Deepseek, a little -known Chinese startup, has released a shock wave through a global technology sector with the release of artificial intelligence (AI) models comparable to the creation of Google and Openai.
The creator of Deepseek-R1 says that the model was developed using more advanced and smaller computer chips than adopted by US high-tech companies.
A research paper published last week stated that the model development team spent less than $ 6 million on computing power to train models. Each of Chatgpt and Gemini, respectively.
Mark and Leisen, one of the most influential high -tech venture capitalists in Silicon Valley, praised the release of the model as “the moment of AI’s SPUTNIK.”
The sudden appearance of a Chinese startup comparable to the Silicon Valley top player is a fear that the Sky High market evaluation of companies such as NVIDIA and META may be separated from reality. I raised it.
On Monday, NVIDIA is close to the production of semiconductors, which has lost nearly $ 600 million in market capitalization after 17 % of generated AI has plummeted.
U.S. President Donald Trump, President Donald Trump, who announced the start of $ 500 million in Texas and SoftBank, a Japanese SoftBank, should be a “morning call” on the need in the US industry. I mentioned. Focusing on the laser by competing to win. “
What is DeepSeek?
Deepseek, based in Hangzhou, was established in late 2023 by Liang Wenfeng, a consecutive entrepreneur who operates hedge fund high flyers.
Although it is rarely known outside China, Liang has a wide range of history that combines fast -growing technology and investment.
In 2013, according to China’s Media Outlet Sina Finance, he co -founded Hangzhou Jacobi Investment Management, an investment company that implemented a trading strategy along with the Zoe University’s joint almunas, along with the Zoe University’s joint almunas.
In 2015 and 2016, Liang continued to establish two companies focusing on computer -oriented investment.
In an interview with China’s Media Outlet Wave in 2023, LIANG has rejected a proposal that it should be too late for startups to participate in AI or that it should be expensive.
“Playback alone is relatively inexpensive. Public paper, open source code, minimal training, or fine -tuning, but according to his comments published by CHINATALK SUBSTACK, Liang: I have stated.
Liang said that his interest in AI was mainly driven by curiosity.
“I want to verify a specific hypothesis from a wider perspective. For example, the essence of human intelligence may be a language, and human thoughts may be essentially a language process. I will assume that, “he said.
“What you think of” thinking “may actually be your brain weaving language. This suggests that human -like AGIs may potentially appear from large language models, “he said, and he added and tries to mimic the cognitive ability of the human mind. I mentioned general information (AGI).
DeepSeek did not respond immediately to the comment request.
On Monday, Gregory Zuckman, a journalist of the Wall Street Journal, wrote a Chinese version of the book written about Lian, who has never heard of it before, about the late American hedge fund’s manager Jim Simmons. I said I knew. 。
“Simmons had clearly had a deep impact,” said Zuckerman and Lian wrote his book, “Unleash many unresolved mysteries before, and bring a wealth of experience to learn from us. I explained how I praised it as a book.
“Even my mother didn’t get much from the book,” Zuckman wrote.
Why did DeepSeek swept the high -tech world?
To put it simply, the company’s success has raised an existing question on the approach to AI, which is both Silicon Valley and the US government.
It is widely assumed that US high -tech companies have a great advantage in AI, especially for a huge size. End the chip.
Arriving at the Deepseek site has challenged the assumption that it costs billions of dollars to be at the forefront of AI.
“Openai was established 10 years ago, has 4,500 employees, and has raised $ 6.6 billion. Deepseek was established two years ago, with 200 employees, less than $ 10 million. It was developed in, “said Adam Kobeissi, the founder of the market analysis newsletter THE KOBEISSI Letter.
“How are these two companies now competing?”
DeepSeek engineers stated in a research paper that they used about 2,000 NVIDIA H800 chips that were more advanced than the most advanced chips to train models.
The team said that the slow chips used multiple specialized models in cooperation so that they could analyze data more efficiently.
For the United States government, DeepSeek arrived at the site questioned the strategy to contain China’s AI progress by restricting high -end chips exports.
Deepseek’s research papers suggest that the most advanced chips are not needed to create high -performance AI models, or that Chinese companies can still procure chips in a sufficient amount or both combinations. 。
NVIDIA’s H800 chip, based in California, designed to comply with US export management, is free to export to China until October 2023, when President Joe Biden’s management was added to the list of restricted items. It was done.
In an interview with the waves in 2023, Liang said he had stocked 10,000 NVIDIA A100 GPU before his company was banned for exports. The GPU or graphic processing unit is an electronic circuit used to speed up graphics and image processing on computing devices.
Tanishq Abraham, a former research director of Stability AI, stated that considering the development of various models such as Alibaba and Baichuan, it was not surprising to the Chinese advanced level in AI.
“There was a limit to China’s ability to acquire the GPU, but China was still innovating anything and was able to narrow down the performance,” said Abraham.
“We think that having many performances that they can still squeeze out are lessons for US companies.”
Tarajavidi, a co -director of the California University San Diego University, and a joint director of the security center, said that Deep Shek was excited about the “rapid progress” of AI development around the world.
“My only hope is that the attention given to this presentation promotes greater intellectual interest in the topic, further expanded the talent pool, and ended, but it is a private investment in AI research in the United States. Javidi says that both public investments are increased. ” Aljajira.
On the other hand, at least in the short term, investors’ trust in the US technical scene was hit.
Apart from the NVIDIA dramatic slides, Google’s parent Alphabet and Microsoft have fallen by Apple and Amazon, but the stock price has fallen by 4.03 % to 2.14 % on Monday.
“If the cost of DeepSeek is realistic, now almost all large organizations can build and host which large -scale organizations of which companies can build and host them. You can do it, “said a professor who specialized in Queensland University AI to Al Jazeera.
“So in this sense, the game has changed completely because of the new” rules “that anyone can play. “
Does this mean that China is over the AI race?
Not always.
The high-techerist agrees that Deepseek-R1 performs a performance at the same level as Chatgpt, or a specific task is more performing, but the field is moving rapidly.
Openai CEO’s Sam Altman said earlier this month that within a few weeks after examining user feedback, the company will release the latest AI model O3 mini.
On Monday, Altman acknowledged that Deepseek-R1 was “impressive” while focusing on larger computing power.
“We provide a much better model, and it is legal to have a new competitors! We will raise some release,” Altman said in X.
“But most of the time we are excited to continue running the research roadmap, and we believe that it is more important than ever to succeed in our mission.”
Rui Ma, the founder of Tech Buzz CHINA, said that when Openai first released Chatgpt in 2022, the Chinese technical department was surprised itself.
“Most entrepreneurs have completely missed the opportunity to represent the generated AI and felt very humble,” MA told Al Jazeera.
“It’s clear that they’ve been working so hard since then. What we have shown last weekend is how seriously they are reflected, and we have tried to catch up with Silicon Valley. Masu. I think Washington should focus on improving Silicon Valley instead of suppressing China in order for the United States to maintain the lead. “
ABRAHAM, a former research director of Stability AI, states that, unlike DeepSeek, the fact that Openai and other companies do not freely use the most advanced models in the public can be distorted. I did it.
“Deepseek has made the best model for free. On the other hand, Openai’s best model is not free,” he said.
“Therefore, most people who use Chatgpt for free believe that if Openai has been shocked by DeepSeek and Openai has already paid the same performance model for several months, the ability jumps. This wage wall of the AI model leads to people who do not really know the progress and ability of AI.
Miller, a professor at Queensland University, stated that DeepSeek’s progress and recent developments suggest that China had at least AI’s United States and there.
“At the end of last year, the next scientific breakthrough of AI made a little prediction that there was a possibility of coming from small players, such as individual university researchers who could not access many computing power.” He said.
“Deepseek’s clear progress is almost an example. Even if you did not have enough calculation ability to build a model that is as large as Chatgpt, the need is the mother of the invention.”