AI is in a plateau phase of stepwise development. Returns from the current research path are converging, and the next leap will require a new paradigm. At the same time, industrial applications are maturing quickly, and 2026 could become a pivotal year for large-scale Agent deployment.

Last night, OpenAI and Anthropic released their latest model updates almost simultaneously: OpenAI Codex 5.3 and Claude 4.6. There were no loud launch events and no sweeping claims of disruption, but developers and industry observers quickly read the two updates as a clear signal: large model capabilities are nearing a stage-level ceiling, and the industry is collectively looking for a new breakthrough.

If one word captures the AI industry in 2025, it is “threshold.” On one hand, the general capabilities of large models have reached a high level, approaching or even surpassing human experts in language understanding, reasoning and code generation. On the other hand, continuing to stack scale and computing power along the existing path is producing rapidly diminishing marginal returns. The technology has not stalled, but the question of where the next qualitative shift will come from has become one the entire industry must confront.

What is the direction of the next paradigm breakthrough? Where does the real gap in U.S.-China competition lie? How can Agents move from concept to real industrial deployment? These questions run through the industry, and in 2026 they will become impossible to avoid.

With these questions in mind, we recently conducted an in-depth interview with Zheng Shuxin, vice president of the Zhongguancun Institute for Artificial Intelligence and associate professor at Beijing Zhongguancun Academy, referred to below as the Zhongguancun institutes. Zheng believes AI is in a plateau phase before its next stepwise leap, and that the next leap will require an entirely new paradigm. He also argues that the core gap between China and the U.S. today is not the technical route, but access to high-quality data and computing power resources.

On the industry side, Zheng said there is always a lag between technical breakthroughs and broad adoption. That is a historical norm, not a failure. Just as the invention of the steam engine did not immediately bring about the full-scale industrial revolution, turning AI capabilities into mass-market applications also depends on the gradual maturity of supporting systems and product forms. In his view, 2026 will be the year Agents land at scale in real-world scenarios, while new paradigms such as Coding Agents are already reshaping the basic logic of traditional software development.

The following is a transcript of the interview, edited and organized by InfoQ:

Opening: Personal Introduction and Research Background

InfoQ: You have worked deeply in AI for many years. Could you share your research journey and main areas of work?

Zheng Shuxin: I began working with artificial intelligence more than a decade ago and have long focused on large models. In the early days, I worked on large-scale distributed optimization and built what was then Microsoft’s largest asynchronous distributed training system. I later shifted to large language model research, proposing training optimization and architectural improvements such as Pre-LN, which improved model training efficiency by roughly an order of magnitude. These results were later widely adopted by mainstream large models, including OpenAI’s open-source model gpt-oss.

During my work on general models and methods, I proposed the Graphormer architecture, which is now one of the mainstream foundation models in graph learning. More recently, I have focused on bringing large models and generative AI into scientific discovery. I proposed a molecular equilibrium distribution prediction framework that broke through bottlenecks in traditional biomolecular simulation, improving molecular dynamics simulation efficiency by hundreds of thousands of times. The related work appeared on the cover of Science and in top journals such as Nature Machine Intelligence.

At the end of 2024, I joined the Zhongguancun institutes. I am now an associate professor at the academy and vice president of the institute, where I lead research and strategic planning for large models in the AI Foundations Division.

InfoQ: You just mentioned that you are now responsible for large model research in the AI Foundations Division at the Zhongguancun institutes. The institutes carry a mission to build Beijing’s, and even China’s, AI innovation ecosystem. Could you introduce their core positioning? What role does the AI Foundations Division play?

Zheng Shuxin: Beijing Zhongguancun Academy and the Zhongguancun Institute for Artificial Intelligence are two sides of the same entity, developing in an integrated way. They are a new attempt to integrate education, technology and talent, and a second-order version of a new type of R&D institution. Beijing Zhongguancun Academy has an important mission to train leading AI talent and serves as a “test field” for China’s integrated reform of education, science, technology and talent. Together, the Zhongguancun Institute for Artificial Intelligence and Zhongguancun Academy carry out future-oriented, industrially valuable and disruptive AI technology R&D, as well as commercialization of research results.

Within this framework, the AI Foundations Division undertakes specific technical challenges and strategic planning. Our strategic goal is to complete the key missing pieces in the second half of AGI, deliver core variables that can truly reshape industry logic, and train leading talent with both engineering capability and scientific intuition.

Overall AI Development Overview

InfoQ: Standing at the beginning of 2026, what do you think is the most critical issue China’s AI development needs to solve?

Zheng Shuxin: AI development is in a plateau phase before a stepwise leap. Marginal returns along the current technical path are declining, so we need to find the direction of the next-generation breakthrough. At the same time, AI itself has two characteristics: it is a technology rooted in industry, and this competition has a clear time window, with the outcome very likely to become clear within three to five years.

Based on those judgments, I think two core issues deserve attention. The first is strategic: behind this paradigm competition is the U.S.-China technology contest, and the question is how we can seize the initiative and develop an independent ecosystem. The second is application: how AI can truly drive GDP and enable high-quality development. AI’s penetration across industries is already high, but its actual contribution to GDP remains limited.

Current State of AI Technology

InfoQ: You just said the key technical issue is gaining the initiative in the U.S.-China technology contest. Could you expand on how you see the current stage of AI technology development? Where will the next-generation breakthrough come from?

Zheng Shuxin: AI development follows a pattern of “stepwise leaps.” The most recent major leap was the scaling law brought by GPT. But now, improvements in intelligence have entered a plateau, and returns along the existing technical path are declining. Several recent signs have confirmed this. First, the pretraining paradigm has hit a bottleneck. The dividends from scaling laws are nearing exhaustion, high-quality internet data available for model training has peaked, and the marginal returns from further expanding model scale have fallen sharply. Second, the post-training paradigm also has limits. The industry has broadly shifted toward refined reward function design, and the complexity of designing reward functions is now comparable to feature engineering in the past. In essence, it is repeated tuning within a fixed framework. Recent research released by Meta also suggests that the incremental space from post-training may be more limited than expected. If “Less Structure, More Intelligence” holds true, then frankly, it is uncertain whether the current strategy can take us all the way to AGI.

So where will the next-generation breakthrough come from? It may come from improving the weaknesses of the current AI paradigm and finding breakthrough points, such as overcoming bottlenecks in memory and continual learning, opening up paths for Learning from Experience and Self-Play, improving long-context support, and exploring new training methods for dynamic data. But it may also require exploration of entirely new technical paradigms, such as neuroscience-inspired hardware-software architectures, new data sources, new modeling methods such as discrete Diffusion, and new theories of intelligence and reward function design. The problem is that next-generation exploration is high-risk and long-cycle. For commercial companies, it is often a lower priority because they must also balance short-term performance and shareholder returns. Most universities have academic freedom, but face real constraints in computing power and engineering resources. That is why the Zhongguancun institutes hope to play a unique role at this moment, doing the hard but right things and making bets both on breakthroughs along the existing route and on entirely new paradigms.

InfoQ: Agents were very popular in 2025. Some people see Agents as an application-layer wrapper around large models, while others see them as an application form for deployment. How do you view the current state of AI Agents?

Zheng Shuxin: People generally understand Agents as either a technical research field or a deployed application form. But in my view, the Agent is the foundation model. It is the main technical route the industry is betting on for improving intelligence.

Why do I say that? The fundamental reason the marginal benefits of the pretraining Scaling Law are declining is that high-quality internet data is close to its limit. One of the core solutions now is to find new data sources: synthetic data. Its essence is search, using pretrained large models to discover new valuable data in an ultra-high-dimensional language space, and then relying on that synthetic data to further improve model performance. Reasoning models represented by o1 generate high-quality chain-of-thought data in language space through search and reinforcement learning. Agents further expand the boundary of the search space by interacting with environments and calling tools to discover entirely new high-value data, where a new Scaling Law may exist.

InfoQ: In 2026, what technical breakthroughs in AI Agents are you most looking forward to?

Zheng Shuxin: Similar to the broader direction of progress in AI, I am looking forward to two things: improvements to the shortcomings of the existing paradigm, and new training paradigms.

In improving the existing paradigm, several directions are worth watching. First is Runtime Learning, which allows agents to continue learning and improving during operation rather than relying only on capabilities acquired during pretraining. Second is memory mechanisms. Agents need to maintain contextual coherence over long-horizon tasks and effectively store and retrieve historical information. Hallucination and reliability, next-generation evaluation methods, and the overall usability and intelligence of agent systems are also key topics.

In exploring new paradigms, self-iterative training methods and reward mechanisms driven by Intrinsic Motivation could both bring step-level breakthroughs for Agents.

These are also key areas of focus for the Zhongguancun institutes in large models.

InfoQ: Compared with overseas players, what do you think are China’s biggest strengths and weaknesses in AI research? In global AI competition, what is the “key lesson” China most needs to make up?

Zheng Shuxin: China has a huge talent base and a deep mathematical and scientific tradition. Many engineers have strong mathematical foundations and outstanding engineering implementation capabilities. At the same time, China has a full range of industries, rich application scenarios and a large market. This unique ecosystem provides a natural test field for AI deployment and has fostered very strong productization capabilities.

As for weaknesses, there are two core issues at present:

The first is data. The technical routes in China and the U.S. have become increasingly transparent, and the biggest gap between China’s top tech companies and their U.S. counterparts is data. This is the main source of intelligence improvement for large models. The U.S. is systematically collecting long-horizon, complex, high-difficulty professional-grade data. This type of data features long reasoning chains, multi-turn interaction and multiple tool calls, and a single piece can be worth thousands of dollars. This is also a focus of companies such as OpenAI. Specialized companies are already helping major firms collect expert-level knowledge and data in programming, finance, law, consulting and other fields. It is foreseeable that there will be significant breakthroughs in these professional domains in 2026. China is still relatively lacking in this area.

The second is computing power. I believe computing power is the first principle for improving intelligence: scientific progress depends on diverse exploration, and diverse exploration depends on sufficient computing power. But we currently face many challenges here. First, chip performance itself is constrained. Second, large-scale networking capability still needs improvement. It is rumored that xAI in the U.S. already has clusters with 800,000 H100-class GPUs, while China’s leading “six little dragons” are basically still around 50,000. Under these conditions, the bar is higher for us. We need especially clever and refined designs and must use resources sparingly to produce results. The U.S., by contrast, can currently carry out large-scale, multi-directional parallel exploration.

Current State of the AI Industry

InfoQ: You mentioned earlier that the current industrial issue is high industry penetration but limited real impact on GDP. Looking across the AI sector, when do you think the real inflection point for industrial takeoff will arrive?

Zheng Shuxin: We are indeed facing a situation where the technology is running ahead: model capabilities have reached “PhD-level” intelligence in many fields, but the felt impact on the industrial side remains weak, and the boost to GDP is limited. But this is normal, because there is a time lag between technology R&D and industrial deployment.

To use an analogy, the emergence of the steam engine was a revolution in power. It redefined manufacturing, transportation, energy and almost every other industry. But it took decades from Watt’s improved steam engine to the full spread of the Industrial Revolution, because railways, factories, coal supply chains and a complete set of supporting systems had to gradually take shape. AI is at a similar stage: the core “power source” has appeared, but to truly reshape industries, it still needs deep integration of data infrastructure, engineering toolchains and industry know-how. The difference is that this round will move much faster, likely in years rather than decades.

In fact, this process is already accelerating. The breakthrough of Agents in 2025 is one example. More broadly, AI has begun penetrating every industry. Many scenarios do not require “PhD-level” intelligence; the key is turning the technology into products that are genuinely usable.

My judgment is that 2026 will be a key year for AI industry deployment. On one hand, product forms such as Agents and Coding Agents will allow more users to actually use AI in work and life. On the other hand, AI applications in vertical industries are also maturing quickly, and many companies in the primary market are already doing quite well.

White-collar and knowledge workers are especially worth watching. Current models are already close to PhD-level intelligence across multiple disciplines. Law, finance, consulting, research and other fields are likely to release productivity gains first, and AI’s pull on GDP may well begin there.

InfoQ: Coding Agents are a hot topic right now. How do you see them?

Zheng Shuxin: Coding Agents are disrupting the traditional software development paradigm. The old logic was that a team carefully polished three products and perhaps one would succeed. Now, with Coding Agents, an individual can quickly develop 100 products, completely changing both the probability of success and the path to it.

I use tools like Codex every day, often running multiple tasks in parallel. Right now, I have four Codex Agents running on my computer at the same time, helping me complete different tasks. Many projects that used to remain at the idea stage can now quickly become runnable products.

What excites me even more is that this capability can be quickly replicated for people with no background. At Beijing Zhongguancun Academy, I launched an AI Agent programming course built around the principle of “writing code from zero frames.” About half a month ago, Stanford also launched a similar course, with the idea of “not writing a single line of code throughout,” which aligns closely with my thinking.

The course lasts only four half-days. The students come from physics, materials science, finance and other disciplines, and many have no programming background. But by the end of the course, every group produced a runnable demo. Some turned Deep Research into “Deep Research with fact-checking.” Others transformed voice-conversation GPT into a “fast-and-slow dual-system version”: the fast system handles immediate responses, while the slow system performs deeper reasoning in the background and then integrates the result. For people with no foundation and different backgrounds to independently build products in four half-days was unimaginable before. That is the value Coding Agents bring.

InfoQ: In your view, which companies or products are doing especially well in Agents?

Zheng Shuxin: This field is very active right now, and Agents are evolving from “conversation” to “getting things done.” If last year people were still discussing concepts, this year we have already seen many real deployment cases that genuinely improve productivity.

Several currently popular products share common traits: they deeply take over systems and files, plan autonomously, execute asynchronously and complete tasks. The open-source Clawdbot, for example, has been called an “AI Jarvis.” Anthropic’s Claude Cowork marks a leap from “chat assistant” to “digital coworker.”

Coding Agents are among the fastest-deploying directions right now. Overseas, Cursor and Claude Code have become standard tools for developers. In China, Kimi K2.5 has delivered strong performance as an agentic model, GLM-4.7 leads among foundation models, and DeepSeek-V3.2, Qwen3 and MiniMax-M2.1 are also good.

InfoQ: You just mentioned some Agent application products as well as some foundation model providers. This touches on an ongoing industry debate: are general-purpose large models only a game for big companies? A previous guest argued that general-purpose large models require huge amounts of people, resources and capital, so they should be left to large companies, while other vendors can look for room to survive in vertical models. What do you think?

Zheng Shuxin: If we are talking about large language models, I tend to think the so-called “room to survive” is more like “room to tell a story.”

The development of general-purpose large models is already very mature. Taking recently released models as examples, Gemini 3 and the GPT-5.2 Deep Think version are both very strong. At present, it is hard to find a vertical model that can outperform these two in any particular field. For legal and education questions, for example, I would be more inclined to use GPT-5.2 or Gemini 3 directly rather than a specialized legal or education model. Although these general-purpose models are more expensive, their performance is already excellent.

If I were developing an application for a vertical domain, I would build directly on GPT-5.2, focus on the user interface, database and basic paradigm, rather than developing a vertical model myself. This view may be somewhat extreme, but it is based on the current state of the technology: the survival space for vertical models is very limited.

InfoQ: But vertical model vendors would say their costs are lower. Is that an advantage?

Zheng Shuxin: I think that argument puts the cart before the horse. First, a model must be able to genuinely solve problems before cost optimization matters. Many embodied intelligence companies are still fixated on cost, but they may not even have found application scenarios that can truly create value. This “cost inversion” mindset is unreasonable.

Research Directions and Talent Development

InfoQ: We have talked a lot about technical directions and industry trends, and you also mentioned many possible frontier explorations. Could you talk specifically about your current research directions and layout? Which direction are you most optimistic about, and why?

Zheng Shuxin: I am responsible for large model research at the academy, and the team is advancing many directions in parallel. One recent project is to make agents “predict the future.”

Every industry, in essence, faces the same challenge: using predictions about the future to support scientific decision-making. That may sound sweeping, but different fields and institutions are all exploring this direction in their own ways. For example, governments need to anticipate market and social feedback before issuing policies. Companies need to estimate industry trends before setting strategies. Financial institutions even use systems to predict U.S. election results or who will win the next game.

The core of this work is to form a closed loop across three stages: information collection, logical inference and simulation. First, agents automatically mine multimodal open-source intelligence across the internet to eliminate information gaps. Then, large models use complex reasoning capabilities for causal modeling and trend judgment. Finally, thousands of agents repeatedly rehearse in virtual environments, producing evolution curves and risk probabilities across different time scales. We have participated in multiple international forecasting evaluations, with our best result ranking second globally, and our latest model is now pushing for first place. Once these three stages are connected, predicting the future is no longer mysticism. It becomes an engineerable scientific decision-making platform.

InfoQ: When introducing the Zhongguancun institutes and the AI Foundations Division, you placed special emphasis on talent development. In the process of tackling hard AI problems, we need large numbers of technical talent. How do you judge whether a young person has the potential to become an outstanding scientist? In your view, what three core capabilities should China’s future AI talent have?

Zheng Shuxin: When I judge whether someone has potential, I value three qualities. First is problem awareness: whether they can discover and define problems themselves, rather than simply waiting for someone else to assign a topic. Second is their response to setbacks. Research is failure 99% of the time, so the key is how they deal with failure. Third is cross-disciplinary curiosity: whether they actively seek to understand things beyond their own field, because many breakthroughs come from intersections between fields.

Outstanding talent should also have three core capabilities. The first is first-principles thinking in mathematics and physics, which is an even more important foundational capability in the AI era. The second is systems engineering capability, meaning the ability to turn an idea from a paper into a runnable system. The third is scientific taste: knowing which problems are worth pursuing. This is the hardest to teach, but also the most important.

InfoQ: With the spread of AI, we have observed a phenomenon: companies, public institutions, universities, primary and secondary schools are all learning AI and programming, but they are also increasingly relying on ready-made tools, from calling APIs and applying templates to directly using AutoML, Copilot and other tools, while paying less attention to mathematical foundations and algorithmic principles. Microsoft CEO Satya Nadella has also said AI is important, but overreliance should be avoided. How do you view this trend of “high tool proficiency but weak scientific foundations”? Are you worried future researchers will become people who only know how to call packages, not innovate?

Zheng Shuxin: My view may be slightly different. I want to explain this through a bit of history in technological evolution.

The earliest programmers had to communicate with computers directly using “0s and 1s,” even punching holes in paper tape to input programs. Later came assembly language, which used simple English commands in place of those 0s and 1s. Then came Python, where writing code became almost like writing English sentences. You can see that every step did the same thing: packaging away tedious low-level operations so people did not have to worry about “how to do it” and could focus on “what to do.”

In this process, every advance came with concerns like the one you mentioned: what if the next generation of programmers no longer understands the underlying principles? But the fact is, precisely because they no longer had to get stuck in low-level details, programmers were able to free up energy to solve more complex and more valuable problems.

Today’s AI tools are the same. They allow researchers to skip many tedious technical steps and focus on the truly important questions, such as proposing new hypotheses, designing new experiments and discovering new patterns. That is the essence of innovation, not writing every line of code by hand.

So my advice is actually this: embrace the most advanced tools boldly, but be clear about the problem you truly want to solve. Tools are the means; the problem is the purpose.

Conclusion

InfoQ: If you were to predict the most influential AI scientific breakthroughs by 2030, what three things would you bet on?

Zheng Shuxin: I would bet on these three directions:

First, AI intelligence surpasses human intelligence and ASI, or artificial superintelligence, is achieved.

Second, AI can independently make discoveries and breakthroughs in scientific research, such as finding a path to cure cancer or solving unresolved open problems in mathematics.

Third, AI enters the physical world and delivers substantial momentum to real industries.