On June 2, 2023, the showroom of the Nvidia Corp. office in Taipei, Taiwan.
i-hwa cheng | Bloomberg | Getty Images
The U.S. government passed Tightening control of chip outletsincluding restrictions on the highest-level products Nvidiadid not prevent DeepSeek from effectively creating its generated AI applications, and At the best level in the United States Provided from companies such as Openai.
Although details about how deep it is Stay incompleteits success does not mean that export control has no seat in markets and national security policies, it does show that focus on competition cannot be kept in sync with innovation. Now, debate is underway on how far the U.S. government should go to prevent access to U.S. chip technology in the future.
President Biden issued rules for the Department of CommerceRegulate global spread“In the days of government decline, AI chips and models. Rules already exist Severely criticized By technology companies including NVIDIA and policy experts. Brookings analysts believe that the AI Diffusion Rules attempt to create a “central planned global computing economy.”
John Villasenor, a senior nonresident researcher, wrote: “A decade from now on, we will review and recognize that the U.S. government attempted to restrict people from 150 countries to perform fast multiplication in the mid-2020s. How strange is the ability. “Professor of Electrical Engineering, Law, Public Policy, and Administration at Brookings and UCLA.
In any technological war, the question of what counter-effects the United States should take is inevitably to the concept of controlling innovation through such measures as restricting exports inevitably can work, or even work – and may even backfire. Among the risks cited by Brookings: stimulates the development of global AI ecosystems outside the United States; promotes more countries to establish stronger technical links with China; and allows non-U.S. chip manufacturers to sacrifice original innovation The U.S. company behind it increases global market share at the cost.
“I’m worried that we’ll have a significant control over the knee response of ratchet before we completely consider the tradeoffs,” said Martin Chorzempa, a senior researcher at the Peterson Institute for International Economics. ”
Unless Trump revokes or revises the rules before this, there is a 120-day comment period that ends the May 15 AI proliferation rules. The president talked about the need to protect U.S. technology leadership in general, but he did not specifically address the rule. What position will the current government take is unknown – expand, cut or overturn the chip export rules already enacted.
Trump signed an executive order on his first day to “identify and eliminate loopholes in existing export controls”, which shows he can take a difficult route. It said the government will “evaluate and propose suggestions on how to maintain, acquire and enhance our country’s technological advantages and how to identify and eliminate vulnerabilities in existing export controls, especially those capable of transferring strategic goods, software, services and transfers. Strategic competitors and their agents’ national technology.”
The tech sector soon promoted the new government and served as several major CEOs at the inauguration, NVIDIA CEO Jansen Huang meets President Trump Discussions have been held in the White House in recent weeks, including restrictions on chips on China.
Trump is also Called DeepSeek “A wake-up call for our industry, we need to win with laser competition.”
In AI diffusion rules, what is particularly relevant to DeepSeek is the surrounding controls Turn off the weight of AI modelis crucial to developing the training process of how AI systems think and respond to queries.
“To some extent, DeepSeek was able to allow speed limits imposed on chips sold to China in 2022, but was banned in 2023, when the United States realized that the limits imposed were wrong,” Chorzempa said.
Chorzempa explained that when the United States takes control of China in 2022, they set a specific parameter for the speed of exchange between the chips. It is believed that if you control the power of personal chips that may not be enough, because if you fuse a powerful enough chip, there is a chance that the U.S. government does not want China to acquire supercomputer-like capabilities. From what to see DeepSeek describes it in R1 paper The company was able to overcome this speed limit.
“At least in early 2023, experts in the technology community pointed out that as technology develops, other restrictions must be effectively controlled,” Chorzempa said.
In 2023, the U.S. government added an additional layer of restrictions, allowing NVIDIA chip DeepSeek to say it trains models that are no longer legally exported. The Trump administration’s subsequent initiatives could further strengthen control measures.
But by combining the use of a limited number of advanced chips and innovations with this limitation, DeepSeek is able to build better ones, And maybe cheapmousetrap.
“DeepSeeks seems to have been heavily optimized with clever software and hardware engineering to make the speed limits designed to block these chips,” Chorzempa said.
AI competitors will continue to do more with less
There are other aspects of the growing AI race that show a narrowing gap for other reasons.
“This story is really about the gap between open source and closed source models,” said Alexandra Mousavizadeh, CEO of AI consulting firm Vevident. “Now, open source models are getting closer to closed functions, and we see Price fell to zero.”
DeepSeek has shown that you don’t need the maximum computing power you can Use open source as an alternative When building viable LLMs, in fact, these factors may be the driving force behind innovation, according to Mousavizadeh.
“We see restrictions force them to use scientific methods and systems that compress data onto a smaller pool that uses less power Hybrid expert model,” she said. AI competitors can do more things “less than doing more,” Mousavizadeh added.
This continued control may also further boost China Develop your own chip.
Mousavizadeh said: “You can’t really be a real gatekeeper.
If DeepSeek’s success leads to Advanced chip export control The intention to slow down China’s AI efforts, which have become more stringent, should also be clear that they are not silver bullets. “They are not a way to avoid competition between the United States and China,” a blog post last week wrote in a blog post. “Finally, if we want to gain the upper hand, AI companies in the United States and other democracies must have more than China.” Good model. But we don’t have to grant the technological advantages of the Communist Party of China without needing it.”
His problem is not about AI researchers in China, but the government they end up being watched. “In the interviews they did, they seemed to just want to make useful technology smart, curious researchers,” Amodei wrote. “But they are dictatorial governments that commit human rights violations, act actively on the world stage,” Amodei wrote. It seems that if they can match the United States in AI, they will be more affected in these actions.”
To be sure, there is There are many reasons to stay alert Do anything to contribute to the progress and success of China’s AI, such as DeepSeek, concerns from national security Share data with the Chinese government,continued Hacker Riskfor Chinese artificial intelligence applications become popular enough to be used by Chinese intelligence to understand Americans and the American industry and sow seeds in the public.
Alex Karp, CEO of Palantir Technologies, told Sara Eisen of CNBC In a recent interview, “We have to work harder, run faster, and do our national efforts.”
“The second is getting faster, especially if we have completed innovations,” Kalp said, describing DeepSeek as the derivative of the American model “improving margins.”
He expects a “huge policy discussion” to ensure there is no export innovation, but Karp adds that ultimately, as long as the first mover works hard, the real advantage can be attributed to the first move. , we have to focus on making sure our opponents will copy whatever they can do. ”
Palantir’s rise – a 340% stake last year leading the S&P 500 – is not trying to stop others, and that could have a lesson. “We’re not focused on the game,” Karp said. “We’re focused on how to execute.”
