Immo Salo, an artificial intelligence consultant, keeps current enthusiasm for Chinese Deepseek exaggerated and believes that expectations are not necessarily a reality.
Shakes the market. Although Deepsek’s technology is still based on many open issues, the company is in any case shaking the market with its pricing strategy related to its use of artificial intelligence models. Joel Maalmi
Founded in 2023 Chinese Artificial IntelligenceTartup Deepseek has become the center of global attention in recent weeks. The quick march of the company’s application can be considered surprising.
Deepseek’s application rose Apple At the App Store app store on Monday, the most downloaded free app in the United States, passing Chatgpt.
Providing artificial intelligence and training Bonairon managing director Immo Salo It points out that Deepsek is not worthwhile to draw too far -reaching conclusions from the App Store success.
“The question is that in recent days Deepseek has passed chatgpt in downloads, but that does not mean that the application is more popular than established competitors,” Salo explains.
Rapid rise. According to Immo Salo, Deepseek’s claims about low training costs and the effectiveness of the models should be critically examined, as the company has not published enough information to support its claims.
He emphasizes that in the long run, Deepseek’s charging volumes are unlikely to reach the same level as ChatgPT, Google Gemin or Microsoft Copilot.
According to Salo, the hype around Deepseek is explained by the company’s ability to develop impressive artificial intelligence models despite limited resources. This shows that small startups can also challenge established technology players.
In Chatbot Arena, a Chatbot Arena comparison of Deepseek models R1 and V3, maintained by researchers at Berkeley University in California, have already risen to the top of the world. This sets them in performance to the same level as some of the high -end artificial intelligence models.
However, Salo is reserved for the measurement results of the R1 and V3 models. He points out that many experts have made doubts that the models are optimized to succeed in certain standardized tests. This type of optimization does not necessarily indicate the actual usability of the models in rich and challenging everyday situations.
Artificial intelligence experts have been amazed that Deepseek would have managed to train his model at a much cheaper one than his Western competitors. According to the company, its V3 training costs were only $ 5.6 million, while in Western operators the corresponding costs often move in hundreds of millions of dollars.
In addition, Deepseek claims to have used NVIDIA’s older H800 chips, which are clearly weaker in performance than the newer technology used by Western competitors.
Salo points out that Deepseek has not published sufficient information to ensure the accuracy of these claims.
– At least I wouldn’t swallow that the total amount would be so low. Of course, it is very interesting for investors, Nvidia and the entire market, whether there is a really a technical solution in China that could change the paradigm of artificial intelligence development, he ponders.
Affordable pricing attracts
Although Deepsek’s technology is still based on many open issues, the company is in any case shaking the market with its pricing strategy related to its use of artificial intelligence models.
Startup entrepreneur Anthony Poo told For the Wall Street Journal, his company moved to use Deepseek’s models last September because of their competitive pricing. According to him, the models offered by Deepseek offer the same performance as the Anthropic Claude model used by the company, but their use is only about a quarter of the competitor’s cost.
– The difference in pricing is tangible, says Salo.
Although Deepseek’s models are not as technologically as advanced as their Western competitors, their affordable pricing can allow for the development of new artificial intelligence applications for a wider user base.
In addition, Deepseek has released its V3 as an open source, which allows developers to download it for free online. The model can be used on your own computer or server, so you may not need to pay any use of artificial intelligence to Deepseek. If the model works locally and does not need to connect to Deepseek’s servers, users’ data will not move to China either.
However, Salo recalls that the use of artificial intelligence over the network over the Deepseek interface raises the key issue of trust.
“There is no certainty where the data ends up ending, which can make you concern for users,” Salo points out.
In addition, the company’s models are accompanied by censorship problems as they refuse to answer many sensitive issues for the Chinese state, such as the 1989 events of the Tienmen Square.
Salo predicts that more will be heard about Deepseek in the coming weeks, but the news may not be as optimistic as before. He estimates that the actual cost of the company’s training costs may be revealed and the costs may not be as low as the company has suggested. In addition, the quality and possible restrictions of the models are likely to be more accurate.
– I think the current enthusiasm for Deepseek is oversized. I believe that the following news can question some of the company’s models and return expectations to a more realistic level, Salo estimates.