Louis Béthune

I learned programming when I was 14, through the (today defunct!) Site du Zéro, initially out of love for video games, but quickly out of love for the activity itself. I got into competitive programming with the French associations Prologin and France-IOI, and served as publication coordinator for the open-science and writing blog Zeste de Savoir. I also implemented my first games along the way, and was quickly drawn to AI development, first through classical methods like Minimax and Monte Carlo Tree Search (two years before AlphaGo!), then toward Bayesian networks and finally machine learning.

I decided to go to Classes Prépa to acquire the mathematical knowledge I was lacking, and needed for deep learning. I joined ENS de Lyon in 2016 in Fundamental Computer Science, with the intent of becoming a researcher in the field of AI. I was lucky to do multiple research internships in various places, including Google Brain, especially around reinforcement learning, which I saw as the ultimate path to intelligence (at the time).

In 2019, driven by curiosity, I decided to complete a Master’s in Complex Systems, with a focus on complex networks, statistical physics, and machine learning, but also biology and philosophy. At the time, I worked on graph neural networks, as I thought they could become good candidates to store and represent information.

Disappointed by my failed attempts with RL, I thought we needed to fix representation learning first. This led me to pursue a PhD at Université Toulouse III / IRIT, supervised by Mathieu Serrurier, on deep learning (and representation learning) under Lipschitz constraints. By controlling the Lipschitz constant, I was hoping we could stabilize training and make deep learning less of a “black box”. Lipschitz networks never lived up to their promise, but they found great applications in robustness, optimal transport, privacy, explainability, and fairness. Serendipitously, I discovered they were also promising tools for computer graphics.

Today, I am a research engineer at Apple MLR (Machine Learning Research), the research group led by Samy Bengio. I work on large-scale AI research systems and model development: scaling laws, distributed training infrastructure, multimodal modeling, evaluations, pre-training and post-training of foundation models. In particular, I co-lead a multimodal masked-diffusion effort, training 3B tri-modal (text, image, audio) models at a 6T-token scale.