In the first half of 2026, funding in China's AI world models sector totaled approximately RMB 90 billion, growing more than fivefold.

Author and source: Leiphone

In the first half of 2026, money and talent in China’s AI community converged on one goal: world models.

In just the first six months, the domestic world models sector disclosed total funding of approximately RMB 30 billion; when the integrated sector of "embodied intelligence + world models" is included, this figure rises to over RMB 90 billion.

The funding growth rate exceeded five times that of the same period last year.

Billions in capital and a vast pool of talent are flooding in.

An unavoidable question is: With so many people rushing in, how many have truly understood what this thing really is?

Practitioners are seeking direction, investors are looking for people, headhunters are mapping organizations, and big tech strategy teams are studying who will survive.

Something strange is that everyone is asking the same question: Has anyone been able to clearly explain the world model?

Yes.

Today, we conducted an in-depth interview with Professor Zhang Lei, a Distinguished Scientist at IDEA Academy and founder of ShiQi Future, to shed light on this matter.

He is an IEEE Fellow with over 77,000 citations on Google Scholar. His DINO series dominated the COCO leaderboard for five months, and Grounding DINO, along with subsequent DINO-X models, outperformed Google and Meta, becoming a standard reference adopted by leading global institutions such as the Li Fei-Fei team, NVIDIA, and Galaxy General.

In 2025, he founded Vision Future with his team through IDEA incubation, securing nearly 100 million yuan in angel funding within a month.

Behind him stand two AI giants: academic mentor Academician Zhang Bo, one of the pioneers of artificial intelligence in China, and industry mentor Academician Shen Xiangyang, former Executive Vice President of Microsoft and former President of Microsoft Research Asia.

But these are not the main reasons we sought him out.

The reason we reached out to him is simple: in 2026, when everyone was debating “what is a world model,” Zhang Lei was one of the very few who provided a clear, actionable answer that did not compromise with any mainstream view.

From Chief Researcher at Microsoft Research, to Distinguished Scientist at IDEA Lab, and then to entrepreneur—his career path has shifted several times, but his research focus has never changed: “enabling AI to understand objects.”

The world model is the key answer to a problem that has been pursued for over two decades in the physical world.

If you're following the world models赛道 and want to understand whether it's truly a bubble, who will win the race between different approaches, and which types of teams are most worth backing—this conversation is worth your time.

Below is the transcript of the conversation, edited by AI Technology Review without altering the original meaning.

Dr. Lei Zhang, IEEE Fellow, founder and CEO of ShiQi Future, is a Distinguished Scientist at the Center for Computer Vision and Robotics (CVR), IDEA Institute, and an adjunct professor at Hong Kong University of Science and Technology (Guangzhou). Previously, he served as a Chief Researcher at Microsoft Research Asia and Microsoft Headquarters Research. He has published over 200 papers in computer vision and related fields, with more than 77,000 citations on Google Scholar and an H-index of 107. He holds over 60 granted U.S. patents. He was elected an IEEE Fellow for his outstanding contributions to large-scale image recognition and multimedia information retrieval.

The biggest bottleneck for embodied intelligence is not the body, but the brain.

AI Technology Review: Since last year, embodied intelligence has surged, with various robotics companies demonstrating increasingly impressive motion control—Unitree’s robots can run on marathon tracks, and Agi’s robots can perform backflips. However, we’ve observed that real-world applications still seem to be missing one crucial push. What key challenges do you believe must be overcome for embodied intelligence to truly become a reality?

Zhang Lei: The two core issues robots must address are success rate and generalization capability. Success rate refers to whether a task can be reliably completed; generalization capability means the robot can still reliably complete the same task even when the environment or the robot itself changes.

There are four specific challenges:

First, generalization across single scenarios is insufficient. Many teams today can achieve success rates of 70% or even 80% in single scenarios, which sounds like a significant breakthrough. But what does 80% actually mean in real-world applications? It means one failure out of every five operations, still requiring substantial human intervention to cover the gaps—making fully autonomous deployment impossible.

Second, the generality of horizontal scenario expansion is insufficient. First, push single-point capabilities to their limits, then broaden the scope of scenario applications so that the robot can perform more and more complex tasks in more demanding environments. This is the greatest gap preventing robots from transitioning from “lab demonstrations” to “real-world applications.”

Third, there is a lack of deeper capabilities in the "brain." Current deep learning primarily focuses on learning the "behavior" of intelligence, not its underlying "mechanisms" (the principles behind intelligence). An embodied brain requires genuine causal reasoning about the physical world—not just statistical correlations, but an understanding of the logical chain "because A, therefore B"—along with common sense and adaptability. We have only just begun to make progress in this area.

Fourth, it is far from reaching "end-user" level applications. Most robots today still rely heavily on external computing power; the fundamental issue of "not needing human backup" must be resolved before they can be widely adopted across countless scenarios and establish a virtuous cycle of "application-data-iteration."

AI Technology Review: To be honest, opinions within the industry are divided on this issue. The hardware team believes China’s supply chain advantages are clear and that we should first reduce the core cost; the algorithm team argues that the brain—the core technology—is truly the bottleneck. What’s your take?

Zhang Lei: The brain is harder, but also more important.

In terms of hardware, China’s iterative advancements are impressive. Companies like Unitree and AgiBot have refined their motion control through simulation-based reinforcement learning—training control algorithms in virtual environments before transferring them to real robots—while continuously improving key components such as motors and transmission systems, significantly optimizing both cost and reliability. They’ve forged a very solid path. However, the capabilities demonstrated by these robots in marathons or performances are still only roughly equivalent to the basic abilities of animals; they have not yet achieved the instinctive reactions, jumping, evasive maneuvers, or environmental interaction skills that animals possess.

AI Technology Review: You said the "brain" is harder—what makes it harder? Is it insufficient data, or do we simply have too shallow an understanding of intelligence itself?

Zhang Lei: The challenge lies in the fact that we don’t truly understand the underlying mechanisms. The human brain has the greatest distinction and highest level of intelligence compared to other animals. All current breakthroughs in deep learning are based on observing the surface-level manifestations of human intelligence and then replicating them with algorithms. AI is essentially learning human behavior, not human intelligence itself. As Professor Zhang Bo often says, large language models can speak like humans, but they understand language in a way that is completely different from how humans do.

Why was the world model neglected for twenty years, and why has it suddenly become popular today?

AI Tech Review: With the surge in embodied AI, the concept of a "world model" has suddenly become popular. But insiders know this isn't a new term—it was proposed by scholars as early as the 1990s. We’re curious: why did this concept fade into obscurity for two decades, only to become a hot trend today?

Zhang Lei: That’s a great question. World models are not a new concept. Back in the early 1990s, researchers attempted to model environments needed for agent interactions—AI systems that perceive their surroundings and take actions to achieve goals—to improve reinforcement learning algorithms. However, at that time, deep learning had not yet emerged, making it difficult to validate the concept after its proposal.

The concept was truly validated around 2018–2019, when papers titled "World Models" emerged and were applied in gaming environments, as games provide the ideal setting for testing reinforcement learning.

Several years after the development of reinforcement learning, AI had already learned to defeat human players in games. At this point, researchers began to wonder: Could an AI agent autonomously "learn" a world model by continuously interacting with the game environment, encapsulating the dynamics of the environment within the model? This idea was fully validated around 2018 to 2020.

AI Tech Review: So what does this have to do with embodied intelligence today? From what we’ve heard, it seems that embodiment is truly the spark that ignited the world model hype?

Zhang Lei: This connection begins with language models. The industry first developed VLA (Vision-Language-Action models), where robots observe visual scenes and predict their next action, adopting the imitation learning paradigm from language models: while language models perform next-token prediction, embodied VLA models predict the next action based on the current visual input.

But this approach quickly hit a bottleneck. The bottleneck isn't just about data—I believe data is only one aspect; the other is that reinforcement learning hasn't been applied sufficiently. I have an analogy: imitation learning teaches language models how to speak, while reinforcement learning teaches them how to speak correctly.

The intelligence level of the bot and the success rate of its actions also require reinforcement learning.

However, applying reinforcement learning to embodied systems faces two critical obstacles: first, data collection is extremely inefficient, far lower than that of language models, as robots must physically perform each action in the real world and wait for physical feedback; second, the cost of failure is unacceptable—learning to wash dishes might result in broken dishes, and learning to avoid accidents in autonomous driving could require experiencing numerous real-world accidents. This is fundamentally different from language models, which can afford low-cost trial and error in the digital world.

The virtual environments provided by world models are precisely the solution to these problems. If a model can predict “What will the environment become if I perform this action?”, then a robot can mentally simulate countless trials and errors in a virtual world without actually breaking ten thousand bowls.

(Note: Simply put, the logic here is as follows. First, robots require extensive trial and error to learn a task. Second, trial and error in the real world is too costly and dangerous. Third, a world model creates a “virtual training ground” for the robot, allowing it to simulate the consequences of various actions in its mind. Fourth, only the optimal solution is then executed in the real world. This is why world models have suddenly become a focal point for the entire industry at this moment.)

Without actions as input conditions, it’s not called a world model.

AI Technology Review: The industry's definition of "world model" is currently quite confused. Video generation models are called world models, 3D spatial modeling models are called world models, and LeCun’s JEPA (Joint Embedding Predictive Architecture) is also called a world model. So, what conditions must a model at least meet to deserve these four characters?

Zhang Lei: I believe a world model must be action-conditioned—that is, it must be able to answer, “What will the environment become if a certain action is executed?”

Agents can interact with their environment through actions: an agent's actions cause changes in the environment, which then produce a new state and feed it back to the agent, indicating whether the action brought it closer to its goal. This forms a complete feedback loop, where each transition in the environment's state depends on the specific action taken in the previous step.

People often say that a world model is simply changing the language model’s next token prediction to next state prediction, but I believe this definition is not rigorous—it lacks a crucial condition. The correct formulation should be: Action-Conditioned Next State Prediction. Only by incorporating the premise of action does state prediction become meaningful and truly align with the requirements of reinforcement learning.

AI Technology Review: To be honest, many people, including some investors, tend to refer to video generation models like Sora as "world models" in their business plans. But according to the criteria you just outlined, these models clearly lack the dimension of action. We’d like to hear how you break down this issue.

Zhang Lei: The world model we are discussing here originates from model-based methods in reinforcement learning, designed to help agents interact more effectively with their environment. Although world models have many applications, embodied intelligence is currently the clearest and most valuable use case.

In this sense, pure video generation models like the early version of Sora cannot be called world models. The core reason is that they do not model “actions”; they essentially generate a coherent video sequence by predicting how pixels change, without modeling the actions of interaction between robots and their environment. Can they indirectly assist robots? They may learn some patterns of how the world changes, but it is difficult for them to help robots interact more effectively with their environment.

AI Tech Review: What about the recently popular World Action Model? We’ve heard it also deals with actions—is it another type of world model?

Zhang Lei: WAM is not a world model in the sense of reinforcement learning. It uses pixel-level supervision from predicted future frames to improve action prediction and has demonstrated better performance than VLA on certain tasks. However, its modeling is effect-before-cause—it generates future frames first and then back-propagates to infer actions—making it unsuitable for reinforcement learning and thus not meeting the definition of a world model discussed here.

Pixel Pioneers vs. Latent Space Advocates: Both sides are essentially doing the same thing.

AI Technology Review: When it comes to the debate over approaches, this is currently the hottest topic in the field of world models. LeCun is the leading figure of the latent space camp, while Sora represents the pixel-based approach. These two paths seem fundamentally opposed—what are your thoughts?

Zhang Lei: First, we should recognize that the two approaches have commonalities.

In fact, many people now perceive them as highly opposing, possibly because LeCun is very clearly and firmly committed to his views; sometimes, a scholar must articulate their position clearly in order to be remembered outside academia.

But the pixel-based approach also relies on the latent space for operations. Stable Diffusion, Sora, and others first compress images or videos into a low-dimensional representation space before processing. Therefore, the representation space is the more fundamental issue. The real distinction between the two approaches is not whether to use a latent space, but rather what to retain and what to discard once inside the latent space.

The difference lies in this: the latent space approach aims to exclude pixel-level details such as lighting and texture that are not driven by physical laws, helping the model learn the more fundamental patterns behind state changes—but it risks representation collapse, where ten thousand different images map to identical representations, causing the model to lose its ability to distinguish between different states. The pixel-based approach strives for meticulous detail, essentially catering to human vision, as humans are highly sensitive to visual nuances.

AI Technology Review: At this point, we have an intuition: Isn’t human reasoning also taking place in a latent space? When you close your eyes and imagine the consequences of an action, who among us renders it pixel by pixel in their mind?

Zhang Lei: That's correct. If the human brain contains a world model, imagining how actions change the environment involves operating within a latent space, rather than rendering every pixel of the entire scene in the mind. Therefore, using human visual judgment of image quality as a standard for evaluating a world model is insufficient. If the pixel-based approach aims to build a world model, the true goal should not be pixel-perfect realism, but rather ensuring that the generated video sequences adhere to physical plausibility.

(Note: This dialogue discusses the debate between two approaches to AI world models. The pixel-based approach, represented by Sora, focuses solely on visual realism but often produces physics-defying bugs—such as floating cups, object penetration, or candles that won’t extinguish. In contrast, LeCun and Zhang Lei advocate for the latent space approach, which first abstracts visuals into underlying rules and then simulates them, resulting in a better understanding of physical causality. Zhang Lei’s key innovation is incorporating “object structure” into the latent space, enabling the AI to first distinguish independent entities—like cups, tables, and people—and then learn the interaction rules between them. This approach avoids the visual artifacts seen in Sora while being better suited than purely abstract latent spaces for robots to navigate, grasp, and plan actions in the real world.)

AI Tech Review: You and LeCun are following the same latent space approach. But we’ve heard there’s a key difference in your method—you emphasize “introducing object structure.” What’s the core distinction?

Zhang Lei: Our approach shares the same first principle as LeCun’s: the human brain performs counterfactual reasoning (“what if I do this?”) in an action-conditioned manner, using abstract representations rather than re-enacting pixels.

Over the past few years, LeCun has validated this axiom using I-JEPA, V-JEPA, and LeJEPA: prediction should occur in the representation space. However, the complexity of embodied approaches in physical space is extremely high; to apply this method effectively at scale in real-world environments, we believe it is necessary to introduce essential structure into the latent space.

This perspective stems from our proven capabilities. DINO-X, developed under the leadership of our core team, has already solved the problem of “seeing objects” in open-world environments. Therefore, we now treat objects as the fundamental units for world model prediction and planning. Intuitively, physical laws act upon objects; by enabling the latent space representation to understand objects, we can more effectively learn physical laws.

AI Tech Review: Wait, we’d like to follow up with a question. LeCun’s AMI Labs has raised billions in funding and is also aggressively pursuing this direction. What makes you think your improvement of “introducing object structure” can outperform within his larger framework?

Zhang Lei: Because we previously developed DINO-X, an open-ended general object perception model, we have already achieved the world’s best performance in “understanding objects.” Our understanding of the mechanics of vision models—and why these models can produce physically plausible results—aligns closely with our earlier work on general object understanding.

In simple terms, LeCun’s approach is the foundational framework; we’ve built upon it by introducing object structures and enabling it to operate in a closed loop within real-world environments. This “closed-loop” capability is something we’ve developed over the past several years.

Why does the DINO series outperform Google and Meta?

AI Technology Review: Your team’s achievements during the vision large model phase have been extraordinary. Grounding DINO and the subsequent DINO-X have outperformed Google and Meta, and are now standard references adopted by top global institutions such as Li Fei-Fei’s team, NVIDIA, and Galaxy General. We’ve discussed this internally and all felt it was somewhat “unusual”—typically, Chinese teams follow behind international leaders. How did you manage to reverse the dynamic and make overseas teams chase after you?

Zhang Lei: Two dimensions stand out prominently.

Paradigm-wise, we were the first team in the visual domain to achieve SOTA (State of the Art) performance in object detection using Transformer—a deep learning architecture based on self-attention and the foundational model behind large language models. Grounding DINO introduced the language model paradigm into object detection for the first time, extending detection to open-domain scenarios: the model is no longer limited to recognizing only pre-trained object categories but can detect any previously unseen object. To date, it remains the most cited and downloaded open-source model, representing a paradigm-shifting breakthrough.

On the data and engineering side, after Grounding DINO, we shifted toward a closed-source approach, scaling validated ideas further. At the same time, I held very high standards for work quality. Although the team was young and had never delivered work of this caliber, I insisted on maintaining strict thresholds, emphasizing that releases must represent major-version breakthroughs. As a result, it took us about a year to launch Grounding DINO 1.5, followed by another six months to release DINO-X. The extra time invested in refining these versions has indeed earned strong recognition within the industry.

AI Technology Review: What is the relationship between the large vision models you've developed and today's world models?

Zhang Lei: Our work is consistent. Previously, in object perception, our core goal was to provide robots with a foundation for perception; later, many teams used our Grounding DINO to solve embodied environment understanding problems.

Technically, world models are more visually native than language models, and nearly all world model research is not based on pure language architectures like GPT.

In terms of team expertise, our understanding of visual model mechanisms aligns very closely with our previous work on general object understanding. This is why we are confident in continuously iterating on our world model.

The team, consisting of over 20 members incubated from IDEA, is highly stable. Even while developing proprietary models, we have continuously attracted external talent with strong expertise in algorithms and engineering. These 20+ individuals are the “seed” of today’s ShiQi Future.

The evolution path from DINO series models to world models

How can we make AI truly "see" objects?

AI Technology Review: You said existing latent space approaches “don’t truly understand objects”—they don’t know which elements in a scene are independent objects or what relationships exist between them. We’re curious: how do you determine this? How can you tell whether a model genuinely “understands” objects or is merely “pretending” to understand?

Zhang Lei: Understanding itself is difficult to verify directly. No one currently understands how language models truly comprehend language; we can only judge by behavioral outputs. If it continuously generates text that seems indistinguishable from human speech, we say it understands.

World models face similar challenges. It is difficult to directly determine whether the model truly understands physical laws or has genuinely captured object structures. We analyze this through visualization of intermediate steps; if the results align with human intuition, it suggests the model may have captured the object structure.

AI Technology Review: Are there any specific testing methods? Can we conduct an experiment to see?

Zhang Lei: Yes. Counterfactual testing is a good approach—it involves having the model try different actions to see if it can produce physically plausible outcomes. If the model still gets it right even when the trajectory shifts slightly, it suggests the model has learned the underlying rules; if it fails, it likely has only learned statistical correlations.

This requires deep insight. As Professor Richard Sutton said, researchers always feel compelled to help the model by injecting their own experience. Language model researchers once tried desperately to parse syntactic structures, feeding it word segmentation and phrases—but in the end, the simple paradigm of “predicting the next token” still prevailed.

I strongly believe in the power of this simple paradigm to evolve forward. Too much human design may be effective in the short term but will hinder iteration in the long run.

AI Tech Review: In the physical world, there are still some challenging objects—quicksand, flowing water, smoke, and fabric—that lack clear boundaries. We’re concerned that your "object-centric" framework might instead be constrained by the very concept of an "object."

Zhang Lei: We’re also thinking about this ourselves. From a technical standpoint, the simplest approach is to use a mask to segment out the pixels of the object region. Over the past few years, the vision field has made significant progress in object semantic understanding and mask prediction, making it possible to handle non-rigid objects like quicksand and water.

Therefore, starting with Mask is a practical approach. Our accumulated expertise in general object perception surpasses that of any other team.

EgoTwin: Teach a robot's hand the way a human hand does

AI Tech Review: In your technical framework, there's a key concept called "action alignment," which maps actions from different configurations—such as human hands, two-finger grippers, and multi-finger dexterous hands—into the same latent space. This is how we understand it: it’s a bit like translating for robots, enabling them to understand actions expressed in different “languages,” correct?

Zhang Lei: The main inputs for a world model are the visual frame (state) and the action, so understanding the action is crucial.

Action alignment is a fundamental requirement for multi-ontology problems. Aligning human hands with robotic arms is more challenging: the joint positions of a robotic arm have precise XYZ coordinates in physical space, but human hand data is typically captured as 2D video—such as a cooking video that contains only two-dimensional imagery without three-dimensional coordinates. Directly combining them is highly inefficient.

Our approach involves extracting 3D key points to convert 2D video of human hands into a sequence of 3D spatial operations, aligned with robotic arm data within the same physical space. This is what the industry refers to as the collection and processing of Ego data—human operation data captured from a first-person perspective.

Previously, we released EgoTwin, which primarily addresses these data collection issues, and we are currently collaborating with Baidu Cloud’s data collection team.

AI Technology Review: EgoTwin solves the data problem of "how the hands move." But we want to ask further: human videos also contain higher-level information—the intention behind "why the hands move this way"—how do we handle that?

Zhang Lei: Higher-level action intent understanding will be left to world models or VLA models to resolve—but only if the data quality itself is sufficiently high. The raw data is critical. The industry has invested substantial human effort and cost into data collection, yet in many cases, data utilization efficiency remains suboptimal. The core insight we gained from our work on Grounding DINO and DINO-X is that you need algorithm experts to deeply engage in data work; otherwise, the collected data may be unusable during training. Algorithm and data pipelines must iterate in parallel.

Zhang Bo and Shen Xiangyang: Two Mentors, One Common Foundation

AI Tech Review: We’ve noticed that two influential figures are standing behind you—Academician Zhang Bo, one of the pioneers of artificial intelligence in China, and Academician Shen Xiangyang, former Executive Vice President of Microsoft and former President of Microsoft Research Asia. One is more academically oriented, the other more industry-focused—we’d love to know what each of them has contributed to you.

Zhang Lei: Professor Zhang Bo has always been my mentor and role model. I first met Professor Zhang when I entered Tsinghua University’s research lab as an undergraduate—he was the lab director. At the time, I was working on AGVs (automated guided vehicles) and spent most of my spare time in the lab debugging hardware and conducting experiments, during which I often saw Professor Zhang guiding PhD students on robotics-related work. Later, I was fortunate enough to become his PhD student. What impressed me most was his keen sensitivity to new technologies and his strong curiosity. Every time I went to him to discuss experimental results, he would sit down enthusiastically and say, “Why don’t you try modifying this and see what happens?” He was deepening his own understanding through this process. You could see how a scholar’s grasp of algorithms and problems is entirely driven by curiosity.

Another lesson that has stayed with me is his emphasis on “understand things thoroughly”—if you don’t understand fully, you’ll get tricked; only when you truly understand can you grasp the underlying principles. This later became my guiding principle when mentoring students: strive for deep understanding and aim for excellence.

Academician Shen Xiangyang has also had a profound impact on me. During my time at Microsoft, he was my direct supervisor and shaped the course of my entire career: it was he who recruited me from China to the U.S., and it was he who encouraged me to return to Shenzhen. He also demonstrated exceptional insight into industry direction, offering me numerous practical suggestions and valuable experiences on team building and fundraising.

AI Technology Review: We heard that Professor Zhang Bo, at the age of 90, still visited your team for an exchange—this move really moved us just hearing about it.

Zhang Lei: Yes. In November last year, Tsinghua University’s Graduate School invited him to give a lecture in Shenzhen. After finishing his morning session, he dedicated his entire afternoon to our team. We spent the whole afternoon exchanging ideas and had dinner together in the evening. His perspective is extremely broad, and he offered us many valuable suggestions on our general visual understanding and robotics initiatives. Although he is older, his thinking remains sharp, and his logical rigor and depth of understanding are no less than those of younger individuals.

Each wave exceeds imagination, but the waves will come to you on their own.

AI Technology Review: Over the past few years, you have deeply engaged in every wave—from deep learning to large models, and then to embodied agents and world models. We’d like to know: Was there a specific moment when you thought, “This is what I’m meant to do”?

Zhang Lei: The object detection work is very close to the state you described. I’m incredibly proud of this project because it has led to breakthrough advancements across the entire field.

The next world model is also a direction that excites me and one I’m eager to fully commit to. The historical interplay between reinforcement learning and world models over the past few decades, along with the successful paradigm of language models, can both serve as valuable references. I believe this path will have a significant impact on the entire AI industry.

AI Technology Review: Researchers' moments of glory are often discussed, but their lows seem rarely mentioned. Have you experienced a technological low point?

Zhang Lei: Yes. Around 2018, I tried using weakly supervised data for object detection, aiming to perform detection using only image-level class labels and avoiding the need to annotate bounding boxes for each object. The approach sounded very appealing, and I spent a considerable amount of time on it, but I never fully resolved it.

Later, at IDEA, we worked on Grounding DINO, integrating language understanding with detection capabilities, which solved the problem from a different angle. This breakthrough was built on the solid foundation we established through innovations in the DINO detection model, and was also greatly enabled by advances in language models. The breakthrough of Grounding DINO occurred at the end of 2022 and beginning of 2023, precisely when ChatGPT was released. There is a saying: “When data is large enough, intelligence naturally emerges.” With sufficient data, algorithms can become simple enough.

I believe encountering setbacks is normal; the key is to have the determination to relentlessly tackle difficult problems. If one approach doesn’t work, try another—think about it while eating, sleeping, and keep circling around it. Truly good problems linger in your mind for years, and then, at some moment, a transformative breakthrough naturally emerges.

AI Technology Review: Many young people are now starting ventures in world models—for example, Chen Boyuan and Ji Jiaming from Inverse Matrix, and Bai Yinqi from Baize Technology, all of whom are still undergraduates. We find this quite interesting—what are your thoughts on these “next-generation” innovators?

Zhang Lei: This shows that the barrier to entering research has indeed lowered. Before the rise of deep learning, newcomers needed years to build a foundation. Now, open-source technologies have simplified the field—understanding the Transformer is enough to enter it. Young people’s advantage lies in having no baggage; they can more easily embrace new paradigms by starting from first principles. But the truly difficult problems still require seasoned researchers with insight, deep understanding, and the ability to guide direction.

Social trends are always changing, but chasing them just for the sake of chasing won’t keep you ahead. The truth is: no matter which direction things shift, every team always needs the most capable people. If you stay curious and build a strong foundation, the most cutting-edge opportunities will come to you.

An insight

What moved us most in this interview was not Zhang Lei’s dazzling technical achievements, nor even the brilliant young talents he mentored—interns Huang Zheng, Li Xuelong, and team members Zhu Long, Xiao Rong, and Yan Shuicheng from Microsoft’s first-generation face detection algorithm project.

What moved us most was his repeated mention of the phrase: "understand thoroughly."

This word, Zhang Lei has remembered for thirty years.

In the early 1990s, at Tsinghua University, artificial intelligence was a niche field that hardly anyone paid attention to, and Zhang Lei was an undergraduate in the Department of Computer Science tinkering with AGV carts in the lab. The lab director’s guidance of PhD students on robotics-related projects left a deep impression on him, and he later became his student.

The lab director’s name is Zhang Bo, a pioneer of artificial intelligence in China. Whenever Zhang Lei invited him to review the experimental results, he would eagerly sit beside Zhang Lei, gaze at the screen for a moment, and point with his finger: “Try changing this—see what happens.” It wasn’t about checking homework; he genuinely wanted to know what would happen after adjusting the parameters.

What Zhang Bo taught Zhang Lei was later condensed into three words: "Understand thoroughly."

If you don’t understand it, you’ll get scammed; only when you truly understand it will you see how the underlying principles work.

The spirit of China's first generation of researchers is passed down through these small, everyday details.

In 2018, Zhang Lei worked on weakly supervised object detection at Microsoft, and although he couldn't make it work, he didn't change his topic.

If this approach doesn’t work, try a different angle—think about it constantly, even while eating and sleeping, circling around it without pause.

This isn't rhetoric; it's his daily life.

Later, at IDEA, he tackled the problem from another angle with Grounding DINO. The breakthrough came at the end of 2022 with the release of ChatGPT—language models had improved significantly, and data had become sufficiently large.

Zhang Lei used the description: "emerge naturally."

From adjusting AGV robot parameters to decoding world models, thirty years passed. During those three decades, he moved from large-scale image retrieval to general visual understanding, relentlessly pursuing a deep understanding and excellence in every problem, until he returned once again to the goal of “building world models for robots.” As Steve Jobs said, “connect the dots”—every task undertaken with care will never be in vain.

By the end of 2025, 90-year-old Zhang Bo flew to Shenzhen to give a lecture. After finishing his morning session, he spent the afternoon at the office of ShiQi Weilai. A group of seven or eight young people gathered around him, and he spoke for the entire afternoon; they had dinner together that evening, during which he shared stories from his student days and his insights on this field. His thinking was clear, his logic precise—no less sharp than that of the younger generation. Zhang Lei also told his teacher about the work he is currently undertaking.

Academician Zhang Bo at Vision Forward: Zhang Lei updates his teacher on research progress

Thirty years ago, in Tsinghua Garden, a teacher sat beside a student and said, “Try revising this and see how it looks.”

Thirty years later, in a Shenzhen office, a student tells his 90-year-old teacher about the world model he is currently building.

What a moving reenactment of this scene.

People grow old, tides recede, and trends fade, but the determination of Chinese researchers to thoroughly understand each global challenge has never changed. Leifeng.com