Last night, OpenAI and Anthropic went head-to-head. Experts say that in 2026, agents will be widespread in the industry.
Artificial intelligence is currently in a plateau phase of stepwise development. The returns from the current research path are converging, and the next leap requires a breakthrough in a brand - new paradigm. Meanwhile, industrial applications are maturing at an accelerated pace, and 2026 is expected to be a crucial year for the large - scale implementation of Agents.
Last night, OpenAI and Anthropic almost simultaneously released their latest model updates - OpenAI Codex 5.3 and Claude 4.6. There was no grand product launch, nor any revolutionary narrative. However, in the developer community and the industrial sector, these two updates were quickly interpreted as a clear signal: the capabilities of large models are approaching a phased ceiling, and the industry is collectively seeking new breakthroughs.
If we had to use one word to describe the artificial intelligence industry in 2025, it would be "critical". On the one hand, the general capabilities of large models have reached a relatively high level, approaching or even exceeding the standards of human experts in dimensions such as language understanding, reasoning, and code generation. On the other hand, if we continue to increase the scale and computing power along the existing path, the marginal returns are rapidly converging. Technology is not stagnant, but "where will the next qualitative change come from" has become a common question faced by the entire industry.
What is the direction of the next - generation paradigm breakthrough? Where lies the real gap in the competition between China and the United States? How can Agents move from concepts to real - world industrial implementation? These questions permeate the entire industry, and in 2026, they can no longer be avoided.
Recently, with these questions in mind, we conducted an in - depth interview with Zheng Shuxin, the Deputy Dean of the Zhongguancun Artificial Intelligence Research Institute and an Associate Professor at Beijing Zhongguancun College (hereinafter referred to as "the two Zhongguancun institutions"). Zheng Shuxin believes that artificial intelligence is in a plateau phase of stepwise leapfrogging, and the next leap requires a breakthrough in a new paradigm. He also pointed out that the core gap in the current competition between China and the United States lies not in the technical routes, but in high - quality data and computing power resources.
In the industrial sector, Zheng Shuxin believes that there has always been a time lag between technological breakthroughs and industrial popularization, which is a normal historical phenomenon rather than a failure. Just as the invention of the steam engine did not immediately lead to the large - scale implementation of the Industrial Revolution, the transformation of AI capabilities into large - scale applications also depends on the gradual maturity of supporting systems and product forms. In his view, 2026 will be the year when Agents are implemented on a large scale in real - world scenarios, and new paradigms such as Coding Agents are reshaping the basic logic of traditional software development.
The following is the transcript of the interview, edited and organized by InfoQ:
Introduction: Personal Introduction and Research Background
InfoQ: You have been deeply involved in the field of AI for many years. Could you share your research journey and main work with us?
Zheng Shuxin: I started to get in touch with artificial intelligence more than a decade ago and have been deeply involved in the field of large models ever since. In the early days, I focused on large - scale distributed optimization and built the largest asynchronous distributed training system at Microsoft at that time. Later, I shifted my focus to the research of large language models and proposed training optimization and architecture improvement methods such as Pre - LN, which increased the model training efficiency by about an order of magnitude. These achievements have since been widely adopted by mainstream large models (such as the open - source model gpt - oss by OpenAI).
During the research phase of general models and methods, I proposed the Graphormer architecture, which is now one of the mainstream base models in the field of graph learning. Recently, I have been committed to introducing large models and generative AI technologies into the field of scientific discovery. The proposed molecular equilibrium distribution prediction framework has broken through the bottleneck of traditional biomolecular simulations, increasing the efficiency of molecular dynamics simulations by hundreds of thousands of times. The related achievements have been published in top - tier journals such as the cover of Science and Nature Machine Intelligence.
At the end of 2024, I joined the two Zhongguancun institutions. Currently, I am an associate professor at the college and the deputy dean of the research institute. I am responsible for the research and strategic layout of large - model directions in the AI basic department.
InfoQ: You just mentioned that you are currently responsible for the research on large models in the AI basic department of the two Zhongguancun institutions. The two Zhongguancun institutions shoulder the mission of building the AI innovation ecosystem in Beijing and even the whole country. Could you introduce the core positioning of the two institutions? What role does the AI basic department play in it?
Zheng Shuxin: Beijing Zhongguancun College and the Zhongguancun Artificial Intelligence Research Institute are two sides of the same coin, integrating and developing. They are a new attempt at integrating education, science and technology, and talent, and a second - order form of new - type R & D institutions. Beijing Zhongguancun College shoulders the important mission of cultivating leading talents in artificial intelligence and is an "experimental field" for the national reform of integrating education, science and technology, and talent. The Zhongguancun Artificial Intelligence Research Institute and Zhongguancun College jointly carry out the R & D of future - oriented, industrially valuable, and disruptive artificial intelligence technologies and the industrialization of research results.
The AI basic department undertakes specific technological breakthroughs and direction layout within this framework. Our strategic goal is to complete the key puzzle pieces in the second half of AGI development, output core variables that can truly reshape industrial logic in the industry, and cultivate leading talents with both engineering capabilities and scientific intuition.
Overview of the Overall Development of AI
InfoQ: Standing at the beginning of 2026, what do you think are the most critical issues that need to be solved in the current development of AI in China?
Zheng Shuxin: The development of AI is in a plateau phase of stepwise leapfrogging. The marginal returns along the existing technological path are decreasing, and we need to find the next - generation breakthrough direction. At the same time, AI has two characteristics: it is a technology deeply rooted in the industry; moreover, there is a clear time window for this game, and it is very likely to be decided within 3 - 5 years.
Based on these judgments, I think there are two core issues that need attention. The first is at the strategic level: behind this paradigm competition is the Sino - US technological game. How can we gain the upper hand and develop an independent ecosystem? The second is at the application level: How can AI truly drive GDP growth and achieve high - quality development? Currently, the industry penetration rate of AI is already very high, but its actual contribution to GDP is still limited.
Current Status of AI Technological Development
InfoQ: You just mentioned that the key technological issue is to gain the upper hand in the Sino - US technological game. Could you elaborate on your view of the current development stage of AI technology? What will be the direction of the next - generation technological breakthrough?
Zheng Shuxin: The development of artificial intelligence follows the law of "stepwise leapfrogging". The most recent major leap was brought about by GPT, which introduced the scale law. However, now, the improvement of intelligence has entered a plateau phase, and the returns along the existing technological path are decreasing, which has been confirmed by several recent signs. First, the pre - training paradigm has encountered a bottleneck. The dividends of the scale law are nearly exhausted, the high - quality Internet data available for model training have reached their peak, and the marginal returns of further expanding the model scale have significantly decreased. Second, the post - training paradigm also has limitations. Currently, the industry has generally shifted to the design of refined reward functions. The design complexity of reward functions is now comparable to that of feature engineering in the past, which is essentially fine - tuning within the established framework. Recent research published by Meta also shows that the incremental space of post - training may be more limited than expected. If "Less Structure, More Intelligence" holds true, then it is frankly questionable whether the existing strategies can lead us all the way to AGI.
So, what is the direction of the next - generation breakthrough? It may be to improve the shortcomings of the current AI paradigm and find breakthrough points, such as breaking through the bottlenecks of memory and continuous learning, connecting the paths of learning from experience and self - play, improving the ability to support long contexts, and exploring new training methods for dynamic data. However, it is also possible to explore brand - new technological paradigms, such as hardware - software integrated architectures inspired by neuroscience, new data sources, new modeling methods such as discrete diffusion, and new theories of intelligence and reward function designs. However, the exploration of the next generation is high - risk and long - cycle, and it often has a low priority for commercial companies, as they need to balance short - term performance and shareholder returns. Most universities, although having academic freedom, face real constraints in terms of computing power and engineering resources. Therefore, the two Zhongguancun institutions hope to play a unique role at this point in time, do the difficult but right things, and lay out in both directions of breaking through along the existing route and exploring new paradigms.
InfoQ: Agents were very popular in 2025. Some people understand Agents as the application - layer encapsulation of large models, and some others see them as an implemented application form. How do you view the current development status of AI Agents?
Zheng Shuxin: People generally understand Agents as a research field in technology or an implemented application form. However, in my opinion, Agents are the base models and the main technological route that the industry is betting on to improve intelligence.
Why do I say so? The fundamental reason for the decreasing marginal benefits of the pre - training scaling law is that the high - quality Internet data are approaching the upper limit. One of the current core solutions is to find new data sources - synthetic data. Its essence is search, using pre - trained large models to discover new valuable data in the ultra - high - dimensional language space and relying on this synthetic data to further improve the model's performance. The inference models represented by o1 generate high - quality thinking - chain data in the language space through search and reinforcement learning. Agents, on the other hand, further expand the boundaries of the search space, interact with the environment, and call tools to discover brand - new high - value data, which may have new scaling laws.
InfoQ: In 2026, what technological breakthrough points in the field of AI Agents do you think are the most worth looking forward to?
Zheng Shuxin: Similar to the development direction of the entire AI field, what I look forward to is first, improving the shortcomings of the existing paradigm, and second, new training paradigms.
In terms of improving the existing paradigm, several directions are worth attention. First is runtime learning, which enables agents to continuously learn and improve during operation, rather than relying only on the capabilities acquired during the pre - training phase. Second is the memory mechanism. Agents need to maintain context coherence in long - cycle tasks and effectively store and call historical information. In addition, hallucinations and reliability, next - generation evaluation methods, and the overall usability and intelligence of agent systems are also key topics.
In the exploration of new paradigms, self - iterative training methods and reward mechanisms driven by intrinsic motivation may bring step - change breakthroughs to Agents.
These are also the key layout directions in the field of large models at the two Zhongguancun institutions.
InfoQ: Compared with foreign countries, what do you think are the greatest advantages and disadvantages of China in AI research? What is the "key lesson" we most need to learn in the global AI competition?
Zheng Shuxin: China has a large talent base and a profound mathematical and physical tradition. A large number of engineers have a solid mathematical foundation and excellent engineering implementation capabilities. At the same time, China has a complete range of industries, rich application scenarios, and a huge market scale. This unique ecosystem provides a natural testing ground for the implementation of AI and nurtures strong productization capabilities.
As for the disadvantages, there are currently two core points:
The first is data. Currently, the technological routes between China and the United States have become relatively transparent. The biggest gap between top domestic companies and those in the United States lies in data, which is the main source for improving the intelligence of large models. The United States is systematically collecting long - range, complex, and high - difficulty professional - level data. This type of data is characterized by long reasoning chains, multi - round interactions, and the involvement of multiple tool calls, and the value of a single piece can reach thousands of dollars. This is also the research focus of companies such as OpenAI. Currently, there are specialized companies helping large - scale enterprises collect expert - level knowledge and data in fields such as programming, finance, law, and consulting. It is foreseeable that there will be significant breakthroughs in these professional fields in 2026. We are still relatively lacking in this regard.
The second is computing power. I believe that computing power is the first - principle for improving intelligence - scientific progress depends on diverse exploration, and diverse exploration depends on sufficient computing power. However, we currently face many challenges in this regard: first, the performance of the chips themselves is limited; second, the large - scale networking ability needs to be improved. It is rumored that xAI in the United States already has a cluster of 800,000 H100 - level chips, while the top "six emerging companies" in China basically have around 50,000. In this case, the requirements for us are even higher - we need to design very cleverly and sparingly use the resources to achieve results. However, the United States can currently conduct large - scale, multi - directional parallel explorations.
Current Status of the AI Industry
InfoQ: You previously mentioned that the current problem in the industry is that the industry penetration rate is high, but the actual driving effect on GDP is still limited. From the perspective of the entire AI field, when do you think the real inflection point for the industry's explosion will come?
Zheng Shuxin: Currently, we are indeed facing a situation where technology is running ahead. That is, the model capabilities have reached "doctor - level" intelligence in many fields, but the perception in the industrial sector is still relatively weak, and the driving effect on GDP is limited. However, this is normal because there is a time lag between technological R & D and industrial implementation.
For example, the emergence of the steam engine was a power revolution - it redefined almost all industries such as manufacturing, transportation, and energy. However, it took several decades from the improvement of the steam engine by Watt to the full - scale spread of the Industrial Revolution because a complete set of supporting systems such as railways, factories, and coal supply chains needed to gradually take shape. AI is in a similar stage: the core "power source" has emerged, but to truly reshape the industry, it still requires the in - depth integration of data infrastructure, engineering toolchains, and industry know - how. The difference is that this round will be much faster, perhaps taking only a few years instead of decades.
In fact, this process is accelerating. The breakthrough of Agents in 2025 is a microcosm - more broadly, AI has begun to penetrate into all walks of life. Many scenarios do not require "doctor - level" intelligence. The key is to polish it into a truly usable product.
My judgment is that 2026 will be a crucial year for the implementation of the AI industry. On the one hand, product forms such as Agents and Coding Agents will enable more users to truly use AI in their work and lives. On the other hand, the AI applications in vertical industries are also maturing rapidly, and there are already many well - performing companies in the primary market.
Particularly worthy of attention is the group of white - collar workers and knowledge workers. Currently, the models are approaching doctor - level intelligence in multi - disciplinary fields. Industries such as law, finance, consulting, and research are expected to be the first to release productivity dividends, and the driving effect of AI on GDP is likely to start from here.
InfoQ: Coding Agents are a hot topic of discussion at present. What's your view on them?
Zheng Shuxin: Coding Agents are subverting the traditional paradigm of software development. In the past, the logic was that a team carefully polished 3 products, and perhaps only 1 would succeed. Now, with the help of Coding Agents, an individual can quickly develop 100 products, completely changing the probability and path of success.
I use tools like Codex every day and often run multiple tasks in parallel. Right now, there are 4 Codex Agents running on my computer simultaneously, helping me complete various tasks. Many projects that were just ideas in the past can now be quickly turned into runnable products.
What excites me more is that this ability can be quickly replicated to people with no foundation. I have opened a course on AI Agent programming at Beijing Zhongguancun College, with the aim of "writing code from scratch with zero frames". About half a month ago, Stanford also launched a similar course with the concept of "not writing a single line of code throughout", which coincides with my idea.
The course lasts only four half - days, and the students come from various majors such as physics, materials, and finance. Many of them have no programming foundation. However, at the end of the course, all groups presented runnable demos: some turned Deep Research into "Deep Research with fact - checking"; some transformed the voice - dialogue GPT into a "version