DeepSeek: 'Don't trust that China or the US will always use AI for the greater good'

As renowned AI experts, were you surprised by DeepSeek's success?

Hans Uszkoreit (HU): We knew that the company existed. But in all honesty, DeepSeek didn't really stand out in this crowd of Chinese model developers. It was quite a surprise that they have now turned up with such a powerful model that achieves so much with so few resources. Nobody could have predicted that.

Have you been able to examine the model thoroughly? Some critics already doubt the legitimacy and authenticity of the results.

HU: Everyone who sees the figures first doubts whether the results make sense and whether they can be verified based on the technology and methods. And that's what we did.

I just spoke with my former team in China today to verify it. The efficiency compared to ChatGPT and others is indeed impressive. DeepSeek is good not only at machine learning but also at machine teaching. It's a bit like a child at school: it's not so important that they read as much as possible, but that they read the right things.

Effective compression method in pre-training

HU: DeepSeek's so-called post-training is very systematically structured from the simple to the complicated, ensuring complex inferences and cooperative response behavior. Even if the numbers have been tweaked here and there, it is still remarkable how much innovation has been brought together and further developed - although most of it has been bolted together from things that already existed. But there are also genuine innovations, particularly a very effective compression process in pre-training, that is, learning from texts, which requires the most processing time and memory.

Does this mean that Chinese engineering has surpassed the rest of the world?

HU: Not directly. Founder Liang Wenfeng simply didn't try to find the most experienced engineers, who are hard to come by as it is, but people who have good ideas, who are young and creative. We should do the same in Germany and find such people at universities in the first place, which shouldn't be as difficult as in China, given that it is a much larger country.

FX: Over the last 20 or 30 years, we have experienced and supported many technology students in China. A great change has taken place there. You can sense a lot of creativity and energy there. If you make the cull, you will find many highly intelligent, highly motivated people. I hope that something like this will reappear in Germany and Europe and that people will work and research with passion and ambition. That's when the magic happens.

Companies should carefully examine every application

The dangers associated with the use of DeepSeek are currently being discussed widely, precisely because it is a Chinese application. What is your opinion on this?

HU: In the coming weeks and months, the model will be tested very thoroughly to see if there are any hidden traps. But things are looking good now, apart from the occasional refusal to answer questions on politically sensitive topics.

FX: You don't have to send your queries to China to use DeepSeek. You can just download the entire open source to your own device or private cloud to experiment.

HU: IBM and Fireworks, for example, already offer DeepSeek on their cloud structures. However, companies must be very careful when dealing with cloud-based applications. It is also interesting to find out whether DeepSeek's R1 model has received the certificate for political correctness. Only systems that have this certificate are allowed to be released in China. Baidu had to wait a very long time for its Ernie model to receive this approval. And despite politically cautious responses, some issues suggest that DeepSeek may not yet have this certificate, or may have been tested too quickly.

First model that strongly combines Chinese and Western culture

Were you surprised that the DeepSeek model was published as open source? After all, this is no longer a common practice in the US.

HU: It isn't common in China, either. Only very few still do that.

FX: But it's also a very clever strategy for an unknown company like DeepSeek, with an unknown team and an unknown CEO in AI, to become known quickly.

Some users report that they have achieved much better results on topics such as Chinese history or philosophy on DeepSeek than with applications such as ChatGPT. Is the AI world becoming more Chinese?

FX: This is understandable. In Common Crawl, a non-profit organization that provides large amounts of freely accessible web data for AI models, there is relatively little Chinese content.

For example, you can also find good Chinese content on Baidu's Ernie 3 model. However, its Western content is not as pronounced. DeepSeek is the first model to combine Chinese and Western culture so strongly.

Development of AI models is only just beginning

What does DeepSeek mean for the chip industry and the ambitious Stargate project just launched in the US?

HU: The timing is most interesting: Just at a time when Trump wanted to pick up the pace, he got a damper. Nevertheless, if Trump has enough good advisors at his side, he will not terminate the US Stargate project.

Should chip companies like Nvidia be worried, as reflected in their stock prices?

HU: The industry has slightly overreacted. The success of the more efficient DeepSeek in no way means that artificial intelligence is no longer chip-hungry - quite the opposite. Just because some models can now be trained with less learning effort and the inferences are cheaper does not mean that powerful chips are no longer needed. On the contrary, we are only at the beginning. We may currently be training with 14 trillion words, but we have by no means exhausted the full potential. We still lack a lot of high-quality knowledge data: Technical literature and feature films as training data or the enormous amounts of images from autonomous vehicles that can capture and understand entire cities for a world model. The really large amounts of data are yet to come. What we are seeing now are preliminary stages, although very lovely preliminary stages, of course.

Development opens up new opportunities

What does this success mean for the AI race between China and the US?

FX: Meta's Chief AI Scientist Yann LeCun says that the competition on a scientific and technological level is not primarily happening between the USA and China, but rather between open source and non-open source approaches. So it's less about countries and more about different development philosophies – and I agree. DeepSeek, as a Chinese company with external constraints, shows that even small start-ups with limited capital but highly motivated and curious talent can be successful. They do not have to be large AI hyperscalers with vast capital and infrastructure focusing exclusively on rapid commercialization. This development opens up new opportunities – and I think this is a good direction and also encouraging for Europe.

In what way?

FX: Even if it may look as if the USA or China are currently dominating large language models and AI ecosystems, a stronger focus on existing research and open-source results could also lead to even better developments for others. You can always be better than others – so the train never has left the station. Perhaps we simply need to build new tracks and develop better trains. Germany, especially, has often proven that it is capable of doing just that.

HU: Unfortunately, we lack elite funding for research and training centers. Instead, we always have to chop up our money carefully. As a result, we have 20 federally funded AI competence centers, service centers and AI labs, and each one has just enough to publish internationally recognized papers. But none of them can build the truly big systems. If funding were concentrated on one to three elite centers, too many federal states would cry out because they got nothing despite their interests and good universities.

A lot of talent in Europe that could be pooled

Do we need a German strategy or do we need a European one?

FX: DeepSeek shows once again that not everything has to be organized by the state. Instead, entrepreneurs with a strong innovation drive should play a key role. After all, there is also a lot of talent in Europe that could be pooled.

However, Europe often lacks energy. After the Second World War, many new things were created here – but today, we have to ask: Why aren't there as many new companies with digital and software-based business models being founded in Europe? This question needs to be answered, perhaps also from a cultural and socio-psychological perspective.

HU: If these core technologies, which can turn our entire living and working environment upside down – and potentially cause harm – are no longer in our hands, this raises a crucial question: Do we really trust that American or Chinese politicians will always use them for the benefit of all humanity? I have my doubts about that.

Dr. Feiyu Xu studied at Tongji University in Shanghai and at Saarland University and habilitated in big text data analytics. She co-founded the AI start-up Nyonic and was Global Head of Artificial Intelligence at SAP until June 2023, where she led the company's AI strategy.

Hans Uszkoreit is Scientific Director at the German Research Center for Artificial Intelligence. Uszkoreit has initiated German and international research networks. His research results on language and knowledge technologies are documented in over 250 international publications. He was elected a member of the European Academy of Sciences and received two Google Research Awards.

HU : A European one would be better. Fortunately, there is a glimmer of hope on the horizon. On Feb. 1, a major European project called OpenEuroLLM will launch, bringing together 20 outstanding research centers to develop open-source language models for European languages and cultures. This initiative will even have access to significant AI computing power, partly provided by the Barcelona Supercomputing Center. Several German partners are also involved. However, getting to this point took far too long once again – considerable effort was needed to convince stakeholders.