Chinese researchers chart 5-stage path toward ‘last AI built by humans’

Chinese researchers collaborating with tech giants have mapped out a five-stage path toward the development of AI systems that can autonomously improve upon themselves without human involvement. Published in a joint paper titled "The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement," the study outlines the roadmap, emphasizing a shift towards automation in training, evaluating, and fine-tuning AI models.

Initially, the AI system follows human-designed improvement procedures. In the subsequent stages, the AI begins selecting its own upgrades and adapting to new information and experiences post-deployment. The ultimate stage involves the AI persistently refining the improvement methods itself.

The authors argue that automating certain aspects of AI research could provide a competitive edge for developers, potentially reducing development cycles and lowering costs. The research comes as the US-China technological race intensifies, with US firms currently holding an advantage in compute resources. However, Chinese institutions are leveraging scarce hardware more effectively to achieve rapid practical advances.

Despite the challenges, Chinese firms are investing heavily in autonomous training infrastructure, with companies like Zhipu AI allocating a significant portion of their recent funding to develop self-training systems. The researchers stress the importance of safety measures, including verified testing environments, to ensure that the autonomous AI improvements are beneficial and safe before deployment. They do not provide a timeline for when the fully autonomous AI system could become a reality.