RSRC (Recursive Self-Referential Compression) is a research framework I wrote arguing that brute-force AI scaling is thermodynamically and financially unsustainable, and that efficiency should be a first-class metric.
the idea
RSRC proposes twin metrics: RSRC_t scores training efficiency and RSRC_i scores inference efficiency. Together they let teams benchmark throughput trade-offs for any agentic system without runaway spend, and they feed a shared governance loop so the two never drift apart.
where to read it
The paper is published with an archival DOI on Zenodo. It pairs naturally with my CPI work on agent oversight: CPI keeps agents safe at runtime, RSRC keeps them affordable at scale.
