DeepSeek Turned the AI World on Its Head, but Don’t Fall for the Hype Just Yet
AI has just broke out.
Entrepreneur Marc Andreessen Bold claims on XIn the past Sunday, the social media platform was previously called Twitter. Silicon Valley and stock markets and online predictions stand out from the seismic activities in AI space.
Deepseek AI, a new AI model from China Jump to the top of the Apple App StoreIt is sending noise in the entire Silicon Valley. Deepseek claims its AI competition, in some cases OverwhelmThe cost of OPNAI’s O1 reasoning model is very small. Not only that, the R1 model of Deepseek is completely open source, which means that the code can be visited publicly, and anyone can use it for free.
The key difference between DeepSeek-R1 and OPENAI is that R1 allows you to see your mind chain. This is an incredible insight into AI’s “thinking”. You can actually see it Try to answer questions about Tiananmen Square Before the reaction was canceled, halfway. NVIDIA is a chip company that provides power for the AI ​​revolution. I saw it Stock plunge 18 % And lose one Creating a record 600 billion US dollars After Deepseek’s weekend rose. This is very meaningful. If DeepSeek is said to be correct, then it has achieved almost O1 -level performance on the older NVIDIA chip, and it costs a small part of the cost.
Online commentators are still trying to understand the sudden appearance of Deepseek in the AI ​​market. Is it actually using O1, the cost is low? To what extent can DeepSeek and China claim for efficiency? Is the savings saved from the main technical unlocking, or is it cheaper to use in other areas in China?
In any case, R1 is impressive.
“This kind of burden has a small company and startup opened the door to use advanced AI technology that could not be accessed before.” You have your corpusAI Research Engine, in a statement to CNET. Morris added that DeepSeek competition for the established AI participants, and “there is a progress that may stimulate AI technology, which leads to more effective and easier solutions to meet the growing needs. “
This may be why OPENAI CEO Reduce prices Saturday’s near -end O3 mini query.
As Big Tech has continuously invested billions of dollars, processing power and energy on AI, the efficiency unlocking of DeepSeek may be similar to the leap that we see when we transform from a carburetor to the fuel injection system. Unlike Openai, the R1 model of Deepseek is open source, which means that anyone can use this technology. This is a major damage to the market. It is currently dominated by Openai’s ChatGPT and Google Gemini. They are closed and requires users to pay for their functional kit.
In the AI ​​competition between the United States and China, due to the large -scale investment investment of Silicon Valley and the government’s blockade of NVIDIA, the United States has always maintained a leading position to sell the latest AI chips to China. However, this blockade may only be Inspire China to make its own chips fasterEssence Money and protectiveism are considered to be the second way to rank China, so that the world depends on American technology. This dynamics may change. Consumers and companies around the world can access high -performance “reasoning” models at a small part of cost. Not only that, Tiktok’s parent company Bytedance released Even cheaper R1 modelEssence
With the response of market and social media to China’s new development, it is too early to say that the United States has been defeated. But at least, China is catching up quickly.
“China has produced GPT-4 quality models, but the lag is longer-just like spending a year, a year and a half, similar time. But now there is a Chinese model, maybe there are only six backward a few behind. One month, I think this is a difference. ” CivaiA non -profit organization using software to demonstrate AI capabilities. “Therefore, the United States is still leading, but it is not as big as before.”
One of the things that R1 will definitely help AI to catch up with OpenAI is the ability to read its mind chain. Even if R1 cannot get all the answers correctly, you can see how the reason can better help develop it. Hansen said that people’s sense of “shock and awe” of R1 comes from the ability to read its mind chain. Its insight, OPENAI cannot use its O1 model, because hidden secret sauce will allow people to sell subscribing fees for monthly for access.
Nevertheless, the performance ratio of R1 should still take certain doubts. this There are more than 100 co -owners published by DeepSeek’s published white paperEssence AI with low -cost training for $ 5.5 million is a lot of brains. The cost of $ 5.5 million may be only to train the energy cost of the model and minus the personal salary of each researcher, but China is not completely transparent in calculating these energy costs. The cost of establishing a data center in China may be different from establishing a data center in the United States. Hansen said that according to Hansen, the cloud provider or the Chinese government itself subsidized the cost.
Deepseek chips that train their models are also doubtful. Actually it is a company Use the older NVIDIA A100 and H800 chips or China visits the latest H100 chip in other waysAs SCALE AI CEO Alexandr Wang said.
Even if we use a number of $ 5.5 million as a highly conservative estimate, it is still much less Training GPT-4’s $ 100 million IT IT spent OpenaiThe company’s previous AI model. Openai has not yet released data about the cost of building O1, but given that its tokens are much higher, it may be more expensive.
“As the data center of the United States is expected to be doubled or three times by 2030, any efficiency savings may have a significant impact,” Mark James, an interim director of the Institute of Permon Law and Environment, is in a statement. Speaking. already, AI’s high -energy demand pressure public businessEssence If DeepSeek’s claim is correct, it may greatly reduce the potential power load, thereby reducing the pressure of consumers and the environment. James said: “On the other hand, more effective models can unlock more growth in the industry, which will reduce efficiency saving and exacerbating pressure on our grid.”
Claims that the United States has Lost the artificial intelligence war It may be too early. At least, the landscape becomes more competitive immediately and has room for continuing innovation. Deepseek does not mean that the world is realizing Artificial intelligenceOr super advanced AI, smarter than humans, can learn by themselves.
“I think DeepSeek has not made us closer to artificially common intelligence (AGI), but I do think that this makes us closer to the business -feasible large -scale language model (LLM) application. This is great.” Ben Gortzel CEO and founder of the Artificial Super Intelligence (ASI) Alliance SingularityNetEssence Deepseek still has the same cognitive limitations as other AI models. Nevertheless, the efficiency of Deepseek can still democrate AI.
