Arthur Mensch

Arthur Mensch is a member of École Polytechnique’s class of 2011 and later graduated from Télécom Paris. He also completed the Mathematics, Vision, Learning master’s program (Mathématiques, Vision, Apprentissage, MVA) at Paris-Saclay, gradually turning his mathematical training toward machine learning research.⁠[1][2]

From 2015 to 2018, Mensch carried out doctoral research in Inria’s Parietal team, focusing on machine learning for functional brain imaging. On September 28, 2018, he defended his doctoral thesis within the Université Paris-Saclay system; the thesis, titled “Apprentissage de représentations en imagerie fonctionnelle” (Representation Learning in Functional Imaging), was supervised by Bertrand Thirion.⁠[2][3]

His doctoral work involved learning representations of brain activity from functional magnetic resonance imaging data, including extracting brain networks through matrix factorization and decoding brain activity using data from multiple studies. The thesis also explored online learning and random sampling to handle high-dimensional, large-scale imaging data.⁠[3]

After his doctorate, Mensch did postdoctoral research in applied mathematics at École Normale Supérieure (ENS Ulm) from 2018 to 2020, focusing on optimization methods and machine learning. This period bridged his earlier research on brain imaging methods and his later research on large-scale models.⁠[2]

In 2020, Mensch joined DeepMind’s Paris team as a researcher working on large language models. The research he contributed to includes RETRO, Flamingo, and Chinchilla, expanding his work from machine learning methods to large-scale language and multimodal models.⁠[2][4]

In March 2022, Mensch co-authored the paper “Training Compute-Optimal Large Language Models” with researchers including Jordan Hoffmann and Sebastian Borgeaud. The research examines how to allocate model parameters and training data under a fixed compute budget, finding that scaling up a model also requires a corresponding increase in training tokens. On that basis, the team trained the Chinchilla model, emphasizing the balance between model size and data size rather than simply adding parameters.⁠[5]

In 2023, Mensch left DeepMind and co-founded Mistral AI with Guillaume Lample and Timothée Lacroix. The company was founded in April of that year; Lample and Lacroix had previously worked on artificial intelligence research at MetaMetaMeta (Meta Platforms, Inc.) is a publicly listed American technology company, renamed from Facebook in 2021, that operates social communication, digital advertising, virtual reality, smart glasses and artificial intelligence businesses.Open the full entry. The three first met during their studies at schools including École Polytechnique and École Normale Supérieure.⁠[2][6][4]

Mensch serves as chief executive officer, responsible for the company’s strategic direction and leading the product and sales teams. Lample serves as chief scientist, and Lacroix as chief technology officer. Building on its in-house models, the team offers businesses tools for model access, customization, and deployment, while also releasing open-weight models for developers.⁠[2][6]

In January 2026, Mensch returned to École Polytechnique at the invitation of a student entrepreneurship organization. He argued that the spread of artificial intelligence tools has not eliminated the value of mathematical reasoning and training in fundamental disciplines; in fields such as physics, materials, and biotechnology, the role of models still depends on data, experiments, and simulation conditions. He also linked the persistence, repeated attempts, and acceptance of criticism in doctoral training to the process of building a company.⁠[1]

In an AxiosAxiosAxios is a US digital news media company founded in 2016 by former Politico executives Jim VandeHei, Mike Allen, and Roy Schwartz and launched in January 2017. It is known for short news items and industry newsletters in the “Smart Brevity” style, especially for its Washington political coverage, and publishes local news newsletters in cities across the United States through Axios Local. Since 2022 it has been owned by Cox Enterprises.Open the full entry interview at Davos, Mensch focused enterprise AI on concrete business outcomes, proprietary knowledge, and deployment conditions. He argued that open models let businesses customize around their internal data and choose operating environments such as on-premises, private cloud, or edge devices; general model capabilities and enterprises’ control over their data and systems need to be considered together.⁠[7]

He also noted that a prototype being able to complete most tasks does not mean a system is ready for stable production. Edge cases, human intervention, evaluation, and continuous improvement often determine usability in the final stage. He argued that the speed of enterprise AI adoption is constrained by organizational integration and staff capabilities, and that actual deployment progress cannot be inferred from model demos alone.⁠[7]

On July 3, 2026, in “Your AI, your growth,” Mensch argued that businesses should retain control over models, business records, and access permissions, reducing their dependence on existing systems when switching AI vendors. He called for combining open models, clear access rules, and continuous learning so that systems gradually improve from interactions with employees and customers, while keeping internal knowledge and operational experience within the enterprise’s own systems. This is his position on enterprise AI architecture and long-term use.⁠[8]

Mistral AI: The company Mensch co-founded and leads as chief executive officer, offering in-house models, enterprise customization tools, and AI applications. Official website: https://mistral.ai/ (opens in a new window).⁠[6][9]

École Polytechnique: His alma mater, which identifies him as an X2011 alumnus. Official website: https://www.polytechnique.edu/ (opens in a new window).⁠[1]

X: Public account: https://x.com/arthurmensch (opens in a new window), used to post company updates and his views on artificial intelligence.⁠[10]