Unlike ordinary model development, where humans design, train, and evaluate at every step, recursive self-improvement (RSI) has the AI system participate directly in improving its successor — rewriting code, generating training curricula, or directing research decisions. Weco AI's 2026 four-level ladder gives the concept a measurable framework: Level 0 (delegation, still slower than human R&D), Level 1 (net positive, beats human-only development under a fair baseline), Level 2 (ignition, the improved system becomes a better improver than its predecessor), and Level 3 (inflection, the rate of improvement itself accelerates). RSI is the mechanism researchers point to when explaining how an intelligence explosion could occur; bounded, narrow demonstrations of the lower rungs (Weco's AIDE², NeoHorse-1) exist as of September 2026, with no public claim of reaching ignition.