Can China's AI Lead the World in the Next Five Years? A Tsinghua Roundtable

The real bottleneck for Chinese AI is not technology, but the spirit of adventure and computing power. In this high-density, 26-minute conversation, five top scholars spanning academia and industry touched upon almost all the key issues facing Chinese AI in 2026.

Program Overview

This is a panel discussion from the AGI Next Frontier Summit, hosted by Li Guangmi, CEO of Shixiang Technology. Participants included: Zhang Bo, the 91-year-old founder of Chinese AI; Professor Tang Jie, founder of Zhipu; Lin Junyang, Chief AI Scientist of Alibaba Cloud’s Qwen; Yao Shunyu, a post-95s scientist who recently returned to China to join Tencent; and Academician Yang Qiang, the pioneer of federated learning. The conversation revolved around four core topics: the divergence path of large models, the next paradigm of AGI, the strategic implementation of Agents, and whether Chinese AI can lead the world in three to five years.

To B vs. To C: Two Completely Different Games

Yao Shunyu provided a very clear framework for differentiation. He observed that global AI super-applications have formed two paradigms: ChatGPT representing To C, and Claude Code representing To B. However, the growth logic of the two is completely different.

In To C scenarios, ordinary users’ feelings about ChatGPT have become flat, and most people just regard it as an enhanced version of the search engine. The improvement of model intelligence has not translated into a qualitative change in experience. In To B scenarios, the intensity of intelligence is directly linked to productivity and revenue. He gave a very practical example: in the US market, Opus 4.5 costs $200 per month, and a model one level lower costs only $20 to $50, but practitioners with an annual salary of $200,000 are willing to pay a premium to choose the best model. The reason is simple: Opus can get eight or nine out of ten tasks right every day, while a slightly inferior model can only get five or six right, and you can’t predict which ones will go wrong, requiring a lot of extra effort to monitor, which is counterproductive.

“不是AI替代人，而是会使用AI工具的人替代不会使用的人。” It’s not AI replacing people, but people who can use AI tools replacing those who can’t.

Tencent’s strategic choices also confirm this differentiation. Yao Shunyu admitted that Tencent has a stronger To C gene, and they found that the bottleneck in the To C field is not a larger model or stronger reinforcement learning, but additional context and environment. He gave a life-like example: if you ask AI where to eat today, no matter which year you ask ChatGPT, the experience is not good, because the core of the problem is not that the model is not strong enough, but that key information is missing, such as weather, location, and family’s dietary preferences. Only by synchronizing WeChat chat history with Yuanbao AI can it bring real value to users.

On the To B side, Tencent chose to first serve the real scenarios of its own 100,000 employees. Yao Shunyu compared this by saying that startups like Anthropic need to rely on data vendors to obtain Coding Agent data, and the limited number of labeling personnel leads to limited data diversity. Large companies can accumulate real data through internal scenarios, rather than relying on labeling vendors or model distillation.

Lin Junyang added an important observation: there is a huge difference in Coding Token consumption between the Chinese and American markets. The United States has formed a scale, but the Chinese market has not kept up. Anthropic’s choice of the Coding track is not a simple strategic bet, but based on real needs discovered through frequent communication with customers.

Autonomous Learning: Already Happening, Just Not Yet Named

Yao Shunyu, as a firsthand participant who worked at OpenAI, brought a key judgment: autonomous learning has become a hot topic in Silicon Valley, almost a consensus in the industry, but everyone’s understanding is different. More importantly, autonomous learning is already happening, but many people have not realized it yet.

He gave several examples that have already happened: ChatGPT uses user data to fit the chat style, making the experience better and better; Claude Code has completed 95% of its own project’s code writing, helping itself to continuously optimize; Cursor’s auto-completion model learns from the latest user data every few hours.

“今天的AI系统本质上由两部分构成：神经网络和代码库。如果未来的Agent能够自己优化代码仓库，那其实就是AGI的一种形态。” Today’s AI system essentially consists of two parts: a neural network and a code base. If future Agents can optimize the code repository themselves, that is actually a form of AGI.

Yao Shunyu believes that the bottleneck of autonomous learning is not in methodology, but in data and task scenarios, as well as imagination. The O1 model improved from 10 points to 80 points on math problems, but if autonomous learning becomes a new paradigm, what task should be used to verify it? Is it to make it a profitable trading system, or to solve scientific problems? The verification standard itself is an unanswered question.

When asked which company is most likely to lead the next paradigm, Yao Shunyu is still optimistic about OpenAI, although commercialization has weakened its innovation gene.

Lin Junyang put forward two key directions from a more practical perspective: one is Scaling during testing, allowing AI scientists to work continuously for 30 hours to solve difficult tasks; the other is the initiative of AI, allowing the environment to directly trigger AI’s thinking and action, rather than waiting for human prompts. But he also expressed concern: he is not worried about AI saying things it shouldn’t say, but more worried about it doing things it shouldn’t do.

“就像培养小孩一样，需要给AI注入正确的方向。主动性是重要的范式，但安全问题必须同步解决。” Just like raising a child, you need to inject the correct direction into AI. Initiative is an important paradigm, but safety issues must be solved simultaneously.

Hallucinations Cannot Be Eliminated, But Can Be Priced

Academician Yang Qiang put forward several very enlightening academic perspectives. He quoted Gödel’s incompleteness theorem, pointing out that when a large model as a system cannot prove its own innocence, hallucinations cannot be completely eliminated. Then the real question becomes: how many resources are needed to exchange for how much reduction in hallucinations? This is like the risk-return balance in economics, there is no free lunch. This requires the mathematics community, the algorithm community, and the industry to study together.

He also shared an interesting observation about continuous learning: when multiple Agents are connected in series, the ability will decrease exponentially. However, humans can clear noise through sleep, avoid error superposition, and continuously improve accuracy. This brings new inspiration to the calculation mode of large models, and in the future, it may be necessary to jump out of the existing framework of Transformer calculation.

Yang Qiang proposed a four-stage theory of Agent development, the core dividing standard is whether the definer of goals and plans is human or AI. It is still in the primary stage: the goal is defined by humans, and the plan is also designed by humans. Agent is essentially just a high-level programming language. Future Agents should be able to observe the human work process and use process data to autonomously define goals and plans.

The Efficiency of Intelligence: Tang Jie’s Paradigm Prediction

Professor Tang Jie cut in from the perspective of resource efficiency. He observed that in the past, the resource gap between industry and academia was as high as 10,000 times, but now many schools have thousands of cards, and the innovation gene of academia has been activated.

The more critical judgment is that the investment in large models is getting bigger and bigger, but the efficiency is not high. Data increased from 10T to 100T, the calculation cost soared, but the revenue did not increase proportionally. The marginal benefit of reinforcement learning is also diminishing. In this case, “the efficiency of intelligence” will become the core, using less investment to obtain more intelligent increments. This will inevitably give rise to new paradigms, which may be breakthroughs in continuous learning, memory technology, model architecture, or multi-modal fields.

In terms of Agent, Tang Jie gave a three-dimensional evaluation framework: the size of the value (many Agents generated by early GPT can be replaced by prompts, and the value is insufficient, so they disappeared), the cost (if the API can solve it, the base model will directly integrate), and the iteration speed (the era of large models is about the time window, and it is key to quickly meet the needs within half a year). Zhipu’s bet in the fields of Coding and Agent has seen results, and the call volume continues to grow.

From Personalization to Embodied Intelligence: The Next Step for Agents

Lin Junyang believes that to achieve automation of long-term human work in 2026, two things are needed: one is that Agent must be deeply integrated with autonomous learning, realize self-evolution in long-term work, and autonomously determine task priorities; the other is to improve the ability to interact with the environment. Currently, the interaction environment of Agent is mainly computers. If it can be extended to the real world, such as commanding robots to do wet experiments in the pharmaceutical field, it can greatly improve efficiency.

He predicts that there will be breakthroughs in computer-side Agent applications this year, but Agents combined with embodied intelligence and entering the real physical world may take three to five years.

Regarding the ownership of general Agent opportunities, Lin Junyang quoted Manus CTO Peak’s point of view: the core opportunity lies in the long-tail market. The Matthew effect of head tasks is obvious, and model companies can easily solve them with computing power and data advantages; but the demand for long-tail tasks is scattered and the scenarios are diverse, which is an opportunity for entrepreneurs.

Lin Junyang also mentioned a practical dividend brought by reinforcement learning: it is now much easier to fix problems than before. In the past, data allocation and quality were difficult problems when customers did supervised fine-tuning, but with reinforcement learning, only a small amount of data points, or even no labeling, can be used to quickly train and optimize with only queries and rewards.

The Probability of Chinese AI Leading the World: Optimism and Sobriety Coexist

This is the most tense topic. The five guests gave completely different judgments.

Yao Shunyu (probability “very high”): Taking manufacturing and electric vehicles as examples, Chinese companies can quickly catch up or even partially surpass once they find a track feasible. But he pointed out three problems that must be solved: computing power bottleneck (lithography machines and software ecosystem), To B market maturity (insufficient willingness to pay in China, going overseas is an important direction), and insufficient adventurous spirit.

He compared the differences between Chinese and American research cultures: Chinese researchers are more inclined to do safe things. Once a technology is proven feasible, it can be conquered within a few months, but there is insufficient motivation to explore unknown fields such as long-term memory and continuous learning. OpenAI began to deeply cultivate reinforcement learning in 2022, and domestic companies did not follow up on a large scale until 2023, and there is a clear difference in the depth of understanding. In addition, the domestic industry pays too much attention to lists and numbers, while Claude is not the highest score on many programming lists, but it is recognized by users as the best model to use.

Lin Junyang (probability less than 20%): He believes that this is already a very optimistic judgment. The core gap is that the total computing power of the United States is one to two orders of magnitude higher than that of China, and companies such as OpenAI and Anthropic will invest a lot of computing power in next-generation research, while most of the computing power in China is used for product delivery, and resources for exploratory research are stretched.

But he also mentioned the possibility of “poverty leads to change”: American companies have sufficient computing power, but lack the motivation to jointly optimize algorithms and infrastructure; Chinese companies have limited resources, and may find breakthroughs in the combination of software and hardware, such as combining the next-generation model structure and chip design end-to-end. He shared a regret: in 2021, the Alibaba chip team consulted him on whether the model would still be a Transformer architecture three years later (because chip tape-out takes three years), but the communication between the two parties was not smooth, and they missed the opportunity for collaboration.

Academician Yang Qiang: Taking the Internet as an analogy, it originated in the United States but China quickly caught up, and applications such as WeChat are world-leading. AI is an enabling technology rather than a terminal product, and China has a natural advantage in the landing of To C products. He also cited Palantir’s ontology method and front-end deployment engineer model, believing that this engineering solution is worth learning for domestic AI companies.

Professor Tang Jie: The most confident, the core reason is the rise of the younger generation. He believes that for Chinese AI to achieve leadership, three conditions are needed: more smart people dare to take risks, a better environment to support (reduce internal friction, so that researchers have the energy to invest in innovation instead of being tired of delivery), and insist on deep cultivation on the right track.

A Little Insight

The most valuable part of this conversation is not any single point of view, but the tense answers given by five scholars in different positions to the same set of questions. Several insights worth chewing on repeatedly:

• The gap between strong models and ordinary models will continue to widen in the To B market, which is driven by the simple fact that “you cannot predict which tasks will go wrong”
• Autonomous learning is not a breakthrough moment in the future, but is already happening in gradual change. Claude Code writes 95% of its own code, which is itself a form of AGI
• Hallucinations cannot be eliminated but can be priced, this perspective transforms the engineering problem into an economic problem
• What Chinese AI lacks most is not talent and execution, but the willingness to take risks in unproven directions. “As long as the technology is proven feasible, it can be conquered in a few months” precisely illustrates the structural problem of strong follow-up ability and weak pioneering intention
• Lin Junyang’s judgment of less than 20% is the most sober: a computing power gap of one to two orders of magnitude cannot be compensated by cleverness, but the “poverty leads to change” style of software and hardware combination may be the only asymmetric opportunity
• The missed collaboration opportunity of the Alibaba chip team in 2021 reveals a deeper problem than computing power: chip tape-out requires three years of prediction, and model architecture evolution also requires prediction, but there is a lack of effective dialogue between the two teams