2024-02-27 12:26:03
Mistral, the French gem of generative AI, unveils a high-performance LLM, a ChatGPT-style conversational agent and a partnership with Microsoft. This last point raises eyebrows, the start-up having served as an argument to reduce the scope of the AI Act in order to preserve the emergence of “European AI champions”.
The French start-up Mistral unveils Mistral Large, a very powerful text generation model displaying performance close to GPT-4 on various benchmarks. The start-up does not indicate the number of parameters of its model, nor whether it is a Mixture of Expert model (> see: Mistral AI unveils a new LLM trained with the “Mixture of Experts” approach ).
Comparison of GPT-4, Mistral Large (pre-trained), Claude 2, Gemini Pro 1.0, GPT 3.5 and LLaMA 2 70B on the MMLU (Measuring massive multitask language understanding) benchmark.
Mistral Large is characterized by its mastery of several languages other than English – French, German, Italian, Spanish – with an “understanding of the cultural context”, says Mistral. The model also stands out for the possibility for developers to define their moderation policies. On the integration side, the platform also allows Mistral to call external system functions, and to require outputs respecting JSON syntax.
It is also possible to interact directly with Mistral Large (or with its Small version) via a conversational agent called “Le chat”, in a manner similar to ChatGPT. “We designed it as an educational and fun application to evaluate our technology,” explains the company. The chatbot is also available in an Enterprise version with the possibility of deploying on a dedicated infrastructure and configuring its moderation system.
Partnership with Microsoft
Mistral is also announcing a multi-year partnership with Microsoft, which will allow it to leverage Azure’s AI infrastructure for the training and inference loads of its models. Microsoft will also make Mistral’s premium models available to its customers in Model-as-a-Service mode within Azure Studio and its catalog.
“We tested Mistral Large via Azure AI Studio in a use case aimed at internal efficiency. Performance was comparable to state-of-the-art models with even better latency. We look forward to further exploring this technology in our company,” says Philippe Rambach, Director of AI at Schneider Electric.
The two partners also intend to collaborate in R&D around the development of specific models for certain clients, particularly in the European public sector. Finally, Microsoft is taking a small stake in Mistral, as the start-up confided to Reuters.
“We are delighted to enter into this partnership with Microsoft. With Azure’s cutting-edge AI infrastructure, we are taking the next step in our expansion by bringing our innovative research and practical applications to new customers around the world,” comments Arthur Mensch, Managing Director of Mistral AI.
Contrasted reactions
Mistral’s agreement with Microsoft sparks mixed comments. In its newsletter, Sifted judges that the partnership demonstrates that “there is real commercial value in creating alternative AI systems to those developed in the cultural bubble that is Silicon Valley.” And the site adds: “As Europe seeks to pick its battles in an AI economy currently controlled by Big Tech, this partnership illustrates the saying ‘if you can’t beat them, join them’ “It also shows that you don’t have to have a ZIP code in the Bay Area to develop cutting-edge technology.”
But the partnership with Microsoft also raises eyebrows, because the Mistral nugget served as a standard for those who wanted to reduce the scope of the AI Act, in the name of protecting future European AI champions. Caught between its desire to protect users and its complex of being the region where we regulate more than we innovate, Europe has decided to largely exempt designers from open source language models, such as Mistral.
The pill is therefore bitter. “I’m quite perplexed. Didn’t the French government and the European Commission tell us for weeks that the chapter of the AI Act on foundational models had to be greatly reduced in its scope in order to safeguard the few “true independent European champions” what we have in this area? Without these changes, we would miss our only chance to catch up, they declared. Mistral AI would be forced to close open access to its models and cooperate with US technology companies, as it would not be able to comply with the AI Act on its own. Result? The French company benefits from a broad exemption from most of its obligations, as decided on December 6, 2023,” comments Kai Zenner, advisor to MEP Axel Voss, in a post on Linkedin.
1709041535
#Mistral #unveils #model #rivaling #GPT4 #partnership #Microsoft #people #cringe