On routing, and what it suggests for AI systems

Sonata42 began with a question about scale. Training and running very large AI models is expensive in money and energy. A common answer in engineering, when a system grows too large, is to break it into smaller connected parts. The 2024 Sonata42 proposal, NeuroNet Synergy, set out the hypothesis that smaller, connected AI systems, coordinated through established network protocols, could outperform large single models in efficiency and scalability without a meaningful loss of accuracy.

The discovery problem

Splitting a system into parts creates a new problem. When a query arrives, how does one system find the specialist system, possibly itself generative, that can answer it best? Retrieval augmented generation and the Model Context Protocol both point at this need. Neither solves the discovery of capability.

What routing offers

The public Internet has a working answer to a similar problem. Border Gateway Protocol lets each network advertise what it can reach and choose routes by local policy, so that no participant needs a global view. Open Shortest Path First shares link state within a single domain. Multiprotocol Label Switching separates the choice of a path from the forwarding of traffic along it. Both founders of Sonata42 have designed and operated communications networks since the 1990s.

The planned elements of a routing layer for AI systems follow the same pattern: a capability advertisement format, a route selection method, a session handshake, and a label switched forwarding plane.

How this relates to measurement integrity

The second Sonata42 research line asks a related question. When one system passes a value to another, or to a person, what must travel with it? In the measurement integrity layer every critical value carries its source and age, and a gate refuses values that were generated rather than measured or retrieved. Both lines of work treat the coordination of AI systems as a protocol problem rather than a model problem.