Self-improving AI the key step towards AGI but safety is key: Z.ai founder

Chinese computer scientist Tang Jie, co-founder of AI model company Z.ai, stated at a Shanghai event that developing more sophisticated AI systems, through recursive self-improvement (RSI), could accelerate the path to artificial general intelligence (AGI). RSI refers to AI systems that not only become more powerful but also improve their own capabilities in an escalating feedback loop.

Tang is a professor at Tsinghua University, where he also co-founded Z.ai. Recent US AI startups OpenAI and Anthropic have reported early signs of RSI, with Anthropic noting a 26% increase in AI-related research and development in August compared to March. While Tang did not disclose the extent of R&D at Z.ai currently being done by AI systems, he emphasized that RSI represents a new paradigm in research and development, potentially surpassing previous advancements.

Earlier this year, Tang identified RSI as the next critical phase for AI development. As AI models transform from basic chatbots to complex AI agents, Z.ai's latest GLM-5.3 model optimized its inference infrastructure for Chinese chips at an unprecedented scale. Tang noted that as AI systems take on more complex tasks, turning to AI for assistance becomes more logical, given the engineering challenges and financial constraints faced by human researchers.

However, concerns about potentially unmanageable acceleration of RSI have been raised. Tang urged researchers to remain cautious, advocating for a shared responsibility in ensuring AI progresses safely, not just for one nation, but for the entire world. Z.ai, also known as Zhipu AI, was spun off from Tang's Tsinghua University laboratory in 2019, with numerous students becoming influential figures in China's AI industry.