The French start-up Mistral AI, founded in June by three researchers from Google and Meta, presented on Wednesday, September 27, its first generative artificial intelligence program, freely reusable and designed to compete with some American competitors despite its small size.
“It’s a first step, we are in the process of developing larger models and developing a platform to make them easy to use,” detailed Arthur Mensch, co-founder of the Paris-based company, which now has 18 employees, among them. 15 engineers. With 7 billion parameters, the Mistral 7B will be more powerful than a similar Meta model, which has twice as many, its developer claims. However, it is still a long way from GPT-3, the tool behind ChatGPT, which had 175 billion.
“It’s not an app, it’s the fundamental component that a developer will use to create their app,” Mensch warns. The model can therefore be used to complete texts, summarize documents or answer certain questions in the form of a “chat”, provided that you host the model yourself on a computing infrastructure.
Polyglot models in development
Mistral AI is one of the few European companies that has set out to chase the American giants Meta, Google and OpenAI (supported by Microsoft) in the race for generative artificial intelligence, which requires, in addition to specialized skills, significant computing power. to train AI. The young company raised nearly 100 million euros during the summer from numerous investors, including the owner of the Iliad group, Xavier Niel, who announced on Tuesday that he had purchased a supercomputer dedicated to AI.
Arthur Mensch does not reveal any details about the data corpus used for training, to protect its competitive advantage, but also because the use of this data obtained from the open Internet raises numerous questions regarding intellectual property.
We just learned that the model “speaks” primarily English and that the “cleaning” of this data does not depend on human workers clicking, as was the case with OpenAI, the Californian start-up that created ChatGPT. Polyglot models are in development and the start-up aims to generate its first revenue thanks to a platform currently in testing that will allow companies to host and improve free or proprietary models.
Source: BFM TV
