After causing Shock waves with artificial intelligence model Through the capabilities competing with the creations of Google and Openai, Deepseek from China faces questions about whether its bold claims stand to check.

The Hangzhou residing announcement is that R1 has developed a cost of the cost of the latest Silicon Valley models immediately, and immediately called for skepticism about the US domination of artificial intelligence and higher market assessments in the most important technology companies.

However, some skeptics have challenged Dembeic’s novel to work on an expected budget, indicating that the company is likely to have access to more advanced and more funding chips than it recognized.

“It is a largely open question whether Deepseek can be taken with the nominal value. Pedro Domingus, an honorary professor in computer science and its engineering at Washington University, said that the artificial intelligence community will dig and discover that.

“It is reasonable for me to be able to train a model of $ 6 million,” Domingus added.

“But it is also possible that this is just the cost of refining the post -processing models that cost it more, and Deepseek could not do it without building on more expensive models by others.”

In a research paper released last week, Deepseek Development said they used 2000 Nvidia H800 GPU – a less advanced chip originally designed to comply with US export controls – and spent $ 5.6 million to train the foundation of R1, V3.

The Openai CEO, Sam Altman, stated that it costs more than $ 100 million to train Chatbot GPT-4, while analysts estimated that the model used up to 25,000 advanced H100 graphics processing units.

The Dibsic Declaration, which was established in late 2023 by the serial entrepreneur Liang Winfing, has widely raised the belief that companies that seek to be at the forefront of artificial intelligence need to invest billions of dollars in data centers and large quantities of expensive high chips.

It also raised questions about the effectiveness of Washington’s efforts to restrict the artificial intelligence sector in China by banning the most advanced chips exports.

California-based NVIDIA shares, which enjoy the presence of approximately friction in the processing of drawing graphics processing units generated from artificial intelligence, on Monday, decreased by 17 percent, which led to a survey of approximately $ 593 billion in the market value of the chips giant-which is No. Compared with GDP (GDP (GDP) (GDP) from Sweden.

While there is a wide consensus that the release of the Deepseek for the R1 is at least a great achievement, some prominent observers have warned against taking their nominal value.

Palmer Lucky, founder of Oculus VR, on Wednesday, was described as Deepseek as “fake” and accused many “useful idiots” of going out for “Chinese propaganda”.

Loki said in a post on X.

“America is a fertile bed for Psyops like this because our media device hates our technology companies and wants President Trump to fail.”

In an interview with CNBC last week, Alexandr Wang, CEO of Scale Ai, threw doubts at the Deepseek account, saying that he “understood” can reach 50,000 advanced H100 chips that cannot be talked about due to US export controls.

Wang did not provide evidence of his claim.

Technical billionaire, Elon Musk, supported one of the closest close associates of US President Donald Trump, skeptical in Debsik, and wrote “clearly” on X under a job about Wang’s claim.

Deepseek did not respond to the suspension requests.

But Zehan Wang, a doctorate candidate who worked on a model earlier Dembsic, returned to critics of startups, saying: “The hadith is cheap.”

“It is easy to criticize.”

“If they spend more time working on the code and reproducing the Deepseek idea themselves, that will be better than speaking on the paper,” Wang added, using an English translation of Chinese people about people who participate in talking.

He did not respond directly to a question about whether Dibsic believed that he had spent less than 6 million dollars and used less advanced chips to train the basic model of R1.

In an interview in 2023 with Chinese media waves, Liang said his company had stored 10,000 A100 chips in NVIDIA-older than H800-before the administration of the President of the United States then prohibited its export.

R1 users also refer to the restrictions they face because of their origins in China, which are controlling topics that Beijing considers sensitive, including the 1989 massacre in Tiananmen Square and Taiwan.

In a sign that the initial panic about the potential impact of the American technology sector has begun to decline, the NVIDIA share price on Tuesday recovered about 9 percent.

NASDAQ 100 in technology increased by 1.59 percent after a decrease of more than 3 percent the day before.

Tim Miller, a professor of artificial intelligence at the University of Queensland, said it was difficult to determine the amount of inventory that should be placed in Dibsic’s claims.

“The same model gives some details about how it works, but the costs of the main changes they claim – I understand – do not appear in the same model very much,” Miller told Al -Jazeera.

Miller said he had not seen any “alarm bells”, but there are reasonable arguments both with trust in the research paper.

“The penetration is incredible – almost” very good to be real. “Miller said:” The collapse of the costs is not clear. “

On the other hand, he said that the breakthroughs occur from time to time in computer science.

“These huge models are a very modern phenomenon, so efficiency must be found,” Miller said.

“Given that they know this will be reasonably clear to others to reproduce it, they knew that they would look stupid if they are everyone. There is a team that is already committed to trying to reproduce the work.”

Low costs

Lucas Hansen, co-founder of the non-profit CIVAI, said, while it was difficult to know if Deepseek has circumvented the US export controls, the training budget backed by the aforementioned Startup, which is almost equivalent to GPT-4 from Openai, not R1 itself.

“GPT-4 has finished training in late 2022. There have been many algorithms and devices since 2022, which prompted the cost of training the GPT-4 category. For training, but you can now train it for $ 20 in 90 minutes. ”

“Making Deepseek R1 by taking a basic model – in this case, V3 – and applying some smart methods to teach this basic model to think more carefully,” Hansen added.

“This teaching process is relatively cheap compared to the basic model training price. Now that Deepseek has posted details about how to get a basic model in the thinking model, we will see a large number of new thinking models.”