A frenzy on an artificial intelligence chatbot made by the Chinese technological startup Deepseek was upheaval of stock markets Monday and fuel debates on economic and geopolitical competition between the United States and China in development AI technology.
The AI assistant of Deepseek has become the free downloaded application No. 1 on the Apple iPhone store on Monday, propelled by curiosity for the chatgpt competitor. Part of what is worried about certain observers of the American technology industry is the idea that the Chinese startup has caught up with American companies at the forefront of the generator to a fraction of the cost.
This, if it is true, questions the enormous amounts of money that American technological companies provide for spending on data centers and the computer chips necessary to fuel AI progress.
But threshing media and false ideas on the technological progress of Deepseek also sowed confusion.
“The models they built are fantastic, but these are not miracles either,” said Bernstein analyst Stacy Rasgon, who follows the semiconductor industry and was one of the many analysts scholarship holders describing Wall Street’s reaction as exaggerated.
“They do not use any unknown or secret innovation or something like that,” said Rasgon. “These are things with which everyone is experimenting with.”
What is Deepseek?
The startup Deepseek was founded in 2023 in Hangzhou, China, and published its first large language model of AI later that year. His CEO Liang Wenfeng had previously co-founded one of the best hedge funds Chinese, High-Flyer, which focuses on the IA quantitative trade. The fund, by 2022, had accumulated a cluster of 10,000 of the high performance A100 graphical processor chips from NVIDIA based in California which are used to build and execute AI systems, according to a Post this summer On the Chinese social media platform WeChat. The United States Shortly after restricted sales of these tokens in China.
Deepseek said that his recent models had been built with the more efficient H800 chips from Nvidia, which are not prohibited in China, sending a message that the most fanciful material may not be necessary for the search for advanced AI.
Deepseek began to draw more attention to the AI industry last month when he published a new model of AI which, which was boastful, was tied with similar models of American companies such as Chatgpt Maker Openai, and was more profitable in using expensive Nvidia chips to form the system on data wicks. The chatbot became more widely accessible when it appeared in Apple and Google Applays stores at the start of this year.
But it was a follow -up research document published last week – the same day as the inauguration of President Donald Trump – who triggered the panic that followed. This article concerned another deep AI model called R1 which showed advanced skills of “reasoning” – such as the ability to rethink its approach to a mathematical problem – and was much cheaper than a similar model sold by Openai called O1.
“What their economy looks like, I have no idea,” said Rasgon. “But I think prices have scared people.”
The “Sputnik” backdrop
Behind the drama on the technical capabilities of Deepseek is a debate in the United States on the best way to compete with China on AI.
“Deepseek R1 is the spoutnik moment of Ai,” said the risk capital Marc Andreessen in a Sunday post on the social platform X, referring to the launch of 1957 satellite which sparked a spatial exploration race of the Cold War between the Soviet Union and the United States
Andreessen, who advised Trump on technological policy, warned that the over-regulation of the AI industry by the US government will hinder American companies and allow China to get ahead.
But attention to Deepseek also threatens to undermine a key strategy of American foreign policy in recent years to restrict the sale of AI semiconductors designed by the Americans to China. Some experts in American-Chinese relations do not think it is an accident.
“Technological innovation is real, but the moment of liberation is political in nature,” said Gregory Allen, director of the Wadhwani AI Center at the Center for Strategic and International Studies. Allen compared the announcement of Deepseek last week to the publication by the Chinese company sanctioned by the United States, Huawei, the publication of a new phone during diplomatic discussions on export controls of the Biden Administration in 2023.
“Trying to show that export controls are in vain or counterproductive is a really important goal of Chinese foreign policy at the moment,” said Allen.
Trump said on Monday that Deepseek’s breakthrough was “good because you don’t have to spend as much money.”
Speaking on Monday among the Republicans at home in Miami, Trump qualified the news as Deepseek “positive” if it is correct because “you will not spend as much and you will get the same result”. He described the development of “alarm clock for our industries that we must focus on the laser on the competition to win”.
Trump signed a prescription on his first day of power last week which declared that his administration “would identify and eliminate gaps in existing export controls”, indicating that it is likely to continue and harden the approach de Biden.
Deepseek’s progress on AI without the same number of expenses could potentially undermine $ 500 billion on Openai, Oracle and SoftBank that Trump was praised in the White House.
Nvidia’s actions fell 17% on Monday, but the company in a press release congratulated Deepseek’s work as “an excellent AI progression” which has taken advantage of “models and a largely available calculation which are fully in conformity export control ”.
What makes Deepseek different?
One thing that distinguishes Deepseek from competitors such as OpenAi is that its models are “open source” – which means that key components are free for anyone who can access and modify, although the company has not disclosed the data that She used for training.
But what attracted the most admiration in the R1 model of Deepseek is what Nvidia calls a “perfect example of time testing” – or when the AI models effectively show their train of Thought, then use it for more in-depth training without having to feed them new sources of data sources.
“It’s just thinking out loud, basically,” said Lennart Heim, a researcher at Rand Corp.
Openai reasoning models, starting with O1, do the same, and it is likely that other competitors based in the United States such as Anthropic and Google have similar capacities that have not been published, said Heim.
But “this is the first time that we have seen a Chinese business being as close in a relatively short period. I think that’s why a lot of people pay attention, ”said Heim. “I used to believe that Optai was the leader, the king of the hill, and that no one could catch up. It turns out that this is not completely the case. »»
