Produced | Author of Huxiu Technology Group | Edited by Wang Xin | Miao Zhengqing's headshot | Visual Zhang Zhongmou, founder of China's TSMC, said: "Globalization is dead." As trade barriers between different countries continue to deepen, this trend of anti-globalization The "co

produced | Huxiu Technology Group

author | Wang Xin

editor | Miao Zhengqing

head picture | Visual China

TSMC founder Zhang Zhongmou once said: "Globalization is dead."

As trade barriers between different countries are increasing, Intensifying, this "cold air" of anti-globalization has also been transmitted to the most cutting-edge technology industry - AI. Competition among major powers in the field of AI is becoming increasingly fierce.

Today, the historical moment that changes the rules and destiny of the global game has arrived. Trump returned to the White House with an "America First" policy.

As the initiator of the strong suppression of Chinese technology companies, Trump’s return has also made the AI ​​competition and cooperation relationship between China and the United States even more confusing.

Xiang Huxiu, the founder of a Chinese AI company, said: "Trump's coming to power may bring more uncertainty and make their R&D investment in the United States more cautious."

Trump once said that he has understood In order to compete with China, the "geniuses" in Silicon Valley need more resources to promote the development of AI. To this end, he has relaxed regulations on AI to cope with competition with China.

Some AI practitioners are also worried that Trump will revisit the national AI strategy in 2019 to achieve the goal of "America's first AI" and intensify the unfair competition between China and the United States in AI.

The United States is putting away the suspension bridge of globalization

Spring River Plumbing Duck Prophet, Caiyun Technology CEO Yuan Xingyuan, who went overseas to the United States five years ago, has long felt the undercurrent. With populism on the rise, American users will also start "qualification reviews" of Chinese companies.

Yuan Xingyuan told Huxiu such a ridiculous story - because the product is free and comes from China, it was mistaken by American users for a spy company.

Yuan Xingyuan launched the AI ​​continuation product dreamily in the US market and accumulated 4 million users worldwide. “U.S. users of dreamily who speak both Chinese and English discovered that we are a Chinese company and translated our official account’s tweets into Chinese and posted them on reddit, an American Tieba forum, which led some users to suspect that we were spyware that steals privacy. Users did not understand why dreamily No charge." What makes Yuan Xingyuan even more amused is that after he changed his product to a paid model, he still received a lot of "meme attacks" from Reddit users. They poured into the app store to give low scores and express their intention to charge for the product. dissatisfaction.

Under Biden’s term, the US government has gradually built up barriers to counter globalization and built small courtyards and high walls.

The US government announced on October 28 that new regulations restricting US companies and Americans from investing in China in the semiconductor, artificial intelligence (AI) and quantum fields will take effect from January 2025.

Lalu Smart CEO He Ming clearly feels this limitation. In 2017, as the winner of the Nvidia Award, he attended the GTC conference in San Jose, the center of Silicon Valley. Many American investors took the initiative to talk to him. At that time, the focus of American VC was still on the technology business itself, and nationality would not be a factor. Their test rate range.

Recently, he returned to Silicon Valley again, looking for investment, but failed.

In fact, since February 9, 2024, after five investment institutions including GGV Capital and Jinshajiang Venture Capital Fund were investigated by the U.S. Congress, American investors have become increasingly uneasy.

After investigation, He Ming discovered that even American companies, as long as they have Chinese shareholders, will be rejected by American investors due to risk aversion.

He gave Huxiu an example. There is an AI imaging product with millions of monthly active users. Because there are Chinese shareholders behind it, it became the only company in the a16z incubator that did not raise funds in that period.

In addition to investment restrictions, with the rise of China in the field of AI, the United States is increasingly anxious about AI, and is also trying to inhibit the development of its opponents through technological blockade.

Subsequently, a bipartisan group of U.S. lawmakers unveiled a bill that would make it easier for the Biden administration to impose export controls on artificial intelligence models to prevent foreign competitors from using U.S. AI technology.

The U.S. government is worried that rival countries may use these artificial intelligence models to launch aggressive cyber attacks or even create powerful biological weapons. Reuters quoted sources as saying that the new US restrictions will target China, Russia, North Korea and Iran.

If approved, the measure would remove regulatory barriers to the export of open source models under the International Emergency Economic Powers Act and give the U.S. Department of Commerce regulatory authority.

shadow reappears?

After Trump came to power, the insecurity of operating AI companies in the United States has been exacerbated.

Chinese investor Steve revealed to Huxiu that an American AI company he invested in has recently been in urgent need of "survival" to prevent risks, and is negotiating with him to turn the shares into personal holdings or buy out.

"I once provided office space and engineers for his project, and I paid a lot." Quitting was not Steve's original intention. But the current situation makes them worry about whether Trump will continue to 'check for leaks and fill in the gaps' and block the paths that were once able to be avoided. He believes that "the worst case scenario is that even individual shareholders' proxy holdings will penetrate the review."

Physical AI companies are also worried that they will be affected even more. "In July 2018, after Trump announced tariff increases, we were forced to reduce costs again through channels and supply chain reforms. In recent years, some friends have also set up factories in Mexico to reduce tariffs." The CEO of a sweeping robot company recalled that in 2018 In 2016, I learned about the scene when Trump imposed a 25% import tariff on Chinese goods including robots, and such trade protectionism may escalate. ——In his speech last month, Trump advocated raising tariffs on imported goods from China to 60%.

The Sword of Damocles also hangs over the heads of Chinese local AI companies. As XAI plans to double from 100,000 cards to 200,000 cards, the threshold for computing power competition has been further raised. “This year, 100,000 cards has become the threshold for domestic and foreign computing power competition,” said a practitioner from a major AI factory. Huxiu calculated: "According to scaling law, the current number of gpt5 cluster cards is about 100,000." According to him, the specific number of cards may be less than 100,000, and it may be between 50,000 and 100,000. The level of parameters will be increased by about 3-5 times.

In this case, computing power may become a greater competitive variable. He believes that after Trump takes office, he will theoretically carry out "maximum pressure" to make it more difficult for domestic AI companies to obtain advanced computing power.

"If Trump introduces restrictive policies for cloud servers, then the TPU training costs we have deployed on Google Cloud may be in vain." 90% of overseas AI entrepreneurs who serve on Google Cloud expressed themselves to Huxiu worries.

But there are also people who hold the opposite view. In this regard, Yin Zhi, chief consultant of the Shanghai Artificial Intelligence Technology Association, believes: "Trump is still a businessman at heart. He may relax export restrictions on low-end GPUs such as A100 out of interests. After all, compared to GB200, the competitiveness gap of A100 Obviously. "

But one thing is certain, in order to maintain "the United States' advanced position in the field of AI", Trump opposes excessive regulation in the field of AI and minimizes obstacles to the development and application of AI technology in order to carry out comprehensive competition with China. .

Liu Yun, an assistant researcher at Tsinghua University Think Tank Center and Law School, told Huxiu: “Since Trump was first elected, he has proposed a ‘one in, two out’ market policy to reduce government intervention in AI supervision and adopt The policy tone of market autonomy."

There are rumors that Trump may repeal the "Executive Order on the Safe, Secure, and Credible Development and Use of Artificial Intelligence" signed by Biden on October 30 - the White House's policy on AI. First regulatory requirement.

In this regard, Liu Yun said that under the separation of powers, the US President’s executive order has nothing to do with legislation. Trump may not repeal the executive order, but issue a new executive order on AI on the basis of inheritance. .

According to the Wall Street Journal, the Trump team is formulating a new executive order - the "AI Manhattan Project", which aims to promote the rapid development of AI, which includes the part of "making the United States the number one in the field of AI", which means special Trump will promote AI policies that are beneficial to technology giants and relax regulations to deal with Chinese competition.Policies like

have also allowed Trump to gain more support from the American technology community compared to the 2016 election.

Breaking the rules

Qiu Chen, an overseas partner of China Film Group Corporation, who continues to pay attention to technology investment in the Chinese and American markets, believes that in the long run, there is no core difference between the red and blue parties’ AI policies, except for some of Trump’s targeted policies towards China. It may move faster, while Democrats are more inclined to ramp up gradually.

As Chris Miller, author of "Chip Wars," said: "Both Trump and Biden launched similar policies targeting the chip industry during their presidencies, and I expect the next president will do the same "

Trump is committed to promoting "decoupling" from China. Harris also said, "In key industries of the future, the United States cannot stand aside and hand over leadership to China. "

But the world is open, and such complacency cannot be tolerated. It often goes in the opposite direction. In fact, some restrictions in the United States cannot have a substantial impact on the domestic AI industry, but are more symbolic.

For example, although the Model Export Restriction Act mentioned above seems to be more lethal, in practice, it is difficult for the US government to implement export supervision of large open source models.

In the field of open source models, China's artificial intelligence models do rely on open source models developed in the United States to some extent. Many domestic companies further refine their own industry models through private data training based on open source models such as llama. A technical expert on large models of

told Huxiu: "llama and gpt are both basic large models, trained based on the transformer architecture, using different methods, including encode-decode structures and decode only like gpt. So start from scratch. In other words, it is necessary to re-develop based on an architecture like transformer and train it yourself. This is the case for large self-developed models by major manufacturers. There is also a category based on existing basic large models, including continue pretrain, finetune and other methods. Retraining or fine-tuning, based on open source architecture like llama, there are many teams working on it. "

In other words, training large models based on llama is the mainstream form of large model entrepreneurship in China.

However, at this stage, it is almost a false proposition to impose export restrictions on open source models like llama.

This is because: Currently, the llama 3 model is completely open source, which means that any user can obtain and use it through open source communities such as Hugging Face, so it is difficult to find effective measures to specifically ban the use of open source models in a certain country or region. In addition, after the "makeover" of retraining and fine-tuning, it is difficult to tell whether the model was developed based on the open source model architecture.

Trump's taking office may bring new challenges, but most Chinese AI companies, especially in the AI ​​to C track, still believe that the United States is a market where "the bigger the storm, the more expensive the fish", and its gold content is still rising. The AI ​​imaging company that did not receive financing is also doing well in the US market because it has access to PMF.

's AI expansion in the United States means, to some extent, that it will directly compete with the world's top large model companies such as openai. Qiu Chen's suggestion is the core value proposition - find your own leveraging point of differentiation, establish your own threshold step by step, and take advantage of your understanding of the supply chain and large-scale user operations. The best response is to have your own underlying technical route. In addition to pre-training, fine-tuning, rlhf, etc., return to the source of technology to make native innovations, and to some extent reduce the technical dependence on transformer and NVIDIA.

Yuan Xingyuan believes that no matter how precarious the external environment is, the way to overcome the situation is to maintain technical competitiveness amid uncertainty. Therefore, while improving the underlying architecture of Transformer, he also improved Dreamily's story creation capabilities to complete more complex tasks.

He finally discovered that in the face of absolute strength, "those influencing factors" were not worth mentioning. Five years after

went overseas, his realization is: "For products with an annual revenue of less than 100 million US dollars, the most important thing is to improve the competitiveness of the software itself. For example, I saw such users in the Apple store Comment - 'Although I think dreamily is a spyware, it is really useful.’”

tips: I am Wang Xin from the Huxiu Technology Medical Group. I focus on the fields of AI and venture capital. To communicate with industry professionals, please add WeChat: 13206438539. Please indicate your identity.

People who are changing and want to change the world are here