Cosa è successo
The South China Morning Post reports that the US-China AI competition is increasingly focused on recursive self-improvement: systems using AI to help research, train and evaluate more capable successors. US companies appear to have an early lead in AI-assisted research, while Chinese labs are developing related self-training and “self-evolution” approaches. The report says neither side has demonstrated an open-ended, fully autonomous improvement loop.
The South China Morning Post reports that leading US and Chinese AI companies are deploying advanced models to write code, design experiments and develop training techniques for stronger models. It describes the long-term objective as recursive self-improvement, in which AI systems autonomously contribute to training and improving successor systems. The report says OpenAI recently described work toward an automated AI researcher, while also acknowledging uncertainty about how to reach fully aligned recursive self-improvement safely.
According to the report, US companies currently appear several months ahead in using AI to conduct AI research independently. It cites Anthropic’s reported work on having Claude models design experiments, write code and develop techniques, as well as Google DeepMind’s description of Gemini 3.8 Flash as suited to long-horizon software engineering and autonomous agents. The report also says OpenAI reported an automated research intern capable of running some experiments that would take a skilled researcher days. These claims are attributed to the South China Morning Post and the companies or researchers it cites; they are not independently confirmed here.
The report says Chinese labs are pursuing related goals under terms such as “self-evolution” and “full self-training.” Z.ai founder Tang Jie reportedly described a roadmap for GLM-6.0 in which AI would manage the training from pre-training through . The report also discusses work associated with MiniMax and DeepSeek, while noting expert criticism that autonomous agents and self-training capabilities do not yet amount to genuine recursive improvement. A key unresolved issue is whether a system can decide when a change is actually an improvement without relying on a separate human authority.
Dettagli della fonte: scmp.com ↗
Perché è importante
If AI systems can reliably improve the methods used to build their successors, progress could accelerate and make oversight more difficult. The report describes potential benefits in software development, experiment design and training efficiency, but also warns that flawed models could pass errors or misalignment into later generations. The practical significance remains uncertain because current systems still rely on human judgment, infrastructure and external evaluation.
Recursive self-improvement could create a feedback loop in which AI helps produce more capable models, potentially shortening development cycles and increasing the strategic importance of , research infrastructure and evaluation. The report says US analysts see an advantage in American access to compute, while Chinese researchers are described as effective at extracting performance from more limited hardware.
Safety concerns are central because training future systems with flawed predecessors could compound errors or misalignment. The report cites warnings from researchers in both countries about acceleration, sandbox escapes and possible national-security consequences. However, it also reports that fundamental architectural breakthroughs and judgments about whether changes are beneficial still depend on human expertise. The scale, speed and reliability of any future recursive improvement remain unknown.
Meccanismo interattivo: come funziona realmente
Esplora la tecnologia alla base di questo sviluppo in modo interattivo.
crm_get_transaction(id='4092').Which component of an AI application is the machine-learning model itself?
Cosa guardare dopo
Watch for independently verifiable demonstrations that an AI system can propose, implement and judge improvements without a separate human-defined verifier. Also watch whether companies publish safety evaluations, deployment limits and evidence that self-improvement systems do not amplify mistakes or bypass safeguards. No access terms, pricing or general availability are documented because the report concerns research and development rather than a public product launch.
The most important evidence would be a reproducible system that independently identifies useful changes, implements them, evaluates the results and improves its own evaluation process. The South China Morning Post reports that public evidence for this open-ended capability remains thin, so claims about “self-evolution” should not be treated as proof of autonomous recursive self-improvement.
The report does not document a public release, consumer access, pricing or deployment schedule for the systems discussed. It also does not independently validate the companies’ performance claims. Future reporting should establish what tasks these systems can complete without human intervention, what safeguards constrain them, how results are measured and whether failures are detected before changes reach production systems.