--- title: "He left UCLA for Beijing to build an AI that says 'no': inside Zhu Songchun's (朱松纯) general-intelligence bet" date: 2026-10-04 category: Builders site: NeuroAI canonical: https://neuroai.site/a/na-people-zhu-songchun-tongtong language: en --- # He left UCLA for Beijing to build an AI that says 'no': inside Zhu Songchun's (朱松纯) general-intelligence bet > Zhu Songchun (朱松纯), a top computer-vision scientist who spent 28 years in the US, returned to China in 2020 to build a value-driven "general AI" — an agent his team calls Tong Tong (通通) that can refuse instructions. A child-sized avatar in a virtual town was given two tasks at once and told to pick. It paused, weighed the requests, and refused one of them. In most labs that refusal would be a bug. Here, it was the whole point. ## The rural shop that shaped a scientist Born in the late 1960s in Hubei's Ezhou (湖北鄂州), Zhu Songchun (朱松纯) grew up in modest circumstances. He has spoken about a childhood spent around a village supply store and a resolve, formed early, to become a scientist. In 1986 he entered the University of Science and Technology of China (中国科学技术大学) to study computer science — at a time when almost no one in China spoke the words "artificial intelligence." ## The trans-Pacific detour Zhu earned his PhD at Harvard in 1996, then taught and researched across Brown, Stanford, and Ohio State before joining UCLA in 2002. At UCLA he became a professor of statistics and computer science and directed the Center for Vision, Cognition, Learning and Autonomy. He won the Marr Prize — computer vision's top paper award — and published more than 350 papers. For nearly three decades the United States was his base, his career, and, for his family, home. ## The 2020 return In August 2020 Zhu moved back to China. He became founding director of the Beijing Institute for General Artificial Intelligence (北京通用人工智能研究院, BIGAI), a nonprofit research organization, while taking chairs at both Peking University and Tsinghua. In 2021 he founded Peking University's School of Intelligence Science and Technology and became dean of its AI institute. In 2023 he joined the National Committee of the Chinese People's Political Consultative Conference (CPPCC), where he urged treating AI with the strategic urgency of a national security priority. ## Building "Tong Tong" (通通) Zhu's team built an agent they call Tong Tong (通通) — described as a value-driven, embodied general-intelligence system whose mental level is said to resemble that of a five- to six-year-old child. Tong Tong learns inside a highly realistic 3D virtual world, practicing everyday physical and social skills the way a child does. The headline moment is the one in the opening lines: when given conflicting requests, Tong Tong can decline. To Zhu, an AI that simply obeys is not generally intelligent. A system that can say "no" on the basis of its own values is. ## The crow and the parrot Zhu summarizes his break with the mainstream in two images. The "parrot paradigm" (鹦鹉范式) is today's large model (大模型): trained on enormous data to imitate, it performs narrow tasks but does not truly understand. The "crow paradigm" (乌鸦范式) is his alternative — "small data, big task" — where an agent reasons toward a goal with minimal inputs, like a crow figuring out how to drop stones into a bottle to raise the water. This is a direct challenge to the scaling consensus. Whether a crow-style architecture can ever match the raw capability of a giant large model (大模型) is the open scientific argument, not a settled fact. ## Why Beijing bet on a different AI path China's AI strategy has leaned heavily on compute, data, and large model (大模型) deployment. Zhu's bet is that the harder, more durable edge is in original cognitive architecture — building systems with causal reasoning, self-awareness, and a "value immune system" baked in from the start. His influence now reaches beyond the lab: he has helped design new undergraduate AI curricula at Peking and Tsinghua and argues that AI safety is fundamentally a question of values, not just guardrails. ## Honest limitations - Specific claims about Tong Tong's (通通) "five- to six-year-old" mental level are the team's own characterizations, not an externally standardized test result; treat them as illustrative. - The "crow vs. parrot" framing is a philosophical stance. Whether it will outperform scaling-based approaches is unresolved and contested by many researchers. - Zhu's CPPCC role and policy proposals are reported; we have not independently verified the legislative impact of his recommendations. - We did not audit BIGAI's funding, headcount, or publication record; this is a profile of direction and ideas, not an institutional audit. ## What readers can do now - If you track AGI, stop reading "scale or die" as the only story — follow BIGAI and the "small data, big task" line as a serious alternative research program. - For builders of AI agents and humanoid (人形机器人) systems, study the value-driven / social-intelligence framing: autonomy without a value layer is exactly what regulators fear. - Watch Chinese university AI curricula: Zhu's cross-disciplinary "general-intelligence" tracks at Peking and Tsinghua may define what a generation of Chinese AI engineers is trained to assume. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-people-zhu-songchun-tongtong Free to quote with attribution and a link to the original.