Does the little -known Chinese startup have caused “artificial moments” for AI? : NPR

On Monday, January 27, 2025, Deepseek artificial intelligence logo on mobile phones arranged on Latvia.

Andrey Rudakov (Andrey Rudakov)/Bloomberg via Getty image

Hidden subtitles

Switch title

Andrey Rudakov (Andrey Rudakov)/Bloomberg via Getty image

Does AI just have a “explosion moment”?

That's it Some Investors, after the little -known Chinese startup Deepseek released a chatbot, experts said that although there are less money and computing power, experts said that although leaders of industries such as OpenAi and Google (such as OpenAi and Google and Google and Google ) Maintaining their own influence on their own industry leaders.

The wave of attention was pierced around Deepseek. It wiped nearly $ 60 billion from the market value of Chipmaker Nvidia.

Not iterative or evolution, but breakthrough

Antonia HMAIDI, a senior analyst at the Mercator China Institute of Berlin, said: “I think this does show the United States' life in the bubble to some extent.”

“Companies like Openai and OpenAi do bet on extensions, and need to buy more and more chips to improve chips.”

She said Deepseek showed different paths.

The company said it used more than 2,000 NVIDIA H800 GPUs to train the robot and trained at a price of $ 5.6 million for a few weeks. According to reports, others have deployed 10,000 or more GPUs and spent $ 100 million More Get similar results.

Marina Zhang, a scholar of the University of Sydney, said that Deepseek also showed a new innovation for China, not iterative or evolution, but breakthrough.

She said: “They do not really follow the existing models.” “It is basically based on algorithm optimization, and use software to destroy the limitation of not enough computing capabilities.”

Has the US chip export control failed?

These restrictions are applied by the United States to China. In 2022, the Biden government banned the export of cutting -edge micro -chip to China, thinking that they could be used to strengthen the Chinese army.

Zhang said that DeepSeek has shown that the chip blockade has not been successful so far. Beijing has doubled the self -reliance driver in the technical field to develop funds into chips and other areas including AI.

Others believe that it is too early to say that the failure of chip export control is too early.

Gregory Allen, director of Wadhwani AI Center Washington Strategy and International Research Center, said that Deepseek may buy all chips before feeling the effect of control measures.

DeepSeek founder Liang Wenfeng said in a widely reported interview in 2023 that the company has inventory of about 10,000 NVIDIA A100 GPUs, which are all in the US export control list. Experts believe that these models may have deployed early versions of these models.

After the beginning of the chip blockade, NVIDIA developed a solution to create a slightly powerful H800 GPU, which is a legal that it has been sold to China for a period of time.

Allen said: “We are currently living in the backward era of the Bayeng government's fire, in the first batch of AI export control.”

Deepseek has a window that can buy H800-before the government eventually prohibits it to China.

He said: “DeepSeek discovered some architectural innovation and some algorithm innovation. These innovations increased the number of IQ points, the amount of intelligence, and the given AI model can be obtained from a given number of computing resources.”

He said, but AI development requires computing power, and high -end GPUs that Deepseek or any other Chinese company can access are limited by export control. In the end, it will bite.

Allen said that this means that the United States has advantages: accessing advanced chips unlimitedly.

He said: “We can copy the advantages of China. They cannot copy our advantages. At least it will not be fast.”

As far as the cheap Deepseek has developed its almost cut marginal mode around DeepSeek, Allen said that there is no doubt that the cost is said to be far more than the $ 5.6 million according to reports. He compare it to the development of drugs.

He said: “The cost of developing new drugs is not only the cost of clinical trials.” “This is the cost of all clinical trials. As the AI ​​model training runs. DEEPSEK published the cost of final successful training. “

He said that it is not yet known how much the company has spent.

HMAIDI said Deepseek is “a very legal victory in Chinese projects.” But she said that this is not a threat for many people.

She said: “I can't see how you get a better model in the current pipeline-there is no more calculation.”

“Personally, I think this is not a threat to the American AI strength.”