At a UN podium in Geneva, a Chinese engineer-turned-academician told the room that safety should be designed into AI from the first line of code — not patched in after an accident. The line drew quiet nods. The hard part, he knew, was making anyone actually do it.
Zhang Yaqin (张亚勤) has spent four decades close to the center of Chinese computing — first as a researcher, then as the president of one of the country's largest internet companies, and now as the founding dean of a Tsinghua institute that tries to turn AI research into industry. His latest cause is less glamorous than a new model: governance.
From Microsoft to Baidu to the classroom
Zhang's path reads like a miniature history of China's tech rise. He entered the University of Science and Technology of China at age 12 in 1978 and became an IEEE Fellow at 31. He spent roughly 1999 to 2014 at Microsoft, where he helped grow the company's Beijing research organization from a handful of researchers into a body of more than 3,000. In September 2014 he became president of Baidu, a post he held for five years until October 2019.
What he did next was unusual for someone at his level. Rather than join another company or start a fund, he returned to academia and, in 2020, founded the Tsinghua University Institute for AI Industry (AIR, 清华大学智能产业研究院), serving as its founding dean. AIR concentrates on three applied lanes — smart transportation, smart IoT, and smart healthcare — explicitly meant to shorten the distance between a published paper and a shipped product.
Building AIR as a bridge
The institute is Zhang's answer to a complaint he has repeated for years: China produces strong AI research but struggles to convert it into deployed industry. AIR's structure reflects that — professors paired with engineering teams, and a mandate to work on problems with clear commercial or social value. He has also chaired the Apollo Industry Alliance since 2018, keeping one foot in the autonomous-driving ecosystem he helped seed during his Baidu years.
That dual identity — former operator, current academic — is what gives his governance views weight. He is not arguing from a podium he has never stood near. He has run the training runs, the product launches, and the regulatory conversations.
The autonomous-driving wager
Zhang has been one of the most consistent public voices predicting that autonomous driving will cross from demo to deployable within this decade. He frames the arrival point as a "DeepSeek moment" for self-driving: a sudden leap in capability and adoption rather than a slow creep. Concretely, he forecasts that by 2030 more than 10 percent of new cars will carry L4 or better autonomous capability.
Behind that forecast is real engineering. AIR has released what Zhang describes as the world's first practical end-to-end open-source autonomous-driving model, Apollo FM, alongside a broader open-source model family. His reasoning rests on a distinction he draws often: the large model (大模型) is becoming the operating system of the AI era. As he put it in a 2024 speech, the large model (大模型) will reshape the technology application ecosystem "just as Windows defined the PC era and Android defined the mobile era." If that is true, then the model that drives the car is infrastructure, not a feature.
Governing the frontier
Zhang's most quoted line of 2025 came not in Beijing but at the UN/ITU AI for Good Global Summit in Geneva on July 10. "Safety is the foundation for AI, not an afterthought," he said, arguing that governance must keep pace with a technology now entering physical domains like autonomous driving and robotics. He laid out a practical agenda: a tiered, classified framework for large models (大模型); scenario-based evaluation for high-risk domains such as healthcare and self-driving; mapping of AI agents to identifiable legal entities so responsibility can be traced; and a red line banning AI systems from self-replicating or self-improving without human approval.
He has also pushed a concrete budgetary norm: dedicate more than 10 percent of large-model (大模型) investment to AI risk research. At the Boao Forum in March 2025, he pointed to China's measures requiring labels on AI-generated content — issued in March 2025 and effective September 1, 2025 — as, in his words, the first such law among major countries, and noted his estimate that roughly 65 percent of internet content is already AI-generated, making traceability urgent.
Reading the trend
The through-line of Zhang's recent work is a single bet: that the winners of the next AI decade will be the ones who treat safety and deployment as the same problem. His five predicted directions for large models (大模型) — multimodal, autonomous, edge, embodied, and biological intelligence — all point toward AI leaving the screen and entering machines, hospitals, and roads.
The governance bet in plain terms
Zhang Yaqin's move from Baidu's presidency to Tsinghua AIR is best read as a pivot from building AI fast to building it responsibly and making it stick in industry. His autonomous-driving forecast (L4 in >10 percent of new cars by 2030) is a bold, checkable claim; his governance agenda — tiered classification, entity mapping, a 10 percent safety-investment floor, and hard red lines — is a template other regulators are studying. For readers tracking China's AI, the markers to watch are not just new model releases but whether AIR's Apollo FM gains real adoption and whether Zhang's "safety-first" framing shows up in actual policy.
Honest limitations
Roles, dates, and quotes are drawn from Tsinghua AIR's official pages (air.tsinghua.edu.cn), reporting by China.org.cn (中国网) on his June 2024 Wuxi speech, the People's Pictorial (人民画报) on Boao 2026, and HKUST/IEEE-adjacent biographical sources for his career arc — cross-checked against each other. The Geneva "safety is the foundation" quote and the governance agenda are from Tsinghua AIR's official account of the July 10, 2025 AI for Good session. The "Windows / Android of the AI era" line is from China.org.cn's report of his 2024 speech; translation follows the original Chinese. The "65 percent of internet content is AI-generated" figure is Zhang's own estimate stated at Boao, not an independently audited statistic, and we report it as his claim. We did not verify AIR's current headcount or the exact number of professors and students, the precise adoption rate of Apollo FM, or the 2030 L4 forecast's underlying assumptions. His governance proposals are personal views, not enacted law.
What readers can do now
- Track Tsinghua AIR's Apollo FM releases and any real vehicle deployments to test Zhang's 2030 L4 adoption forecast against evidence.
- Read the official Tsinghua AIR write-up of his July 2025 Geneva talk for his full tiered-classification and entity-mapping governance proposal.
- Watch China's AI-generated-content labeling rules (effective September 1, 2025) as a live test of the "safety-first" framework Zhang advocates.