Commentary: China’s Strategy to Win the AI Race — Giving Machines a ‘Heart’
Listen to the full version
* Mainstream AI models’ limits are increasingly clear.
* China is taking its own rigorous path.
* China is building an ecosystem for the post-language-model era.
For the past two years, large language models have been heralded as the undisputed ticket to artificial general intelligence. As companies race to build ever-larger data centers, the brute-force era of artificial intelligence is hitting a wall. Behind the scenes, the top echelons of academia and industry are quietly abandoning the scaling laws of the past in favor of a new paradigm — one that could define China’s second growth curve in the global AI race.
Unlock exclusive discounts with a Caixin group subscription — ideal for teams and organizations.
Save an extra $50. Introductory offer for new readers. Subscribe now.
- DIGEST HUB
Digest Hub
Back
- Mainstream LLMs face limits like lacking world models, hallucinations, and diminishing scaling returns due to data/compute constraints.
- China’s Beijing Institute proposes CUV architecture for value-driven, autonomous AI agents emphasizing cognition, utility, and ethics.
- Shanghai’s Zhangke Tongzhi Incubator, launched August 2025, fosters 200+ post-LLM startups sharing resources for fields like protein mining and railway inspection.
1. The article highlights three key points: mainstream AI models' limits are clear, China is pursuing a rigorous path, and it is developing an ecosystem for the post-language-model era [para. 1].
2. Large language models (LLMs) were seen as the path to AGI for two years, but scaling is hitting limits due to data centers; academia and industry are shifting paradigms, positioning this as China's second AI growth curve [para. 2].
3. Mainstream models' flaws are evident: Yann LeCun criticizes lack of world models for reasoning and planning; Fei-Fei Li notes absence of physical perception, making them "brains in a vat" without sensory common sense [para. 3].
4. AI industry is splitting; mainstream black-box methods cause hallucinations and contradictions, with diminishing returns as compute and high-quality text dwindle [para. 4].
5. Innovators respond: David Silver raised $1.1 billion for Ineffable Intelligence, using reinforcement learning over human data to challenge LLM dominance [para. 5].
6. China advances with Beijing Institute for General AI's CUV (Cognition, Utility, Value) architecture, aiming to "give machines a heart" [para. 6].
7. CUV differs from big data/small tasks by creating value-driven cognitive systems; it fosters autonomous agents with intrinsic motivations, ethical decisions via heart (value), environment, and cognition interactions [para. 7].
8. LLMs are AI's first curve limited by compute/data; CUV offers China's second via small data/big tasks, enabling explainable AI for ethics-sensitive fields like medicine and law [para. 8].
9. China builds a Y Combinator-like ecosystem: Zhangke Tongzhi Incubator launched August 2025 in Shanghai by Beijing Institute and Zhangjiang Technology Venture Capital [para. 9].
10. In under a year, it nurtured startups like Nuka Intelligence (protein mining), Gengzhi Technology (3D modeling), Xinshi Fengyu (railway inspection), and Infiniflow (RAGFlow platform); they share resources in a community of 200+ vertical AI firms [para. 10].
11. Incubator GM Li Wei contrasts AI inefficiency with human brain's 20-watt power for multi-modal processing, reasoning, creation, and value judgments [para. 11].
12. Amid U.S. President Donald Trump's semiconductor restrictions, China pivots to cognition redesign; incubators commercialize breakthroughs, predicting 2030 AGI unlike current models, led by Shanghai startups [para. 12].
- ...
- {"chinese":"北京通用人工智能研究院","english":"Beijing Institute for General Artificial Intelligence"}
- MOST POPULAR
Sign in with Google
Sign in with Facebook
Sign in with 财新