The conference hall was full of people selling the next breakthrough. He walked up and told them most of it was a story they were telling themselves. At this summer's World AI Conference in Shanghai, Zhu Songchun did not pitch a product — he questioned the premise of the entire race.
"The chokepoint is above the neck"
Zhu Songchun (朱松纯) is dean of the Beijing Institute for General Artificial Intelligence (北京通用人工智能研究院) and of Peking University's School of Intelligence. Speaking at the WAIC 2026 Thinkers Forum on 19 July 2026, he aimed his sharpest line at the country's own habit of measuring progress in hardware.
"我们常说的'卡脖子'，卡住的不是脖子，而是脖子以上的思想与认知。" "The 'chokepoint' everyone talks about — it isn't the neck. It's the thought and cognition above the neck."
His point: the real contest has moved past chips and compute into who defines the theory and the narrative. He was blunter about China's posture:
"美国提出概念，我们解释概念；美国形成趋势，我们论证趋势。" "America proposes the concepts; we explain them. America sets the trends; we justify them."
For Zhu, staying inside someone else's framework — even while producing more papers — is the deeper dependency.
Three stories the market keeps telling
He named three "cognitive traps" that, in his view, the industry has dressed up as investment theses:
- Equating the large model (大模型) with AGI. Fluency in language, he argues, is not understanding or causality.
- Worshipping the Scaling Law. More compute and data, on their own, will not autonomously produce general intelligence.
- The AI-extinction narrative as a capital story. He pointed to US labs that restrict model access citing risk while simultaneously chasing multibillion-dollar valuations — a script written for investors as much as for safety.
The line that travelled furthest from the forum was his compact verdict on division of labor: "AI handles computation; humans steward civilization" — carried by China News Service from his remarks. His warning underneath it: the smarter AI gets, the more clearly humans need to keep their own judgment.
His alternative: a value-driven architecture
Zhu's answer is not "smaller models" but a different architecture. The institute's CUV framework — Cognition, Utility, Value — is presented as the first mathematically rigorous definition of AGI, published in the journal Engineering. It shifts research from task-driven behavior to value-driven action: an agent that chooses its own goals because it has an internal value system, not because a prompt told it to.
From that framework came "通通" (TongTong), a general agent now in its third generation. Per the institute's own GSA evaluation, TongTong ranked first overall and beat GPT-5 in six sub-tasks, with capabilities the team compares to a 5–6-year-old child. The same logic is meant to migrate into bodies through "通脑" (TongNao), a general "brain" for embodied robots — putting the "heart" Zhu talks about into machine platforms.
To test social behavior at scale, the team built a Large Social Simulator (LSS) anchored in China Optics Valley (中国光谷): a simulated urban area of roughly 200 square kilometers and 1.18 million anonymized residents, organized in three layers — individual, organization and society.
Why "social intelligence" is his bet for China
Zhu's strategic claim is that the next frontier is not language or physics but social intelligence — an agent's ability to read duties, rights and relationships, and to operate inside human institutions. He argues this is both the last barrier to true AGI and a structural weakness of Western stacks, which optimized for benchmarks rather than social context.
"社会智能" is, in his framing, China's chance to stop following and start defining a赛道 (track) on its own terms.
It is a pointed contrast to the Listing-floor excitement of the model labs: where others race to ship bigger models, Zhu is arguing the winner will be the one who solves how machines behave among people.
Honest limitations
Quotes are from Zhu's WAIC 2026 Thinkers Forum speech (19 July 2026) as reported by Guangming (gmw.cn), Jiefang Daily / Shanghai Observer (shobserver), and China News Service. The CUV framework and TongTong results are drawn from the institute's own publications and benchmark (GSA); the claim that TongTong "beat GPT-5 in six tasks" is a self-reported result, not an independent third-party test, and should be read as the institute's framing rather than settled fact. WAIC is an industry conference, not a peer-reviewed venue. Zhu is an advocate for a specific research route (value-driven, social intelligence), and his critiques of Scaling Law and large models reflect a competing paradigm, not consensus science — other leading labs explicitly disagree. The LSS population figure (1.18M) is the institute's stated dataset size. English renderings of the Chinese quotes are translations of the reported remarks.
What readers can do now
- Read the primary source before picking a side. The CUV paper in Engineering lets you judge "large model = AGI" — and its opposite — from the math, not the headline.
- Wait for independent benchmarks of TongTong (通通). Institute self-reports are a starting point, not proof; watch for outside evaluations.
- If you follow China's AI policy, track the "social intelligence" and brain-computer interface (脑机接口) threads — they hint where state research money may tilt next, and where the next genuine divergence from US approaches could appear.