BytesAndBrains: A Substrate for Networked Machine Learning

Machine learning increasingly runs outside the data center, and the strategies the field has produced for that setting (federated learning, gossip learning, split learning, peer-to-peer inference) do not compose. BytesAndBrains is the substrate they can share. Each strategy ships with its own reference code, its own RPC stack, and its own training loop, and the substrate that would let them sit on a common foundation does not exist in the open ecosystem. BytesAndBrains records a workload into an ONNX-backed intermediate representation, partitions the recording across nodes, and owns the structured byte layer between them. The runtime is sans-IO: a state machine driven by the host program, with every distributed behavior testable in-process. The extension surface is composed of Components, typed plug-ins that satisfy Roles, runtime contracts the substrate dispatches against. This paper positions BytesAndBrains relative to Flower, Ray, Substra, libp2p, TensorFlow Federated, Hivemind, and Spark, and identifies the points at which the substrate occupies a different layer of the stack.