The Address Space Is Meaning: Embeddings as a Connection Primitive for Serverless P2P Networks

For three decades, finding the right information (or the right person) has meant querying a central index: a search engine, a recommendation model, a follow graph. Ranking happens on an operator's servers, under rules the user cannot inspect, optimized for the operator's incentives. There have historically been only two ways to decide who connects to whom in a network: a central index that ranks for everyone, or a social graph of explicit follows. This paper develops a third: semantic proximity computed on the device. We ask whether the embedding of a piece of text can serve directly as its address in a peer-to-peer network, so that two strangers whose writing is near in embedding space find each other with no central index and no follow graph. We call this semantic addressing. Its enabling condition has only just arrived: a useful text-embedding model now runs entirely on a commodity device, so the coordinate can be computed by the device that owns the data, and relevance becomes a purely local, fully reproducible cosine. We give an announce-then-pull gossip protocol that makes globally exact local ranking affordable under edge device/phone/IoT constraints by separating a cheap ranking signal (a signed embedding announcement gossiped to all) from an expensive transfer (the post body, pulled only for the bounded set a node admits), with an adversarial re-scoring step that defeats a node lying about its own embedding to be selected. We document the design evolution that led us to remove an earlier locality-sensitive-hashing bucketing layer in favor of exact local ranking, explaining precisely why coarse semantic routing fails.The architecture needs only two capabilities: an on-device embedding model and a gossip/DHT overlay, and is otherwise substrate-agnostic. Finally, because the protocol never inspects whether a node is a person or a program, we argue the same substrate is a natural fabric for the agentic era: a serverless alternative to rebuilding agent-to-agent discovery on a central registry.