According to a Reuters report, the Chinese emerging artificial intelligence (AI) company DeepSeek stated in a peer-reviewed paper published on 17 in the scientific journal "Nature" that it spent $294,000 to train its large language model "R1". This is significantly lower than the training costs of American models, and the debate over whether China is gaining an edge in the AI development race is likely to reignite.
Since Deep Seek stunned the world with its low-cost generative AI model in January, founder Liang Wenfeng has rarely appeared publicly.
However, in this paper, Mr. Liang is listed as one of the co-authors, and it was written that, in addition to research expenses, they also used 512 of NVIDIA's "H800" from the US semiconductor giant. The previous paper in January did not include this information.
Meanwhile, OpenAI CEO Sam Altman said in 2023 that the so-called "foundation model training" cost "far more than" $1 billion.
However, DeepSeek's explanation of the costs and technologies used in AI development has raised doubts among U.S. companies and government officials.
The H800 mentioned by DeepSeek this time is a product designed by the company for the Chinese market after the U.S. government effectively banned the export of NVIDIA's high-performance GPUs "H100" and "A100" to China in October 2022.
However, several U.S. government officials told Reuters in June that Deep Seek was using a large number of H100s purchased after the start of the export controls on China.
In a supplement to this paper, DeepSeek for the first time admitted to owning A100s and stated that they had already been used during the development preparation phase.
Original article: https://www.toutiao.com/article/1843647934807040/
Statement: The article represents the views of the author. Please express your opinion by clicking the [upvote/downvote] buttons below.