I'm a computational-quantum-physicist in the field of AI4Science. I exist in a superposition between the academic and industrial world, holding the chair of Artificial Intelligence and Quantum Physics at the Center for Theoretical Physics of Ecole Polytechnique and being an AI Scientist at Mistral.

I develop computational tools leveraging Machine Learning to solve problems in quantum mechanics, ranging from quantum optics to chemistry, from superconductivity to particle physics. Being a scientist, I care a lot about designing useful and efficient algorithms that can be reliably used by colleagues, and I want to understand the underlying mechanism that makes such techniques so powerful: how do neural networks learn wavefunctions? When do they fail? And can we circumvent those limitations by synergically combining even more ideas coming from the world of computer science and ML?

I also try to apply my classical-optimisation knowledge to develop new variational quantum algorithms to simulate quantum systems. On quantum systems. Ain't that funny? In general, however, Quantum Computers ain't there yet, and we need to complement them with a lot of classical compute to do anything interesting. I don't care about proving that one is better than the other. I want to find how to design the best possible algorithm, and i am convinced that in needs to combine classical and quantum computing.

Within Mistral, I am part of an AI4Science team working on exciting new applications of those research tools, while also contributing to the fight that is making LLMs better scientists. I worry about the ethical and societal implications of AI, but I am also excited about the implications that this technology can have. I hope the worldwide political class can act resonsibly in front of this opportunity and wish to support them.

In social circles I pretend I don't understand what MLIR is and hide my nerdiness, but at times I can get hardcore and read assembly. Though, I'd rather not. I'd rather write qasm and embed it in LLVMIR. But to be honest, the reality is that nowdays I will mostly ask agents to do that for me, sip a coffee, work on some other project, and carefully read their slop in order to lower the slop-tropy they generate.

I am a convinced leader of open-science and open-software. I lead the development of NetKet, a machine learning toolbox for many-body quantum physics. It's a Python package based on Jax. I used to think that Julia was nicer than Python, and contribute to plum, a multiple-dispatch Julia-like system for it, but in the age of agents who really cares about carefully crafted code anymore?