AI for Science and the Thermodynamics of Generative AI - with Max Welling (UvA, CuspAI)

Max Welling of the University of Amsterdam and CuspAI discusses AI for science and the physics underlying generative models. He argues that the next advances will come from physical AI and science itself, and describes how machine learning works with real laboratories, including materials discovery, balancing simulators against experiments as an exploration-exploitation trade-off, and active learning for experimental design. The conversation also asks whether priors such as geometry and symmetry should be built in or learned, and whether transformers will remain dominant.