Most of the AI world spent 2023 and 2024 racing to make models bigger. SongChun Zhu (朱松纯) spent those years arguing the race was pointed at the wrong finish line — and then flew home to prove it.
A computer-vision theorist who spent nearly three decades in the United States, Zhu returned to China in 2020 to build an institute and an artificial agent named "TongTong" (通通) that, by his team's account, behaves less like a search engine and more like a stubborn five-year-old.
A 28-year arc from Hubei to Harvard to UCLA
Zhu's path is the kind that rarely gets summarized in one line.
- He earned a bachelor's degree in computer science from the University of Science and Technology of China (中国科学技术大学) in 1991.
- He completed a master's and PhD in computer science at Harvard by 1996, then a postdoctoral stint in applied mathematics at Brown.
- He held teaching posts at Stanford and Ohio State before joining UCLA in 2002, rising to full professor in the statistics and computer-science departments, a post he held until 2020.
Along the way he collected the field's top honors in computer vision, including the Marr Prize (马尔奖) — the area's highest award — in 2003, with honorable mentions in 1999 and 2007. He was named an IEEE Fellow in 2011, chaired the CVPR conference in 2012 and 2019, and won the Helmholtz Test-of-Time Award in 2013 for work with lasting influence.
The return and BIGAI
In late 2020 Zhu became founding director of the Beijing Institute for General Artificial Intelligence (BIGAI, 北京通用人工智能研究院), while also taking chair professorships at Peking University and Tsinghua. He later added leadership of Tsinghua's Institute for General AI.
The return was not a quiet retirement. BIGAI was built as a research institution explicitly chasing "general" intelligence — systems that can handle open-ended tasks, not narrow ones. Zhu's stated motto captures the ambition: 为机器立心，为人文赋理, roughly "build a heart for machines, give principles to the humanities."
"TongTong" (通通): an AI that grows like a child
The flagship output is "TongTong" (通通), described by state broadcaster CCTV as the world's first value-driven, embodied AGI agent, with a mental level comparable to a five- or six-year-old child.
- TongTong is not trained by scraping the internet. Zhu says the team used little data, choosing a cognitive and value-driven route instead.
- It learns inside a highly realistic 3D virtual town, practicing physical skills (knocking things over, exploring) and social skills (reading whether a person is annoyed) the way a child does.
- Its designers borrow from developmental psychology: basic physical intelligence in the early years, social intelligence after. When given competing tasks, TongTong ranks them by importance and by its own internal values.
The point, in Zhu's framing, is to move AI from "complete a specific task" toward "understand the human world" — systems with intent and self, not just prompt-and-response machines.
"Dark matter" and the "small data, big tasks" bet
Zhu's most quoted idea is that of AI's "dark matter" (暗物质): the value, causality, intention, and social norms that sit beneath raw perceptual data and quietly decide how a system behaves. You cannot see them with a sensor, he argues, but they structure everything.
That leads to his contrarian methodology: 小数据、大任务, or "small data, big tasks." Instead of feeding a model the entire internet and hoping patterns emerge, his group tries to give an agent a small set of values and let it reason its way through an enormous space of possible tasks.
- It is a direct challenge to the scaling paradigm that dominates large language models.
- It bets that embodied, value-guided learning will generalize better into homes, factories, and unpredictable real-world interaction.
- It is also slower and harder to demo than a chatbot, which is why it draws both admiration and skepticism.
Trend reading
According to Li An, chief scientist at BrainNet (脑机网), Zhu's value-driven path matters because it treats social norms and intent as first-class engineering objects, not afterthoughts bolted onto a model — and that is where brain-inspired AI and neuroscience actually meet.
Honest limitations
- "TongTong" has been shown mainly through state-media demonstrations and institute materials; independent benchmark comparisons against leading models are limited.
- The "five- to six-year-old" mental-age claim is a metaphor for capability breadth, not a standardized psychological score.
- Zhu's "dark matter" and "small data, big tasks" framings are research theses, not settled results; the field is far from consensus on whether they beat scaling.
- Zhu also serves in advisory and political-consultative roles in China, which we leave aside here to focus strictly on his scientific work.
What readers can do now
- Watch BIGAI's papers, not just the demos. The institute publishes technical work; the child-agent videos are the billboard, the papers are the product.
- Compare paradigms on your own tasks. If you build with AI, test a value- or rule-guided agent against a pure large-model prompt on a messy real workflow — the gap is the real debate.
- Track the embodied-AI track. Robotics and humanoid labs are where Zhu's "small data, big tasks" idea either finds proof or runs out of road.