Huatai Securities: Recursive Self-Improvement Enters Real-World R&D; AI Self-Improvement Is Now Measurable
Summaries are public. PDF access requires an active membership; all members have the same access. Sign in
Huatai Securities assesses the arrival of recursive self-improvement (RSI) in real-world model R&D. Since August 2026, leading developers including OpenAI, Anthropic and Z.AI have disclosed measurable results from automating research: Claude now accounts for 26% of AI R&D tasks, over 90% of tasks involve AI collaboration, and AI agents at OpenAI's research organization run for 3.1 times the daily working hours of human researchers. Z.AI applied RSI to its inference infrastructure, lifting end-to-end throughput of GLM-5.3-Flash roughly threefold, while MiniMax's M2.7 autonomously iterated its Research Harness, improving internal evaluations by about 30%. Autonomy nonetheless remains constrained: more than half of 4-8 hour research tasks require at least one human intervention, and scaling RSI hinges on verification reliability, long-horizon autonomy and retaining gains across model generations. Huatai argues RSI will restructure R&D investment, benefiting foundation models, compute and CPUs as well as AI safety and observability solutions. Named focal points include Z.AI and MiniMax among model developers, and Hesheng Special Material, Hygon Information Technology and China Greatwall Technology Group on the CPU side.