For decades, the development of artificial intelligence has been a labor-intensive process of human trial and error. Engineers spend months tweaking architectures and refining datasets, acting as the necessary scaffolding for machine learning. However, researchers at Shanghai Jiao Tong University have introduced a framework called ASI-Evolve that aims to remove the human bottleneck, allowing AI to iterate on its own design through a recursive, autonomous loop.
ASI-Evolve functions by simulating the experimental rigor of a human researcher. The system is built on two core pillars: a cognitive base that integrates accumulated human expertise and a dedicated analyzer that distills complex experimental results into reusable insights. Unlike previous automation attempts, this framework is unified, capable of simultaneously optimizing the three fundamental pillars of AI development: data inputs, model architectures, and learning algorithms.
The initial results suggest a significant shift in efficiency. In controlled benchmarks focused on optimizing attention mechanisms—the critical components that allow models to process context—ASI-Evolve achieved a performance gain of 0.97 points. Human researchers, tasked with the same optimization, managed an improvement of only 0.34. While the technology is still in its nascent stages, the project signals a future where the most effective architects of intelligence may no longer be biological.
With reporting from Olhar Digital.
Source · Olhar Digital
