Entering ChinaIn just one month, the Chinese AI startup DeepSeek (DeepSeek) has released two large models, DeepSeek-V3 and DeepSeek-R1, which have low cost and performance comparable to OpenAI, which shocked Silicon Valley and even caused panic within Meta, and engineers began to try to replicate DeepSeek’s results overnight.
Scale AI founder Alexander Wang said in a Jan. 24 interview that DeepSeek was the best performer in their tests, on par with the best models in the United States.
Previously, Alexander Wang commented that DeepSeek-V3 is a bitter lesson that the Chinese tech community has brought to the United States. “While the U.S. is resting, China (the tech community) is working, catching up at a lower cost, faster and with greater strength.”
In addition, China’s AI “swipes” foreign major media outlets, which believe that the new progress of China’s large model has sounded the alarm bell for Silicon Valley.
At the time of the announcement of the $500 billion Stargate program, DeepSeek built a groundbreaking AI model at a fraction of the price and without the use of cutting-edge chips, raising the question of whether the huge investment of hundreds of billions of dollars in capital in the AI industry is really the most effective approach.
Meta went into panic mode in an attempt to replicate DeepSeek
On January 24, a post posted on the anonymous platform Teamblind went viral. A Meta employee said that Meta has now entered panic mode internally because of DeepSeek’s model.
The Meta employee wrote:
“It all started with the advent of DeepSeek-V3, which dwarfed Llama 4 in benchmarks. Even more embarrassing is the fact that a ‘Chinese company with only a $5.5 million training budget’ has done just that.
Engineers are racing against time to analyze DeepSeek, trying to replicate every possible technology in it. This is no exaggeration.
Management is worried about the huge investment in GenAI’s R&D department. When the salary of one executive in the department exceeds the cost of training the entire DeepSeek V3, and there are dozens of such executives, how do they account to the executive?
The advent of DeepSeek-R1 has made the situation even more dire. The details are confidential and will not be revealed, but they will be made public soon. ”
On December 27 last year, DeepSeek launched the open-source model DeepSeek-V3. At the time, Chatbot Arena showed that DeepSeek-V3 ranked seventh among all models and first among open-source models. Moreover, DeepSeek-V3 is the most cost-effective model in the top 10 in the world.
Less than a month later, on January 20 of this year, DeepSeek officially open-sourced the R1 inference model, allowing anyone to distill R1 to train other models under the MIT License.
On January 24, DeepSeek-R1 ranked third on the chatbot arena composite list, tied with the top inference model o1.
In highly technical areas such as difficult prompts, code, and mathematics, DeepSeek-R1 came out on top, ranking first.
In terms of style control, DeepSeek-R1 tied for first place with o1, meaning that the model did a great job of understanding and following user instructions and generating content in a specific style.
DeepSeek-R1 also tied for first place with o1 in the test of combining difficult prompts with style control, further demonstrating its strong ability to control complex tasks and fine-grained control.
Artificial-Analysis’ initial benchmark results for DeepSeek-R1 also show that DeepSeek-R1 achieved the second-highest score in the AI Analytics Quality Index, at about one-thirtieth the price of o1.
The AI boss exclaimed: China’s AI has caught up with the United States
After the release of DeepSeek-V3 in December last year, Alexander Wang, founder of AI data services company Scale AI, posted that DeepSeek-V3 was a bitter lesson for the United States from China’s tech community. “While the U.S. is resting, China (the tech community) is working, catching up at a lower cost, faster and with greater strength.”
Mark Anderson, the founder of the well-known investment firm A16z, posted on January 24 that Deepseek-R1 is one of the most amazing and impressive breakthroughs he has ever seen, and that it is also open source, and it is a gift to the world.
On January 24, Anjney Midha, partner at A16z and member of the board of directors of Mistral AI, said: “From Stanford to MIT, DeepSeek-R1 has become the model of choice for researchers at top U.S. universities almost overnight. ”
As for why China’s AI has made such rapid progress, Nobel laureate and “AI Godfather” Jeffrey Hinton said in an exclusive interview with blogger Curt Jaimungal on January 21 that China’s STEM (science, technology, engineering, mathematics) education is better than that of the United States, and it has more well-educated talents, which will provide a solid foundation for the development of AI. Although the United States is trying to slow down China’s development by restricting (such as Nvidia chips), this will only prompt China to accelerate the development of its own technology, “they may be a few years behind, but eventually they will catch up”.
Does DeepSeek or a game-changer
still work?
Researchers at Stanford University and Epoch AI published a study in the middle of last year showing that the largest models will cost more than $1 billion to train by 2027. Gartner predicts that hyperscalers like Google, Microsoft and AWS will spend up to $500 billion on AI servers alone by 2028.
But DeepSeek is completely different, and it’s not expensive to train. Noah’s Arc Capital Management said the DeepSeek-V3 model has the potential to revolutionize the game in the field of training and inference.
Especially after the announcement of the $500 billion “Stargate” program, DeepSeek makes people wonder if this kind of “miracle by force” method is really the most effective way.
U.S. stock market big V “THE SHORT BEAR” posted on X on January 24 that DeepSeek has brought a painful moment to AI giants, and investors must sound the alarm bells about it.
“If it only takes $55 million to beat OpenAI, the commercialization of this industry will be much faster than many people expected,” he said. ”
“According to Sequoia, U.S. AI companies must generate about $600 billion a year in revenue to pay for their AI hardware,” he noted. Now it seems that this risky behavior is becoming more and more unprofitable. ”
Holger Zschaepitz, a well-known financial journalist, said on January 25 that DeepSeek has built a breakthrough AI model at a very low price and without the use of cutting-edge chips, raising questions about the utility of the industry’s hundreds of billions of dollars in capital expenditure.
Some investors even believe that the stock price of U.S. chip stocks will also face challenges.
Investor Geiger Capital says Deepseek is as good, if not better, than OpenAI, and at a price of only 3% of the latter…… And American companies are investing hundreds of billions of dollars. So…… What will happen to the NASDAQ?
Notably, Nvidia shares fell 2% after the release of DeepSeek-V3. After DeepSeek-R1 sparked overseas discussions, Nvidia’s stock price fell by 3.12% on January 24.
Foreign media collectively swiped the screen: sounding the alarm for
Silicon Valley
If DeepSeek-V3 just made waves, DeepSeek-R1 is making a splash. In the past four days, foreign media have focused on DeepSeek, and unanimously agreed that the new progress of China’s large model has sounded the alarm for Silicon Valley.
On January 22, American media Business Insider reported that the DeepSeek-R1 model adheres to the spirit of openness and is completely open source, which has caused problems for American AI players. Open-sourced, advanced AI could challenge companies that are trying to make huge profits by selling their technology.
On January 24, the US media outlet CNBC launched a 40-minute program, inviting Perplexity CEO Aravind Srinivas to analyze why DeepSeek raises concerns about whether the US global lead in AI is shrinking.
The Financial Times reported on January 25 that DeepSeek, a small Chinese AI startup, shocked Silicon Valley. The report focuses on whether the more resource-rich U.S. AI companies can defend their technological superiority.
The report quoted Ritwik Gupta, an AI policy researcher at the University of California, Berkeley, as saying that DeepSeek’s recently released model shows that “AI capabilities have no moat.” Gupta added that China’s talent pool of systems engineers is much larger than in the U.S., and they know how to make the most of computing resources to train and run models cheaper.
