Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
Deepseek, the Chinese artificial intelligence company Deepseek, shook the markets this week, as its demands outperform the new artificial intelligence model and cost a small part of the price.
The assurances – specifically that the Grand Language Model in Deepseek only has raised $ 5.6 million for training – raised concerns about the amounts of frequency that technology giants currently spend on the computer infrastructure required to train and operate AI’s advanced AI’s work burdens.
The investor fears of the subversive Deepseek impact are erased approximately 600 billion dollars in NVIDIA’s market value on Monday-the largest decrease for one day for any company in the history of the United States.
But not everyone is convinced of Dibsic’s allegations.
CNBC has asked industry experts about their opinions about Deepseek, and how it actually compared to Openai, Chatbot Viral Chatgpt, which sparked artificial intelligence revolution.
What is a dipcique?
Last week, Deepsek R1 released, The new thinking form that competes with Openai’s O1. Thinking model is a large language model that is shattered down to a smaller cut and looks at multiple approaches before creating a response. It is designed to address complex problems similar to humans.
Deepseek was founded in 2023 by Liang Wenfeng, the co -founder of the quantitative hedge fund that focuses on artificial intelligence, to focus on large language models and access to artificial general intelligence, or AGI.
AGI, as a loose concept, refers to the idea of artificial intelligence that equals or transcends human thought in a wide range of tasks.
Many technology behind R1 is not new. However, what attracts attention is that Deepseek is the first to publish it in a high-performance AI model with-according to the company-significant discounts in energy requirements.
“The ready -made meals are that there are many possibilities for the development of this industry. The intense method of advanced chips/capital is a single technological method,” said Xiaomeng Lu, director of geological technology at Euroasia Group.
“But Deepseek proves that we are still in the emerging stage of developing artificial intelligence, and the path that Openai has developed is the only way to very artificial intelligence.”
How is it different from Openai?
Deepseek has two main systems that have a sensation of the artificial intelligence community: V3, the great language model that surrounds its products, and R1, the thinking model.
Both models are open source, which means that their basic code is free and available to the public for other developers to customize and redistribute.
Deepseek models are much smaller than many other large language models. V3 has a total of 671 billion teachers, or variables that the model learns during training. Although Openai does not reveal the parameters, experts appreciate its latest model, at least a trillion.
In terms of performance, diplisk He says Its R1 achieves a comparative performance with Openai’s O1 on thinking tasks, pointing to the standards including AIME 2024, Codeforce, GPQA Diamond, Math-500, MMLU and Swe-Bunc.
In a technical report, the company said that its V3 model had a training cost of only $ 5.6 million – a small part of the billions of dollars spent by prominent West AI laboratories such as Openai and Anthropic to train and operate artificial intelligence models. It is not clear after the amount of Deepseek costs to run.
If the training costs are accurate, this means that the model has been developed on a small part of the cost of competing models by Openai, Anthropor, Google And others.
Daniel Newman, CEO of Tech Insight The Futurum Group, said these developments indicate a “massive penetration”, although he had made some doubts about precise numbers.
“I think Dibsic’s breakthroughs indicate a meaningful turn to expand laws and a real necessity,” he said. “After saying this, there are still a lot of questions and doubts about the full picture of the costs because it is related to Deepseek.”
Meanwhile, Paul Trioelle, the first vice president of China’s policy and technology at DGA Group, indicated that it was difficult to make a direct comparison between the cost of the Deepseek model and the cost of the main American developers.
“The number was 5.6 million for DeepSeek V3 only to buy one training, and the company stressed that this does not represent the total cost of research and development to develop the model,” he said. “The total cost is likely to be much higher, but it is still less than the amount spent by the American artificial intelligence companies.”
Deepseek was not immediately available for comment when contacting it by CNBC.
Comparison Deepseek, Openai on the price
Deepseek and Openai reveals the prices of their models’ accounts on their websites.
Deepseek says that R1 costs 55 cents per million symbols of inputs – “symbols” that indicate each individual unit of text processed by the form – and $ 2.19 per million symbols of production.
By comparison, Openai’s pricing page shows that the company receives $ 15 per million input symbols and $ 60 per million output symbols. For GPT-4O MINI, the smaller and low-cost linguistic model in Openai, the company receives 15 cents per million input symbols.
Doubt on chips
The Deepseek detection of R1 has already resulted in a hot public debate about the validity of its demand – not the least of it because its models were designed despite export controls from the United States that restrict the use of advanced artificial intelligence chips to China.
Deepseek claims to have made her penetration using mature NVIDIA clips, including H800 and A100 chips, which are less advanced than the advanced H100s in the chips maker, which cannot be exported to China.
Since then, NVIDIA came out and said that Deepseek graphics units have been completely compatible with export.
The real deal or not?
Industrial experts seem to agree widely that what Dibsic has achieved is impressive, although some urged doubts about some Chinese company claims.
“Dibsic is legally impressive, but the level of hysteria is an indictment for many,” wrote American businessman Palmer Loki, who founded Oculus and Anduril on X.
“The number of $ 5 million is false. It is paid by a Chinese hedge box to slow the investment in American startups, service their shorts against American fennel such as NVIDIA, and hide the evasion of sanctions.”
Seena Rejal, the Senior Commercial Employee at NetMind, a London -based startup that provides access to Deepseek models of artificial intelligence via the distributed GPU, said he had not seen any reason for Deepseek.
“Even if it is turned off with a specific factor, it still comes largely effective,” Reagal told CNBC in an interview over the phone earlier this week. “The logic of what you explained is very reasonable.”
However, some have claimed that Deepseek technology may not have been built from scratch.
“Deepseek makes the same mistakes that O1 makes, a strong indication that technology has been separated,” said Vinod Khosla, billionaire investor in X, without giving further details.
It is required that Openai itself has indicated, as CNBC told a statement on Wednesday that it is reviewing the reports that may have Deepseek “using inappropriate output data from its models to develop its artificial intelligence model, a method referred to as” distillation “.
“We are taking aggressive and pre -emptive counter measures to protect our technology and we will continue to work closely with the United States government to protect the most capable models that are built here,” Openai spokesman told CNBC.
AI
However, the scrutiny surrounding Deepseek excels, artificial intelligence scientists agree widely that it represents a positive step for this industry.
Yan to be, the chief artificial intelligence scientist in DeadHe said that the success of Deepseek represents a victory for open source artificial intelligence models, and not necessarily a victory for China for the American definition is behind an open source AI model called Llama.
“For people who see Deepseek and think:” China is superior to the United States in artificial intelligence. “You read this error.
“Deepseek has benefited from open research and open source (for example Pytorch and Llama from Meta). They have come up with new ideas and built them at the head of others. Since their work is published and an open source, everyone can benefit from. This is the open search power. And the open source.
He watches: Why does Deepseek put AI in America at risk
– Katrina Bishop and Hayden Field, CNBC, contributed to this report
