A modern AI system is built by learning billions of numbers, the weights of its neural networks, whose meaning we can rarely decipher.

Mathematics and physics, where the ground truth is known, offer an ideal laboratory for developing tools to read them. In this talk I’ll introduce JEPAWG, a hypernetwork that generates other neural networks, built on a joint-embedding predictive architecture (JEPA). JEPAWG turns trained networks into a new (meta-)observable of the system.

Applied to trained neural samplers for lattice field theories, it locates the phase transition and recovers the critical exponent.

Applied to parametric dynamical systems, the same method finds Hopf bifurcations and period-doubling cascades and estimates the Feigenbaum constant. The talk is based on recent work with T. Göbel, J. R. Ebelt, Z. Mensch and M. Gerdes.