Mistral AI's CEO on Microsoft and Europe’s AI Ecosystem

🌈 Abstract

The article discusses the rapid rise of Mistral AI, a Paris-based AI company that has become a prominent European AI champion. It covers Mistral's success in attracting top AI talent, its funding and business model, its approach to open-sourcing AI models, and its stance on the EU's AI regulations.

🙋 Q&A

[01] Attracting AI Talent

1. How has Mistral managed to draw so many talented researchers away from well-resourced companies like Meta?

• Mistral initially hired their friends who were interested in the company's mission and contributions to the field.
• Later, Mistral started hiring people they knew less, attracted by the company's strategy to push the field in a more open direction.
• Mistral estimates they have hired around 15 people who knew how to train the types of AI models Mistral develops, which is around 10% of the people with that expertise at the time.

2. Does Mistral believe it has hired a significant proportion or even all of the people who know how to develop the types of AI systems it creates?

• No, Mistral has not hired all of the people with this expertise. There are still some at companies like Google, OpenAI, and Meta.

[02] Funding and Business Model

1. What is Mistral spending its $645 million in funding on?

• Mistral is spending the majority of the funding on compute resources, as the upfront investment needed to develop frontier AI models is quite significant.
• Mistral believes it can be more capital-efficient than its competitors, spending a fraction of what they are spending on compute.

2. Does Mistral expect to spend $100 billion on compute like some of its competitors?

• No, Mistral expects to spend much less than $100 billion on compute. In its first 12 months, Mistral spent a little over €25 million, and believes it can continue to be more capital-efficient than its competitors.

3. What is Mistral's business model?

• Mistral's business model is to build frontier AI models and provide a developer platform that allows customers to customize and deploy the models, including on-premises rather than just on public cloud services.
• Mistral aims to offer models that are as capable as its competitors' but at a lower cost to Mistral and its customers, while also being more openly available.

[03] Open-Sourcing and Regulation

1. Why did Mistral move some of its most capable models behind an API, after initially making all of its models open-source?

• Mistral always intended to have leading models available in open-source, but also offer premium features through monetized services.
• The open-source models enable developers to adopt the technology, and the paid platform provides additional value in terms of performance, efficiency, and manageability.

2. How does Mistral view the EU's AI Act and the potential impact on the company?

• Mistral believes the constraints in the AI Act are already things the company is doing, such as documenting model usage and evaluation.
• There are discussions to be had around transparency of training data, as Mistral wants to enable this but also needs to protect its intellectual property.
• Mistral expects to provide input on the technical specifications and is confident they can find an acceptable solution for all parties.

3. Would Mistral ever decide not to open-source a model due to concerns about its capabilities?

• Mistral does not foresee a scenario where they would choose not to open-source a model in the foreseeable future.
• Mistral believes open-sourcing is the safest way to govern software and technology, as it has been for cybersecurity and operating systems.
• Mistral sees no risk in open-sourcing models, only benefits, as the technology itself is neutral and the focus should be on regulating the applications built with it.