China Artificial Intelligence Corporation’s Deepseek rock market this week claims that its new AI model is better than the OpenAI market, and it takes a small part of the price.
These assertions-especially the large language model of DeepSeek only spent $ 5.6 million to train-caused people’s concerns about witnesses, that is, technology giants are currently calculating the computing infrastructure required for training and running senior AI workloads.
However, not everyone is believed by Deepseek’s claim.
CNBC asked the industry experts’ views on Deepseek, and its practical comparison with OPENAI, the founder of Stagral Chatbot ChatGPT, which triggered the AI ​​revolution.
What is Deepseek?
Last week, Deepseek released R1 New reasoning model That match OPENAI O1. The reasoning model is a large language model that breaks through smaller parts and considers multiple methods before generating responses. It aims to deal with complex issues similar to humans.
Deepseek was established in 2023 by the high -end co -founder of the quantitative hedge fund with AI as a high -level co -founder, focusing on large language models and reaching artificial intelligence or AGI.
AGI, as a conceptual concept, refers to the idea of ​​AI, which is equal to or surpassing human intelligence in extensive tasks.
Many of the technologies behind R1 are nothing new. However, it is worth noting that DeepSeek is the first person to deploy it in a high -performance AI model. This model (according to the company) greatly reduces power demand.
“The main point is that there are many possibilities that may develop the industry. High -end chips/capital -intensive methods are a technical method,” Said Xiao Lu, director of the Geographical Technology Practice of Eurasia Group.
“But Deepseek proves that we are still in the new stage of AI development. The road established by OpenAI may not be the only way to highly powerful AI.”
How is it different from Openai?
DeepSeek has two main systems. These systems have buzzed from the AI ​​community: V3, that is, a large language model from the product, and its reasoning model R1.
Both models are open source, which means that their basic code is free, and can be used to openly use other developers to customize and redistribute them.
Deepseek model is much smaller than many other large language models. V3 has a total of 671 billion parameters, or variables learned during the training period. Although OpenAI did not disclose parameters, experts estimate that its latest model has at least trillion US dollars.
In terms of performance, deepseek explain Its R1 model can obtain equivalent performance as OPENAI in reasoning tasks, and uses the benchmarks of AIME 2024, Codeforcess, GPQA Diamond, Math-500, MMLU, MMLU and SWE-BENCE.
The company said in a technical report that the training cost of its V3 model is only $ 5.6 million, which is a small part of the billions of dollars, namely the famous Western AI laboratory (such as OpenAI and Anthropic) flowers in training and. Run on its basic AI model. However, what is the current cost of running.
However, if the cost of training is accurate, it means that the development of the model is a small part of the cost of the cost of the competitors model of OpenAI Google There are others.
Daniel Newman, chief executive of Tech Insight Futurum Group, said the development of these situations implied a “huge breakthrough”, although he raised some doubts about the exact number.
He said: “I believe that the breakthrough of DeepSeek indicates that the meaningful inflection point of the expansion method is real necessity.” “Nevertheless, this is related to the development of Deepseek, and there are still many problems and uncertainty.”
At the same time, Paul Triolio, a senior vice president of China and the senior technical policy of the consulting company DGA Group, pointed out that it is difficult to directly compare the model cost of DeepSeek’s model with the model cost of the main developers in the United States.
He said: “The 5.6 million numbers of DeepSeek V3 are only trained at one time. The company emphasizes that this does not represent the total cost of development of the model.” The expenditure of the big company “”
When CNBC is connected to Deepseek, no comments were published immediately.
Compare DeepSeek, price Openai
Both DeepSeek and Openai disclosed their models on their websites.
DeepSeek said that the price of R1 is 55 cents per million token (referring to the text unit processed by the model) per 1 million token, while the output of each 1 million tokens is 2.19 Dollar.
In contrast, the O1 pricing page of OpenAI shows that the company’s fees are $ 15 per million dollars per 1 million US dollars, and $ 60 per million dollars per 1 million US dollars. For the low-cost language model GPT-4O mini with a smaller OpenAI, the company charges 15 cents per 1 million input tokens.
Doubt about chips
DeepSeek’s exposure to R1 has aroused the fierce debate of the authenticity of its claim-especially because although the United States restricts the use of advanced AI chips for China’s export control, its model is still established.
Deepseek claims that it has made breakthroughs with mature NVIDIA editing (including H800 and A100 chips). These chips are not as advanced than the advanced H100 that cannot be exported to the top H100 that cannot be exported to China.
However, in Last week’s comment on CNBCSCALE AI CEO Alexandr Wang said he believes that Deepseek uses the banned chip, which is the claim of Deepseek denied.
Since then, NVIDIA came out and said that the GPU used by DeepSeek was completely in line with exports.
Whether it is a real transaction?
Industry experts seem to be widely agreed, and the achievements of DeepSeek are impressive, although some people urge some of Chinese companies’ claims.
American entrepreneur Palmer Luckey wrote on X: “Deepseek is very impressive, but the level of hysteria is the prosecution of many people.”
“The number of $ 5 million is forgery. It is driven by the China Hedie Fund to slow down the investment in the American AI startup, provide its own shorts for the US Titan team like NVIDIA, and avoid approval.”
NetMind’s chief business officer Seena Rejal is a startup of London Headquarters. He visited Deepseek’s AI model through a distributed GPU network. He said he thought that there was no reason not to believe Deepseek.
Rejal said to CNBC in a telephone interview earlier this week: “Even if it is a certain factor, it is still very efficient.” “The logic they explained is very wise.”
However, some people claim that DeepSeek’s technology may not be built from scratch.
Billionaire investors Vinod Khosla said on X: “Deepseek has made the same mistake O1, which is a strong sign.”
Claiming that OpenAI itself has mentioned that in a statement on Wednesday, CNBC is told that it is reviewing the report of DeepSeek’s report may be “inappropriate” from its model to develop its AI model from its model. This is called “” The method of distillation.
A OpenAI spokesman told CNBC: “We have adopted active and proactive countermeasures to protect our technology and will continue to work closely with the US government to protect the most capable model established here.”
Artificial intelligence
However, the review around DeepSeek was shocked, and AI scientists widely agreed that this marked the positive step of the industry.
Yann lecun, chief AI scientist YuanHaving said that the success of Deepseek represents the victory of the open source AI model. It is not necessarily the victory of China to the US yuan. This is a popular open source AI model, called LLAMA.
“For those who see DeepSeek’s performance and thinking: ‘China surpasses the United States in AI.’ You are reading the correct reading.
“Deepseek is profitable from open research and open source (for example, PyTorch and LLAMA from Meta). They put forward new ideas and established new ideas in other people’s work. Because their works are published and open, they are published and open source, and they are published and open source, and they are published and open. So everyone can profit from it.
-CNBC’s Katrina Bishop and Hayden Field contributed to this report
