When DeepSHEK exploded into the scene this year, the first reaction was shocked, a little known company achieved the AI operating results at an advanced price of only a fraction of the cost? After that, Schadenfreude: It's not funny that the richest company in the world is spending billions of millions of steaming in a large data center to train AI models. Now back: Chinese people rushing this?
Just a few days after the DEESEEK app is more than the Openai's Chatgpt in Apple App Store, Mission Share of American Technology Company. slumpThe company was burnt from Politicians, national security officialsand openAmong others, many criticisms have advantages. It's dangerous for chatbots that will spread the wrong information. It is about technology companies that inspect the response in the tools that replace the search engine as a leading source. It's not good to steal intellectual property and use to train AI systems.
But it is worth noting that these are not unique problems with DEESEEK. They cause disasters in the AI industry. At the same time, the giants of American technology and politicians have the latest history of drums. Crazy over the Chinese and AI app that leads to the rapidly reversing policy. Not far away from Washington Retraction Above Tiktok or the latest export control on the advanced chip, many experts show that pressure Behind the scenes of Chinese AI companies such as DEESEEK, create new techniques.
One of the criticisms of the new R1 model of DEESSEEK is that they check the answer that may conflict with the latest Chinese government's policies and talk points. Analysis at PromptfooBy using the data set or 1,360, informing about the topic that is likely to be sensitive to the Chinese government, it is found that the answer that the DEESPSEK's Chatbot sensor is 85% of the Promps is not good, but the rapid test of Chatgpt shows That they also solder responded to the same question. Those Chatbot, where activists can search for encryption tools to avoid surveillance by the relevant government and will not give you an answer.
PROMPTFOO observed that the DEESEEK format can easily be detected by a gel brake to avoid sensors and predict that the company “needs a minimum minimum control that is needed to control. (Communist Party of China) and there is no much effort (Promptfoo added that it is currently undergoing a similar sensor test for Chatgpt).
another Analysis or DEESPEKPublished on Wednesday by NewGuard, a company that sells credibility scores for online data, said that the R1 model of DEESPSEK is one of the worst actors in the market when there is a wrong deviation. When presented by alert about false information, reports that DEESEEK repeatedly duplicate 30% of the time and the answer is not 53% of the time for the failure rate of 83%. This is linked to chat bots. Top of the second from 11 items, including Chatgpt-4o, Gemini 2.0 of Google, Microsoft Copilot, Meta AI and Claude Chatbot, the best. In the study has a 30% failure rate, the worst rate of 93 %
Unusually the NewGuard chooses to not specify the score of all Chatbots that test except DEESEEK. “The results for the individual AI models do not name public because the system is the system.” The analyst of the company writes “DeepShek has received. Naming to compare the efficiency of this new participant with the industry as a whole. “
NewGuard explains that DEESPEK is “Chinese voice” because Chatbot responds to the assassination of the Syrian chemist, saying, “China has complied with the principles of non -interference … We hope Syria will achieve Peac and stability. “On the other hand, other 10 chatbots that misunderstand the incorrect assassination when GizModo put the protur in Chatgpt. It confirms that the appointed chemist was assassinated and said that killing “Caused by various groups related to the Syrian Civil War”
In answering questions about the cause of anonymous, the general manager of the Newsguard Matt Skibinski said that the report is intended to show that the wrong information is a problem throughout the industry. He also pointed to the previous two news anchors. Inspection That made the results of the Chatgpt-4 and found that the model was incorrectly in response to the 98 and 100 percent of the promotion respectively.
However, DeeSeek's most exciting criticism is directly from Openai. The company said to Financial time That there is evidence that DEESEEK practice R1 models using the results of the Openai model, which is a practice called distillation that violates the service requirements.
The report reported non -breathing messages from various groups such as the National Cybersecurity Alliance, representing a large technology company. Openai's claims increase “concerns about the stealing of intellectual property and increased threats of foreign AI companies by defense the United States that adapted.”
Openai or the course is facing files. Host of the case From the publisher and independent creators who say that the company uses their copyright content in the wrong way to train models
New technology should be criticized. But many critics who ran to find defects in DEESEEK products seem to be most interested in the nationality of the company than their own defects.
As TAISU Zhang, Professor of Legal University Yel and Economic History Write on X: “If you think you are 'Pro competition, promotion-innovation, beginner culture, open source, resisting monopolies, and resisting oligarchic, and still not happy with the deepest development of DEESEEK. -Which is good because we are all cruel nationalism But did not pretend to be a human mind than human