DeepSeek R1 is currently the dominant topic in the AI sector and is attracting more attention than even models such as Claude[*] and Perplexity[**]. As a competitor to OpenAI and Google Gemini, it is not only convincing in terms of results, but especially as an open source variant that can be run locally. The key question here is whether it is a game changer or just another building block in the technical evolution.
Writing an article about AI for a printed magazine feels like competing in a 100-meter race with an iron ball on your foot. Even as the sentences are being typed, they are aging in fast motion because the market is changing faster than a goldfish loses its attention. That's why the following pages focus on the recent past and a look into the not-too-distant future.
The AI market is changing faster than a goldfish loses its attention.
Earthquake from the Far East
The release of DeepSeek R1 caused an uproar. The language model can be used free of charge and performs about as well as the expensive subscription models from OpenAI. But that's not all: the Chinese start-up announced that it only needs a fraction of the resources of its competitors. There are also several variants available as open source on GitHub. The latter has already existed with numerous other models, but never at this level.
DeepSeek R1 should be seen as a frontal attack. On Google, OpenAI, Microsoft and Nvidia, i.e. some of the biggest US tech companies of our time. The fact that the Chinese AI does not include all the functions - such as image generation - of the competition is currently of secondary importance to most users.
AI substructure
DeepSeek operates one of the largest AI infrastructures in the world, which comprises around 50,000 Hopper GPUs - including 10,000 H100/H800 - and also includes orders for H20 chips. The total investment amounts to 2.54 billion dollars. Although training costs of $6 million are often quoted for the V3 model, this relates solely to GPU utilization; the actual total cost for research, architecture and infrastructure is over $500 million.
The V3 model impresses with the use of more efficient algorithms, which require four times fewer computing resources per year and therefore outperform GPT-4o. With synthetic data and reinforcement learning, the R1 model achieves comparable performance to OpenAI's o1, although benchmarks do not take into account the differences in resource consumption. A significant architectural innovation is the use of Multi-Head Latent Attention (MLA), which has led to the development of more efficient models that are already being adapted by Western labs.
A song about censorship
The two main arguments of the West against Chinese AI are 'data protection' and 'censorship'. We can disregard OpenAI's accusations that DeepSeek trained itself with ChatGPT and thus 'stole' data at this point due to numerous lawsuits and accusations against OpenAI as a result of copyright infringements.
Data protection is a major issue. From a European perspective, however, it is difficult to judge whether it is better for the data to end up in the USA or in China. And anyone who wants to process sensitive data can at least run DeepSeek locally, which is a step forward.
For professional use, censorship - due to legal requirements, ethical considerations, company reputation and alignment - plays a rather subordinate role. If you ask ChatGPT about sensitive topics in US politics, you will get answers comparable to those from DeepSeek about China or Taiwan - possibly with less propaganda.
Open source code
Open source is an important foundation of software development. Without it, we would have no Linux and therefore no Android, no alternative browsers, office or email programs. In the field of AI, it has been shown that open source solutions can deliver better results in the long term than commercial products, such as Stockfish for chess. It remains to be seen what impact DeepSeek will have. Although it is open source, the community has no influence on the training data from China. Enthusiasts could certainly train their own AIs, but in most cases this would exceed the capacities of hobby developers and small companies. However, it would be conceivable to extend the existing model with 'specialized knowledge'.
It is difficult to predict what future developments we can expect from the open source version. However, we are already seeing a direct impact on OpenAI: prices are tumbling and now users have free access to ChatGPT models, for which they previously had to pay $200 per month. DeepSeek is thus leading to a market-driven democratization. The pressure on the aforementioned US companies is enormous.
Practical benefits
Strictly speaking, the so-called AI models are not real AI. The outputs are based on statistical assumptions [***]. As you can more or less watch the programs think with the latest language models, it quickly becomes clear that they arrive at numerous false conclusions and logical errors and can only eliminate these by iteration. Nevertheless, the results are often astonishing and have permanently changed our world. This change will be unstoppable, as people and companies that use such models sensibly will generally be more successful.
The key word is 'sensibly'. There have been numerous examples of the opposite in recent years. Some magazines and entire publishing houses have ruined their reputation and thus their existence through excessive, ill-considered use of AI.
The slump in Nvidia shares seems somewhat surprising. It may be that not quite as much hardware is needed as assumed to reach the current level. However, one could also conclude that with more resources, much more magnificent language models can be created. Even if you "only" achieve the same result in less time, it is also progress - after all, we use jet airplanes instead of propellers for longer flights.
We immediately realize that DeepSeek R1 is another tool. In practical tests, it is sometimes better than ChatGPT, sometimes worse. However, since it can be run locally and therefore keeps the data at home, it seems to be more than just an alternative - with limitations. On most home computers, only smaller and therefore less powerful models can be run locally, and you can wait quite a long time for usable responses. Online, on the other hand, the data is sent to the Far East, but processing is very fast if it works despite overloading. During tests in February 2025, it could take up to thirty attempts before DeepSeek R1 responded. It may also be a deliberate contribution to slowing down the digital world.
Despite the hurdles: AI-supported solutions are already gaining importance in manufacturing - from quality control to process optimization. Open source models such as R1 could open up new possibilities here.
Negative consequences
Strictly speaking, so-called AI language models are not real artificial intelligences - their outputs are based on statistical modelsIf youwant to use the current models sensibly in operations, you must also be aware of possible negative consequences and take countermeasures if necessary. Anyone who has spent a certain amount of time with such language models knows that you cannot blindly trust the output. This is particularly critical when it comes to specialist areas in which you are not an absolute expert yourself, because you cannot recognize inaccuracies or even hallucinations. The problem often starts with the task at hand, especially as the programs still tend to reinforce or justify statements made by the user, even if they are fundamentally wrong.
You have to be very aware of the weaknesses. For example, the fact that the models do not know anything that is not digitally available. You always notice this when it comes to information that can only be found in old textbooks. The models are not or only insufficiently able to draw the right conclusions from current knowledge about past findings. This is particularly true for technologies that are underrepresented in the digital world - keyword: statistical assumptions.
One problem with the technology currently in use is the relative lack of training data, which makes it difficult to adequately train a model in niche specialist areas such as PCB or electroplating technology. This problem is also encountered in programming, for which the current models are actually praised. If the user uses a rather rare programming language for which too little training data exists, this inevitably leads to a digital catastrophe despite popular syntax (such as C). Whether ChatGPT or DeepSeek: The models start to invent functions and commands that do not exist in the respective programming language. If the code cannot be compiled, it is fine because the error is obvious. It becomes problematic when serious but not so obvious errors are hidden. For example, a crash after several hours of using a program. And then we imagine that such errors are hidden in the information on chemistry, electrical engineering or medicine. However, the core problem is not the language models themselves, but the expectations of the users.
It is almost impossible to assess possible long-term damage to people, industry and society. There will be numerous negative results, but their relevance is unclear. There are plenty of examples from history, such as the impact of calculators on the ability to do mental arithmetic or digitization on handwriting. The two most prominent dangers currently appear to be
- Poorer basic knowledge: New generations are no longer learning the necessary basics in areas such as mathematics, physics and electronics.
- Lack of traceability of results: The outputs of AI can be understood less and less, but are considered correct.
We already have the phenomenon of point 2 in games such as Go and chess. The AI indicates the best move, but cannot say why. This will also threaten us with language models. Although it tries to give reasons, at a certain level the explanation is no longer understood by humans.
Computer intelligence took much longer than chess to beat humans at Go - but it can no longer explain to humans why it is now able to do so
Step instead of jump
Even if DeepSeek R1 essentially does hardly anything better than its competitors, the development should be seen as progress. It is a new, disruptive player in the AI market that will drive the competition to new heights and democratize the industry.
However, due to the level and limitations of its capabilities, it cannot be seen as a leap forward. This could happen in future versions, possibly through the open source community. R1 is a new tool in the AI sandbox and its free usability gives hope that smaller and medium-sized companies will also create better, more customized models to master specialized tasks, provided the training data is sufficient.
References
[*] https://claude.ai/
[**] https://www.perplexity.ai/
[***] PLUS columnist Dr. Jan Kostelnik pointed out in his 'PlattenTektonik' (issue 5/2024) that in the 1990s, 'expert systems' were considered the latest craze and early representatives of artificial intelligence (without being so.)
