On January 24, the DeepSeek-R1 benchmark has risen to the third place in all categories of large models on the foreign large model ranking Arena, which is compared with OpenAI in the style control model (StyleCtrl) classification O1 tied for first place. And its arena score reached 1357 points, slightly surpassing OpenAI 1352 points for O1. This is the successor to DeepSeek-V3 in culling OpenAI After closed-source models such as o1 ranked first in the open-source model category, DeepSeek-R1 once again attacked the world's most powerful AI model.
|DeepSeek震惊硅谷
Recently, the AI community has been dominated by a startup from China — DeepSeek, stirred "The world is turned upside down"! In just one month, DeepSeek-V3 and DeepSeek-R1 have released two large models one after another, the key is that they are not only low cost, but also can compete with OpenAI in performance, which can surprise Silicon Valley, and even Meta engineers are scared overnight "Overtime" in an attempt to replicate the success of DeepSeek.
Scale In an interview on January 24, Alexander Wang, the founder of AI, did not hesitate to praise DeepSeek, bluntly saying that in their tests, DeepSeek performed the best, on par with the top models in the United States. Before, Alexander Wang commented that DeepSeek-V3 was given to the United States by the Chinese technology community "Bitter lesson", he also sighed: "While the United States is resting, China (the scientific and technological community) is burying its head in hard work, catching up with lower costs, faster speed, and stronger strength." ”
This wave of AI in China "God Operation" has also successfully attracted the attention of major foreign media, who have reported that they feel that the new breakthrough of China's large model is like a wake-up call to Silicon Valley.
Even more surprising is that at $500 billion When the "Stargate" plan was announced, DeepSeek created a breakthrough AI model at an ultra-low cost without using cutting-edge chips, which makes people wonder: Is investing hundreds of billions of dollars in the AI industry really the most effective way to develop?
|DeepSeek-R1: The new star of the chatbot arena
In the latest chatbot arena comprehensive list, DeepSeek-R1 stood out with excellent performance, tied for third place with the top inference model o1. However, what is even more remarkable is that in some specific areas, DeepSeek-R1 has shown absolute superiority.
When it comes to technical challenges such as difficult prompts, coding and math problems, DeepSeek-R1 leads the way. This demonstrates the unmatched ability of DeepSeek-R1 to handle complex tasks and areas of expertise.
Not only that, but in terms of style control, DeepSeek-R1 and o1 go hand in hand to occupy the first position. This means that the model is not only able to accurately understand the user's instructions, but also to generate content that fits a specific style based on the user's needs. Whether it's a formal presentation or a casual conversation, DeepSeek-R1 can handle it with ease.
In a test that combines difficult prompts with style control, DeepSeek-R1 once again tied for first place with o1, further demonstrating its strength in performing complex tasks and fine-grained content control. This capability allows DeepSeek-R1 to not only solve difficult problems, but also present results in the way the user expects.
Artificial-Analysis conducted an initial benchmark of DeepSeek-R1 and the results showed that it received the second-highest score in the AI Analysis Quality Index. What's even more appealing is that the price of DeepSeek-R1 is only about one-thirtieth of that of o1, providing users with a very cost-effective option.
|DeepSeek以低成本挑战巨头
Stanford University and Epoch A study published by AI researchers in the middle of last year suggests that the cost of training the largest AI models could exceed $1 billion by 2027. Gartner predicts that by 2028, hyperscalers like Google, Microsoft and AWS will spend a staggering $500 billion on AI servers alone. However, in this capital-intensive sector, a company called DeepSeek has taken a very different path.
Noah's Arc Capital Management notes that the DeepSeek-V3 model has the potential to be a game-changer in the field of training and inference. Unlike other companies, which invest billions or even tens of billions of dollars, DeepSeek is relatively inexpensive to train. This cost-effective solution has made the industry rethink whether the huge investment of "big miracles" is really the most effective way?
The well-known big V "THE" in the U.S. stocks SHORT BEAR" tweeted via Platform X (formerly Twitter) on January 24: "DeepSeek has created a painful moment for AI giants, and investors must sound the alarm bells about it. He further explained, "If it only takes $55 million to beat OpenAI, the commercialization of the AI industry could be much faster than many people expected." ”
He also mentioned: "According to Sequoia Capital, AI companies in the U.S. need to generate about $600 billion in revenue annually to pay for their AI hardware. Now it seems that this high investment is becoming more and more unprofitable. ”
Holger, a well-known financial journalist Zschaepitz also said on January 25 that DeepSeek has built a groundbreaking AI model at a fraction of the price and has not relied on cutting-edge chips. This raises questions about the utility of hundreds of billions of dollars in capital expenditures across the industry. Some investors believe that the stock price of chip stocks in the US stock market may also face challenges as a result.
Investor Geiger "DeepSeek is not only as good as OpenAI, but even better, at a cost of only 3% of the latter," Capital said. At the same time, hundreds of billions of dollars have been poured into American companies. So...... What will happen to the NASDAQ? "It's worth noting that Nvidia's share price has fallen 2% since the release of DeepSeek-V3. And on January 24, after DeepSeek-R1 sparked widespread overseas discussions, Nvidia's stock price fell by 3.12%.
Technological strength and commercialization challenges coexist
In the eyes of industry insiders, DeepSeek is particularly fortunate compared to other large-scale model startups in China. Not only does it not have the pressure to raise funds, it also does not need to prove its value to investors, and it does not have to struggle between technology iteration and product application optimization. This freedom allows DeepSeek to focus on technological innovation without being tied down by short-term business goals.
However, as a commercial company, sooner or later, DeepSeek will face the same pressures and challenges as other model companies. While it seems to be comfortable at the moment, the future path to commercialization is still fraught with uncertainty.
For the domestic AI model industry, the addition of a company with real technical strength like DeepSeek is undoubtedly a good thing. It not only improves the overall technical level of the industry, but also sets an example for other companies. As one industry veteran put it, "If a company like DeepSeek exists, the whole industry will benefit from it." ”