The breakthrough of artificial intelligence, which broadcasts shock waves through the stock markets, will scare the giants of Silicon Valley and generating breathless breath occupies the end of the technological dominance of America with an undemanding, reversible title: “Motizing LLMS thinking ability to strengthen learning.”
The 22 -page paperPublished last week Skarppy by Chinese start-up AI called Deepseek, immediately did not start the alarm bells. The scientists took several days to spend the claims of paper and the consequences of what he described. The company has created a new AI model called Deepseek-R1, built by a team of scientists who claimed to have used a modest number of AI chips to correspond to the performance of leading American models AI for a fraction of the cost.
Deepseek said he did it using smart engineering to replace raw computing power. And it did in China, a country, in a country that many experts thought to be second in AI.
Some industry observers initially responded to Deepseek’s breakthrough with distrust. They certainly thought that Deepseek was cheating on to achieve R1’s results, or upset their numbers to make their model look more impressive than they were. Perhaps the Chinese government promoted propaganda to undermine the narrative of American dominance AI. Maybe he was Deepseek Hidden a stash of illegal chips NVIDIA H100forbidden under US export checks and lied about it. Perhaps the R1 was in fact just a clever re -re -re -re -re -re -skinning of the American AI models that did not pose too much in the way of real progress.
In the end, when more people dug into Deepseek-R1-KTE details, unlike most AI front models, they were released as open source software, allowing foreigners to examine their inner functioning in more detail-skepticism has changed in concern.
And at the end of last week, when a lot of Americans started using Deepseek for themselves, and Deepseek’s mobile application hit the Apple’s App Store, leaned into a fully blown panic.
I am skeptical to the most dramatic I have seen in the last few days – for example, a claim that has been made One Investor Silicon ValleyThis deep search is a sophisticated conspiracy of the Chinese government to destroy the US Technology Industry. I also think it is likely that the budget of the company that is breeding was heavily exaggerated, or that it was advocating the advances of the American company AI in a way that did not publish.
But I think Deepseek’s breakthrough R1 was real. Based on the conversations I led with the initiates of industry, and weekly experts who have established themselves and tested the finding of paper for myself, it seems to cause several main prerequisites that the US Technology Industry has made.
The first is that you have to spend a huge amount of money on powerful chips and data centers to create top AI models.
It is difficult to overestimate how this dogma has become the foundation. Companies such as Microsoft, Meta and Google have already spent tens of billions of dollars by building infrastructure, which they thought needed to create and operate new generation models. They Plan to spend tens of billions – or, in the case of Openai, up to $ 500 billion via A common business with Oracle and SoftBank This was announced last week.
Deepseek seems to have spent a small part of this building R1. We do not know the exact cost and exist a lot of objections About the characters they have published so far. It is almost certainly higher than $ 5.5 million, the number that the company claims to have spent the training of the previous model.
But even if R1 costs 10 times more for training than Deepseek assertion, AI when you take into account other costs, they could exclude, such as utilities or basic research costs, it would still be less than what US AI company spends on development their most capable models.
The obvious conclusion that needs to be inferred is not that American technology waste money. Once they are trained, it is still expensive to operate the powerful AI models, and there are reasons why you think hundreds of billions of dollars will still make sense to companies like Openai and Google that can afford to pay expensive to keep the head of the package.
But Deepseek’s breakthrough about cost challenges, however, “bigger” narrations, which have managed the AI armament in recent years by showing that relatively small models, when they are properly trained, can correspond or exceed the performance of much larger models.
This, in turn, means that AI can be able to achieve very strong capabilities with much less investment than it was before. This suggests that we can soon see a flood of investment in smaller start -up AI and much more competition for Giants of Silicon Valley. (Which is mostly competing because of the huge costs of training their models.
There are other technical reasons that everyone in Silicon Valley pays attention to Deepseek. In the research article, the company reveals some details of how R1 was actually built, which includes some of the top techniques in the model distillation. (Basically it means compressing large models AI down to smaller, which makes them cheaper without losing much in the way of performance.)
Deepseek also contained details that proposed The fact that it was not as difficult as it was previously thought that it transformed the language model of “vanilla” into a more sophisticated model of thinking using technology known as strengthening learning. (Don’t worry if these conditions pass over the head – it depends on the fact that the methods for improving AI systems that have previously been protected by American technology companies are now on the web, free of charge for anyone who could marry and replicate.)
Although the prices of stocks of American technology giants will recover in the coming days, Deepseek’s success raises important questions about their long -term AI strategies. If the Chinese company is able to build cheap, open-source models that correspond to the performance of expensive American models, why would anyone pay for us? And if you are a meta-unique American technical giant that releases its models as free software with open source code-that prevents Deepseek or other start-up to simply take your models on which you spent billions of dollars and distill them to Smaller, cheaper models that can offer for pennies?
Deepseek’s breakthrough is also undermined by some of the geopolitical assumptions that many American experts recorded about the position of China in Rase AI.
First, it challenges the narration that China is meaningfully beyond the border of building strong AI models. For years, many AI experts (and the creators of politicians who listen) assumed that the United States had had leadership for at least several years and that copying the progress of American technology companies was disproportionately difficult for Chinese societies.
However, Deepseek’s results show that China has advanced AI capabilities that can match or exceed models from Open and other American AI companies, and that the breakthroughs that US companies have made in a few weeks.
(The New York Times has sue Openai and his partner, Microsoft, accuse them of violating copyrights of the content of intelligence related AI systems. Openai and Microsoft have denied these demands.)
The results also raise questions about whether the US government has reduced the spread of powerful AI systems to our opponents – namely export checks used to prevent the fall of strong AI chips into the hands of China – they work as designed or whether these regulations must adapt to to take into account new and more efficient training ways.
And of course, there are concerns about what it would mean for privacy and censorship if China took over the management of powerful AI systems used by millions of Americans. Deepseek users they noticed that they routinely refuse to respond to questions about sensitive topics in China, such as the square on the square and detention camps Uyghur. If other developers build on top of Deepseek, as is common with open source software, these censorship measures can be built into the entire sector.
Personal data protection experts also have fears The fact that data shared with Deepseek models can be accessible to the Chinese government. If you feared that Tiktok would be used as a tool of supervision and propaganda, he should also take care of Deepseek’s rise.
I am still not sure what the full impact of Deepseek’s breakthrough will be, or whether we will consider the release of R1 as a “sputnik moment” for industry AI because some have claimed.
However, it seems wise to take seriously the possibility that we are now in the new AI Brinkmanship era – that the largest and richest American technology companies no longer have to win by default and that the content of increasingly powerful AI systems can be harder than it can be harder than we thought.
Deepseek at least showed that the AI armament race is really turned on and that after a few years of dizzying progress, there is an even more surprise.
