Z.ai cofounder Tang Jie has called recursive self-improvement the main research problem for achieving artificial general intelligence, putting AI-assisted AI development at the center of his research outlook.
In a report published November 1, 2026, the South China Morning Post described Tang’s remarks at an event in Shanghai. Tang, also a computer science professor at Tsinghua University, argued that using AI to build smarter systems could accelerate progress toward AGI. He also said the work needs to proceed safely.
“[Recursive self-improvement] is the main research problem for [achieving] AGI,” Tang said, according to the report.
The remarks describe a research direction, not a new Z.ai product. The available account is a news report, not a full transcript or technical proposal.
A September engineering post from Z.ai provides a useful way to examine that direction. It describes an AI agent helping build the infrastructure that serves the company’s models. That is a concrete example of AI contributing to AI development, although it leaves a substantial distance between infrastructure automation and an autonomous system that designs its own successor.
What Recursive Self-Improvement Means
Recursive self-improvement, often shortened to RSI, describes a feedback loop: an AI system helps create a more capable system, which becomes better at carrying out the next round of improvement.
The recursive part matters. A model writing one useful software patch is AI-assisted engineering. A continuing cycle in which improved systems become increasingly effective at improving subsequent systems is the larger ambition.
Possible contributions include writing training software, designing experiments, developing model architectures and improving the infrastructure used to run models. These tasks affect different parts of AI development, so progress in one should not be treated as proof that the entire loop works.
In its September 17, 2026 research post, Z.ai describes the endpoint as a system capable of designing and training its successor entirely autonomously. The company immediately qualifies that ambition:
“We are not there yet, but early forms of it are already emerging.”
That distinction is worth keeping. AI can make a meaningful contribution to building another AI system while people still define the objective, supply the computing resources and determine whether the result is acceptable.
My reading is that Tang is identifying a potential engine of future progress, not claiming Z.ai has already completed an autonomous improvement cycle. The company’s own description supports that narrower interpretation.
Why Tang Sees an AGI Research Path
Tang’s argument changes the role assigned to AI inside a research lab. Instead of being only the object researchers develop, AI also becomes a tool for doing the development.
The potential acceleration comes from a compounding effect. If a stronger model helps researchers produce better experiments or software, that work could contribute to another model that is more useful for research. Successive generations could then improve both the product and the process that creates it.
That is the logic behind calling RSI a new research-and-development paradigm. It is also a hypothesis about how progress might compound, not evidence that every improvement will produce another one.
The idea extends beyond Z.ai. A September 13 SCMP report described leading AI companies in the United States and China using advanced models to write code, design experiments and develop training techniques.
The difficult step is showing that these contributions reliably produce broader capability gains. Faster engineering can shorten a development cycle without resolving harder research problems. Tang’s statement establishes the importance he assigns to RSI; it does not provide a timetable for AGI or a measured relationship between automated research and general intelligence.
Z.ai’s Infra Agent Offers a Concrete Example
Z.ai’s engineering account centers on an Infra Agent powered by GLM-5.3. The company says the agent carried out much of the work required to bring GLM-5.3-Flash’s inference infrastructure into production.
Inference means running a trained model to generate outputs. Training changes a model’s learned parameters. Improving an inference service therefore improves how a model operates, but does not itself demonstrate that the model has trained a smarter successor.
The deployment involved more than 100,000 Chinese-made AI accelerators, according to Z.ai. The company describes limited chip memory capacity and bandwidth, incomplete kernel support, a new model architecture, a one-million-token context window and multimodal requests as engineering constraints.
Those details explain why infrastructure work matters. Serving a model efficiently requires coordination across memory, computation, communication and request handling. A change that saves memory can add computation; a faster component can expose a bottleneck elsewhere.
Z.ai reports the following results:
| Area | Z.ai’s published claim | How to interpret it |
|---|---|---|
| Deployment scale | GLM-5.3-Flash’s production inference runs on a cluster of more than 100,000 Chinese-made accelerators | The work concerns a large serving system, not a small coding demonstration |
| Development time | The model went from initial adaptation to production readiness in less than two weeks | This is the company’s account of its development timeline |
| Serving performance | End-to-end throughput roughly tripled relative to the initial baseline | The comparison is against an earlier version of the deployment, not a competing model |
| Usage | Under the anonymous name Ox-Alpha, the model processed more than 62 trillion tokens in six days across OpenCode and OpenRouter | The company reports substantial usage; that volume does not measure intelligence or autonomy |
These figures come from Z.ai’s own post, not independent measurements. The reported threefold improvement should be read within that deployment and its initial baseline.
The company lists several optimizations, including parallel execution, quantization, memory-management changes and separating encoding, prompt processing and output generation into distinct serving stages.
For interpreting the RSI claim, the important point is the agent’s contribution to the surrounding engineering system. The published example concerns AI helping improve the machinery used to run AI.
The Feedback Loop Is More Revealing Than the Speedup
Z.ai’s most consequential technical observation is that code generation alone was insufficient. The Infra Agent needed useful feedback that could be traced to specific causes.
The blog explains that an agent can understand a codebase and still struggle when a change causes a numerical error or performance regression. An overall throughput number reveals that something changed. It may not identify whether the cause lies in a computational kernel, communication between accelerators, memory management or serving orchestration.
The company gives examples such as a 30% increase in time to first token and a 20% drop in output throughput. These illustrate the kinds of feedback an agent might receive, rather than a reported benchmark result for the deployment.
Time to first token, or TTFT, measures how long a request waits before the model begins producing its answer. Output throughput concerns the rate at which output tokens are produced. A system can improve one metric while worsening another, making a single performance score an incomplete guide.
Frequently Asked Questions
3 questions
1Is Tang Jie Announcing a New Z.ai Model?
No. The reported remarks describe his research priorities and the potential role of recursive self-improvement in achieving AGI.
