The global race to achieve recursive self-improvement (RSI) in artificial intelligence, the theoretical point where an AI system can autonomously improve its own code and capabilities, is intensifying in China, spurred by recent claims from US-based Anthropic that its frontier model is nearing this critical threshold. The pursuit of RSI, long considered the “holy grail” of artificial general intelligence (AGI), is increasingly viewed by Chinese researchers and policymakers as a decisive battleground in the broader US-China tech war, according to an analysis by the South China Morning Post.
The concept of an “intelligence explosion,” originally coined by British mathematician I.J. Good in 1965, posits that an ultra-intelligent machine could design even better machines, leaving human intelligence far behind. While once relegated to science fiction, the rapid advancement of large language models (LLMs) has brought the concept into the realm of near-term engineering.
The Mythos Catalyst
The urgency in China has been catalyzed by recent developments at Anthropic. The San Francisco-based company recently claimed that its most advanced model, Mythos, is “edging toward” the capability for recursive self-improvement. While Anthropic has not provided specific technical details or timelines, the assertion has sent ripples through the global AI community.
When Mythos was first announced in April, it was withheld from general release due to its unprecedented capacity to spot and exploit cybersecurity vulnerabilities. The subsequent release of its public-facing version, Claude Fable 5, included strict safeguards preventing its use for model distillation, further highlighting the strategic sensitivity of frontier capabilities.
For Chinese researchers, Anthropic’s claims represent both a benchmark and a warning. The prospect of a US company achieving RSI first raises the specter of an insurmountable technological gap, in which the leading AI system accelerates its capabilities faster than competitors can match.
China’s Domestic Push for Self-Evolution
In response, China’s AI ecosystem is aggressively pursuing its own pathways to “self-evolution.” The topic has become a central focus at major industry gatherings. At the Zhongguancun Forum in Beijing this past March, Luo Fuli, the lead developer of Xiaomi’s highly successful MiMo model, said that “self-evolution” would be the biggest trend in AI over the next year. “We’ve already begun to see an implementable and actionable path to achieve AI model self-evolution,” she told an audience of major policymakers and investors. “A year ago, I thought this might take three to five years, but recently, I’ve started to think that the time frame should be shortened to one to two years.”
This focus is not merely theoretical. Several prominent Chinese researchers and startups are actively developing architectures designed to facilitate autonomous improvement. Notably, Tian Yuandong, a former director of research at Meta’s Fundamental AI Research (FAIR) team, recently co-founded Recursive Superintelligence, a startup dedicated to building self-improving AI systems.
The approach taken by Chinese researchers often emphasizes iterative refinement through synthetic data generation and automated feedback loops. By training models to evaluate and correct their own outputs or those of smaller models, researchers hope to create a virtuous cycle of continuous improvement that reduces reliance on human-annotated data.
The Hardware Constraint
However, China’s pursuit of RSI faces a significant structural hurdle: the availability of advanced computing hardware. Recursive self-improvement, particularly in its early stages, is expected to be extraordinarily compute-intensive, requiring massive clusters of high-performance GPUs to train and test successive model iterations.
US export controls have severely restricted China’s access to the most advanced chips from companies like Nvidia. While domestic alternatives are emerging, they currently lag behind their American counterparts in performance and in the maturity of their software ecosystems. This hardware constraint means that Chinese researchers must often rely on more efficient algorithms and novel architectures to compensate for the lack of raw computing power.
Despite these challenges, the strategic imperative remains clear. The successful realization of recursive self-improvement would fundamentally alter the balance of technological power. As the SCMP analysis suggests, the race is no longer just about building the smartest model today, but about building the model that can build the smartest model tomorrow. In this high-stakes competition, China’s AI sector is mobilizing its resources to ensure it is not left behind when the intelligence explosion occurs.
The Policy Dimension
China’s pursuit of recursive self-improvement is not purely a private sector endeavor. The government has identified AGI and self-improving AI systems as strategic priorities in its national AI development plans. Research grants, talent programs, and dedicated computing resources are being directed toward labs working on foundational AGI research.
This state backing gives Chinese researchers a degree of institutional support that their counterparts at private US companies may lack. However, it also introduces a different set of constraints, including the need to align research priorities with national security objectives and the potential for bureaucratic friction in a field that requires rapid iteration and open scientific exchange.
The international dimension of the RSI race also raises profound governance questions. If a single AI system achieves recursive self-improvement, the speed of subsequent capability gains could be so rapid that existing regulatory frameworks—which operate on timescales of months to years—would be entirely inadequate to manage the transition. For China, the strategic calculus is clear: it is better to be the country that achieves RSI first than to be subject to the decisions of a system developed elsewhere. This logic, mirrored in Washington, is one of the key drivers of the accelerating pace of frontier AI development on both sides of the Pacific.