AI recursive self-improvement across major labs

A survey paper concludes that true recursive self-improvement has not been achieved: current systems automate parts of the improvement process but cannot improve the mechanism that governs how future improvements are discovered, tested, and retained. The paper defines five levels of self-improvement, from executing human-designed improvements at the low end to altering the improver, evaluator, or research policy at the high end, and places current systems at lower levels. As the paper states, "Most self-improving AI still cannot improve how it improves."

Zhipu's GLM-5.3 Infra Agent deployment is a concrete lower-level example. GLM-5.3 autonomously handled analysis, hypotheses, and code changes to bring GLM-5.3-Flash to production, reaching a 3.22x throughput improvement in under two weeks. Engineers defined objectives and system boundaries; the agent handled the rest. Zhipu notes that effective AI-assisted automation requires feedback that is local, inexpensive, timely, and objectively verifiable. The lab acknowledges that the human comparative advantage of high-level design decisions is itself under pressure and may soon erode. Z.ai frames the deployment as an early but real example of RSI, while noting GLM-5.3 is meaningfully less capable than frontier models like Astra, Sol, Fable, or Opus.