Researchers at Bar-Ilan University in Israel have published work arguing that the capabilities of artificial neural networks arise less from their sheer size than from the way their internal components organize themselves during training, with individual nodes taking on specialized roles that combine into abilities no single part possesses.
The study, led by physicist Ido Kanter, examined how the units inside a network change as it learns. According to the university’s announcement, the researchers found that nodes gradually become specialists in particular features or sub-tasks, and that the interaction among these specialists is what produces the network’s overall performance. The team draws a comparison to the brain, where neurons and regions handle distinct functions and complex thought emerges from their coordination.
“More is different” in AI
Kanter frames the result as a contrast with a principle from statistical physics. In many physical systems, adding more identical components does not create new information; the behavior of a large collection can be predicted from the behavior of a small one, a pattern the researchers summarize as “more is the same.” In neural networks, they argue, the opposite holds: adding components leads to new specializations and new collective behavior, so that “more is different.”
The phrase deliberately echoes a well-known 1972 essay by physicist Philip Anderson on emergence in complex systems. The Bar-Ilan group is applying that idea to machine learning, using tools from physics to analyze what happens inside trained networks rather than treating them as black boxes.
Implications for how models are built
If the findings hold more broadly, they suggest that progress in AI need not depend only on building ever-larger models. Better organization of a network’s internal structure, encouraging useful specialization among its parts, could yield capability gains at lower cost in computing power and energy. The researchers point to the human brain, which performs extraordinary feats on roughly the power of a light bulb, as evidence that efficient organization can substitute for scale.
The work is a theoretical and experimental contribution from a physics group, and its conclusions are drawn from the network architectures the team studied rather than from the largest commercial language models. Whether the same dynamics govern systems with hundreds of billions of parameters remains an open question, but the research adds to a growing body of evidence that understanding the internal structure of AI systems is as important as expanding them.
