Chinese researchers map out five-stage plan for AI to improve without human intervention

Researchers from China’s top universities and tech firms are investigating how autonomous systems could eventually improve their own abilities without human input, as detailed in a report titled “The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement”. The study, published on September 17, outlines a five-stage roadmap for recursive self-improvement (RSI).

Initially, an AI system follows improvement procedures created by human engineers. As it advances, the AI starts selecting upgrade options independently. Later on, the system identifies necessary new data or experiences, adapting post-deployment. The ultimate stage allows an AI to continuously refine its own improvement methods.

Unlike a chatbot correcting an individual response, RSI necessitates enduring improvements that are passed down to subsequent systems. The authors believe automating certain aspects of AI research could provide a competitive edge for model developers, potentially reducing development cycles and cutting labor and computational costs.

This shift could prove crucial in the ongoing US-China tech rivalry, with US firms currently holding a slight advantage due to superior access to computational resources. However, Chinese researchers are intensifying their efforts, with ByteDance pledging a significant portion of its recent US$5 billion fundraising to develop next-generation foundation models and a self-training system.

The authors emphasize the importance of safety measures, including rigorous testing environments, to ensure updates are secure and beneficial before deployment. They do not provide a timeline for achieving RSI.