In Europe the real race for artificial intelligence (AI) begins now. The first EU law to regulate the scope of this technology and its applications on the continent sets the rules of the game. The first “champion” of that race has already applied. “We created Mistral AI because we cared about two things: helping a European champion emerge and promoting an open approach to technology,” says one of its founders. From now on they are not only in terms of purposes but also at the level of capital.
Mistral AI launched in Paris last May. This weekend it closed its second round of financing, in which it raised 385 million euros. The main investors are two of the large American venture capital funds most specialized in technology startups, Andreesen Horowitz and Lightspeed Venture. Also on the list are names like Salesforce, BNP Paribas or AI hardware champion Nvidia.
The round establishes the value of Mistral AI at $2 billion, one of the highest in history for European startups in their first year of life and the absolute record for the emerging field of AI. The company is already a “unicorn” with only 22 people hired. Its three founders are around 30 years old: two have worked in Google’s AI laboratories and another in Meta’s.
Mistral specializes in the development of large natural language models. It is the same technology that ChatGPT uses, which is why France is already celebrating having achieved “its own OpenAI.” Mistral wants to allow other companies or institutions to integrate their artificial intelligence to generate their own chatbots, assistants, search engines or automatic content generators adapted to their own context, which would allow them to automate processes and reduce costs.
This potential is what has made Microsoft invest more than 10 billion dollars in OpenAI, which offered a blank check to its boss or the trigger for its battle with Google in this field. If the AI revolution manages to add one more layer to the digitalization of society through these types of advances, technology such as OpenAI and Mistral will be as instrumental in the process as Microsoft’s was for personal computers or the from Google to browse the Internet.
Mistral’s bet is that these models are open source. This is what differentiates it from Google or Microsoft (and also from OpenAI after its civil war), which put technology at the service of other companies but with restrictions. Their artificial intelligences are restricted, among other things to prevent them from using or reproducing what they have learned about racism, machismo or conspiracy theories.
Venture capital funds have applauded Mistral’s approach of offering technology without restrictions. “These models are much more efficient than most alternatives, have less restrictive filtering and demonstrate rapid improvement in the underlying architectures,” the Andreesen Horowitz fund celebrated in the statement in which it reported its investment in Mistral. “AI should be open,” he emphasizes.
Mistral has announced its new round of financing just one hour after the new European AI law was signed. “The exciting thing is not so much the amount raised as what it says: that some of the best investors in the world are now willing to invest significant amounts in a European startup, and that this money will allow the development of an open technology,” celebrated Cédric Or, former French Minister of Digital Transformation.
One of the first investments of the small startup was to hire Cédric O to defend its interests before the Elysée and remember the key factor that a more restrictive European law for AI could result in its business. “Depending on what the AI law is like, it could kill Mistral or it could allow Mistral to grow, and there is a lot of gray area in between,” the former minister warned in the media before the final negotiations.
Negotiations in which France, Germany and Italy presented themselves with a new proposal that broke the line established until then by the 27. They asked not to regulate the terrain in which Mistral or OpenAI operate and that the law focus on how it is implemented. its technology in specific products. For them, the only rules would be a “code of good practices” written jointly by companies in the sector and that there would be no fines if they broke it.
Have they achieved it? “Not entirely,” responds Kris Shrishak, a researcher at the Irish Council for Civil Liberties who has advised several governments on the negotiation. “The code of good practices is temporary and there will be sanctions in case of violation. What they achieved was that the obligations for large AI models are above all transparency. There is no risk assessment. There are no cybersecurity requirements. There are no evaluations “Even for those models with ‘systemic risk’, external evaluations are not required,” he adds.
There are another series of relaxations in the measures included in European law intended for companies that develop open source models, something that Mistral does not do at the moment, says the expert, since its current products “are far from being open source.”
Mistral will not be the only European startup that will benefit from these measures. Germany’s Aleph Alpha, founded in 2019, is also developing large natural language models. In November it closed a financing round that valued it at $500 million.
“Linguistic models are formidably powerful insofar as they mediate our relationship with the information sphere,” Cédric O wrote this Monday. “It is likely that the interfaces we use today, such as the browser in which you may be reading this message, will tomorrow be largely replaced by conversational robots, whether in the private or professional sphere.
“These models are very important for European culture,” he continued: “A linguistic model can be considered a form of representation of the world. If all the providers are American and proprietary, then the tools that mediate our relationship with the world will all be of ‘American inspiration’. That is not bad in itself, but it carries the risk of a very powerful cultural formatting. That is why it is not only essential that there are suppliers of European (and African, Asian, Latin American…) models, but also that they exist from open models that allow everyone to control and model behavior.”