Mistral AI, the French hope in the race for generative artificial intelligence, released its first large language model (LLM) this Wednesday. In June, the founding team composed of Arthur Mensch (ex-DeepMind), Guillaume Lample and Timothée Lacroix (both ex-Meta) raised a record amount of 105 million euros, after barely a month of existence. Since then, the government has not missed an opportunity to promote the startup, starting with the Minister for Digital Transition Jean-Noël Barrot.
Mistral AI advances on Meta’s flowerbeds
For its premiere, Mistral AI is aimed solely at AI specialists, and does not hide it. Moreover, his first communication is none other than a lien « torrent » (a download link) to his model, a nod to the developer community. And it’s no coincidence: the startup is flirting with the open source world and its thousands of passionate developers, ready to contribute voluntarily to interesting projects. In exchange, the young growth authorizes the free use of its language model in almost any scenario. In other words, the business world and paid offers will wait. For now, Mistral AI wants to unite around its offer and create a real community.
« A community-supported model is the safest path to avoid introducing censorship and bias into the technology that will shape our future », defends Mistral AI in his press release. The startup aims to create an alternative to “ the emerging AI oligopoly », without expressly naming the OpenAI-Microsoft, Anthropic-Amazon or even Google duos. Better, she projects herself into “ head of the line » open models, a place currently occupied by Meta, which opened its Llama-2 model in July.
On his account (ex-Twitter), Arthur Mensch also specifies that the startup’s first model, composed of 7 billion parameters, exceeds in all respects the performance of the 13 billion parameter version of Llama. Implied, the French company would be able to do better with less than Meta, but this still needs to be demonstrated on a larger scale.
OpenAI has no reason to tremble, for now
While the number of LLM parameters increases with each new iteration, Mistral AI is launching with a very small model by current standards. But this small size nourishes Arthur Mensch’s discourse, which he develops in his interview with Le Figaro : the startup wants to multiply models of different sizes, to provide the best price/performance ratio for each scenario. It thus promises to build a generative AI offer focused on the search for savings, in a sector that is currently extremely resource-intensive.
Conversely, if OpenAI or Google also multiply models of different sizes, they above all focus their communication around their largest models, a sort of flagship. OpenAI’s GPT-4 thus serves as a benchmark in the industry, and a calling card to defend the company’s superiority over the competition. Google is expected to release an even more impressive model by the end of the year, called Gemini, while Meta has promised that Llama will expect GPT-4 performance from next year.
Before worrying the AI ogres, the road will be long for Mistral AI. And to achieve this, it may be necessary to go through a demonstration of force: Figaro, the manager explains that other models will be made available in the coming months, some of which “ much larger and more efficient » on difficult tasks, but “ more expensive to use “. The startup must act quickly: while it is building its community and producing its first models, the American giants are already at the next stage, where they are commercializing generative AI using new tools. It’s up to Mistral AI not to arrive too late.
#dethrone #giants #Mistral #dredging #world #open #source