A listed Chinese lab publicly putting AI-built AI at the center of its roadmap makes Z.ai's safety commitments, including its planned interpretability spending, a disclosure item for investors and regulators.
Tang Jie, co-founder of Beijing-based model developer Z.ai, has put recursive self-improvement at the center of the race to artificial general intelligence (AGI). "[Recursive self-improvement] is the main research problem for [achieving] AGI," Tang said in a keynote speech at the World Laureates Forum in Shanghai on October 31, the South China Morning Post reported. Recursive self-improvement describes AI systems that help build better versions of themselves, in a loop that could speed up over time.
Tang, a computer science professor at Tsinghua University, called it a "new paradigm" of research and development that would accelerate AI progress further. As frontier models move from chatbots to agents assigned complex scientific and research projects, he said, it makes sense to turn to AI for help, because human researchers face growing engineering difficulties and financial constraints. He did not disclose how much of Z.ai’s own research and development is now done by AI.
He paired the ambition with a call for caution. Tang urged researchers to be "humble" about recursive self-improvement and said responsibility for its safety should be shared. "Probably … the world should organise something together so that we can guarantee the safety for not just one country but for all the world," he said, according to the Post. DigiTimes reported the same message on November 2: an important direction for AGI, with safety still first.
The topic ran through the forum. Several Chinese industry and academic speakers described AI speeding up the design and manufacture of the chips that power AI, the Post reported. US labs are pointing the same way: Anthropic reported in September that AI was leading 26% of its internal AI research and development in August, up from about 1% in March, according to the Post.
Z.ai’s first result came from its own infrastructure
Z.ai’s concrete evidence dates from September. Z.ai says a GLM-5.3-powered agent built the inference system for GLM-5.3-Flash on more than 100,000 Chinese-made chips and tripled its throughput in under two weeks. The figures were reported by The Paper when Z.ai disclosed them on September 17. The agent analyzed performance bottlenecks, located accuracy defects, changed low-level code and ran layered tests, iterating on the results.
Z.ai said hardware efficiency and cost per token matched mainstream Nvidia GPUs, and that GLM-5.3-Flash, released anonymously as Ox-Alpha on OpenCode and OpenRouter, handled more than 62 trillion tokens in six days. The company said the system had not reached the stage of designing and training its own successor, but that an early form of recursive self-improvement had appeared. EastFrontier reported the result at the time: Z.ai says GLM helped build its inference stack.
The division of labor in that project is what Tang’s safety line refers to. Z.ai’s blog post said engineers defined objectives and system boundaries, while the agent handled analysis, hypotheses and code changes, according to researchers writing in Asterisk Magazine.
The roadmap pairs self-training with safety spending
Tang laid out the strategy in an internal letter in July. The "Touch High" plan commits Z.ai to two years of research toward AGI over short-term commercialization, focused on four areas: long-horizon tasks, autonomous agent systems, full self-training and what the letter called extreme safety governance, according to IT Home. The letter said Z.ai would commit resources on the scale of 10 billion yuan (百亿级) to mechanistic interpretability, research into how a model reaches its decisions.
Tang’s vocabulary has shifted along the way. He described "self-evolution" as a promising path in May, called it "AI training AI" in July, and adopted the term recursive self-improvement by September, the Asterisk researchers wrote.
China’s official safety framework now names the concept too. The AI Safety Governance Framework 3.0, released on September 14 by the National Cybersecurity Standardization Technical Committee (TC260) under the guidance of the Cyberspace Administration of China, updates its risk categories and governance measures, People’s Daily reported. According to the Asterisk analysis, the framework notes a "self-accelerating trend" that includes recursive self-improvement, but treats it as a trend to watch rather than an urgent risk.
Disclosure of AI’s share of research will be the test
Tang’s statements leave open how far Z.ai’s loop has gone. The company has not said what share of its model development is now done by AI, how it checks that an AI-made change actually improves a model, or how its interpretability spending is progressing.
The signals to watch are a second published result that goes beyond infrastructure into model training itself, any figures on AI’s share of Z.ai’s research output in its next results, and whether TC260 or the Cyberspace Administration turns the framework’s passing mention of recursive self-improvement into specific testing or reporting requirements.