Artificial intelligence Why Deepseek is shaking up the AI world

Related Vendor

FAULHABER Drive Systems

The Chinese AI application Deepseek has climbed to the top of the app charts and has incidentally triggered a stock market tremor. But what sets Deepseek apart from the competition—and how does it perform in practice?

The mere fact that the app from the one-year-old startup Deepseek is mentioned in the same breath as the AI models of Open AI, Google, and Meta is remarkable. China, after sanctions imposed by US President Donald Trump during his first term, has been cut off from the supply of high-performance chips like Nvidia's H100 and has to make do with older and less powerful chips.

According to its own statements, Deepseek has spent just under six million US dollars to train its AI language model. In contrast, OpenAI has invested over 100 million in Chat-GPT. Meanwhile, Microsoft and OpenAI express the suspicion that the Chinese have unlawfully tapped into US competitors. However, potential data theft is not the only reason why Deepseek achieves its goal with less effort than established AI applications.

In the same league as Chat-GPT

The fact is: Deepseek can answer complex questions and solve complicated problems. And apparently, it does so just as well as the US market leader OpenAI with Chat-GPT or Google with its AI system Gemini. The Chinese startup not only targets AI chatbots that generate texts or create program code but also competes with AI image generators like Dall-E and Stable Diffusion.

Among the innovations implemented by Deepseek in its AI model R1 is the concept of breaking down tasks into individual reasoning steps. Deepseek's algorithm works with several small AI systems that are activated only when needed. Users also notice this because Deepseek takes more time for responses with this multi-step approach compared to Chat-GPT or Google Gemini.

Transparency through open source—but also censorship

Deepseek relies on an open-source license for its system and has made the code for its AI models publicly accessible to all interested parties on the programming platform GitHub. In contrast, major US players like OpenAI, Google, Microsoft, or Anthropic treat their AI code as closely guarded trade secrets. Among the big US tech companies, only Facebook's parent company Meta has released its AI model Llama as open source. The open-source approach is intended to promote transparency and flexibility.

Developers are free to use, modify, and integrate the model into their applications. However, there are significant transparency deficits with Deepseek as the startup is subject to censorship in China. As a result, information not approved by the Chinese government, such as details about the Tiananmen Square massacre in 1989, is suppressed by the chatbot.

Incompatible with European data protection

It is also problematic that with Deepseek, the data is stored on servers in China. The European General Data Protection Regulation (GDPR) only allows data exchange with countries whose data protection is at the level of the European Union. For China, there is a lack of an agreement with the EU to enable data exchange on a legally secure basis.

There are also differences in the business models. With Deepseek, all offerings for end users are currently free of charge. Figuratively speaking, Deepseek entices with "free" as in "free beer," not as in "free speech."

In contrast, OpenAI and Google's most powerful AI models can only be used with a paid subscription, not for free. For instance, OpenAI currently charges $20 per month for Chat-GPT Plus, offering faster response times and prioritized access even during peak times. The Chat-GPT Pro subscription costs $200 per month, providing users with access to the most powerful models and allowing them to generate longer videos with the AI Sora.