What happened
The South China Morning Post reports that researchers from Chinese universities and technology companies published a paper describing five stages of recursive self-improvement, from executing human-designed upgrades to persistently refining the methods used to improve AI systems. The paper presents this as a research direction, not a demonstrated product.
The South China Morning Post reports that researchers from ByteDance, Tsinghua University, the Shanghai Artificial Intelligence Laboratory and other institutions published “The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement.” The paper sets out five progressive stages for recursive self-improvement, or RSI. At first, an AI would execute improvement procedures designed by human engineers. It would then select upgrade strategies, determine what information or experience it needs, adapt after deployment, and ultimately refine the methods used to improve AI itself.
The report distinguishes RSI from a chatbot correcting a single response. For an improvement to qualify as RSI, the change would need to persist beyond one task and be inherited by successor systems. The researchers argue that automating parts of training, evaluation and fine-tuning could shorten development cycles and reduce labor and computing costs, but these are the authors’ claims rather than independently demonstrated results in the supplied report.
The article also describes related efforts by Chinese companies. Z.ai reportedly plans to direct about 60 percent of proceeds from a US$5 billion fundraising round toward next-generation GLM models and a self-training system. Researchers associated with MiniMax 2.7 reportedly described memory updates and reinforcement-learning experiments, while DeepSeek reportedly developed an agentic harness for multi-step tasks, code execution and external software interaction. These examples do not establish that any company has achieved the paper’s final RSI stage.
Why it matters
Automating parts of AI research could affect how quickly and cheaply developers train future models, while shifting more control over model improvement from people to AI systems. The proposal also highlights a central safety challenge: systems that change their own improvement processes could require reliable testing and oversight before deployment. The report provides no evidence that the final stages have been achieved.
If reliable, systems that automate parts of AI research could become a competitive advantage by allowing developers to run more experiments and improve models with less direct human labor. That could influence the pace and cost of foundation-model development, although the report supplies no independent measurements of those effects.
The proposal is also consequential for AI safety. The researchers say genuine RSI would require strict safeguards and verified testing environments so that updates are shown to be safe and beneficial before deployment. The supplied report does not independently verify the roadmap, the related company claims, or the effectiveness of any proposed safeguards.
What to watch next
The paper gives no timetable for reaching genuine recursive self-improvement. There is no documented product access, pricing, or general availability. Key unknowns include whether the proposed stages can work reliably outside software engineering, how much compute they require, and whether safeguards can detect harmful or ineffective updates.
The researchers provide no timetable for achieving genuine RSI, and the South China Morning Post does not report a working system that has reached the final stage. Access to the research described is not specified, and no product pricing or general availability is documented.
Progress may differ substantially by field. The authors reportedly see software engineering as a clearer path, while robotics and scientific discovery present harder technical challenges. Compute availability is another limitation: the report quotes an expert saying US companies remain several months ahead and have greater deployment compute, while Chinese researchers continue to optimize under hardware constraints.
Future reporting should establish whether claimed self-training systems produce persistent, reproducible improvements, how updates are evaluated, and whether human approval remains required. No independent test results are provided in the supplied source.