The neural network analyzed images resembling a top-down view of a square where thousands of people hold signal flags — up or down. The program compressed the picture into a few numbers and learned to reconstruct it. Without any instructions, it first detected the overall direction of the flags, and then how much neighbors agreed with each other. When two nearby people show different things, tension arises; when the same, harmony. This simple physics model is called the Ising model, and Ludwig Boltzmann derived the equilibrium law for such systems.

An unexpected twist: in complex neural networks, during rapid training, the program gets stuck at the first stage, never noticing conflicts. Such a breakdown resembles a phase transition studied by Lev Landau. Samples were generated using Roy Glauber's method.

This experiment demonstrates how artificial intelligence independently rediscovers fundamental concepts — a step towards programs capable of discovery. During learning, the network struggles with information chaos, akin to entropy in quantum thermodynamics.