A shocking waves through the stock market, sucking up Silicon Valley giants, and breathing out of breath, has arrived at the end of the American technical advantage.
A 22 -page paper on page 22, released last week by a poor Chinese emerging company called DeepSeek, did not immediately cause an alarm bell. It took a few days for researchers to digest the claims of the dissertation and the meaning of what they explained. The company has created a new AI model called Deepseek-R1. This was built by a research team that uses a modest number of AI chips to match the performance of a major American AI model at a slight cost.
DeepSeek stated that it used to use a clever engineering to replace raw computing horsepower. And it did it in China. A country where many experts thought they were second in the worldwide AI racing.
Watchers in some industries were initially reacted to DeepSeek's distrust. Certainly, they thought that DeepSeek was deceived to achieve the R1 results, or defeated the numbers to make the model look more impressive. Perhaps the Chinese government was promoting propaganda to impair the American AI rule. Probably, Deepseek hides the hidden location of the illegal NVIDIA H100 chips, banned under US exports, and lied about it. Probably, R1 was actually a clever rebat of an American AI model, which was not much in the actual progress method.
Eventually, as more people delve into the details of Deepseek -R1 -Unlike most major AI models, they are released as open source software, and outsiders can investigate the internal mechanism more closely. You can -their skepticism has turned anxiety.
Late last week, many Americans started using DeepSeek models themselves, and when the DeepSeek Mobile app hit the number one spot in the Apple App Store, it fell into a full -fledged panic.
I am skeptical of the most dramatic take on the past few days, including claims made by one Silicon Valley investor. Deepsquek is an elaborate conspiracy for the US government to destroy the US technical industry. Also, I think it is due to the fact that the company's shoelfish budget is very exaggerated, or that it has been saved by an American AI company's progress in a way that has not been revealed.
However, I think DeepSeek's R1 breakthrough was real. Some major assumptions have been conducted by the US technical industry based on conversations with the industry insider and one week's worth of paper survey results for themselves and testing. He seems to be asking questions.
The first is that building a state -of -the -art AI model needs to spend enormous money on powerful chips and data centers.
It is difficult to exaggerate how basic this doctrine has become. Companies, such as Microsoft, Meta, and Google, are already spending hundreds of billions of dollars to build infrastructure that they needed to build and execute next -generation AI models. They plan to spend $ 500 billion in Openai, in the case of Openai, through a joint venture between Oracle and SoftBank, which was announced last week.
Deepseek seems to have spent some of the building R1. I don't know the exact cost. There are many warnings about numbers released so far. Almost surely exceeds $ 5.5 million. This claims that the company spent training for previous models.
However, even if R1 is 10 times more training than Deepseek claims, even if other costs, such as the salary of engineers and the cost of basic research, it will be more magnificently less than the American AI. Companies spend the most capable models to develop.
The obvious conclusion is that the US high -tech giant is wasting money. It is still expensive to execute a powerful AI model after training. There is also a reason that spending hundreds of billions of dollars makes sense for companies such as Openai and Google. pack.
However, Deep Shek's breakthroughs on costs have been promoting AI Arms Race in recent years by indicating that relatively small models can match or exceed the performance of a much larger model when they are properly trained. I'm trying the story of “Bigger is Better”.
This means that AI companies can achieve very powerful abilities with much less investment than before. And we immediately suggest that there is a possibility that there will be much more competition with Silicon Valley giants, which will quickly cause a flood of investment in small AI to emerging companies. (This is a huge cost for training models, so most have competed with each other until now.)
There are other technical reasons that everyone in Silicon Valley pays attention to DeepSeek. In the research paper, the company has clarified some details about how R1 was actually built. This includes state -of -the -art technology in the distillation of the model. (Basically, it means compressing a big AI model into a smaller one, and it is cheaper to run without hindering the performance).
Deepseek also included the details that suggested that it was not as difficult as converting the “vanilla” AI language model into a more sophisticated inference model. (Don't worry even if these terms go beyond your head. The important thing is that there is a method on the web on how to improve the AI system, which was previously protected by high -tech companies in the United States, and can be duplicated for free. It is.
Even if the stock price of a high -tech major in the United States has recovered in the next few days, Deepseek's success makes important questions about their long -term AI strategy. Why do Chinese companies pay our price if they can build an open source model that matches the performance of expensive American models? And if you are the only high -tech company in the United States who will release a model as a meta -free open source software, DeepSeek or another startup simply spend billions of dollars and distill a smaller distillation. What to prevent, inexpensive models that can be offered to penny?
DeepSeek's breakthrough also overturns some of the geopolitical assumptions that many American experts have been in China's position in the AI race.
First of all, for the construction of a powerful AI model, we will challenge the story that China is meaningful behind the frontier. For many years, many AI experts (and policy proprieters who listen to them) assume that the United States has at least a few years of lead, and the progress has been made by a US high -tech company. Copying was very difficult for Chinese companies to do it quickly.
However, the result of DeepSeek is that China is promoting AI functions that can match or exceed the models of Openai and other American companies, and the breakthroughs that US companies have made are Chinese companies or at least one China. It indicates that it may be trivial for companies. In a few weeks.
(New York Times complains of Openai and his partner, Microsoft, accused them of copyright infringement of news content related to AI Systems. Openai and Microsoft denied these claims.)
The results are also designed by the US government to restrict powerful AI systems to enemies, that is, export management used to prevent powerful AI chips from falling in Chinese hands. Questions about whether or not it is designed, or designed. Are these regulations?
And, of course, there are concerns about what China takes the initiative to build a powerful AI system used by millions of Americans. DeepSeek model users have noticed that they have always rejected answers to delicate topics in China, such as Tianannmen Square Massacre and Uyghur detention camps. As is common in open source software, if other developers are built on a model of DeepSeek, these censorship can be incorporated into the entire industry.
Privacy experts also raise concerns about the fact that the Chinese government could access data shared with the DeepSeek model. If you are worried that Tactoku will be used as a monitoring and propaganda tool, you will be worried about the rise of deep safe.
I don't know if the R1 release is considered to be “SPUTNIK MOMENT” in the AI industry, as some people have claimed, as the complete impact of Deepseek's breakthrough is claimed.
However, it is wise to take the potential to be in a new era of AI now. We thought that the largest and wealthy American high -tech companies could not win by default, and it might be more difficult to include the spread of more powerful AI systems.
At least, DeepSeek shows that AI Arms Race is really on, and after a few years of progress, there are even more surprises.