Chinese AI Researchers Target Recursive Self-Improvement
Researchers from prominent Chinese universities and Big Tech are concentrating on a new objective in the artificial intelligence competition against the United States. This involves creating systems capable of generating enhanced versions of themselves without human intervention. A recent paper describes a five stage framework for recursive self improvement, termed RSI.
According to SCMP, a joint paper published by researchers from ByteDance, Tsinghua University, and the Shanghai Artificial Intelligence Laboratory among others outlines this five stage road map for recursive self improvement. The study, titled "The Last AI Built by Humans: Toward Genuine Recursive Self Improvement", shows a rising industry focus on automating the labour intensive life cycle of training, evaluating, and fine tuning artificial intelligence models.
The document details five progressive stages of autonomy for these systems. Initially, an AI simply executes improvement procedures specified by human engineers. It then advances to choosing how to upgrade itself, rather than only carrying out preprogrammed instructions. But subsequent stages allow the system to determine what new information or experiences it needs to acquire, adapting to changes after deployment.
The ultimate stage would permit an artificial intelligence to consistently refine the very methods used to improve artificial intelligence itself. Recursive self improvement differs from a chatbot correcting a single response; it necessitates improvements that persist beyond one task and are carried over to successor systems.
The authors contend that automating parts of AI research could offer a competitive edge for model developers. This shift, if realised, has the potential to shorten development cycles. It could also significantly reduce the labour and computational costs associated with building foundation models, according to the paper's creators.
The automation of AI research has become a central area of contention in the US China technological competition. Chinese institutions continue to make rapid practical advancements. However, experts observe that American firms retain an early advantage, primarily due to better access to computing resources. And Erich Grunewald, a senior researcher at the Institute for AI Policy and Strategy, stated that US companies appear several months ahead of China, possessing more compute for deployment.
Grunewald also noted that Chinese researchers are proficient at extracting performance from limited hardware, but these compute constraints remain a challenge for them. Despite these difficulties, Chinese firms are increasing their investment in autonomous training infrastructure.
Z.ai, known in China as Zhipu AI, announced a recent decision to allocate approximately 60 per cent of the net proceeds from its latest USD 5 billion fundraising round. This funding is intended to support the creation of its next generation GLM foundation models and its "fully self training system".
Researchers involved with MiniMax 2.7 reported that their model could update its memory and build complex skills while conducting reinforcement learning experiments. A prior article titled "Early Echoes of Self Evolution" described this process, where the resulting experience fed back into its learning. So DeepSeek developed an agentic "harness" that provides models with greater autonomy for navigating multi step tasks, executing code, and interacting with external software.
The authors of the paper indicated that the speed of progression through the five stages would vary significantly across different artificial intelligence fields. Software engineering, for example, presented a relatively direct path forward. Robotics and scientific discovery, conversely, faced more substantial obstacles.
Safety considerations also remain a significant concern. The researchers emphasised that genuine recursive self improvement demands stringent safeguards. These include verified testing environments designed to confirm that updates are both safe and beneficial before being implemented. But the authors did not provide a timeline for when RSI could potentially be realised.
Chinese researchers from universities and Big Tech have proposed a five stage road map for recursive self improvement.
This framework details a progression from AI executing human designed procedures to refining its own improvement methods.
Automating AI research is a key area of competition between the US and China, with American firms currently holding an advantage in compute resources.
Companies like Zhipu AI, MiniMax, and DeepSeek are already working on aspects of autonomous AI training and agentic capabilities.
The authors highlight that achieving genuine RSI will necessitate strict safety measures and verified testing environments.
Source: SCMP
