Harvard physicists propose physics-style theory for how neural nets learn

This is a classic example of theory chasing practice. For years, practitioners have leaned on empirical scaling laws—make the model bigger, give it more data, get better performance—without a satisfying account of why extreme over‑parameterization doesn’t simply memorise the training set. By importing renormalization tools from statistical physics and showing how high‑dimensional fluctuations can actually stabilize learning, the Harvard team offers a plausible mechanism behind the magic. ([techxplore.com](https://techxplore.com/news/2026-05-simple-physics-ai.html))

For the race to AGI, a real theory of learning matters more than it might seem. Today’s frontier labs are effectively doing very expensive curve‑fitting over architectures and training recipes. A principled understanding of when and why generalization emerges could make model design less brute‑force, reduce the need for planet‑scale compute, and help us reason about failure modes in new regimes. It also provides a bridge between abstract safety work—like scaling‑law‑based risk forecasting—and the concrete behavior of actual models. If AGI is going to be more than just ‘bigger GPT‑x’, we’ll need theories like this to guide what ‘different’ should mean.