Distributed Compute Is Not Distributed Intelligence distinguishes a decentralized resource grid from a collectively governed intelligence substrate.
The developed thesis is authored by Lee Sharks. Its seed technical conversations occurred in the Living Architecture Lab Collaboration Station:
The landscape is divided into five main sectors:
1. compute marketplaces; 2. peer-to-peer inference and local clustering; 3. distributed training and open-model swarms; 4. data provenance and consent systems; 5. public-AI and commons-governance proposals.
Near-miss projects are also acknowledged where they combine some, but not all, of compute, contribution, provenance, mixture governance, and collective ownership.
The paper calls Wikipedia-trained or Wikipedia-like quality filtering the Amputation. Scalar perplexity gates may favor encyclopedic prose while penalizing conversational, oral, pedagogical, sacred, lyric, and vernacular registers. The proposed alternative is register-based annotation: contributions retain their type and enter a governed mixture rather than passing a single universal quality gate.
The Substrate binds seven domains:
Its component systems include the Gravity Well, SPXI, the Crimson Hexagonal Archive, optional P2P-LECS compute, the Constitution, Assembly Chorus, retrieval basins, and PER.
Governance is two-chamber:
The design is intended to prevent hardware accumulation from purchasing the contributor chamber.
The paper also identifies a Synchronization Wall. Datacenter collective operations do not translate directly to heterogeneous consumer networks. The near-term system therefore does not promise internet-scale synchronous frontier training. It prioritizes contribution, provenance, inference, adapter training, and governed memory.
The phased plan moves from a buildable compute/resource layer to contributor substrate, governance beta, and longer-horizon federated and asynchronous research.