Techonology

Why deepsek can change what believes about Silicon Valley AI

Artificial Intelligence Breakthrrow who is sending shock waves through stock markets, overthrowing the silicon Valley veterans, and a uninterrupted, victory about the end of America’s technical dominance in breathing: “LLMS To encourage logic in. “

22 page paperLast week, a scrapy Chinese was released by AI Start-up, called Deepsek, did not immediately set the alarm bells. It took the researchers a few days to digest paper claims, and its implications described by it. The company had created a new AI model called Dipsec-R1, created by a team of researchers who claimed that the second rate of AI Chips to match the performance of the leading American AI model at an excerpt of the cost. A minor number was used.

Deepsek said that it did this using clever engineering for the option of raw computing horsepower. And this was done in China, a country many experts thought that the global AI was at each other in the race.

Those who monitor some industries initially reacted to Deepsek’s success with mistrust. Certainly, he thought, Deepsek had cheated to achieve the results of R1, or rejected its number to make his model more impressive than his model. It may be that the Chinese government was promoting publicity to reduce the story of American AI dominance. Maybe was Deepsek Illegal Nvidia h100 hide a stand of chipsUS export control, and lying about it. It may be that R1 was actually just a clever re-scining of the American AI model that did not represent much in the way of real progress.

Finally, as more people dug in the details of the Deepsek-R1-which was released as open-source software, unlike most of the major AI models, to examine its internal functioning more closely to outsiders. Got permission from-Their skepticism was converted into anxiety.

And at the end of the last week, when many Americans began using the model of Deepsek for themselves, and the Dipsek Mobile App hit the number one place on Apple’s app store, it got completely nervous.

I have been seen most dramatically in the last few days – as have been claimed, as it has been. One silicon valley investorThis is a detailed conspiracy by the Chinese government to destroy the Deepsek American technical industry. I also feel that it is admirable that the company’s shosting budget has been badly exaggerated, or that it has been piggiback on progress made by American AI firms, which has not been revealed.

But I think Deepsek’s R1 success was real. Based on the conversation, I have done with the internal sources of the industry, and a week experts to test the findings of paper and test the findings of paper for it, it seems to question many major beliefs. Which is creating American technical industry.

The first belief is that to create a state -of -the -art AI model, you need to spend huge amounts of money at powerful chips and data centers.

It is difficult to beat how basic this dogma has been made. Companies such as Microsoft, Meta and Google have already spent billions of billions of dollars, which they felt that they felt the next generation AI models and felt that they felt to build infrastructure. They plan to spend tens of billions more than billions of billions – or in the case of openi, $ 500 billion through a joint venture with Oracle and SoftBank, which was declared last week.

Deepsek has spent a small fraction of that building R1. We do not know the exact cost, and there are Lots of warnings to make About the data released so far. This is almost certainly more than $ 5.5 million, the company claimed that it spends training a previous model.

But even if the cost of R1 is 10 times more to train than deepsek claims, and even if you are a factor in other costs, they can be excluded, such as engineer salary or basic research. Cost of, it will still order lower quantity than American AI. Companies are spending to develop their most capable models.

The clear conclusion to draw is not that American technical giants are wasting their money. Once trained, it is still expensive to run a powerful AI model, and it is because of thinking that spending hundreds of billions of dollars will still be understood for companies like openi and Google, which Pack can pay dearly.

But the success of Deepsek at the cost “is” great better “challenges the story that has operated the AI ​​weapons race in recent years, which is relatively small models, when properly trained, the performance of very large models is done. Can match or cross.

In turn, this means that AI companies may be able to achieve much powerful abilities with much less investment than before. And it suggests that we may soon see the flood of investment in small AI start-ups, and can compete much for the stalwarts of Silicon Valley. (Which, due to the huge costs of training their models, are mostly competing with each other so far.)

Other, there are more technical reasons that everyone in Silicon Valley is focusing on Deepsek. In the research paper, the company explains some details about how the R1 was actually created, including some sophisticated techniques in the model distillation. (Originally, this means compressing the large AI model into small people, became cheaper to run without losing them too much in the way of performance.)

Details also included details suggested It was not as difficult as a “vanilla” AI language model was thought to convert into a more sophisticated logic model, applying a technique to learn reinforcement on top of it. (Don’t worry if these conditions go to your head – what matters that ways to improve the AI ​​system that were previously protected by American tech companies, are now out on the web, to take and repeat anyone. Are independent.)

Even if the prices of American tech giants stock are cured in the coming days, the success of Deepsek raises important questions about their long -term AI strategies. If a Chinese company is capable of creating cheap, open-source models that corresponds to the performance of expensive American models, why will anyone pay for us? And if you are a meta-so the only American tech veteran who releases his model as a free open-source software stops the deepsek or any other start-up only prevents his model, which you spent billions of dollars Are, and have distilled them in small, cheap models that they can offer for Penny?

The success of the Deepsek has also outlined some geopolitical perceptions that many American experts were making the situation in China in the AI ​​race.

First, it challenges the legend that China is meaningful behind the frontier, when it comes to making a powerful AI model. Over the years, many AI experts (and policy makers who listen to them) have assumed that the United States had at least several years of leading, and copying the progress made by American tech firms for Chinese companies Was quickly difficult.

“In a few weeks.

(New York Times sued Openai and his partner, Microsoft, accusing them of violation of copyright violations of news material related to the AI ​​system. Openai and Microsoft denied those claims.)

The results also raise questions about whether the US government has taken steps to limit the spread of powerful AI systems for our opponents – namely, used to prevent export control powerful AI chips from falling into China’s hands Going – is working as designs, or whether those rules need to be adapted to keep in mind the new, more efficient methods of the training model.

And, of course, it is concerned about what it would mean for privacy and censorship if China led the creation of a powerful AI system used by millions of Americans. Users of the models of Deepsac have noticed that they regularly refuse to answer questions about sensitive themes inside China, such as Tianmen Square Massacre and Uyghur Nirodh Camp. If other developers manufacture at the top of the model of Deepsek, as common with open-source software, those sensorship measures may be embedded throughout the industry.

There are also privacy specialists raised concerns Regarding the fact that the data shared with the Deepsek model may be accessible by the Chinese government. If you were worried about Ticketkok, it was being used as a means of monitoring and publicity, the rise of Deepsak should also worry you.

I am still not sure what will be the full impact of Deepsek’s success, or will we consider the release of R1 that the “Sputnic moment” for the AI ​​industry, as is something Claimed,

But it is intelligent to take this possibility seriously that we are now in a new era of AI Brinkmship – that the largest and richest American technical companies can no longer win by default, and that the spread of a rapidly powerful AI system We can be more difficult than we thought.

Very at least, Deepsek has shown that the AI ​​weapons race is actually, and that after many years of affair, there are still more surprises in the store.

)China
#deepsek #change #believes #Silicon #Valley

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *