The European plan to give AI start-ups access to supercomputers first appears to face another challenge. It appears the start-ups lack the knowledge to harness the capabilities of a supercomputer.
In September, Europe launched the initiative to make supercomputers accessible for AI start-ups from the region. Start-ups should gain the means to train AI models with this initiative.
That plan now appears to have become a reality, and France’s Mistral AI was allowed to participate in an early test. The European Commission shared this information during a press conference on Tuesday. Mistral AI is seen as one of the most important AI promises from Europe.
Supplement plan with knowledge-sharing
However, the tests gave the European Commission insight into the fact that the September plan still needs refinements. “One of the things we observe is the need to provide not only access but also facilities. This is, in particular, the case for sharing the skills, knowledge and experience we have. We need to think not only about improving this access but also about developing training algorithms that make the most of the architecture and computing power currently available in each supercomputer centre and in our machines.”
The European Union plans to establish an “AI support centre” next year. This is where start-ups should find the guidance they need.
Different approach needed
According to the Commission, the lack of knowledge lies in the fact that European AI start-ups have by now become accustomed to using computer hardware from U.S. hyperscalers to train AI models. “In many cases, the AI community has tremendous knowledge about how many GPUs you can get in one box. In that area, they are very good. But what we have on supercomputers are a lot of boxes with GPUs, and it takes some extra skills and some extra help to scale and use the supercomputer to its fullest.”
Supercomputers are interesting for training purposes because of their greater computing power and parallel processing capacity. Among other benefits, this reduces training time and allows larger data sets to be processed simultaneously. The parallel processing capability indicates that the hardware is capable of performing multiple tasks simultaneously.
There is a strong desire to help AI start-ups use supercomputers, as Europe also wants to become a player in the AI market. According to many experts, that desire has been difficult to reconcile with legislation around the technology in the making. The agreement on the AI Act should have eliminated those risks by allowing AI start-ups to set up “regulatory sandboxes” and “real-world tests”. This will allow smaller companies to safely test the rules without immediately risking a large portion of their revenue.
Also read: AI Act: OpenAI and Google may not violate copyrights
21 hours ago
