Liquid Substrates Essential for Decentralized AI Mesh Intelligence

This research delves into the fundamental requirements for "mesh intelligence," a concept where sovereign AI agents operate autonomously without any central clock, shared model, or coordinator. The challenge lies in how each agent can integrate diverse, irregularly timed observations from peers into a single, evolving internal state using a fixed-weight substrate. The paper identifies two crucial necessary conditions for such a substrate. First, because the latent state changes over time, an adaptive timescale is essential for optimal estimation; fixed-gain filters are inherently suboptimal. Second, due to the asynchronous nature of data arrivals, the optimal estimate must depend on the elapsed time between observations, a capability that networks blind to these time gaps cannot achieve, regardless of their scale. These two conditions converge in the "continuous-time liquid class" of networks. While LSTMs satisfy the first condition and fixed continuous-time filters satisfy the second, multi-timescale liquid networks are shown to satisfy both. Synthetic experiments confirm these theoretical findings, highlighting that scale alone cannot compensate for the missing temporal dependencies. This characterization provides structural conditions for building truly decentralized AI systems.