.png)
DeepSeek, ChatGPT and the future of AI
Distillation model revolution and European stumbling blocks
Everyone has an opinion about DeepSeek. Of course, so does Wouter Verlinden!
Thanks to DeepSeek's innovation, the cost of training advanced AI models has dropped dramatically. Whereas previously only wealthy tech giants such as OpenAI, Meta and Google had the resources to develop state-of-the-art AI systems, it is now becoming increasingly easy to train new models. This has the potential to cause a Cambrian explosion of AIs, resulting in a diversity of models spreading and innovating at lightning speed. Presumably, not everyone will continue to work with an OpenAI model. Diversity in the workplace, including for our AI colleagues! For humanity this is a positive development, but for the established American big tech companies it is a nightmare: the exclusive control they thought they retained is slipping through their fingers.
After the old industries like steel, shipping and automobiles, China is now surpassing the West in robots and AI as well. Fortunately, we in Europe are still good at making handbags.
The emergence of distillation models
One reason for these cost reductions is the emergence of distillation models. These are models that are trained by other models, a process reminiscent of the education system in which biological neural networks train each other. DeepSeek, a Chinese AI developer, already appears to be using this technique to train new versions of their own models. Of course, this violates the terms of service of OpenAI and Meta. But this means that an AI model can refine and improve itself without human intervention. This puts us at the point of a technological "liftoff," where AIs train AIs and development accelerates exponentially.
What a time to be alive! Rockets, AI, electric bikes, flat screens, cell phones, ... Who would have ever thought that science fiction, would become non fiction ?
Why is this not happening in the EU?
In the United States and China, new AI developments are being implemented at lightning speed, while Europe is sidelining itself. The European AI Act, intended to ensure transparency and security, imposes strict requirements on training data and the use of General AI models. Not only must training data comply with copyright laws (how do you begin that when training data includes the entire Internet and all books?), but all training data must be tracked. This makes distillation quasi-impossible in practice, because the origin and processing of training data is no longer easily traceable. AI teacher-learner training happens directly at the level of digital neurons and has no immediate human-readable equivalent. While the rest of the world experiments and innovates, Europe blocks itself with regulations that are essentially already obsolete at the time of implementation.
Lessons from history
The situation in Europe is reminiscent of Dutch patent law between 1869 and 1912. During this period, the Netherlands abolished patent law under the influence of free-market liberals. This gave companies such as Philips (b. 1891) the opportunity to develop rapidly and become an industrial heavyweight. ASML, On semiconductor, Signify and NXP are direct results of this.
A similar situation at Nordisk (and its competitor Novo, which would later merge). Drugs such as insulin were not patentable (the manufacturing process, however, was). The non-patentability of drugs in the early 20th century gave Nordisk room to produce insulin without legal obstacles and position itself as a market leader. US patent holder Eli Lili and Novo Nordisk still dominate these giant markets today, more than 120 years later.
Europe can take a similar course and remove all copyright restrictions and training requirements on AI training. Distillation makes it possible to build competitive models even at a disadvantage, but current legislation prevents these developments. It is not that U.S. and Chinese modelers are complying with European legislation.
If we want everything to remain as it is, everything must change.
Author: Wouter Verlinden
Date: 29/01/2025
