Deepseek: As good as a Ferrari, but built from spare parts-Microsoft expert explains the basis of the immense success

A sputnik moment-The ex-Microsoft employee Dave Plummer talks about the Chinese Ki-Chatbot Deepseek on his YouTube channel Dave's Garage. He explains what makes Deepseek as special for his competition like Chatgpt.

Link toYouTube content

A ferrari from spare parts

Deepseek is imas a chatt and comparable LLMS. What factor it should actually be cheaper is, howevercontroversial.

The ex-Microsoft employee Plummer speaks on his YouTube channel as well as other sourcesthat are said to have flowed in Deepseek. This is a fraction of the billion dollar investments in the models of the competition.Nevertheless, the AI ​​can keep up with flagships like Chatgpt.

In addition, the developers of the AI ​​are said to have had no access to the latest Nvidia chips. After all, their performance should be so integral for the AI ​​boom that. And during the core product for which many Nvidia know, namely.

Deepseek is according to the plummer like

A Ferrari that was built from spare parts - just as good, but much cheaper.

Like master and apprentice

This is possible through a different kind of training. Deepseek, just like the chatt, is a model on the basis of a large-length model.However, this is adistilled model(distilled model).

This means that a smaller model is trained with the help of large models in such a way that it provides the most similar results as the big ones - but with much less resources.

So it happens that the huge models still have a greater knowledge of knowledge,The smaller model performs almost as well in most applications.

Plummer compares this in his video like this:

It is as if a master trains his apprentice - the apprentice doesn't have to know everything, but he can do the work as well.

One of theseMeisterWas the open source model Llama from Meta, but also openais chat.

This knowledge distillation makes Deepseek significantly more resource-saving. For operation, it no longer needs the immense hardware with hundreds of GPUs in huge data centers such as the large models.

This also leads to the question:

If you can build a Ferrari in your garage from Chevy parts, what does that mean for the value of a ferraris?

Memory of the PC revolution

Of course, this means nothing good for the expensive original ferrari.

However, it is an advantage for users that theyThe model can also run locally on your home hardware. Of course, Deepseek cannot run locally on any small work notebook. Plummer needs for the largest Deepseek modelan AMD thread ripper with an NVIDIA RTX 6000 GPU (48 GB VRAM). Smaller variants even ran on oneMacBook Pro.

He feels reminded of the time of the PC revolution of this development.

It reminds me of the early days of the PCs - they weren't as good as mainframes, but they changed the world.

Than major computing systems under computer (Mainframes) Nobody could imagine that everyone would have something at home at home at some point.

A sputnik moment

In contrast to the PC revolution, however, the geopolitical implications must also be observed today. Deepseek is a Chinese model that particularly competes with the US Silicon Valley Groups.

Plummer describes this asSputnik-Moment. This is an allusion to the Soviet satellite Sputnik, whose start in 1957 the beginning of the space competition (Space Race) andA new phase of systemic competition between the Soviet Union and the West in the Cold War marked.

Similarly, the development of Deepseek leads to geopolitical tensions that reflect the competition between the USA as a technological hegemon and China as an emerging world power.

This technological competition is not only a struggle for innovations, but also a symbol of the systemic rivalry between the capitalist democracy of the United States and China's capitalism.