AI Reimplements Large Codebases

Preliminary MirrorCode experiments show current large models (Claude Opus 4.6) can reimplement substantial, multi‑command codebases — e.g., a ~16,000‑line Go toolkit — achieving tasks that would take an unassisted human engineer weeks. The experiments were done with execute‑only access to a program and its tests, suggesting models can infer functionality and produce working independent implementations. — If reliably replicable, this capability changes labor demand in software, raises questions about code provenance and IP, and concentrates bargaining power around compute providers and model vendors.