Brains and artificial neural networks are often modeled as collections of static nonlinearities, yet biological neurons operate in a fundamentally dynamical world shaped by instability, controlling which relies on continual prediction. In this colloquium, I present a physics-motivated framework in which neurons act as detectors and controllers of unstable directions in high-dimensional dynamics. Starting from local linearization near saddle points, I show that short-time past–future correlations identify modes whose variance grows most rapidly, providing a data-driven notion of predictive structure. Projecting activity onto these unstable modes maximizes predictive information, connecting neural computation to information-theoretic principles while extending them to time-irreversible regimes.
This perspective leads to a new computational primitive — the Rectified Spectral Unit (ReSU) — that replaces static nonlinearities with dynamical operators learned from data. ReSUs naturally unify prediction and control: when the objective shifts from stabilization to attract–repel or state-transfer goals, the same mechanism yields motor-like control policies. I will illustrate how these ideas link connectomics, population recordings, and theoretical neuroscience, and discuss implications for biologically grounded AI and backpropagation-free learning. More broadly, the work suggests a shift in viewpoint: intelligence may emerge not from stable representations, but from the selective amplification and regulation of instabilities in complex dynamical systems.