Most talk about recursive self-improvement treats it like the model just keeps getting smarter. This paper draws a sharper line. True RSI is not AI getting better at tasks. It is AI getting better at getting better. The authors map a path most labs are still mixing up: L1 — the system executes improvements humans already designed L2 — it starts choosing its own improvement strategies L3 — it generates the learning experiences it needs next L4 — it adapts from live deployment, not just benchmarks L5 — it modifies the mechanisms that produce future improvements That last step is the real claim. The goal is not another human-built model that scores higher. It is an AI whose gains persist, feed the next generation, and then improve the improvement process itself. If that lands, this generation of systems really is the last one we build by hand. Worth reading if you work on agents, evals, or post-training loops: https://t.co/scFLD1h9Ow