In fewer than 18 months, Mistral AI has rapidly emerged as a breakout in the competitive AI landscape, challenging the dominance of tech giants with its innovative approach to language model development. Founded in 2023, the team has garnered attention for its efficient, open-source AI models that promise to democratize access to advanced AI technologies. Mistral AI not only represents a technological powerhouse in AI globally, but also embodies Europe's ambitions to carve out its own path in the evolving tech landscape.
With its commitment to responsible AI development, and the potential to reshape industries across the board. Listen in as Headline Partner, Jonathan Userovici, talks to Mistral's Lead Scientist and Founding Team member, Devendra Chaplot about their pace of development, how they attract talent, and how the team is finding new applications for AI.
Mistral AI has raised billions in funding over a short period. As a company of about 100 people, where do you invest the most to differentiate yourselves?
Despite the significant funding we've received—$1 billion in less than 18 months—we remain a small and agile team. Our key differentiator is focus. We've managed to gather an incredibly talented team dedicated to building the best models in the world. This focus allows us to compete effectively, even with our relatively small size and funding.
How does Mistral attract top talent, given the high salaries offered in the AI industry?
While we can't always match the eight-digit salaries offered by some companies–which is shockingly, not uncommon–we attract talent through other means. Many researchers in our field have enough liquidity that short-term financial gains aren't their primary motivation. We offer the opportunity to work on ambitious projects with more ownership and room for impact than some of the larger tech companies. The prospect of developing models that could potentially be used by millions or billions of people is a significant draw for top talent.
What makes Mistral stand out from other high-performing AI models?
Mistral stands apart because of our capital efficiency. We focus on using our capital for compute, which remains our biggest expense. We also prioritize new model innovation. Our success comes from aligning incentives within the company. Everyone is working towards improving the model with the least resources possible. We place a strong emphasis on data quality, ensuring it's clean and free of noise, which translates to better data efficiency and, consequently, better compute efficiency.
How does Mistral balance research with real-life applications?
In today's AI landscape, balancing research and applications isn't as challenging as it might seem. Improvements in the model often directly translate to business value. We manage this balance through a separation of responsibilities: our research team focuses on improving the model and adding capabilities, while our engineering and business teams concentrate on new applications.
What are some of the most interesting applications of Mistral's technology?
One of the most interesting applications is customization for enterprises. For example, Harvey, a legal startup, uses Mistral models to create specialized models for lawyers. We're also seeing significant interest in on-premise usage, particularly in sectors like banking, finance, and defense, where data privacy and security are crucial. These industries can deploy Mistral on their own servers, maintaining full control over data security. The models are used for various tasks such as summarization, question answering, and data processing.
What do you see as the next big leap in AI?
We're seeing a lot of progress in agentic applications, particularly in coding agents. The next big leap is likely to be in multi-modal agents—AI that can use computers like humans, booking flights and hotels, for instance. This development will greatly expand what enterprises can automate. We expect to see significant advancements in this area within the next 1-2 years.
How does Mistral address issues like drift, hallucinations, and toxicity in AI?
We've observed that these problems tend to improve automatically as we enhance the model's overall capabilities. While you can address these issues individually, we find it more effective to focus on improving the model as a whole. As the model becomes better, we see corresponding improvements in bias, hallucinations, and other problematic areas.
What factors could potentially slow down the current rapid pace of AI innovation?
Data is more likely to be a limiting factor than compute in the short-term. We're already approaching the limits of available text data, which is why there's increasing interest in synthetic data and multimodal data (like images and video). It's possible that we may face a data bottleneck in the future.
As an investor, we look for opportunities in AI. What framework would you suggest for identifying investment potential in AI?
I'd advise against investing in areas that will improve automatically as model performance increases. Instead, focus on vertical applications and domain-specific investments in areas like legal, finance, and healthcare. These sectors require specialized data and expertise to apply AI models effectively. Investments in improving a model's fundamental reasoning capabilities are less likely to be fruitful if they apply to foundational models, as these improvements often happen organically.
What do you see as critical for enterprise deployment of LLMs?
We're likely to see a proliferation of AI applications, particularly in knowledge work sectors like legal, finance, and healthcare. In the next few years, we expect to see more agentic applications where AI is used not just as an assistant but for completing tasks autonomously. For example, in the legal field, AI might draft entire documents or handle bookings in other industries.
Do you foresee a future where AI can assign its own tasks?
While AI will likely become more adept at creating and managing its own sub-tasks, I hope and expect that the highest-level tasks will always be assigned by humans. The level of abstraction will continue to increase, with AI handling more complex subtasks, but the overarching direction should remain human-guided.
Why should investors consider Mistral now, given the high valuations of companies like OpenAI and Anthropic?
Mistral offers something unique: openness and transparency, which allows us to license our model for on-device and on-premise deployment. We have partnerships with companies like Qualcomm, something that OpenAI and Anthropic can't replicate due to their closed-source approach. Our competitive advantage lies in our hard work and data cleaning processes, not in keeping our model architecture secret. This openness, transparency, and model portability, sets us apart in the market.
For more information, visit Mistral AI.