Francesco Cagnetta

I am a research scientist at Simplex, working to build a scientific understanding of (artificial) intelligence. My research characterises how deep learning methods acquire and represent the latent hierarchical and compositional structure of natural data. I use toy models to isolate these phenomena and turn them into quantitative predictions about real-world systems, an approach that recently led to the first theory of neural scaling laws that applies to modern LLMs trained on real data.

Before joining Simplex, I was a Marie Skłodowska-Curie Fellow at SISSA and a postdoc in Matthieu Wyart’s lab at EPFL. I hold a PhD in statistical physics from the University of Edinburgh, and I am an affiliate researcher of the Simons Collaboration on the Physics of Learning and Neural Computation.