The candidate for the US Secretary of Commerce Howard Lutnick accused Deepseek, an artificial intelligence company (IA) based in Hangzhou, to steal American technology and bypass American export controls to obtain high -end Nvidia fleas.
During the hearing on the appointment with the US Senate on January 29, Lutnick said Deepseek could create his “cheap” AI models because he was able to buy a large quantity of Nvidia chips and steal data from the Meta open platform.
“I take a very yellow vision of China,” he said. “They only think of themselves and seek to hurt us, so we have to protect ourselves. We have to conduct our innovation and we have to stop helping them. Meta’s open platform lets Deepseek count. The tokens of Nvidia – of which they bought tons, and they found their ways (export controls) – lead their Deepseek model. It must end.
Lutnick said that it would coordinate and empower controls and export prices of the Bureau of Industry and Security (BIS) to prevent China from using American tools to compete with the United States.
Beijing has so far not responded to Lutnick’s comments when he celebrates Chinese New Year from January 28 to February 4.
David Sacks, former chief of the Paypal exploitation and advisor to the White House on AI and cryptocurrency issues, said Fox News that there was “substantial evidence” that Deepseek used a method called “distillation” to extract data from the OPENAI models of Microsoft to its use.
IT Experts said “Distillation” or “distillation of knowledge” is commonly used in AI formation. It is a technique where the outputs of a larger AI model are used to train and improve a smaller one.
Deepseek, in this process, can be understood as a student who continues to ask questions to a competent teacher, for example Chatgpt, and uses the answers to refine his logic. At one point, Deepseek will be as smart as Chatgpt.
The “distillation” process requires much less computing power than Optai used to form the Chatppt.
OPENAI said The financial times where she had seen evidence suggesting that Deepseek could have exploited her data by “distillation”. He criticized Deepseek for violating his intellectual property.
Some Chinese IT experts agree that Deepseek has been created by “distillation”.
Wang Zhiyuan, a computer columnist based in Beijing, written in a article That it is obvious that Deepseek V-3, published on December 26, 2024, had used the technique of “distillation” in training. He said he arrived at this conclusion after analyzing the characteristics of Deepseek.
He says that many other Chinese AI models have also used distilled Chatgpt O1 data, published on September 12 of last year. He says that a newspaper Published by a group of Chinese researchers on November 25, 2024, has already explained the distillation process and its effectiveness in detail.
He says that an AI model made with distilled data may not be able to answer very difficult questions, but is sufficient to solve secondary level problems. In his opinion, all small models of AI should improve with distilled data before entering the markets.
“Don’t laugh with those who have taken a shortcut!” Said Wang. “Deepseek used a special method to save IT power. After all, its training cost represents only $ 5.58 million, or 1.1% of US $ 500 million from Meta’s Llama 3.1. »»
After publishing the Deepseek-R1 on January 20, 2025, a group of deep researchers published a paper On January 22, saying that its latest AI model reached comparable performance to Chatgpt-O1.
They said that the Deepseek-R1 training used the distilled data from Tongyi Qianwen d’Alibaba (Qwen) and Meta lama. They said that the Deepseek-R1-Distill-Qwen version surpasses Chatgpt-4o.
50,000 flea H100?
Deepseek said that he had used only 2,000 NVIDIA H800 chip units to form its AI model. His parent High Flyer, a Chinese hedge fund, said that he had accumulated a group of 10,000 A100 chips before the United States did not prohibit flea exports to China in October 2022.
But now Lutnick suspects that Deepseek has bypassed American export controls by important Nvidia fleas high -end via third counties, like Singapore.
The Wall Street Journal reported Last July, certain questionable institutions of Singapore bought NVIDIA A100 fleas and paid Chinese students to bring them back to China.
Without providing any evidence, Alexandr Wang, director general of the United States-based scale, said CNBC that Deepseek has 50,000 H100 chip units, the most advanced Nvidia chips on the market.
Xiang Zhinging, a computer writer based in Hubei, finds it plausible. “It is not surprising that Deepseek has 50,000 H100 chips. Any Chinese Internet giant could have accumulated many Nvidia tokens ”, Xiang said.
Even if Deepseek has many chips, he says, he will use software and new IT executives to win the game, instead of following the traces of American companies to increase endless calculation power.
The United States has prohibited exports of A100 and H100 flea markets to China in October 2022, then the slower A800 and H800 fleas in October 2023. After that, Nvidia made the equal flower H20 Chips for Chinese markets.
On January 14, 2025, the Biden administration announced a regulatory framework to restrict exports of American chips and friendly models. The framework entered into force on January 31.
Some observers have said that the step -by -step strengthening of the flea ban has given China too much time to accumulate high -end AI fleas.
Yong Jian contributes to Asia Times. He is a Chinese journalist specializing in Chinese technology, economics and politics.