Han Sirui, an assistant professor at the Hong Kong University of Science and Technology and senior business development manager of the Hong Kong Generative Artificial Intelligence R&D Center, said that this incident will promote the increase in the market share of domestic large-scale model companies in the short term.
However, in the long run, whether the domestic large model can completely replace Open AI will ultimately depend on the model capability.
In this regard, the domestic large model has indeed made a lot of progress compared with last year.
In 2023, when Baidu released Wenxin Model 4.0, Robin Li said that its comprehensive level was no less than that of GPT-4.
In the early morning of June 27, Clem, co-founder and CEO of huggingface, a world-renowned open source platform, announced on social platforms that Alibaba's latest open-source Qwen2-72B instruction fine-tuned version has become the first place in the open source model rankings. This list includes more than 100 mainstream open source models around the world.
Specifically, in terms of model Chinese capabilities, many domestic models perform better than OpenAI.
"Daily Economic News" united more than 30 outstanding reporters, editors, and engineers, and spent 2 months evaluating the performance of mainstream large models in the market in financial news work scenarios, showing that Yi-Large, which was incubated by Kai-Fu Lee, ranked first in the total score of the four major application scenarios of "financial news headline creation", "Weibo news writing", "article error proofreading" and "financial data calculation and analysis", while ChatGPT 4.0 performed poorly, and even ranked first in "financial news headline creation" Ranked at the bottom of the scene.
Source: National Business Daily
In addition, in terms of price, domestic large models are also more affordable. In Alibaba's free migration plan, the API pricing of the qwen-plus large model is only 1/50 of OpenAI's.
However, from an objective point of view, Han Sirui also admitted that OpenAI's model capabilities are still relatively leading in advanced tasks such as complex document parsing and complex document generation in text models, and the demand for these complex model capabilities will not change greatly in the short term due to the discontinuation of API services.
The American model started earlier. In May 2023, the number of basic large models with more than 1 billion parameters in the United States has exceeded 100; In addition to the well-known ChatGPT, representative general model companies in the United States include Anthropic, Cohere, and Google.
Among them, Anthropic was once known as the "fierce rival of OpenAI". According to the list released by it, its newly released AI model Claude 3.5 Sonnet has better results in graduate-level reasoning, coding ability, text reasoning, etc. than GPT-4o, which was a smash hit not long ago.
Source: Weibo screenshot
Wang Peng, a researcher at the Beijing Academy of Social Sciences, believes that the differences between China and the United States are mainly reflected in three aspects: the level of financing, the level of development of the basic model and the level of development of the application layer.
According to dealroom.co data, as of July 10, 2023, global generative AI companies have raised $15 billion, of which 89% has gone to U.S. startups.
In terms of the development level of basic large models, there are still problems such as lack of total data, lack of computing resources, and limited scenario penetration rate of domestic large models. After all, in terms of the amount of public data, English data itself dominates, and the United States is still taking various ways to restrict China's access to core resources of computing power.
As for the application layer, China is also in a state of following; Among them, it lags behind the United States in the office, finance and medical fields.
This has made the outside world have some concerns about the domestic large model. Bloomberg News mentioned in a report that Liang Zhongwei, CEO of Singapore's Dorje AI, said that he was worried that Meta's Llama and other open source models could also cut off access for Chinese developers, which would bring greater uncertainty.
What cannot be ignored is that China has a large market scale and rich application scenarios, which provides a broad space and conditions for the landing and application of large models; And the more data and scenes, the more practical the large model can be. This gives China a chance to catch up with and surpass the United States even though it is slightly inferior to the United States in terms of underlying R&D technology.
However, in the process of catching up, more and more obstacles will be encountered, which may be a definite trend and a challenge for domestic manufacturers.
Author | Zhang Kaijing, editor | Fun Business Technology Group
This article was written by Everyone is a Product Manager Author【Fun Solution Business】, WeChat public account: [Fun Solution Business], original/authorized Published in Everyone is a product manager, without permission, it is forbidden to reprint.
Image from Unsplash, based on the CC0 license.