孟子立: 15岁，清华，HKUST，教授，家庭计算中心，云游戏，WiCi，网络系统，创业

[00:00:00] I was in a student club at the time. We went to work on a construction site. But they wouldn't let me work. They said I was underage. Wow, you're so quick with the jokes. You're leaving me, the host, with nothing to say. Haha, exactly. Your insight and your sense of timing can often be far more important than simply working hard. Right. Society is changing so fast. I study every day and still can't keep up. The standards really may be different every single year. How could students put on glasses, look at a question, and—just like in a sci-fi movie— have the answer instantly appear on the lenses? All the computing could be offloaded to your device. And we hope it won't be limited to robots. Every device could offload its computing. You have to swallow your pride and ask people for help. If you don't ask, no one will volunteer. Interviewing 100 Tsinghua alumni in Silicon Valley. This is Uncle Moon. Today, we're joined by Zili Meng. He entered Tsinghua at 15, completed both his bachelor's and PhD there, and at 23 became the youngest assistant professor at HKUST. This year, he also founded his own company, WiCi. Welcome, Zili Meng. Hi, hello. Uncle Moon. Hello, everyone. I'm delighted to be here to chat about my journey so far. Let's get straight to it. [00:01:00] First, how did you manage to enter Tsinghua at 15? It wasn't actually Tsinghua's special class for gifted students. My middle and high school had their own accelerated-track program. From elementary through high school, there was a special admissions route. They recruited students in fifth grade. Once admitted, you skipped the middle-school entrance exam and the high-school entrance exam. You just kept studying all the way through the Gaokao. So basically, what we saved was the year of the middle-school exam and the year of the high-school exam. Otherwise, we learned much the same material. I also started school a little early, so that's probably how it happened. Of course, I heard that the program was shut down last year. Apparently, it no longer recruits many students. The education bureau opposed selecting top students early, so programs like this may be less common now. But they still existed in our time. Right. We recently saw that Professor Hong Wang, who won the Fields Medal, also entered Peking University at 16, right? Your stories have something in common. I can't compare myself with her. Professor Wang's talent is obvious to everyone, and her achievements have earned worldwide recognition. All I can say is that I aspire to learn from her. Over the past while, we've interviewed many Tsinghua alumni. Some had photographic memories, and some could read ten lines at a glance. They all seemed to have [00:02:00] some kind of superpower that got them into Tsinghua. Throughout your education, what made you most different from everyone else? That's a hard question. Even at Tsinghua, I think people with genuine superpowers usually end up on The Brain. They're the people you see on TV. Maybe they can do those things. But I think most students simply studied diligently, practiced problems seriously, and earned admission one step at a time. I don't think there was anything special. When I got into Tsinghua, I was only one point above the cutoff. The cutoff was 682, and I scored 683. I somehow stumbled my way in. So I'd call it a lucky coincidence. See? Successful people really are this humble. There's another theory I heard on a podcast recently: people who get into Tsinghua or Peking University have a kind of dark force inside them. They're extremely driven to win. It supposedly has nothing to do with their upbringing or anything else. They're simply born highly ambitious, with exceptionally strong internal drive. Do you agree? I agree with half of it. Let me put it this way: if you mean strong internal drive, I think most Tsinghua students do have quite a lot of it. If you had no internal drive, in China's Gaokao system, [00:03:01] I don't think you'd survive all 18 years. But are they all fiercely competitive or determined to do everything perfectly? Not necessarily. Some classmates around us got into Tsinghua almost effortlessly. That really does happen. So I think everyone may have perfectionist tendencies, but they aren't necessarily hypercompetitive. Overall, it's still a diverse group, right? Yes, definitely. It's very diverse. When you first arrived at Tsinghua, you discovered your roommates were all three or four years older. How did you get along? Yes, that really was the case. As I mentioned earlier, I had been in an accelerated high-school class, so everyone in my class was about the same age. At Tsinghua, it was my first time interacting with classmates older than me. How should I put it? There were indeed some difficulties. To be honest, under Chinese laws and regulations, age can sometimes be a real barrier. For example, other people could open bank accounts, but I couldn't. We also had a social-practice field study. I was in a student club at the time, and we went to work on a construction site. But they wouldn't let me work. They said I was underage, so I wasn't allowed. Things like that really did happen. But we gradually learned to get along, one step at a time. After about six months, I had basically adjusted. [00:04:01] Things worked out well. Most people are kind, after all. Wow, you're so quick with the jokes. You're leaving me, the host, with nothing to say. Haha, exactly. Then tell us more fun stories from Tsinghua. I read that, in your freshman year, you took an upper-level course, and scored only in the seventies in calculus. Can you share that story? Those were actually two separate things. The upper-level course I took was university physics. Mm-hmm. How should I put it? I did physics competitions in high school, but unfortunately I didn't get very far. I only ended up with a provincial first prize. So after I entered college, several classmates said in our freshman chat group, "Hey, let's take that course." If you only got a provincial first prize, you hadn't made it. Making it meant joining the provincial team, then going on to win a gold medal, right? Our cohort had an Asian Physics Olympiad gold medalist and an IPhO gold medalist. Those people had truly made it. They were genuinely incredible. Mm-hmm. So people in the freshman group chat wanted to take sophomore-level university physics. I glanced at the idea and was pretty fearless back then. I had just entered Tsinghua and knew nothing, so I enrolled. That's how it happened. As for calculus, that was really a tradition in the EE department. It wasn't just that I scored in the seventies. Calculus scores in Tsinghua's EE department [00:05:02] are like that every year. It almost serves as a wake-up call. Our professor had failed more than a hundred students two or three years earlier, in the class two cohorts ahead of us. Mm-hmm. So I think it was really part of the adjustment process. Everyone received similar scores, so we weren't too upset. Still, the top students were truly exceptional. Our average was in the fifties, and I scored in the seventies. But the very best students nearly got full marks— 98 or 99. So I think the first lesson was that there is always someone better, and always another mountain beyond the one you see. But for me, I had barely crossed the admission cutoff. I already understood that. So it wasn't a big shock for me. Some students, however, had always been at the top in high school and throughout their education. They might go through a period of emotional adjustment. So entering Tsinghua only one point above the cutoff actually became an advantage. Haha. How should I put it? I had more resilience to setbacks. It depends on whether the result matches your expectations or comes as a surprise. I remember that in my senior year of high school, my grades had been decent at first, but as the Gaokao approached, [00:06:02] the closer it got, the more sharply my scores fell. I remember my final mock exam, one month before the Gaokao. By then, there were only thirty students in our class, and I had fallen outside the top ten. What that taught me was: before you do something, don't set your expectations too high. Then almost any outcome can feel like a happy one. That sounds similar to something Jensen Huang said at a Stanford commencement speech. Lower your expectations, build greater resilience, and both your achievements and your happiness in life may increase. Is that the idea? Absolutely. The further I went, the more I realized it— especially when many things are beyond your control. With more resilience and lower expectations, you can at least make it safely through every low point. We've talked a lot about lowering expectations, but someone in the live comments will surely accuse us of humblebragging. Still, everyone has their own perspective. We also learned that, as an undergraduate, you minored in business administration. What led to that decision? Honestly, it's a little embarrassing. I originally pursued a second degree in business administration. [00:07:02] But I didn't earn enough credits to graduate, so it became a minor instead. That's how the business minor came about. Back in my sophomore year, Tsinghua allowed students to choose minors and similar programs. My experience reflects the fact that I've always loved trying new things. So my background is fairly diverse. I basically wanted to try everything. When a dual-degree option appeared in my sophomore year, naturally I wanted to give it a shot. People in our dorm pursued economics and management, or even dual degrees at the Academy of Arts & Design. Others studied various second languages. Everyone wanted to learn and try new things. I think this was, in a way, our personalities being set free after entering college. Even now, and throughout all the years since then, whenever I encounter something new, I want to try it and find out what it's about. That's also one reason I decided to start a company. Wow, I've met many Tsinghua alumni, but you're the first person to say that Tsinghua set your personality free. Most people say that happened at Peking University. Okay. Let's move on to your PhD journey. After finishing your undergraduate degree, you chose to pursue a PhD. What was behind that decision? How should I put it? I think the main reason was that I joined a lab quite early. [00:08:04] I entered a lab in the second semester of freshman year and began doing research with a professor. I first worked on research related to neural signal processing. Now the live comments will surely say you're working everyone to death. You started research as a freshman! It really happened by chance. Truly, it was a coincidence. Tsinghua's EE department offers a course called Fundamentals of Electronic Circuits and Systems. Every student in Tsinghua's EE department probably remembers that course, because it has quite a reputation. It's difficult and complicated. I remember when it first began— this was pretty funny— people said the professor printed the textbook himself. It was really a set of lecture notes. The course combined analog circuits, digital circuits, and circuit theory. The two textbooks together were this thick. One classmate of ours— we were just joking among ourselves— put a student ID card against them and said they were thinner than one student card. Mm-hmm. He meant they were thinner than the height of a student ID card. At the time, Lei Jun happened to have that famous one-yuan coin joke, so we were riffing on it. Take the joke or leave it. That was the gag. The point was simply that the course covered a huge amount. [00:09:06] At the time, I think I may have enjoyed a challenge. The professor teaching the course posted an announcement on the online learning portal: another professor was recruiting students interested in research. I thought, "That sounds great." Other professors might not take me, but this one was actively recruiting. So why not apply? I went and worked there for three months. That's basically what happened. This was my first research experience, in freshman spring. After three months, I felt that perhaps— I'll explain in detail later why I moved into systems and networking. In any case, from the first semester of sophomore year, I began doing computer-networking research, and I've continued ever since. Right. So pursuing a PhD was a natural choice. Once I started, I found it interesting and genuinely liked it, so I just kept going. I followed my instincts. If you enjoy something, just do it. Mm-hmm. Later, you visited MIT as an undergraduate researcher. Mm-hmm. At that time, I had already read your papers. About half the papers I read were by Junchen Jiang, and the other half were your MIT papers with Hongzi Mao. So our paths were closely connected. Why did you choose MIT instead of Stanford? I think it came down to the advisor. [00:10:06] When students look for summer research opportunities, there is a lot of randomness. I remember an older student telling us— I think he was from the student association— that when you first send cold emails, you might send a hundred and receive only eight or ten replies. Perhaps a few will lead to interviews, and getting one position is already a good outcome. My grades were fairly good. I happened to rank first in my year at Tsinghua, so I was able to connect with several professors fairly smoothly. At that point, it came down to how well the research fit. To be honest, Stanford didn't have that many networking professors. MIT's networking group had a long history. Its roots went back a long way. I especially wanted to join Professor Mohammad Alizadeh's group. They had produced famous work, including DCTCP. They had developed it while at Cisco. So naturally, I wanted to learn from leading researchers and see how they did research. That's why I went. Why do you think those top professors chose you? From their perspective— and now that I'm a professor myself— taking on a summer research student is a very low-cost decision. They may not have specifically chosen me. [00:11:07] They may simply have thought, "This kid looks promising. Let him come and try." They weren't admitting me as a PhD student. A PhD is a long-term commitment. You need funding to pay the student's stipend. For a summer research student, although in our case MIT did pay us as RAs, it was inexpensive— extremely inexpensive. The cost of trying was very low. Let the student come and give it a try. I'd probably do the same now. Once you joined the group at MIT, you did some work using reinforcement learning to improve network performance. That's right. I worked with Hongzi, who was at MIT as Professor Mohammad Alizadeh's PhD student. Honestly, I mostly helped with miscellaneous tasks. My main contribution was running baselines and other experimental support work. But I learned a great deal. It was also the first time I encountered machine-learning research. I had just started my own PhD then, and those papers profoundly influenced me. They used reinforcement learning for cluster management. Today, everyone knows that foundation models and reinforcement learning are also closely connected. Back then, [00:12:07] how did you use reinforcement learning? Could you explain it in the simplest possible terms? We published early. I think the key was simply being early. At that time, developments in networking lagged AI by a year or two. I think we were essentially arbitraging that time gap. We took techniques that were popular in AI, such as A3C and PPO, and tried applying them to networking problems. Meanwhile, many networking researchers were still tackling the field's hardest problems directly, such as congestion control. We simply tried these new methods. At the time, the ideas all came from Hongzi. By the time I joined the group, the ideas had already taken shape. I contributed more on the experimental side. But I think the most important factor was that we started early. Of course, there were many systems challenges. Anyone who works on systems will understand this. You can't simply take an algorithm and expect it to work out of the box. When you scale it up to a large cluster and a real system, all kinds of challenges emerge. You solve them one by one. What we did was the unglamorous, labor-intensive work of fixing each issue. Once that was done, the results followed. Wow, that's almost too honest. [00:13:07] Anyone who has done a PhD can immediately tell that you've truly been through it. You take the most advanced algorithms from machine learning and apply them in your own field. Right. During that process, you inevitably discover where they don't quite fit. Exactly. Then you make those mismatched parts work in the new setting. That becomes your innovation. And you have a strong paper. If you publish early, the paper is also likely to have greater impact, because the problem is still at an early stage. So the biggest lesson I learned from this was that timing matters. Overall, your insight and your sense of timing can often be far more important than simply working hard. Timing is everything. Exactly. Put that in the live comments. Haha. At that point, you had already worked with reinforcement learning. What was your personal view of those algorithms? That's another interesting story. To be honest, AI was extremely popular in 2017 and 2018. Every researcher wanted to try it. Everyone thought it might be the next major trend. I was no exception. Of course I tried it too. [00:14:07] Reinforcement learning was one example. One thing I felt at the time— and my career shows that I didn't stay in AI— as everyone can see— was that AI had a major issue. This applied to reinforcement learning too: everyone was chasing leaderboard rankings. Whether you worked in computer vision or NLP, and perhaps even today, there was a fixed benchmark or a fixed dataset. Everyone tried to raise the accuracy and optimize the metrics. To me, it sounded more like solving a given textbook problem. Networking or systems research takes a different approach. There is no single benchmark and no fixed question. For every paper, you must define the question yourself. As I continued, I gradually felt that this style suited me better. So I chose this direction. My next paper after working in AI was about interpretability. I applied it to networking and published another SIGCOMM paper. That was the focus of the work. This was something I personally didn't particularly prefer about reinforcement learning or the broader AI community at the time. But after working on interpretability for a while, you eventually realize [00:15:07] that people don't really care about interpretability either. Why not? If performance is good enough, if you rank high enough, and the technology solves real-life problems, then it has practical value. Concrete applications will naturally emerge. You just described another common part of the PhD journey. Many machine-learning problems and research projects are about building benchmarks, climbing leaderboards, and solving a given problem. But when you apply those ideas to networking, you must first define the problem— in other words, tell the story. Exactly. The most important PhD training in networking systems is storytelling. It's storytelling. It truly is all about storytelling. I don't think anything else comes close. It depends on which skill you want your PhD to train. Looking back, each field has its moment. Networking used to be the hottest area. From 2000 to 2010, the hottest PhD topic was certainly networking systems. Back then, there were all kinds of video platforms— Tudou, YouTube, Youku— a hundred flowers blooming. But from around 2012 or 2015 until today, machine learning has dominated. [00:16:07] Exactly. Every industry has its own cycle, and every field has its own characteristics. People should choose the kind of PhD they want to pursue. Exactly. That's absolutely right. Huaizheng, I think you put it perfectly. Why? We used to boast that networking-systems researchers could do more than tell stories. They had another ability: low-level programming. We wrote code close to the hardware. We used to write kernel code every day and write driver code. The barrier to entry was high, and the learning curve was extremely steep. Now that AI has arrived, people can simply have a model write it. That skill is no longer quite so special. So storytelling and technical instinct may now be the only two important qualities for a PhD in networking systems. Right: storytelling and research taste. Exactly. That said, regarding the low-level work in networking that we just mentioned, based on what I've observed in Silicon Valley, those skills are still difficult to develop. You can ask Codex or Claude Code to write it, but directing them to produce good code still takes years of practice. That's why, in machine-learning systems today, [00:17:08] the people earning the highest salaries are often those with deep mastery of these skills. I think it's only a matter of time. That's my personal view— just my opinion. It's only a matter of time. Today, programmers no longer write machine code. They don't write assembly, or other low-level languages either. Eventually, higher-level tools such as web front ends will replace lower-level programming languages. Development always moves toward more convenient methods. AI simply may not yet have reached that level. We're in an intermediate stage. But ultimately, these skills will be replaced. Right. Everyone should seriously think about what they should do next. Haha. I saw many reports about you saying— yes, many reports said— that when you first read networking and systems papers, you found them extremely painful. Oh yes. It was incredibly painful. It's hard to compare directly with other fields, but I'll give two examples. Papers at conferences such as NeurIPS or ICML might be six or eight pages long. In networking, every paper was twelve double-column pages. I got a headache just looking at them. Then I talked with some classmates. I studied electrical engineering as an undergraduate, so I knew classmates in communications research. Their papers could be even longer— [00:18:08] maybe twenty pages. But their advantage was that the papers were full of equations. You could follow the math, derive each step, and eventually work through the paper. With our twelve-page papers, you had to enter an entirely new setup and understand its story. You also had to grasp what was truly clever about it. You had to understand why reviewers thought the work was good. How was I supposed to know? I was only an undergraduate. How could I know what reviewers were thinking? Program committees then were highly specialized, filled with strong professors. How could I learn their research taste? As someone without their expertise or experience, how could I acquire their taste? That was the truly difficult part. Could I understand the literal text? Of course. But could I understand why a particular paper was accepted by SIGCOMM? That was the gap. It took a long time to understand. I work in the same field as you. I completed a PhD and started a company, and only this year did I finally understand why. I genuinely feel it took me ten years to understand this question. I completely agree. So throughout your PhD, the central focus of your papers— and your thesis—was the intersection of video transmission and deep learning. Could you explain it [00:19:09] in the simplest possible language? Honestly, deep learning was only a tool. What we did was simple— just one thing: reduce video stalls. When people make video calls, whether on Zoom or WeChat, or WhatsApp, there are times when the network freezes. The conversation no longer feels smooth. That is the problem we addressed. My entire PhD thesis took a top-down approach, gradually eliminating the factors that caused stalls. Anyone familiar with computer networks knows that networks have a layered architecture: the application, transport, network, and link layers. My PhD work essentially moved from the application layer to the transport layer, then to the network layer, and finally to the link layer. I worked all the way down the stack, addressing latency at each layer. That was the central idea. We simply used some machine-learning algorithms and techniques along the way. During your PhD, you chose not to study abroad. Instead, you stayed at Tsinghua. What led you to that choice? How should I put it? To be honest, a large part of it [00:20:09] came down to the era we were in. I began my undergraduate studies in 2015, and applied for PhD programs in 2018. Several things happened at the same time in 2018. First, there was the trade war. Our own summer research applications went fairly smoothly, but many Tsinghua students in the cohort after us had their visas revoked. This was the first year when China–U.S. relations began turning toward friction. At the same time, I had to consider: if I pursued a PhD in the United States, there would be a cost. Anyone who has studied economics knows the phrase, "Rational people think at the margin." For any decision, you must understand its marginal cost and marginal benefit. If I pursued a PhD in the U.S., what would the marginal cost be? I would have to study for the GRE and prepare for a language exam. That might take nearly half a year. And what would the marginal benefit be? People might say MIT or the U.S. offered a better research environment. The marginal cost was clear, but the marginal benefit wasn't clear to me. So I wanted to see for myself. That's why I did a summer research visit. I wanted to determine whether the research itself— [00:21:11] for example, publishing papers— would truly be better there. Around 2017 and 2018, Chinese institutions had published almost no full papers at top networking conferences. You could count them on one or two hands. So I wanted to find out why. After I went, my overall impression was that it was indeed better, but perhaps not enough to justify giving up six months of research to prepare for language exams and other application requirements. That was the tradeoff. As I mentioned, the cluster-scheduling project and the interpretability project were both already underway. So I decided to keep working on them and stayed. Looking back, how should I put it? There may be two lessons. First, when making a decision, consider the marginal cost and marginal benefit. If, at that moment, the marginal cost is lower than the marginal benefit, then it is a reasonable decision. But the second lesson is that you should also consider other factors. One factor we didn't consider at the time was that a new university [00:22:11] would give you a new alumni network. To be honest, once people begin working, they increasingly recognize the importance of alumni networks. But as a student, it's difficult to appreciate that. By staying at Tsinghua, you remain within the same set of connections. Going somewhere new gives you an entirely new network. Those are the two lessons I see in hindsight. From a research perspective, my choice was fine. The development of networking research over the past decade confirms that. At top networking conferences in recent years, Chinese universities may not dominate, but they now have a very significant presence. So the research itself wasn't insurmountable. At the time, I felt that the research atmosphere in my advisor's group in China wasn't very different from the group at MIT. If anything, people in China seemed even more driven. It felt like everyone was tackling one huge assignment, all in on the same project. Many people worked on it together. I really enjoyed that feeling. Wow, the best advertisement for Tsinghua admissions has unexpectedly come from you. Wonderful. Haha. You've just shared a great deal about your thought process. First, you applied principles from economics. [00:23:11] That's an excellent way to think through a problem. You compared the options and realized that staying at Tsinghua would still allow you to do excellent research. Your reasoning was: as long as you produce outstanding work, in the end, your work will speak for itself. Right. Whether you go to MIT or anywhere else doesn't matter. What matters is producing excellent work wherever you are. At the same time, you later reflected that outstanding work alone isn't enough. You also need connections. In life, first you must be excellent, and second, others must recognize your excellence. That was something you hadn't considered then. Exactly. You keep talking about lowering your expectations and similar ideas, but your reasoning and behavior are both exceptionally well grounded. Honestly. I think that came naturally. Studying and pursuing a PhD is a critical life decision. You have to think carefully. It isn't like choosing a course. Students can easily drop one course and add another. That's an easy decision. A PhD clearly requires serious thought. You just said that while preparing for [00:24:11] and pursuing your PhD, you didn't fully consider the importance of connections. Besides that, what is the biggest mistake PhD students tend to make? Another major mistake is doing nothing but research. Only research. To be honest, at least from what I saw at CMU, that isn't a particularly good decision. I recognized this at the time. It was also another reason I chose Tsinghua. Why? During my summer internship, I realized that in the U.S.— at least back then— international students faced visa restrictions. This meant it was difficult to intern during a PhD. Summer internships were manageable; you could use CPT. But at other times, you faced more restrictions. I'm someone who loves exploring. I always want to try something in one place, then go somewhere else and try another thing. So I thought staying in China, without those concerns, might be better. I completed my PhD in four years. Altogether, I may have spent only six months to a year at Tsinghua. I spent most of my time elsewhere. I interned at Tencent for a year, [00:25:11] at Alibaba for six months, and at a startup for several months. I also spent a year at CMU. All of that fit into a four-year PhD. I think this environment gave me more freedom to explore. Looking back, I benefited greatly from it. So what you're saying isn't necessarily that students make a mistake, but that they should remember: don't do research alone. Travel widely and pursue more diverse experiences. Try different ways of living. A PhD is really your final period as a student. Exactly. I repeatedly tell my students this: a PhD is your best opportunity to keep experimenting. Whatever you try, the PhD degree is your safety net. Once you begin working, you may rarely have the courage to leave your comfort zone and try something new. A PhD gives you that freedom. But honestly, I've met many PhD students, and based on my own experience, many universities and advisors do not support students going elsewhere. That situation is very common. To find an advisor who truly supports you, and a team that gives you freedom to explore the wider world, [00:26:12] from my perspective, requires exceptional luck and exceptional ability. That's absolutely true. My advisor was very supportive of me. We'll talk more about that later. You interned at Tencent and many other companies during your PhD. At each place, what was the most important lesson you learned? Each place taught me something different. Let's take them one at a time. All right, I'll go through them. At Tencent, I joined a production team, a business unit. We worked on cloud gaming. The most important lesson I learned there was this: if you want to build something that can actually reach production, you must consider every detail. You have to think through every aspect carefully, because there will always be things your A/B tests don't cover. There will always be things you failed to consider when you first designed the idea. That level of meticulous attention is extremely important. It is different from publishing a paper. For a paper, if something works in one case, you can say it works. But to put something [00:27:12] into production, you must ensure no corner case can break it. That requires craftsmanship— the spirit of an artisan. That was my first major experience. Then I went to Alibaba, where I joined DAMO Academy. That group has since been shut down. I don't know why, but many groups I've joined no longer exist. Honestly, the Tencent group is largely gone too. The group I joined at Alibaba was more research focused. The most important lesson I learned there came from an interesting story. Tell everyone about it. All right, I'll explain. Go on, tell us. We want the story. It was quite interesting. What did we do at Alibaba? We built a Wi-Fi router that could provide more stable latency. When we use Wi-Fi, latency may be fine 99% of the time, but fluctuate badly the other 1%. We built something to address that. We had an idea. It started as a paper that hadn't been published yet— only an early prototype. We wanted to find a company where we could test it. But honestly, [00:28:13] a young researcher can't simply contact Xiaomi or another router company and expect a response. We needed a connection. A senior student I had worked with as an undergraduate happened to be at Alibaba. We spoke, and he said, "Great, come work on it here. Alibaba may have a need for this too," because the company was exploring hardware then. So I went there to work on it. Once there, we found a potential use case. When would Alibaba build a router? We thought it through. After all, you can't expect users to replace their router just because a Taobao livestream buffers. They have no incentive. The business model doesn't work. DingTalk, on the other hand, might create an incentive. But DingTalk was too large; our small experiment couldn't move it. The use case we found was online education in 2021. If an online class kept freezing, parents might be motivated to buy this router. So I went to Alibaba. Why did this teach me that timing matters? Three months after I arrived, China introduced the "double reduction" policy. The online-education industry effectively disappeared, and our product never launched. It had begun as a research project, so we later published a paper that did fairly well. [00:29:13] But overall, that was my biggest lesson from Alibaba. Every one of your stories involves finding connections. You met all kinds of people in different places and pitched the idea to them. Then there was timing— being in the right moment. You saw how good or bad timing can profoundly affect the projects in your life. Exactly. Don't you see? Whether in your PhD or your internships, you repeatedly searched for the right intersection of ability, connections, and timing. You kept looking for those moments. Exactly. That's how I see it. At least now, I feel the same even after entering the workforce. Take finding a job: for most people, it involves chance. So in these situations, ability is still essential. You need a solid baseline that earns people's recognition. But beyond that, other factors become more important. It's all incredibly random. Everyone's job and current circumstances— mm-hmm— when you look forward from the beginning, you attribute them to effort and luck. But looking backward, you see that it was all timing and chance. Exactly. Yes. You also said that at Tsinghua, [00:30:14] the overall atmosphere in your group was already very similar to MIT's. There were no major shortcomings, and it continuously helped you grow. Your advisor was Professor Jun Bi. That's right. I noticed that you still feature him on your personal website, even though he has passed away. Yes. What was Professor Bi's greatest influence on you? I think it was his spirit of scientific research. How should I describe it? After he passed away, the Institute for Network Sciences published an article about him. I believe the title was something like "The Man Who Raced Against Time." It described how extraordinarily hard he worked. Students are generally rebellious. In the past, when I read heroic stories praising an individual, I would usually smile and move on. I didn't feel they contained anything especially important. But when I read the article about Professor Bi, I immediately remembered all our moments together. Every detail was true. For example, he truly could spend most of the week sleeping in his office and go days without returning home. If a deadline was at eight in the morning, he might bring us coffee at 2 a.m. [00:31:15] and another round at 5 a.m. Things like that. That spirit was something I deeply admired. How old was he when he was bringing you coffee? He was 46 when he passed away. So he must have been 44 or 45 at the time. He truly loved the work. Exactly. He worked incredibly hard. So another lesson is to take care of your health. That's absolutely true. Especially today, with so many tempting new waves and trends, everyone should keep working toward their dreams. But never forget your health. Still, perhaps— investors may not want to hear that. Hahaha. I think you're absolutely right. This is a long-term endeavor, whether you're building a company, doing research, or pursuing anything else. It's about thinking long term. Exactly—long-term thinking. Whether entrepreneurship or academic research, make it a way of life. Take it one step at a time, over ten, twenty, or thirty years. Exactly. First, it has to be sustainable. If you work only in bursts, it won't last, no matter what. [00:32:16] Exactly. Mm-hmm. We've now covered your undergraduate and PhD experiences. There was so much packed into them. Incredible. Your journey is extraordinary. Huaizheng, I think your summary was spot on. I said all those things, and you immediately distilled them into "long-term thinking." I thought that was excellent. I could never have found that phrase myself. My wife and I discuss these factors every day: ability, timing, and connections. Ability, timing, and connections. We just covered your undergraduate and PhD years at Tsinghua. After graduating, you became the youngest assistant professor at HKUST, at age 23. I was actually almost 24. I was born in 1999, so I was 23 and change. You had many options then. You had interned at major tech companies and visited both CMU and MIT. Why did you ultimately choose to become a professor? Honestly, my students have asked me the same question. We have to view the decision from the perspective of that time. Banghua spoke here a few days ago about seeing AI's rise in 2022. Unfortunately, I didn't see it. I even tried ChatGPT then. [00:33:16] I spent all of 2022 at CMU, and went through the entire academic job market in the United States. I prepared my portfolio here, submitted all the application packages, and did onsite interviews. Yet I still didn't see the trend. What did I see? 2022 was the pandemic era, and big tech was laying people off. If you read 1Point3Acres or other forums in 2022, every day the news was that one company had laid off a certain percentage, and another would do the same tomorrow. In that situation, did industry have openings? Yes, but could you do research? That was very difficult. Right. I completely understand. My wife and I received our PhD degrees in July 2022. I know exactly how you felt. The landscape was truly barren. Exactly. Industry was bleak then. Honestly, from the beginning of my PhD, I had aimed to join a research lab— an industrial research lab. Somewhere in that direction. Exactly, a place like that. But in 2022, the landscape was barren. Looking back, we now know OpenAI was rising. It would take off. But I couldn't see it at the time. [00:34:17] So that path didn't seem viable. I didn't see it either. Given the situation, why did you choose HKUST and an academic faculty position? That was also largely a matter of chance. I thought it over and spoke with Justin, my advisor at CMU. I told him, "I'd like to try applying for faculty positions. If that doesn't work, I'll do a postdoc and hang on for a while. When the market improves, I'll try again." So I gave it a shot and applied to around twenty or thirty universities. I was fortunate: HKUST offered me a position. As soon as I received the offer, I canceled all my other onsite interviews and accepted immediately. I thought, "Finally, someone wants me. Wonderful!" And off I went, delighted. Why did I accept HKUST's offer? Another factor that attracted me was HKUST's strong entrepreneurial culture. During my interview, here's what struck me. Networking, to be honest, has many legendary researchers. As we discussed, people from the Internet's first twenty or thirty years are giants. But not many of them have work that connects directly with what we do today, because much of that work was constrained by its era. When I interviewed at HKUST, [00:35:17] one professor had co-founded Atheros, the wireless-chip company. It was later acquired by Qualcomm, and he then joined HKUST as a professor. I had used that driver. Everyone who works on wireless knows it. I thought, "Wow!" Of course, you prepare for interviews by researching the faculty members. When I saw his name, I thought it was incredible. I had to give this place a try. After arriving, I found the atmosphere truly diverse and energetic. Some people founded companies, some pursued public visibility, some focused on research, and others on education. Every path was represented. That strongly appealed to me. The way you describe it— you're far too modest. You weren't just scraping by. What are all the students in the live comments supposed to think? Seriously. So from your perspective, the entire process of joining HKUST was that random? Yes, it was quite random. Once HKUST accepted me, I skipped all my remaining onsite interviews. I felt that one offer was enough. But you also said [00:36:17] that HKUST has a strong entrepreneurial culture. Yes. For those of us who haven't spent much time in Hong Kong, it may seem surprising. Is Hong Kong's tech-startup scene really that vibrant? Let me give you an example. DJI was founded by a student of a professor in our department. Oh, I know. You mean Professor Zexiang Li. Professor Li. Exactly. Right. HKUST really is different. Exactly. HKUST and Shenzhen have incubated countless hardware companies beyond DJI. Exactly. Right. At the time, among Hong Kong's top three universities, I didn't even apply to HKU or CUHK. I only applied to HKUST. This was one reason. So when you applied for the faculty position at HKUST and received the offer, had you already considered starting a company? Honestly, definitely not. But I wanted more options. I see. I didn't know what I would choose, but HKUST offered many possibilities. Whatever faculty members choose to do, the university is open to it. Your way of thinking has been remarkably consistent. You choose places where fish can roam the vast sea and birds can soar the open sky— places that give you more options. That's what you seek. Mm-hmm. Yes. You became HKUST's youngest professor at 23. If your PhD students were older than you, how did you interact with them? [00:37:18] We were all around the same age. Every new assistant professor faces this issue. Even if you're older than your students, it's usually only by a few years. The more important question is: when you come to pursue a PhD, what is your goal? We are here to collaborate, and collaboration matters far more than age. What standards do you use to select students? Honestly, my standards change every year, because society is changing so quickly. In 2023, we had only just entered the Copilot era. AI coding wasn't yet especially powerful. For systems researchers, what mattered most? Programming ability. On the "Join Us" section of my website, I wrote that to join our lab, you should at least have completed a project with more than 1,000 lines of code. Many computer-science undergraduates had never done that. So at first, we looked for systems ability. Later, we considered other qualities. As AI advanced, raw systems-programming ability became less important. You may no longer need to write 1,000 lines yourself, but you still need systems intuition. You need to sense, "I think this approach can solve it," [00:38:18] or, "That other approach can solve it." This intuition supports work in fields without benchmarks— open-ended fields where you define the problem yourself. It is an essential quality. I think it has become increasingly important. Every era values different qualities. Society is changing so fast. I study every day and still can't keep up. The standards really may be different every single year. There you go, being modest again. Seriously. I believe most people feel the same way. That's how it is. Every three months, AI feels completely new. The field changes every day. Who could have imagined that someone graduating in 2022 would, by 2025, write all their code by speaking? Exactly. Mm-hmm. So, during your three years as a professor, what was the greatest shift from how you thought as a student? The greatest shift was— this may not be the best way to phrase it— the difference between being the advisor and the student. Once you become a professor, you realize you no longer think mainly about publishing one paper or completing one project. You think more about what [00:39:19] would be interesting to pursue, or what work could create what economists call an externality— specifically, a positive externality— and bring value to other people. Faculty members consider that more often. That's the first difference. The second, of course, is that faculty have to raise funding. Much of a professor's work is similar to entrepreneurship: recruiting projects, applying for grants, submitting proposals, securing funding, and using it to support students. That process differs greatly from being a student. The people you're pitching are no longer your paper's reviewers, but senior experts across the field. So there are two points. First, as you said, you move beyond publishing papers to build your own reputation. Mm-hmm. You begin building a broader territory. Right. You develop a larger, more influential research group. Exactly. Then you must keep that large group continuously funded. Exactly. How would you describe your own mentoring style? Some advisors say, "I set the direction, [00:40:19] and everyone follows it." Others say, "Let's discuss it together. You propose ideas and explore them. I'll help you accomplish what you want." Which are you? I think I'm somewhere in between. Neither approach is exactly what I prefer. The first, where the advisor handles everything, creates a problem: once students leave their advisor, they no longer know what to do. If I devise the idea, write the paper, and plan everything, the student is effectively just an agent. Why recruit students at all? In 2026, you could simply recruit an AI agent to execute the tasks. The student may never develop a systems mindset. What's the problem with the second approach? Students struggle greatly at the beginning. If you start by discussing loosely and leaving them alone, many students' initial test ideas are still immature. Their taste and intuition may not be developed. If you simply let them lead or explore freely, they may fall into many traps. So I adapt to each student. Some students already have strong abilities and good prior training. You can let them go and explore what interests them. [00:41:20] That's perfectly fine. Other students may still feel lost and know very little. They've simply arrived to pursue a PhD and wonder, "How do I publish a paper?" They need relatively clear, specific instructions. The way you explain this has such a sense of quality. For anyone who has completed a PhD, your answer feels incredibly thoughtful. You've considered everything. It's so logical. That's truly how it is. These are all lessons learned the hard way. Exactly. There's another view expressed by some professors. One professor I deeply respect said, "Stop publishing papers. In five years, the paper as a format may no longer exist. We should simply communicate with agents, run experiments with agents, and publish the entire conversation and experimental process for everyone to see." What do you think about publishing papers? I think there are two sides to this. From an academic perspective, I partly agree with that professor. Today, authors use AI to write papers, reviewers use AI to review papers, and readers may use AI to read them. [00:42:20] In that situation, a paper as a text-based medium may no longer be so important. So earlier this year, I submitted a proposal to SIGCOMM, the main community in our field. I suggested: why submit papers at all? Stop submitting papers. Submit the codebase instead. Give every reviewer an official AI agent. Whatever reviewers want to ask, they can ask directly against the repository. If something isn't in the paper but is in the code, then the result exists. If it isn't there, it doesn't. You simply accept the criticism. Authors, meanwhile, should maintain the codebase properly. They should make the README and the repository's contents easier for AI systems to crawl. It would be similar to SEO, but more like generative engine optimization. That might actually be more valuable. So I agree with that part. But papers have another very important value: they help students develop confidence and identity. Getting a paper accepted is exciting. Think about when we first started. When my first paper was accepted, I was absolutely thrilled. My first visit to the U.S. was in 2017, when I was a sophomore. A poster of mine was accepted at a conference held here in the Bay Area— in Santa Clara. I was incredibly excited to come. It wasn't only that others recognized my work. [00:43:21] There was also a very direct benefit. At Tsinghua, the National Scholarship paid RMB 8,000 per year, and each department had only a few recipients. For this week-long conference trip, I received a $100 daily allowance. One week was almost half a National Scholarship. So I received both a financial reward and a sense of achievement. It felt fantastic. I committed fully to this path and began doing research. That value of papers is still irreplaceable. Yes. For that moment, it felt like you gained both recognition and reward. Exactly. Your description made me think of another point: writing papers trains your mind. Mm-hmm. Whether speaking or doing anything else, only when you can explain it clearly in writing have you truly understood it. Mm-hmm. At the beginning of a PhD, people spend a great deal of time training their logical and academic writing. I feel that without that writing practice, your thinking would be less clear. I think that is separate from the paper itself. If the goal is to train writing, [00:44:21] teaching students to write an abstract or introduction may be enough. I see. Right? Tell the story well. Exactly. The storyline is already clear in those sections. Everything after that is mostly formulaic. Exactly. That's true. There's another view. Since you're now a professor, let me share something said by another professor I greatly respect. He believes universities have lost their purpose and no longer wants to remain in one. He's a very prominent professor. What do you think? How should I put it? Last year, we built a case study on using AI glasses to cheat. A journalist interviewed me and asked a similar question. This was my answer. As educators, we see two main purposes for universities: the first is development, and the second is selection. What does development mean? It is the process of taking students from not understanding to understanding. Selection reflects the fact that opportunities after university are always limited. MIT, for example, recruits only a few students from Tsinghua each year. Without a selection mechanism, how would MIT decide whom to admit? When social resources are limited, some kind of selection system is necessary. [00:45:23] Those are the two functions. Professor Hong Wang also said that, according to many online comments, her undergraduate experience involved more selection and less development, while overseas she received more development. I agree that the university's purpose is development. But as AI advances, that function is becoming weaker, especially for the very best students. Top students don't need professors to teach them. Perhaps twenty or thirty percent of students need no instruction. They may even feel that class wastes their time. In my classes, students sometimes work on other assignments. That's perfectly normal. Or they skip the lecture. That's also completely normal. The professor's job is to help the students further behind. Some students know AI is powerful, yet never think to try it. Some students may use AI to copy homework, or wear AI glasses into an exam. But those students are at least trying to solve a problem. Their method may not be accepted by current social norms, but they are attempting a solution. You don't need to worry about those students. They'll find a place in society. Other students see AI sitting right there but won't even register an account to try it. [00:46:24] Those are the students we educators need to reach. We need to develop them so they can keep pace with the times, learn independently, and fully explore things they have never tried before. They must carry that ability into their future work. I understand your point. A university has an entrance gate, and that gate necessarily involves selection. Mm-hmm. Students enter with a certain level of ability. Then the goal is to raise their ceiling— to maximize how much they grow before they leave. Is that what you mean? Or perhaps raise their floor. That's really what I meant. Right. Okay. Wow. The image of a rising great university teacher comes vividly to life. Indeed. After about three years as a professor, you decided to start a company. What motivated that decision? I think it happened naturally. Networking systems is inherently very close to products. [00:47:24] It isn't necessarily like today's new AI labs and similar organizations. Look at my past research. Even during my PhD— remember the Alibaba story— what did I build? A router. That was already something like a product. It was simply a deep-tech product, not a straightforward product. That's the only difference. So we always had various ideas we wanted to bring out and test. This year, we felt our research project had reached the right moment, so we decided to give it a try. We'll discuss later what exactly you're building. The three years since you became a professor in 2023 have coincided exactly with the explosive growth of foundation models. Watching all that excitement outside, did your emotions ever fluctuate while you remained a professor? I'd be lying if I said they didn't. Especially when my former classmates were buying houses, and buying cars, while university faculty still struggled. Of course that affected me. If people from your cohort [00:48:24] had joined certain industry research labs, their current annual compensation might be a hundred times yours. Oh. Honestly, that also comes down to timing. Mm-hmm. Each person's judgment is genuinely valuable. Those earning million-dollar compensation packages earned it through their own ability. Mm-hmm. They dared to go all in at an early stage. As people often say, they had conviction. They dared to trust and believe in it. You can only admire that. Our next question is: what is it like to be a professor and an entrepreneur in Hong Kong? Can you share what your daily life is like? Personally, I feel that being a professor in Hong Kong is actually more relaxed than in mainland China or North America. Let me give a few examples to explain why. When professors apply for grants, connections often play a major role. That can create some less positive impressions of mainland academia. North America has similar dynamics. Various NSF workshops may serve a similar purpose. But Hong Kong has one advantage: [00:49:25] it is too small. Hong Kong is so small that every grant proposal we write is sent around the world for review. Even if you wanted to rely on connections, you wouldn't know whom to connect with. This forces everyone to focus on the research itself. That's one aspect of doing research in Hong Kong that I enjoy. So overall, Hong Kong professors are fortunate in that regard. Entrepreneurship is a different story. Hong Kong, as everyone knows, is primarily a financial center. Neighboring Shenzhen has a much more active tech-startup scene. I think starting a company in Hong Kong is well suited to businesses closely tied to finance. That can work well. Otherwise, you may be better off in the Bay Area, Shenzhen, or Beijing—places with a higher concentration of technology. A higher density of tech, and more adventurous capital willing to fund deep tech. Or, even more importantly, talent that appreciates the value of technology. Mm-hmm. Hong Kong may better appreciate the value of finance. That's a different story. I think your assessment of being a professor and entrepreneur in Hong Kong is extremely objective. [00:50:26] Mm-hmm. Let's begin the next chapter and explore your entrepreneurial journey. First, you previously built AI glasses for taking exams. The project received extensive media coverage. Can you introduce it? Honestly, this was also about timing. We did it early. What exactly did we build? Let me explain for viewers who aren't familiar. In 2025— last autumn— the story was actually simple. The previous spring, I was proctoring an exam. I teach Principles of Computer Networks and Advanced Computer Networks, for undergraduates and graduate students. While proctoring, I noticed a student wearing sunglasses in the exam room. My research is in this area. We work on real-time communication, VR and AR, and many wearable devices. So I wondered, "What kind of glasses is that student wearing? Could they be AR glasses?" I went over to take a look and discovered they weren't. I thought, "All right, no problem. If they aren't, I'll build my own and see whether someone could wear AR glasses into an exam." We began the project around July or August last year. That was when it started. [00:51:27] The central idea was simple: how could students put on glasses, look at a question, and—just like in a sci-fi movie— have the answer instantly appear on the lenses? So someone who had always succeeded in exams built a product that helped everyone cheat. We mainly wanted to prove feasibility. More importantly, if such a product truly works, then we need to reconsider how people should learn and how education should work. The goal wasn't to destroy the existing exam system. Even if we didn't build it, someone else would. It was only a matter of months. YC especially likes founders like this: "What system have you hacked?" Recently, a startup built a system that helps candidates cheat in interviews and raised tens of millions of dollars. The ideas are very similar. Right. The point is, if you don't build it, someone else will. And once you do build it, you may discover value that others cannot see. Exactly. Mm-hmm. We approached it as a research project. It involved many systems challenges. Think about it: with glasses, it's very different from taking a phone photo, uploading the question, and asking GPT to solve it. That is easy. But glasses have a very wide field of view— [00:52:27] up, down, left, and right. The exam paper may occupy only a small part of the image. The angle may be wrong, and the resolution may be poor. So there were genuine research questions involved. We also had to return the answer quickly and align it with the question on the page. That required very low latency. So we treated it as real-time communication research— like the video calling we discussed— involving ultra-low-latency video transmission. We didn't expect its social impact to be far greater than its research impact. But regardless, I believe the research itself was very valuable, especially because Silicon Valley now has many wearable AI devices. Mm-hmm. It's easy to see that the technologies you studied will gradually migrate into those wearable AI devices and eventually generate enormous value. Exactly. Mm-hmm. Were there any other interesting stories from building the product? Honestly, that project— we also considered turning it into a startup. We simply didn't step forward publicly. Why? Investors clearly could see its value. But as we developed it, we realized the bottleneck in the pipeline [00:53:27] was the hardware package. Glasses are highly integrated electronic devices. The challenge is extending battery life. For example, after evaluating several options— including their openness and capabilities— we found only three pairs of glasses that could support our system: Rokid, Meta Ray-Ban Display, and a pair from the startup Brilliant Labs. Those were the options. Only those three worked. But their battery life was terrible. Glasses may have only a 200 mAh battery, which is tiny. So how do you solve that? It clearly wasn't my area of expertise. It wasn't what I was good at. So we stopped pursuing the project. But from a product perspective, looking back, it was valuable to identify a product, find its bottleneck, and understand where your own strengths lie. I found that process fascinating. Failure is also a form of learning. Exactly. Failure teaches you too. That's an impressive level of reflection. While building the glasses, you discovered that— setting battery life aside— they could very clearly help someone cheat. One hundred percent. [00:54:27] No question. You were already excellent at exams, and then discovered that AI in this new era could make cheating incredibly easy. From that perspective, what do you think is the most important ability for learning today? What is the most important capability people should develop? Adaptability. Let's set exams aside. I'll give another example. As a child, I was excellent at remembering routes. I could look at a map or other directions, then travel without checking it. I knew how to get from point A to point B and which streets to take. But once navigation apps appeared, did that ability still matter? Today, everyone uses Google Maps and other navigation tools. So of course it no longer matters. Technological development inevitably causes some existing abilities to lose their social value. What should we cultivate instead? We should focus on how developing those abilities shapes our overall mindset. For example, exam-taking skill may no longer matter. You may not need to memorize answers. But when placed in a new environment, how quickly can you learn new knowledge? That ability remains essential. I completely agree. [00:55:27] This comes back to evolution. Why have humans remained so successful? Life moved from fish in the water to amphibians, continually developing the ability to adapt to new environments. Exactly. Now, with AI changing at such extraordinary speed, human adaptability has become more important than ever. Mm-hmm. Later, you built another product or tool called CallArena. Mm-hmm. What does it do? That also began as research. We saw that many evaluation arenas had become popular. So we built one too. I continued following my previous research storyline: real-time communication. We realized that there were many benchmarks for text interaction with foundation models, image and video generation, and text understanding. There were many benchmarks. But we discovered there was no benchmark for voice or video calls with foundation models. The idea also emerged from an unexpected feature of HKUST. We ran a test. [00:56:27] What are HKUST's office buildings like? Let me explain. The buildings have extremely long corridors, like something from Inception. Very long corridors with dozens of office doors. We created a test case. We opened several general-purpose apps that supported AI video calls, walked from one end of the corridor to the other, and then asked, "How many doors did we just pass?" It was a simple counting question. At the time, most AI systems performed poorly. They probably treated each frame as an independent image uploaded to an LLM or VLM for processing. Across a sequence of frames, a door number might be tiny or missing entirely. The model may not see it. So how could it combine everything into one coherent video? We didn't know either. I don't train foundation models. But we thought we could at least evaluate what counted as good and what counted as poor. So we built CallArena. On the website, users could make video calls with various models and compare their responses. They could decide which was better through A/B testing and help produce a ranking. That really is valuable. Did you build it before or after [00:57:27] the GPT-4o video-call demo? GPT-4o had already launched. It was one of the earliest. Several products in China and overseas had also appeared, including Gemini Live. Those were already available. We tested them all. Our goal was simply to provide a more quantitative metric. I've always felt that is important. And recently, LM Arena— Exactly. —raised $150 million. You previously built DeepQoE. Anyone who studies QoE knows this: every metric— people in QoE all feel this way— every metric's underlying design can influence how an entire field's algorithms evolve. You're giving me far too much credit. I'm a little embarrassed. Enough about me. Let's continue your story. Next came your new company, WiCi. WiCi— it's hard to pronounce. What does WiCi do? What we do is simple. We want to build a wireless GPU that makes it easy for users and all their devices to access GPU computing. When we began the project, we hadn't considered using it for AI. [00:58:27] We were focused on gaming. Video games. The motivation was simple. A game became very popular in 2024. Viewers may know it: Black Myth: Wukong. It was huge. Absolutely. I finished it. I beat the game. It was incredibly popular. At Tencent, I had worked on cloud gaming. But cloud gaming— as people who follow the field know— has faced limits to its growth. We reflected on why. One important factor may be this: cloud gaming, cloud PCs, and cloud phones all share one characteristic: both the front end and back end run in the cloud. The cloud sends video back to the user. That places extremely high reliability demands on the Internet. But the Internet is a best-effort network. It cannot guarantee that reliability. So performance cannot be guaranteed. In 2024, we asked: could we keep the front end local and offload only the computation? That was the project's original motivation. You would still play Steam games on your computer and run GPU applications such as video editing or rendering locally. But somehow, we would intercept the compute instructions— [00:59:27] for example, driver-level function calls or API-level function calls— and capture them, send them to a remote machine, and execute them there. That was the original idea. We had noticed that a laptop simply cannot contain an RTX 5090. A 5090 draws nearly 600 watts and weighs three kilograms. That can't fit in a laptop. And beyond laptops, it's even more impossible for phones. Mobile devices always face limits on power and weight. Their compute capacity is therefore limited too. Under those constraints, how could I use the best GPU? Only by doing what I just described. That's why we started the project. So it began with gaming. Exactly. At its core, it is a wireless GPU. Our goal, in one sentence, is that any endpoint— a phone, laptop, or even glasses— as long as it supports Wi-Fi, can connect to the wireless GPU and believe it has a GPU locally. I had always wanted to ask you something: if you built this, would you need to adapt every application? [01:00:29] You've already answered. You don't need to modify the apps, because each app still believes it is running on a real local GPU. Exactly. We currently use an RTX 5090. To the app, it looks like a local 5090. The app may be surprised: how could a MacBook Air or a lower-end ThinkPad have an RTX 5090? But it doesn't care. It simply sees a 5090. So you've modified the driver layer. You've created a Wi-Fi driver. A Wi-Fi driver plus a GPU driver. That's one way to understand it. All the demanding computation that would run poorly locally is offloaded to this device in the home. Exactly. That was another motivation for the project. How remote is "remote"? Right? In cloud gaming, it means an edge data center. Mm-hmm. But we found that once traffic crosses the public Internet, performance degrades greatly. So our ultimate solution is to make it more like a Wi-Fi router. Within a nearby physical area, a device connects to it over perhaps a single hop and a very short distance. It can then easily use the compute inside the router-like device. This lets us offload [01:01:29] weight and power consumption while still guaranteeing performance. I edit videos all the time, so I believe this product could serve certain users. A video project may contain hundreds of gigabytes. Mm-hmm. Uploading it to the cloud or to some remote Codex environment is simply unacceptable. But if I had a small home data center that could run open-source models to enhance video quality and select and cut the shots, I would buy that product. After all, I'm already a NAS and local-storage enthusiast. Right. You're already a fairly advanced user. Another observation we had was about decoupling productivity from local compute. What does that mean? Today, we use many agents. At first, people used chatbots: a back-and-forth conversation with little context and little memory. Now, agents are proliferating and beginning to process local files. But that processing remains shallow: file names, perhaps accessibility trees, or some metadata. That's it. [01:02:30] In the future, as AI works more deeply with local data, it will need to process more and more local files. For example, Cursor, Claude Code, and Codex have succeeded in coding. No question. What distinguishes every non-coding domain from coding? In coding, every file is plain text. You can process it locally with grep and similar tools. But what about non-text data, such as PowerPoint and Word files, or the videos and images you mentioned? What do you do? Must you create embeddings for every single file? In that case, the embedding overhead would be enormous. Who doesn't have 100 GB of local files? We therefore believe there will be enormous future demand for workloads that must run locally, just as with robots. While you were explaining, I kept thinking about robotics. When I researched the product on your official website, I saw connections to robot vacuums and all kinds of household devices. Your explanation made me imagine a home with several robots: [01:03:30] a cooking robot, a cleaning robot, and many others. They could outsource their "brains"— all their computation— to your device. And we hope it won't be limited to robots. Every device could offload its computing. There are two reasons. First, any mobile device is constrained by weight and power. But if a device isn't mobile, its power consumption can keep rising. Consider the 3090, 4090, and 5090: power consumption rose from 300 to 600 watts. The room for compute growth is far greater than on mobile or on-device chips. If that's true, why can't we give users and their mobile devices the best possible compute? That was our motivation. In Silicon Valley, I've seen many world-model, robotics, and embodied-AI startups. I can sense that integrating our technology with those systems could genuinely accelerate the deployment of robots in homes. That's also what I hope. Today, a robot may carry a Jetson Orin board, or in the future a more powerful Jetson Thor. [01:04:31] But because of power limits, it cannot match an RTX 5090, 6090, or other professional GPUs. So we strongly hope to work with robotics companies and explore: if their available compute became ten times larger, or a hundred times larger, what gains could they achieve? Absolutely. Are you currently discussing partnerships with such companies? We're beginning to contact some. We've had a few preliminary conversations. I don't know what investors will think, but as a pure researcher, I believe this has a real future. I hope you're right. I hope investors and others who watch this video will reach out so we can discuss it seriously. Right. I saw that your website already shows a prototype. Yes. But from my perspective, building a prototype requires a supply chain, assembly, and a long chain of work. As a professor, how did you put that entire chain together? By learning. And, as I said earlier, with help from key people. I truly believe that if you want to accomplish something real, rather than merely follow a daily routine, [01:05:33] you need help from many people in many different areas. First, you must dare to reach out. You have to swallow your pride and ask people for help. If you don't ask, no one will volunteer. The same was true for this project. In my first two years, I wrote twenty proposals in a row and only one was accepted. One out of twenty. It was extremely painful. I eventually learned that if you just sit there, no leading expert or person with resources will help you. You must actively reach out. The same applies here. We didn't have a supply chain. Some investors helped us connect with supply-chain resources. Hardware, software, operations, advertising, marketing, and even our co-founders— all came together during this process. If people believe in the mission and agree that your work is worthwhile, then, as long as you ask, someone will help. I once watched an interview with a prominent leader: "Have the courage to ask for help." Exactly. That's absolutely right. Wow. One thing I especially want to explore with everyone through this podcast is that a person's social feed or interview captures only one [01:06:34] brilliant moment. But behind every success and every tangible result is a continuous process of asking for help, seeking collaboration, and being rejected. Exactly. You need a thick skin. That's the biggest thing I've learned. Or, in today's popular phrasing, you need a small ego. A small ego and strong resilience. Exactly. I recently saw an interview with Nadal— no, not Nadal— an interview with Federer. Mm-hmm. As the world's top tennis player, he won only 54% of his points. Fifty-four percent was enough. He lost the other 46%, repeatedly learning and improving. Exactly. I hope everyone can learn from— never mind, let's not go there. Let's continue and return to your— we've gone off topic. Back to your product. Will you pursue B2B or B2C? Here's how we see it. Long term, we want to build a B2C company. Why? The name WiCi reflects our ambition. We replaced the "Fi" in Wi-Fi [01:07:34] with "Ci," meaning Computing Infrastructure. Ultimately, we want this product to enter millions of homes and replace household Wi-Fi routers. It would act as a compute router deployed everywhere. Naturally, it would enter many households as well as offices. That's what we hope to build. In the short term, however, business strategy and practical considerations require a fast entry point. That entry point may not be the general consumer. It may be more professional users: video editors, as you mentioned, or technically sophisticated enthusiasts willing to try it. I see. You'll begin with specialized consumers, and perhaps later expand toward B2B. People sometimes call this "large C, small B": professional users who are willing to buy and try new products, or, as with the video-editing example, a small video-editing studio— a small business customer. Those are our main targets at this stage. As YC says, first find a small group of devoted power users, [01:08:35] then expand gradually. Exactly. In North America, I've seen supermarkets deploy many systems for price comparison, smart shopping carts, and robots. I think your product could provide compute for them too. Restaurants as well— restaurants have all kinds of food-delivery robots. Exactly. If you think openly, there are countless scenarios. My definition is very clear: any device that can connect to Wi-Fi can connect this way and use a local GPU. But many scenarios don't necessarily mean many applications, and many applications don't mean willingness to pay. As solution providers, we must explore gradually, experiment together with go-to-market, and solve problems through iteration. How long have you been a founder? "Many scenarios don't mean willingness to pay"— how did you reach that insight so quickly? Well, we've simply fallen into enough traps. Wow. This may sound like another humblebrag, but people like you learn incredibly quickly. As the host, [01:09:35] I've been asking you questions and can feel the tremendous scale and speed of your growth. It's truly remarkable. I have another question. From my technical perspective, your product requires systems software, drivers, and hardware. In the future, might you even need custom CPUs or chips to control the system? The scope is enormous. Mm-hmm. Your funding and time will let you focus on only one area. How will you approach the next eighteen months? I see two factors. First, systems researchers are naturally built to coordinate multiple components. Almost every paper we write is like that. But you're also absolutely right: we cannot do everything. We need priorities. As I said earlier, our core advantage lies at the lowest layers: the driver layer and the inference layer— for example, inference kernels, the inference engine, the Wi-Fi driver, [01:10:35] and the GPU driver. That is our greatest technical moat. For everything else, we offload whatever we can and outsource whatever we can. For hardware, we work with ODM partners. We don't have a large hardware team. For applications, we currently work with many developer communities. Take robotics, which we just discussed. Robotics developers understand far better than I do what they could do with more GPU compute. Why should I pretend to know their field better? We can simply invite them to build it with us. So our core focus remains the low-level, infrastructure-like components. I only explained half the name WiCi. The second "i"— the "C" stands for Computing, and the final "I" stands for Infrastructure. What we aim to build is Wireless Computing Infrastructure. Your description makes it sound like you must grow together with communities— robotics communities, OEMs, and all kinds of hardware manufacturers. That presents a major challenge for you as CEO. Yes. You need to develop an ecosystem. But I don't think [01:11:35] you can deliberately manufacture an ecosystem. It emerges only when people find something useful. For now, we simply hope that more use cases will recognize that our technology may be useful. That alone would satisfy us. Perhaps among ten scenarios— our vision is certainly broad— but honestly, if even one of those ten visions proves useful, I will feel we've contributed something to society. Exactly. If you contribute to society and create value, the outcome is unlikely to be bad. Mm-hmm. What if a large company— perhaps a NAS company or Wi-Fi router company— begins copying your product? How would you respond, avoid losing ground, and win the competition? That's a good question. I think almost every startup faces it. Honestly, the most realistic answer is to move quickly. Starting early is our only real advantage. For an established company, entering this area would require a strategic decision. It isn't something a bottom-up team can easily pursue. So we must move fast, drive performance as high as possible, [01:12:35] and build the largest possible lead. Even if others copy us, they may copy our hardware— after all, we use an ODM. We don't build the hardware ourselves. But if our system's performance is far better than theirs, that will be enough. That is our short-term goal. In the long term, as you said, we must rely on ecosystem partners and the value we bring to applications to sustain our moat. But in the short term, performance is everything. Why is this so difficult? Let me explain. The challenge is that, normally, how does a GPU connect to a CPU or a mobile device? Sorry— mm-hmm. Normally, how does a GPU connect to any device? Through PCIe. PCIe provides hundreds of gigabits per second of bandwidth and microsecond-level latency. We are effectively replacing PCIe with Wi-Fi. Both bandwidth and latency are orders of magnitude worse. So what do we do? That is where our deepest technical expertise lies. In our 2024 project, this was the first problem we solved. As I said, we initially focused on gaming. Gaming is one of the most intensive applications. It demands very high PCIe bandwidth and extremely low latency. [01:13:35] So what could we do? We essentially defined a new computer architecture: the CPU and GPU communicate over a weak link rather than a strong one. Has anyone solved this problem historically? Yes. It existed more than twenty years ago, before PCIe. Once PCIe appeared, communication largely stopped being a major problem. So researchers gradually moved away from it. It remains an issue in data centers, but I'm talking about consumer devices. On consumer motherboards, it was no longer a problem. So fewer people studied it. We have effectively reintroduced the problem. That also suggests a path to the solution. We can take decades of knowledge and insight from computer architecture and re-optimize it for this new setting. That's the idea. We can use caching, prediction, scheduling, and so on. Apply all the established techniques, and the system may work. That's what we're most proud of. So in the short term, you'll compete on speed and on the years of expertise you've accumulated in this direction. Exactly. You'll optimize development speed and performance. [01:14:36] In the long term, you'll gradually develop the community. Over the long term, a great product ultimately doesn't win the market through technology alone. You've already realized that too? But in the short term, technical advantages open the market. Exactly. And you must make sure people know about your technology. Right. You have to keep demonstrating it. Exactly. Keep promoting it. Exactly. Even the best product can remain hidden. Especially in this era. I hope everyone watching this video will support Zili. As we've talked, another question came to mind: your supply chain. Mm-hmm. Right. Whether in Shenzhen or elsewhere, how did you build the supply chain step by step? I'm genuinely curious. Honestly, as I said earlier, I didn't understand it either. When we first built the glasses, our biggest problems were hardware and supply chain. Why would a supplier know me, a university professor trying to build this? We had no investors and no mature organization. Our volume was especially small, so it was extremely difficult. Even now, we're still learning continuously. But if you have the determination to do it, and you keep asking for help, [01:15:36] you will find friends with hardware and supply-chain experience. You simply ask them. Perhaps one factory has shipping capacity, and another can serve as your ODM. We connect all the pieces, and eventually make it work. That's Shenzhen's advantage. What is that advantage? Well, suppose I want to build hardware. Our device contains a GPU. Inside the box, we may need to handle all kinds of cable compatibility, board form factors, and so on. In Shenzhen, you can iterate quickly and directly find the right person in the supply chain. If a cable doesn't fit or there isn't enough space, a courier can deliver a new cable model within an hour. For our development, that makes hardware iteration extremely efficient. The entire product contains many components. How do you manage them all and manage the overall process? We actually don't have that many components, because we aren't a hardware company. Our main work is in drivers. From a component perspective, you can think of it as a mini PC plus a GPU plus a Wi-Fi router. [01:16:36] Combine those, and you're nearly there. We won't build very low-level components such as our own CPU. So for us, as a startup, it is relatively approachable. It also gives us an opportunity to understand the supply-chain system. Roughly how long did it take to assemble the supply chain? Honestly, I won't pretend otherwise: we are still optimizing and iterating. We can now say we have a system that can provide reasonably reliable supply. But could it improve? There is no endpoint. If you ask large hardware manufacturers, their supply chains are also continuously evolving and upgrading. Take Apple. Everyone knows about the recent iPhone developments. Supply-chain migration and relocation are always happening. This isn't just a zero-to-one problem. Going from zero to one is difficult, but what follows is also dynamic. I've never worked in hardware, so I'm especially curious about the process. While building your supply chain, what was the biggest trap you encountered? Honestly, the greatest difficulty was our low volume. [01:17:37] Looking back, that made everything difficult. Okay, let me rephrase. While building the supply chain, what were your greatest difficulty and biggest pitfall? The greatest difficulty was that, as a startup, our volume was small. Low volume means it is difficult to drive anything from the bottom up. So we had to rely on top-down resources. As I said earlier, we depended on help from many people: friends, relatives, and their introductions. Someone might say, "Perhaps this resource could help," or, "Perhaps that resource could." If you dare to ask for help, eventually you can assemble a working system. The biggest pitfall was related. Looking back, we now think the greatest pitfall was fluctuation in GPU prices. It directly affects our product, which includes a top-tier GPU. As I said, we want to give everyone the best compute. In the consumer market, the best available compute is the RTX 5090. Everyone recently saw Nvidia adjust prices again. [01:18:38] That has a major impact on our product. In hindsight, what could we have done? Perhaps stockpiled more units and signed long-term agreements. Wow. Building hardware truly isn't easy. That concludes our entrepreneurship section. Let's look toward the future. Mm-hmm. From the past hour or two, I've found you exceptionally good at learning, reflecting, and growing. Your life seems to have gone almost entirely smoothly. Looking back, what was your greatest hardship or biggest challenge? Honestly, at the major milestones, my journey has been relatively smooth. I won't deny that. But I believe everyone has a similar feeling: even when the major milestones go smoothly, the process between them is never easy. Most of the hardship in my past— perhaps not since childhood, but since entering university— has involved continually developing [01:19:38] my personality. This returns to our first question. When I entered university, I was younger than most people, so my mentality and personality were certainly immature. Through everyone's patience, understanding, and support, I gradually reached where I am today. For that, I am personally very grateful to all my family and friends. That has been incredibly important. Can you explain more about your personality? Imagine your own adolescence. When I entered university, I was still in that stage. That meant many parts of my personality were more outspoken or aggressive. In that situation, how do you learn to align with other people and work toward a shared goal? That isn't easy. Even today, many people may wonder, "Why won't that person work with me?" Often, a large part of the problem is that you haven't aligned with them on your common interest. You haven't considered their perspective. [01:20:39] I gradually learned these lessons over more than a decade. They apply not only to entrepreneurship, but also to relationships and research. When doing research, how do you collaborate with others? How do you reflect on a piece of past work? What was the author's position? Why did the work get published? When writing from your own perspective, you cannot dismiss their work as worthless. They might review your paper. Every PhD student worries about that, especially in a field as small as computer science. So you must put yourself in their position. Where exactly is your work better? Where does that advantage come from? Everything ultimately comes down to seeing the other person's perspective and building a healthy mindset and personality. Only then can you achieve your ultimate goal. You shouldn't be— as we say in Northeast China—quick to take offense. You can't be easily swayed by emotion and make the wrong decision. Unite everyone you possibly can and get the job done. Exactly. Know how to divide the rewards. Yes. Share the rewards. Right. That summary was incredibly sharp. [01:21:39] I couldn't have said it so well myself. You're praising me again. I've spoken with many founders and investors. Everyone— this is difficult to put into words— is a little like a hedgehog. Right. When dealing with other people, you must always find the right distance. Mm-hmm. Then find the right point where you can work toward a shared goal. Mm-hmm. Then capture and share the value together. You just described some of the hardship in your past— the hardship of developing your personality. If you could say one thing to your younger self, what would it be? Even though I suffered, I think those lessons had to come sooner or later. Mm-hmm. Even if I spoke to my younger self now, he probably wouldn't listen. So perhaps I would borrow the final line from a television drama: "Look forward. Don't look back." I love that. Truly. You can't change your former self. Exactly. Your past self made you who you are today. Exactly. There are some losses you must experience yourself. [01:22:41] No matter how often elders, friends, or mentors warn you, until you truly feel the pain, you cannot learn the lesson. Exactly. We may all have very high intelligence, mm-hmm, but in the end, wisdom is accumulated through mistakes. Exactly. That was beautifully said. If both of us died tomorrow, what single piece of advice would you give everyone still striving? What would you say? Just do it. Mm-hmm. Keep moving forward. Don't look back. Let me answer too. If it were me, honestly, my advice would be— two or three years ago, I might have said, "Live in the present," or, "Look toward the future while cherishing who you are today." But now, I think it truly is this simple: every minute and every second, try to keep yourself in a state of enjoyment. If you feel burned out or deeply unhappy, [01:23:42] stop and give yourself room to recover. Otherwise, even if you force yourself onward, the results may not be good. At the very least, you won't enjoy the process. Whatever you do is a long campaign. First, enjoy who you are today. Beautifully said. I truly enjoyed this podcast with you. I enjoyed it very much. One conversation with you is worth years of reading. Thank you. And many years from now, after we have both passed away, what would you want your epitaph to say? That's another excellent question. I hope it would name one defining piece of work. I don't yet know what it will be. But I hope people remember me through that work, rather than remember me as a person. For example— not about you specifically— we may recognize that a particular piece of work is excellent without realizing who its author was. That's the state I hope to reach. People don't need to know who I am. But if one day they see my gravestone and say, "So this was the person who invented WiCi," [01:24:43] that would be enough for me. The achievement matters more than my name. As long as I can create one great thing. Exactly. Contributing to society is enough. Or, put differently, the value of the work itself matters more. That was wonderful. Thank you, Zili Meng. That concludes today's podcast. Thank you, Uncle Moon. I was delighted to speak with you today. These past two hours— oh my goodness— how did we end up discussing metaphysics? I don't even know what to say. I wasn't prepared! Hahaha.