Chinese-led paper maps five-stage path to self-improving AI

A 35-author paper posted to arXiv on September 10 introduces a Headroom-Closed Index to diagnose existing LLMs and lays out a five-stage roadmap toward what its authors call 'genuine' recursive self-improvement: AI systems that 'turn experience and feedback into persistent changes that improve both their capabilities and the process of future improvement.'

Titled "The Last AI Built by Humans", the paper progresses from improvement-execution autonomy, through improvement-strategy autonomy, experience-acquisition autonomy, and environment-adaptation autonomy, to recursive meta-improvement. The authors examine RSI across scientific discovery, embodied intelligence, and software engineering, noting each domain's 'distinct requirements and development speeds.'

Multiple outlets covering the release describe it as a 75-page effort from researchers at Shanghai Jiao Tong University, Tsinghua, ByteDance and the Shanghai AI Lab. The South China Morning Post reported the roadmap focuses on 'automating the labour-intensive life cycle of training, evaluating, and fine-tuning AI models.'

The abstract itself is hedged. The authors draw on 'preliminary empirical evidence' and identify 'key challenges to achieving genuine RSI' without asserting any of the five stages has been reached.