More is different in AI—More is the same in physics

One of the most influential scientific and philosophical viewpoints is “More is Different” (MiD) by Philip W. Anderson, highlighting the limitations of the reductionist approach in describing complex macroscopic systems. Decades later, the emergence of machine learning applications requires examining the relevance of MiD to artificial intelligence (AI). Here, we show that both macroscopic disordered physical systems and finite AI models undergo spontaneous symmetry breaking (SSB), a prominent MiD feature. However, from an information viewpoint, physics represents “More is the Same” (MiTS), while for finite AI models MiD. Time-dependent states of a spin in ordered physical systems reveal the equilibrium state of the entire system. Furthermore, for disordered systems like spin-glasses, time-dependent states of a spin or many spins do not contain information concerning the macroscopic equilibrium state, MiTS. Conversely, for AI models, even a single-nodal SSB contains information on the global task, which increases with the increasing number of nodes because of cooperation, representing MiD. Learning, which is beyond the physical framework, splits its task among nodes using the essential SSB mechanism, avoiding nodal replication, which prevents complex learning.