US and China race to build self-improving AI
OpenAI says it wants an automated AI researcher; Z.ai points GLM-6 toward 'self-evolution.' The race is now about AI that builds AI.
In short
The contest between the United States and China has shifted from shipping the strongest model to building models that can do the research which produces the next one.
At a glance
- OpenAI said in a blog post that it aims to build an automated AI researcher.
- OpenAI calls GPT-6 Astra, released in early September 2026, the world's most intelligent and aligned model.
- Tang Jie, founder of Beijing-based Z.ai, spoke of self-evolution for the GLM-6 successor model on August 31, 2026.
- RSI, or recursive self-improvement, describes AI that trains stronger successor models on its own.
- The reporting rests on a single South China Morning Post article dated September 13, 2026, partly behind a paywall.
The race between Washington and Beijing has moved past the question of who ships the strongest model. What both sides now chase is AI that performs the research work producing the next model. A September 13, 2026 report in the South China Morning Post describes leading labs in the United States and China pointing advanced systems at writing code, designing experiments and developing training methods.
The idea behind RSI
RSI stands for recursive self-improvement. The premise is a system that runs the research loop itself and turns out a more capable successor, which then repeats the step. That compounding is the appeal — and, for now, also the uncertainty, because the loop is described as a direction of travel rather than a demonstrated result.
What OpenAI has said
OpenAI used a blog post to state that it wants an automated AI researcher. It calls GPT-6 Astra, released in early September 2026, the world's most intelligent and aligned model. In the same breath it frames the hard part as unsolved: how to reach full, aligned RSI safely. Read plainly, that is a goal statement, not a claim that the capability already exists.
The Chinese side of the ledger
On the Chinese side the report names Z.ai, headquartered in Beijing. Founder Tang Jie told an earnings briefing on August 31, 2026 that the company's next model, GLM-6, was heading toward self-evolution. As with OpenAI, the wording points at a roadmap rather than a shipped feature.
What we could not confirm
This story rests on one independent outlet, and that article was cut off behind a paywall. Compute volumes, export-control measures, timelines and funding figures were therefore not verifiable. The readable portion names no Chinese company other than Z.ai, so no market-share or training-cost numbers appear here. Those gaps stay open rather than getting filled with estimates.
FAQ
What does self-improving AI actually mean?
It refers to recursive self-improvement: systems that take over the research and training step themselves and are meant to produce stronger successor models. At this point it is a stated ambition, not a measured capability.
When was GPT-6 Astra released?
In early September 2026. OpenAI describes it as the world's most intelligent and aligned model; no independent assessment of that claim was available for this report.
Who is Tang Jie?
The founder of Z.ai, an AI company headquartered in Beijing. On August 31, 2026 he told an earnings briefing that the GLM-6 successor model was moving toward self-evolution.
