When AI Research Starts Moving Faster Than Human Research - Zhengyao Jiang

Recursive self-improvement in artificial intelligence involves autonomous systems capable of enhancing their own research efficiency and codebases. Zhengyao Jiang, CEO of WeCoAI, details the development of AIDE, a research harness that utilizes a recursive loop to optimize its performance. By separating tasks into inner and outer loops, the system performs hill-climbing toward better research methodologies while mitigating reward hacking through multi-layered evaluation protocols. While current implementations demonstrate net-positive improvements—outperforming human-led R&D—a technical intelligence explosion remains speculative. Autonomous agents generate complex code that generalizes well across diverse benchmarks like WeatherBench2, yet human collaboration remains necessary for defining initial abstractions, creative primitives, and robust constraints. These advancements represent a shift toward automated meta-learning, where systems iteratively refine their own search policies to overcome diminishing returns in scientific discovery and software engineering.