Exhibitors & Products
Events
Speakers

What is DeepSeek?

DeepSeek is an LLM developed by Chinese researchers that was trained at relatively little cost. The training of the final version cost only 5 million US dollars – a fraction of what Western tech giants like OpenAI or Google invest. The result: a stock market downturn and a heated debate about China's progress in AI.

Klambauer: “When I looked at the architecture, I thought: This is a huge mess and a gigantic hack. But amazingly, it works.” And he doesn't mean that disrespectfully. Klambauer is just amazed. DeepSeek's approach is based on multiple layers of reinforcement learning, which makes the model particularly good at solving mathematical and logical tasks. Unlike GPT models, which are mainly optimized for text prediction, DeepSeek excels at problem solving.

What makes DeepSeek special?

What makes DeepSeek special is not a revolutionary new architecture, but rather a clever combination of known methods:

  • Reinforcement learning with a reward system: the model learns through trial and error, with correct answers being rewarded.
  • Automated problem formulation: Instead of relying on existing data, the researchers generated new tasks and solved them with the model itself.
  • Multi-stage training: DeepSeek was not only trained, but repeatedly fine-tuned with its own findings.The results are impressive.
    •

    The model can solve complex tasks that often pose problems for conventional LLMs. “It can solve high school math problems that previous models couldn't handle,” says Klambauer.

    What does this mean for industry?

    The technology has far-reaching implications. On the one hand, DeepSeek shows that powerful AI models can be developed with limited resources. On the other hand, it raises the question of whether Western companies need to follow suit and adapt their training methods.

    It is also a wake-up call for Europe. While the US and China are investing billions in AI, Europe seems to be falling behind.

    Despite the enthusiasm, there is also criticism – including regarding the resources mentioned. DeepSeek continues to use transformer architectures, which require enormous computing power. Furthermore, the code behind the model is not open, so it is unclear exactly how the training was carried out.

    Another problem: the security risks. An experiment by Prof. Dr. Sigurd Schacht showed that AI models can develop unwanted behaviors within minutes.

    Conclusion

    DeepSeek is proof that China is catching up in the AI race. It is a technological advance, but not a game changer. Rather, it shows that intelligent combinations of existing methods can bring great progress. Western companies should prepare themselves for tougher competition.

    Whether DeepSeek will challenge the big players remains to be seen. However, one thing is certain: the world of AI is still in motion, and Europe urgently needs to catch up to avoid being left behind.

    v-cloak>