"This transformation also means that OpenAI has abandoned the principle of 'platform neutrality,' which is very, very, very important."
Text / Ba Jiuling (WeChat public account: Wu Xiaobo channel)
"The entire northwest of Shanxi has become a pot of porridge."
This classic line in the film and television drama "Bright Sword" is extremely appropriate in the AI world in the past two days.
The source of all this is a pop-up message from OpenAI in the early morning of June 25 (the picture below is a translated version).
OpenAI has been trying to make GPT more warm in conversations, but in this pop-up window, it "indifferently" apologized and "regrettedly" informed the other party that as of July 9, all countries and regions that are not included in the list of supported countries will not be able to use API transmission.
It is reported that OpenAI supports dozens of countries to access its services, but Chinese mainland, Hong Kong, Russia, North Korea, Syria, Iran and other places are not listed. According to the company's guidelines, developers from the aforementioned regions may have their products accessed to their accounts blocked or suspended.
Friends who are familiar with foreign trade may find that OpenAI's support list is the same as the list of "friendly countries" that the U.S. government has been emphasizing.
Bang, late this night, everyone heard the heavy closing of the door.
Just like everyone's first reaction to the news, Hu Yanping, chief expert of FutureLabs, said: "This is obviously a geopolitical topic. ”
Surprisingly, this time it will be OpenAI standing in front of the stage, and Wikipedia still has a vision for its founding: "OpenAI's purpose in conducting AI research is to promote and develop friendly artificial intelligence for the benefit of humanity as a whole." ”
Doesn't this group of human beings include those on the "list of unfriendly countries"?
OpenAI for "genetic mutations".
Although the media did not mention the reasons for OpenAI's sudden decision, the answer is not difficult to guess.
Just a week before the "closure", on June 13, another inconspicuous OpenAI news inevitably aroused people's reverie:
OpenAI announced on its official website that it has officially appointed Paul M. Nakasone, a retired U.S. Army general, former commander of Cyber Command, and director of the National Security Agency (NSA), to its board of directors.
The official statement is that Paul Nakasone will help OpenAI "strengthen AI's ability to respond to increasingly sophisticated cybersecurity threats."
According to the official website, Paul Nakasone played a key role in the establishment of the U.S. Cyber Command during his career as an army officer. He is the longest-serving leader of U.S. Cyber Command and previously led the National Security Agency, which was responsible for protecting the nation's digital infrastructure and advancing the country's cyber defenses. He has held command and staff positions at all levels of the U.S. Army and has been deployed to elite cyber units in the United States, South Korea, Iraq, and Afghanistan.
Meanwhile, on June 22, the U.S. released a draft rule that would require regulation of certain U.S. investments in semiconductors and microelectronics, quantum computing, and artificial intelligence, and prohibit or require restrictions on Chinese investments in AI and other technologies.
Therefore, a member with a heavyweight military background enters the core of the company + a highly targeted bill that has just been released, everything goes without saying.
According to Zhang Xiaorong, president of the Deepin Technology Research Institute, the joining of Paul Nakasone shows that the role and goals of OpenAI have changed. At present, OpenAI will not only pay more attention to commercial interests, but also pay more attention to national security considerations, and this company is no longer just pure scientific research.
Compared to the original non-profit organization, OpenAI is likely to be more focused on the commercial application of its technology, security, and contribution to national strategies.
Zhang defined its new role with a "harsh word": a U.S. national intelligence agency dressed in the garb of a commercial company.
In other words, OpenAI and the big model in its hands are no longer just a commercial tool, but also a strategic resource that the government wants to master.
Hu Yanping believes that he is not surprised by this sudden "supply cut". As GPT grows from 3.5 to 4 to 4o, and then to 5, every leap in its capabilities means that OpenAI will be more concerned about whether its abilities will be "abused" by users it does not want to, so it is a matter of time before the limit is made.
In fact, he said, since mid-June, OpenAI has "stopped pretending", that is, it no longer pretends to be a non-commercial organization, but has decided to adjust its system and structure for profit.
According to The information, OpenAI CEO Sam Altman and shareholders have proposed that they are considering changing OpenAI's governance structure to a for-profit business that is no longer controlled by the board of directors.
It has been speculated that the for-profit company may have been set up to pave the way for OpenAI's initial public offering (IPO).
There is still a debate within OpenAI about whether to transform or not, but the move of "closing the door to China" itself may be an example of OpenAI's large-scale genetic modification with a knife edge inward.
"What's more, this transformation also means that it abandons the principle of 'platform neutrality,' which is very, very, very important," Hu lamented. ”
How to capsize the boat of friendship?
The character design will collapse, and the boat of friendship will also be overturned, and many developers of domestic AI models will choke on cold water.
For OpenAI in the past, the process of opening the API was a bit like an ordinary person going to a restaurant, the customer told the waiter which dishes to order, and the waiter would tell the back kitchen, and then the back kitchen would hand over the dishes to the waiter after a series of preparations for ingredients, cutting and preparing vegetables, frying and cooking, and then the waiter would give the dishes to the customer.
In this metaphor, the relevant developer is the customer, the "dish" is the data or service that the developer wants, the "waiter" is the API, and the kitchen is OpenAI's large model ChatGPT.
Because this "back kitchen" level is comparable to Michelin, developers around the world are flocking to it. The dishes ordered by the developers, i.e., GPT's large model capabilities, are used to feed the AI applications they develop. Especially for some Chinese developers who started from zero, and for those who do not have deep pockets, they also tend to use OpenAI's tools to train AI systems and applications.
When exaggerating, some people directly change the packaging of the dishes made by OpenAI and transfer them to their own "restaurants" to sell, that is, the so-called "self-developed large model", which is actually a shell large model company.
According to Zhang Xiaorong, president of the Deepin Science and Technology Research Institute, the first to suffer from this "supply cut" are mainly those AI application developers who rely on OpenAI technology and companies that use GPT models to develop large models.
Affected points include:
▶ ▷ Technology dependency: These companies may be technically highly dependent on OpenAI's APIs, and their products or services may not function properly once the service is terminated.
▶ ▷ Market adjustment: They may need to adjust the technology base and find alternative technologies or partners in China.
▶ ▷ R&D investment: Accessing new large models requires increasing R&D investment, which will also enable some companies to accelerate the development of their own technical solutions.
In Hu Yanping's view, the direct shell loses the most, followed by relying on OpenAI's API to be some application developers with different characteristics, such as knowledge management, dialogue, robots, customer service, etc., the deeper the ability of this part of the population is bound to OpenAI, the greater the loss and the more difficult it is to turn around.
Rescue the overboard
Purely from the perspective of practical application, Zhang Xiaorong does not agree with the so-called expression of "supply interruption". In his opinion, OpenAI has not announced the opening of the Chinese region since the launch of ChatGPT, so it has not generated a "supply".
From a technical point of view, according to The Paper, some developers will use OpenAI's API interface in China through proxy servers or deploy reverse proxy mechanisms overseas, which is not "compliant" and has always been disconnected.
There are also some transit service providers that previously provided APIs to domestic developers, and after the news was exposed, in order to stabilize the morale of the military, they directly shouted "Our source server was originally set up in an area supported by OpenAI, and it is through the relay request to achieve the effect of using ChatGPT in China, everyone can rest assured, it can be used as usual now." ”
It is not known whether it can be used as usual, because such an operation may involve the issue of data export, and its survival time is still unpredictable.
Some people close the door, some people fall into the water, but for some domestic large model manufacturers, it is a great benefit. According to media reports, as soon as the news of OpenAI came out, many large model manufacturers began their own publicity plans, trying to win the market that was ceded by the former.
For example, Alibaba Cloud, Baidu Intelligent Cloud, Zhipu AI, etc. have all said that they can provide exclusive migration services for OpenAI API users. At the same time, there are generous preferential policies, for example, for users migrating from OpenAI, Baidu promises to provide free fine-tuning of AI models and expert guidance on the flagship Ernie model, as well as 50 million tokens that developers can use to query bots.
Baichuan Intelligence, which is jointly supported by Alibaba and Tencent, provided 10 million free tokens. SenseTime provided 50 million tokens. Zhipu provided 150 million tokens and a series of training courses to ease the transition. Even Microsoft, the biggest proponent of OpenAI, posted a step-by-step guide on WeChat on how to migrate to a local service run by a local partner.
*In large models, "token" usually refers to the smallest unit in text, which can be a word, a character, or other substring. For English text, tokens are usually delimited by words separated by spaces.
After the impact
However, while the technological impact is the most immediate, the concern is more long-term. Zhang Xiaorong said that the incident may cause a certain level of worry and unease in China's AI industry and public opinion, and they worry that there may be more technical blockades and restrictions in the future.
For such a change, Hu Yanping is a little "happy to see it".
He believes that this wave of psychological and technological shocks will help stimulate everyone's thinking, that is, "How can we develop AI?" "How can we make China's AI more competitive?"
In the past two years, there has been an unwritten consensus in the mainstream of the industry, that is, many people emphasize that China's AI development should pay attention to the application level, which follows a convention in the Internet era - developed in the United States and applied in China.
Dominated by this mentality, domestic AI development is somewhat "partial and comprehensive", does not pay attention to research and development, and is more committed to running a killer application.
On this basis, relying on OpenAI to "free ride" has become the norm, and the biggest losses in this incident happen to be this kind of group.
OpenAI's "act of cutting off supply" is choking and sobering at the same time. It reminds developers to return to the development of core capabilities and to the underlying research and development, which is a long-term process, but Hu Yanping said that in the field of artificial intelligence, long-termists focus on intelligence, and short-termists focus on energy.
In fact, many observers of the AI industry do not agree overwhelmingly that this move will widen the so-called technology gap, and from experience, the ultimate spur effect of such events is far greater than the suppression effect.
"For China, this event could help bring many of the small startups created during the 'Battle of the Hundred Models' to the fore after ChatGPT debuted in late 2022."
After all, if we keep staring at the gap under our feet, we will never have a chance to look into the distance.
The author of this article | and the wind and the moon half丨Mei Haoyu | Editor-in-Charge | Xu Tao
Editor-in-Chief | He Mengfei | Image source | VCG