'The Last AI Built by Humans': Chinese Tech Firms Race To Build AI That No Longer Needs Human Engineers

Chinese researchers have outlined a five-stage path towards artificial intelligence systems that could eventually take on more of the work involved in training and improving future systems, with progressively less human involvement.

The roadmap, published this month as an arXiv preprint, sets out a framework for how AI could move towards recursively improving both its capabilities and the process used to improve future systems.

The paper, titled 'The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement', was written by 33 researchers affiliated with Shanghai Jiao Tong University, Tsinghua University, ByteDance, Shanghai Artificial Intelligence Laboratory and several other institutions.

It arrives at a moment when so-called recursive self-improvement, or RSI, has become a closely watched area in the broader US-China AI race.

Five Stages of Less Human Intervention

The paper sets out an 'autonomy hierarchy' running from improvement execution to what the authors call recursive meta-improvement. At the bottom rung, an AI carries out upgrade instructions written by human engineers.

Further up the scale, systems begin choosing their own improvement strategies, then deciding what new experience or data they need to adapt after deployment. At the top is what the authors describe as an AI capable of continually improving the methods used to improve AI itself.

The authors stress that, unlike a chatbot correcting one mistake mid-conversation, genuine RSI requires improvements that persist beyond a single session and can be passed on to subsequent generations of models.

They note that progress will vary sharply by field, with software engineering, scientific discovery and embodied intelligence presenting different requirements and development speeds.

America's Head Start

Despite the ambition of the roadmap, researchers outside the paper have suggested that Chinese labs remain behind their American counterparts on this specific front. Erich Grunewald, a senior researcher at the Institute for AI Policy and Strategy, said American companies appeared ahead of China and had access to more compute for deployment.

Grunewald added that Chinese researchers were 'very capable at squeezing performance from scarce hardware', but that compute shortages 'do still bite'. Access to advanced chips has been a major constraint on China's AI ambitions, owing in part to US export restrictions.

Money Behind the Race

Chinese firms are not waiting for the compute gap to close before investing heavily in automated training systems.

Z.AI, formerly known as Zhipu AI, said about 60 per cent of the net proceeds from a US$5 billion share placement and convertible bond sale would go towards research and development of its next-generation models and what it called a 'fully self-training system'.

Separately, MiniMax researchers said the company's M2.7 model could update its own memory and build new skills while running reinforcement-learning experiments.

The company said the model was also used to improve its learning process and agent harness based on experiment results, while DeepSeek has released an agentic harness that allows models to use tools and multi-step workflows, although the system remains in developer preview.

The Wider Self-Improvement Debate

The Chinese roadmap arrives as similar concerns are surfacing at some of the world's best-funded AI labs. Anthropic researcher Anna Wang, writing in a personal capacity, said there was 'not yet a viable scientific plan to solve risks from recursively self-improving AI'.

The comments highlight that recursive self-improvement is not simply a concern confined to Chinese AI research.

Researchers at companies such as Anthropic and OpenAI are also debating the risks and possibilities of increasingly autonomous AI systems, even as they compete for the same technological lead.

If AI systems genuinely reach the higher stages of the roadmap, the labour and computing costs of building new models could potentially fall, reshaping who can afford to compete at the frontier of AI development.

It could also raise the stakes for oversight, since the paper identifies verification, safety and controlled environments among the challenges that must be addressed before genuine recursive self-improvement can be achieved.

For now, the researchers offer no timetable for when any stage might be reached, leaving the roadmap as a proposed framework rather than a demonstrated blueprint. The paper therefore sets out a potential direction for AI development rather than evidence that genuinely recursively self-improving systems have already been achieved.