The Chinese artificial intelligence company Deepseek swinging markets this week, and claims that its new AI model surpasses Openi and cost a fraction of the price for construction.
Claims – specifically that Deepseek’s big linguistic model cost only $ 5.6 million for training – causing concerns about the sum of the eyes that technological giants are currently spending on a computer infrastructure needed to train and start advanced AI loads.
But not everyone is convinced Deepseek’s claims.
CNBC asked an industry expert for his attitudes on Deepsek and how he actually compared to Openi, the creator of a viral chatbot chatgpt who launched the AI Revolution.
What is Deepseek?
Last week Deepseek posted R1, ITS New reasoning model which Openai’s O1 rivals. The explanation model is a large linguistic model that breaks the extensions into smaller pieces and considers multiple approaches before generating an answer. It is designed to process complex problems in a similar way to people.
In 2023, Deepseek founded Liang Weenfeng, co -founder of a quantitative hedge fund aimed on AI, to focus on major linguistic models and achieve artificial general intelligence or agi.
AGI as the concept is slightly related to the idea of Ai that is equal to or outweighs the human intellect on the wide range of tasks.
A large part of the technology behind R1 is not new. What is, however, is that Deepseek is the first to distribute it in the highly successful AI model of the S-PREMA company-significant reduction of energy requirements.
“The company is that there are many opportunities for the development of this industry. The top -notch chip/capital intensive way is one technological approach,” said Xiaomeng LU, director of the EURAASIA Group Geotechnology Practice.
“But Deepseek proves that we are still in the initial stage of the development of AI, and the path that has established an Open and may not be the only way to high capable of AI.”
How is it different from the opening?
Deepseek has two main systems that have collected buzzing from the AI community: V3, a large language model that releases its products and R1, its reasoning model.
Both are open code models, which means that their fundamental code is free and publicly available to other adjustment and redistribution developers.
Deepseeek models are much smaller than many other models of great languages. The V3 has a total of 671 billion parameters or variables that the model learns during training. And while the opening does not detect parameters, experts estimate that its latest model has at least a trillion.
In terms of performance, deepseek he says Its R1 model achieves performance comparable to OPEN’s O1 on the tasks of explanation, citing reference values, including Aime 2024, Codeforces, GPQA Diamond, Math-500, MMLU and Swe-Bench checked.
In the technical report, the company said its V3 model had a training cost of only $ 5.6 million – a fraction of billions of dollars that Western AI laboratories such as Openi and Anthropic spent on training and running its basic AI models. It is not yet clear how deep it costs.
If training costs are accurate, it means that the model has been developed in a fraction of the costs of rival models by the opening, anthropy, Google and others.
Daniel Newman, Executive CEO of Tech Insight Futurum Group, said these events suggest a “huge breakthrough”, although he has led to suspicion of accurate figures.
“I believe the breakthroughs of Deepseek indicate a meaningful fold of the scalar law and that they are the right need,” he said. “Having said that, there is still a lot of questions and insecurity about the whole picture of the cost, as far as Deepseek’s development is concerned.”
Meanwhile, Paul Triolio, Senior Vice President of China and Technology Policy at DGA Group Advisory Company, noted that it was difficult to withdraw a direct comparison between Deepseek’s model and the costs of the main American developers.
“5.6 million figures for Deepseek V3 was only one training session, and the company emphasized that this is not the total cost of research and development for the development of the model,” he said. “The total costs were probably significantly higher at the time, but still lower than the amount consumed by the large US AI companies.”
Deepseek was not available for comment when he was contacted by CNBC.
Comparison of Deepseeka, Openai at the price
Deepseek and Openi reveal prices to calculate their models on their websites.
Deepseek says R1 costs 55 cents per million input tokens – “token” relating to each single unit of text processed by a model – and $ 2,19 per 1 million production tokens.
For comparison, the Openi Prices for O1 shows that the company charges $ 15 at one million input tokens and $ 60 for one million production tokens. For GPT-4o Mini, a smaller, cheap model with cheap language, the company charges 15 cents on 1 million input tokens.
Skepticism of chips
Deepseek’s discovery of R1 has already led to a turbulent public debate about the truth of his claim – especially since his models were built despite export controls from the US that limits the use of advanced AI chips in China.
Deepseek claims that he had his breakthrough with ripe Nvidia clips, including the H800 and A100 chips, which are less advanced than top H100s, which cannot be performed in China.
However, in CNBC comments last weekAII AII AII Alexandr Wang Executive, he said he believed Deepseek uses forbidden chips – a claim that Deepseek denies.
Nvidia has since been released and said that the GPUs used to use Deepseek fully in accordance with exports.
The right job or not?
It seems that experts in the industry are widely agreed that what Deepseek has achieved is impressive, although some have called skepticism because of some claims of Chinese companies.
“Deepsek is legitimately impressive, but the level of hysteria is an indictment of so many,” American entrepreneur Palmer Luckey, who founded Oculus and Andaril wrote on X.
“The number of $ 5 million is false. The Chinese Hedge Fund pushes him to slow down investing in American AI startups, served his own shorts against American titans like Nvidia and hide the sanctions.”
Seeda Rejal, the Certain Commercial Director of NetMind, a startup headquarters in London, which offers access to Deepseek’s AI models through a distributed GPU network, said he had not seen the reason not to believe Deepsek.
“Even if a certain factor is excluded, he still comes so effectively,” Rejal told CNBC in a telephone conversation earlier this week. “The logic of what they explained is very reasonable.”
However, some claimed that Deepseek’s technology may not have been built of scratch.
“Deepsek makes the same mistakes that O1 makes, which is a strong indicator that the technology has been torn off,” said the investor of billionaire Vinod Khosla on X without giving more details.
This is the claim that Openi himself alluded, speaking to CNBC on Wednesday to review the reports that Deepseek may “use the output data from its models to develop its AI model, a method called” distillation “.
“We take aggressive, proactive countermeasures to protect our technology and continue working closely with the US government to protect the most capable models that are being built here,” said Openija spokesman for CNBC.
Commodity AI
However, deepseek surveillance shakes, scientists AI mainly agree that this is a positive step for the industry.
Yann Lecun, the main Ai scientist in TargetHe said Deepseek’s success is a victory for AI models of open code, not necessarily a win for China over the American Meta behind the popular AI model of an open code called Llam.
“People who see the performance of Deepseek and think,” China surpasses now in Ai. “You’re reading this wrong.
“Deepseek profited from open research and open code (eg Pytorch and Llam from Meta). They came to new ideas and built them on top of other people. Since their work was published and an open source, everyone can profit from it. It is the power of open research and open code. “
Meanwhile, Matt Calkins, the CEO of US S
“In my opinion, we will see the comfort of AI. Many companies will achieve competitive AI, and the lack of differentiation will be bad for the first movements of the big consumption,” Calkins said by email.
– CNBC -ov Katrina Bishop and Hayden Field contributed to this report
