New proof: Decentralized mesh AI requires liquid substrate

A new paper from Hongwei Xu tackles a fundamental problem in decentralized AI: how can a mesh of independent agents—each without a shared clock, shared model, or central coordinator—collectively learn from irregular, unscheduled observations? Each agent must fold incoming projections from peers into a single internal state, online, on a substrate whose weights it cannot retrain. Xu proves that under these constraints, any fixed-weight substrate must satisfy two necessary conditions that no standard neural network can meet.

The first condition is an adaptive timescale: because the underlying latent state changes over time, the optimal estimator must be time-varying. Fixed-gain filters (like standard LSTMs with constant forgetting rates) are strictly suboptimal. The second condition is gap-awareness: since arrivals are clock-free, the estimate must depend on the elapsed time between observations—no gap-blind network can recover this dependence at any width or depth. This second condition is capacity-independent; scaling up model size cannot substitute. The solution lies in the continuous-time liquid class, where multi-timescale liquid networks satisfy both conditions simultaneously. Xu confirms these results with synthetic experiments, providing the first mathematical characterization of necessary architecture for mesh intelligence.