Panic about Chinese AI: “complete nonsense”

Because it is supposed to deliver similarly good or even better results for a fraction of the money as the competition from the USA, AI shares crashed on Monday. Hochreiter speaks of one “Panic reaction” And that leads back to the fact that “people have no idea”.

Just like for example ChatGPT Can answer questions, create content independently, present information and analyze data. The recently published recent version of the app had recently overtaken its big US competitors in the number of downloads and excited the industry.

Picked up on the basic model

In any case, it is a “complete nonsense” that the training of large AI models through the methods used at Deepseek and only a fraction of a fraction cost a comparable US models, says the scientist, referring to the research papers to the Chinese model. At Deepseek you have on one Existing basic model Tackled and only improved mathematics.

The costs for the development of the basic model were apparently not included. That, as rumbled by Deepseek, a basic model for scarce 6 My. Dollar (5.7 Mio. Euro) can be developed and trained, in any case, is unrealistic. In order to build such a model from scratch, several hundred million dollars are necessary: ​​”Nothing has changed,” says Hochreiter.

“Many things simplify”

The research papers of the Chinese researchers pays respect. “You have simplified a lot of things.” The way the hardware was put together, and the programming of the software can certainly help to get cheaper. The Chinese model is also smaller than the great American competitors. But even if electricity and the working hours of engineers in China are cheaper, the resource effort remains in a similar dimension.

Also Stefan Neumann assistant professor funded by the WWTF in the Machine Learning area You wien assumes that far more money was invested in the basic work in the basic work than specified. There are far more than 50 names on the research paper alone: ​​”They also have to pay them the salaries,” says Neumann.

In any case, the technology is based on ideas that existed before. Bringing them to work is a performance. One has to wait and see whether and how the work of the Chinese researchers will affect the industry, says Neumann: “In this area is Lots of hype inside. ”

„Sputnik-Moment”

There was respect for the work of the Chinese researchers of industry sizes. The model is impressive Especially if you consider what they can offer for the price, said Sam Altman Head of the Chatgpt developer Openai.

The risk capital provider Marcs Andreats A “sputnik moment” for the industry spoke under alluding to the first satellite, which was shot by the Soviet Union in the 1950s in the 1950s. The Sputnik satellite had led to a thrust in US space research at the time.

The excitement in the industry was largely again on Tuesday. Deepseek’s performance is great, but Europe also does not have to hide, said about Christian KleinHead of the German software provider SAPthat uses numerous AI models in its products.

Hochreiter also works with his company Nxai on a AI model. One is used by a technology that he developed by him called LStM (stands for: Long Short-Term memory). It is considered the basis for speech recognition and was used at Apple’s Siri. It should also need far fewer resources and be much faster than conventional models.

However, Hochreiter’s company does not build basic models, but focuses on Applications for industry. Also because the development of basic models is very cost -intensive and the competition is great, as the researcher says: “The market is closed.”