Recursive self-improvement (RSI) is the process by which an AI system improves its own capabilities - directly editing its architecture or training, or indirectly accelerating the research that builds its successors - so that each generation makes the next one easier to create. Long a theoretical concern of the AI safety literature, it moved into engineering reality in the mid-2020s as frontier labs began using their models to write research code, design experiments, generate training data and debug the systems that train the next models.
Theory
The idea traces to I. J. Good's 1965 "intelligence explosion" argument: a machine able to improve the design of machines like itself would trigger a runaway process, making it "the last invention that man need ever make". The concept underpins hard-takeoff scenarios in the singularity literature and motivates much of alignment research: a system improving itself faster than humans can audit it concentrates exactly the risks that oversight is meant to catch.
Analysts distinguish weak RSI (models accelerating human-led research: coding assistance, data generation, experiment triage) from strong RSI (systems autonomously redesigning themselves with humans out of the loop). The compounding argument applies to both, but the governance implications differ sharply.
In practice (2025-2026)
By 2025, lab leaders described weak RSI as operational: Sam Altman's The Gentle Singularity argued that even "larval" self-improvement compounds dramatically, and internal tooling at major labs relied on frontier models for a growing share of research engineering. In 2026, Dario Amodei cited accelerating RSI across the industry - including at Anthropic - as a primary catalyst for We Must Pace the Frontier, his call to deliberately pace frontier development.