The youngest guest's breakthrough in both mathematical modeling and scientific and technological innovation

Introduction

AI tools are becoming increasingly powerful, allowing one person to accomplish what used to require a small team. However, very few have actually turned AI projects into commercial success. Has the barrier to entry lowered or raised?

This episode's guest is the youngest member since the inception of "The Ambitious Salon": Li Mingyang, a sophomore majoring in Artificial Intelligence at North China Electric Power University. While winning an award in the Beijing Mathematical Modeling Competition, he is also leading a national-level innovation project—equipping garbage sorting equipment with "eagle eyes" to identify recyclable waste using spectral information. His conclusion is simple: no matter how powerful a large model is, it cannot understand the implicit experience of experienced workers on the front lines.

This dialogue ranged from mathematical modeling competitions to the practical application of AI, and then to the real-world business balance: how should students and entrepreneurs view the barriers to entry in AI, the industry's hidden data, and the trade-off between technical capabilities and commercial results? The following is a transcript of the dialogue. Let the guest introduce himself first. Tina: Without further ado, welcome our guest today, Ming Yang. Please introduce yourself to everyone. Li Mingyang: Hello everyone, I'm Li Mingyang, currently a sophomore majoring in Artificial Intelligence at the School of Control and Computer Science, North China Electric Power University. I'm very happy to participate in today's "AI Advancement Forum." Academically, I've consistently ranked in the top 10 of my grade and have even won a provincial award in the Beijing Mathematical Modeling Competition. However, more than grades, I'm concerned with how technology can truly solve real-world problems. Today, I'd like to focus on sharing my current area of ​​expertise: the interdisciplinary application of AI and environmental technology. Specifically, as the project leader, I'm leading a team to advance a national-level innovation and entrepreneurship project—a recyclable waste identification system based on spectral information. The starting point of this project is very practical: traditional cameras easily get confused when encountering damaged, oily, or similarly colored garbage during waste sorting, resulting in a high misjudgment rate. Our solution is to equip the machine with "eagle eyes" by introducing spectral analysis technology, allowing the device to not only see the shape but also discern the chemical composition of the substances. We use spectral information for wide-area rapid screening, and then use a high-precision spectrometer for point-to-point verification in complex areas. The two types of data are fused for decision-making, with the main goal of providing an online, real-time identification solution for waste transfer stations or sorting lines. Li Mingyang: In this process, I was mainly responsible for technical research, hardware selection, and algorithm framework design, while also coordinating with students from interdisciplinary fields such as energy and power engineering, and mathematics to advance this project. This experience has given me a lot of insight into the difficulties and logic of applying AI in real-world scenarios. I am particularly happy to be here today to exchange ideas. What exactly is mathematical modeling? Tina: Let me ask on behalf of myself, who has been away from school for many years, what exactly does mathematical modeling mean? Is it a profession or a tool? Li Mingyang: Mathematical modeling is mathematical modeling. The competition format gives you some practical problems. A set of problems usually has four sub-questions, progressively increasing in difficulty. It not only gives you problems but also some real-world data. You need to analyze this data, then abstract the problem into a model—building the model using formulas—and then solve and predict problems. Tina: I understand it as something that visualizes methods and quantifies solutions. Is that a correct understanding? Li Mingyang: Yes. What is truly scarce in the AI ​​era? Tina: Recently, there has been a lot of discussion: AI tools are becoming increasingly powerful, and the barrier to entry for starting a business seems to have suddenly lowered. One person can accomplish the workload of a small team in the past. However, the number of projects that can actually run successfully and become commercial successes is not as high as imagined. What do you think is truly scarce in AI startups? Li Mingyang: As a sophomore, I haven't really entered the workforce or started my own business yet, but through participating in innovation and entrepreneurship projects and mathematical modeling competitions, I've made some observations. Now, one person can accomplish the work of an entire team with the help of AI. In mathematical modeling competitions, it used to require the collaborative efforts of a modeler, a programmer, and a writer. Now, with AI, as long as you have a general idea of ​​the model, AI can refine it according to your ideas, program it itself, and sometimes even run the entire process directly according to your thought process, finally outputting a paper. Previously, three people would have to stay up all night for three days to barely finish; now, as long as you have an idea and work with your teammates, you can finish it in a morning. The same applies to innovation and entrepreneurship projects. To complete the system I just mentioned, three people are far from enough—system building requires collaboration with students from the computer science department, data analysis requires students from the mathematics and physics department, and a project requires collaboration with other teams. But now, with AI assistance, system building and data analysis can both be completed, greatly reducing the need for manpower and resources. Therefore, the barrier to entry for AI startups is indeed getting lower and lower. Li Mingyang: But I think what AI startups truly lack may not be the model itself, nor computing power, but a deep understanding of real-world scenarios and high-quality scenario data. Tina: Yes. Li Mingyang: And the patience to delve into accumulating this data. Take the undergraduate research project I was in charge of—a recyclable waste identification system based on spectral information—for example. Initially, we thought it was simple: using spectral technology for identification would definitely be more accurate than traditional cameras. But after actually researching waste transfer stations and communicating with frontline workers, we found that the real-world scenario was far more complex than the laboratory. The waste was fragmented, contaminated with oil, and some pieces were even stuck together; the conveyor belt moved at high speed, creating a chaotic and unstable environment. At this point, simply pursuing high algorithmic recognition accuracy was meaningless, because the data itself was not clean. We had to redesign the data acquisition scheme and even consider proactively adding lighting and implementing anti-interference measures at the hardware level. Li Mingyang: This made me realize that many AI projects struggle to be implemented because teams lack respect for the implicit understanding of traditional industries. When we conducted on-site research, we saw workers prioritizing waste disposal and considering equipment maintenance and actual costs—these experiences aren't written in academic papers; you have to be on-site and have in-depth conversations with practitioners to understand them. Therefore, I believe what AI startups truly lack is a kind of "cross-disciplinary translator": someone who can understand AI, comprehend industry jargon, and translate business logic into technically solvable problems. At the same time, what's scarce are people willing to do the dirty work and accumulate scenario data. No matter how powerful a large model is, without vertical scenarios and detailed data, you can't create a product that truly solves business pain points. This is something I, as a student entrepreneur, have always reminded myself: don't just get caught up in research, go and experience the smell of garbage firsthand. Tina: That's interesting. Did you discover this bug in the early or middle stages of the project? What prompted you to realize that garbage is difficult to identify? Li Mingyang: In the middle stages. We had already completed the entire system in the lab, but my advisor said that if we really wanted to collaborate with commercial enterprises, we needed to go to real-world scenarios. So, we contacted a garbage sorting plant in Langfang to test it. The result revealed many problems: the garbage was contaminated, resulting in a very low recognition rate; the data processing latency on the high-speed conveyor belt was very high; the environment was dim, not as bright as the lab. We discovered quite a few problems like these.

Tina: So we still need to go deep into the front lines and experience the realities firsthand. This is something worth considering: many people think that having models and tools equals having projects, but the real gap isn't in the tools themselves.

Industry Hidden Experience: AI Can't Understand the Feeling of a Master Craftsman

Tina: Continuing on this topic: Now everyone is talking about "AI + Industry," and many entrepreneurs excitedly enter traditional industries only to find that they contain a wealth of hidden experiences not written in books and documents. In your opinion, what truly needs to be re-understood after AI enters a traditional industry? Li Mingyang: This question hit a nerve with many of our pain points at the time. When working on our innovation project, we learned a valuable lesson from these implicit experiences. I think that after AI enters traditional industries, what we really need to re-understand is the definition of the problem itself—what constitutes a problem being solved. This standard is often held by experienced workers on the front lines, not in papers or operating manuals. For example, we initially designed only a spectral recognition algorithm with a clear goal: to improve the accuracy of material classification. But no matter how high the scores were in the experiments, it failed miserably at the waste transfer station. Later, we learned from a veteran worker with over ten years of sorting experience and discovered that he judged whether an object could be recycled not just by looking at the material—he would squeeze the material to check its hardness, tap it to listen to the sound, and even smell it. For example, a plastic bottle covered in heavy oil stains might be misidentified by a spectrometer due to surface contamination, but an experienced worker can tell by touch whether it contains PET or PC. These tactile, auditory, and physical sensations are typical examples of implicit experience, not written into any system, yet directly determining the efficiency and safety of sorting. Li Mingyang: After that experience, I realized that AI entering traditional industries shouldn't just be about using technology to transform a scenario, but about re-understanding the boundaries of human-machine collaboration. AI cannot replace the implicit experience of experienced workers in the short term, so we need to redesign the system: not to completely replace humans with AI, but to let AI do what it does best—rapid scanning, initial screening, and marking suspicious items, leaving situations requiring in-depth judgment to human experts. At the same time, the experts' decisions will, in turn, serve as reward signals for reinforcement learning, allowing AI to gradually learn those implicit rules. Li Mingyang: I also have this experience from the perspective of interdisciplinary collaboration. Our team has students from Energy and Power Engineering who reminded me that the speed and vibration of the conveyor belt in a waste recycling station, as well as the humidity of the environment, directly affect the quality of spectral imaging—these physical constraints are completely unthinkable when writing code. Therefore, what truly needs to be re-understood is the systemic nature of technology implementation in traditional industries. AI is not an isolated algorithm; it must be embedded in existing equipment, processes, and even workers' operating habits. Overall, I believe the most crucial cognitive shift needed for AI to enter traditional industries is to move from "using AI to disrupt an industry" to "using AI to understand an industry." We must respect those implicit experiences that are not written into the system, treating them as valuable dark data to be mined and integrated, rather than ignoring them. This is also the biggest lesson I learned in this project: no matter how cool or advanced our technology is, we still need to understand what the people on the front lines need. Tina: What you said—"It's not about letting AI replace humans, but letting it do what it's good at"—is a very insightful statement. Industry tacit knowledge is indeed a major obstacle to the implementation of AI. Many entrepreneurs with technical backgrounds easily fall into a misconception: focusing solely on refining models and optimizing product capabilities. But customers don't care how advanced your underlying technology is; enterprise customers care very directly—can it help me reduce costs, increase efficiency, and make more money? Technology is a means, not an end. Tina: How should entrepreneurs balance technical capabilities and business results? Li Mingyang: We really felt this way back then because the whole team almost fell into this trap. As students, it's easy to have a kind of technical heroism when doing projects: thinking that the more complex the model and the more impressive the metrics, the better. But when we were working on the National Undergraduate Innovation Program's spectral garbage classification project, we were given a harsh lesson from reality. Li Mingyang: Enterprise clients don't care whether you're using a Transformer or a CNN; they only ask three questions: How fast is it? How accurate is it after installation? How much money can it save? Li Mingyang: I think there's only one balance point: using the lowest technical cost to solve the most core business pain points. Technical capability is a means, not an end. For example, our project initially pursued extreme accuracy, planning to use a high-precision industrial spectrometer for full-coverage scanning. However, after some research, we discovered that installing such a camera would cost tens of thousands of yuan per unit, and high precision meant high data processing latency, making it unusable on high-speed conveyor belts. Li Mingyang: Later, we adjusted our approach, using a low-resolution spectrometer for wide-area initial screening, followed by a high-precision spectrometer for targeted verification of suspicious targets. This collaborative solution, while not as sophisticated as the original, reduced costs to one-third, increased speed, and maintained the accuracy required for the sorting line. I feel this compromise wasn't a technological regression, but rather a commercial advancement—we used our limited computing power where it mattered most. Li Mingyang: Another time, we were obsessed with optimizing the algorithm, wanting the model to recognize hundreds of materials. After talking to several waste transfer stations, we realized that the client's real headaches were only a few types of waste: mixed plastics, black plastics, and some contaminated materials. So we went back and decided to cut off the recognition of a large number of low-frequency materials, focusing our efforts on those core scenarios to improve robustness. After system testing, the client was actually more satisfied—the model recognized fewer materials, the false positive rate decreased, the number of worker interventions decreased, and the overall production line efficiency increased. Therefore, I believe that AI entrepreneurs must find a balance and adopt the mindset of "optimal solution under constraints." Just like participating in mathematical modeling competitions, resources are always limited. Before AI, time was never enough; we had to find the fastest, most functional solution. The business world is even more so—perfection is the enemy of excellence, because customers want speed, accuracy, and affordability, not technological perfection. My experience is: first, think from the customer's production line perspective, then go back to the lab to write code. Technical ability determines what you can do; business results determine what you should do. As student entrepreneurs, the most important thing we need to cultivate is the shift from a technical mindset to a business mindset.

What's the difference between entrepreneurs who navigate the cycle

Tina: AI trends iterate very quickly. Today it's Agent, tomorrow there's a new model. Many people chase the trend, but after the tide recedes, very few entrepreneurs remain. What is the biggest difference between entrepreneurs who can truly navigate the AI ​​wave and create long-term value, and those who chase trends today? How do you distinguish between these two types of people during the competition? Li Mingyang: Entrepreneurs who chase trends may be somewhat blindly optimistic about the latest productivity advancements. AI trends come one after another, and they might think the more powerful the model, the more they want to use it to start a business—but this seems to be just a technological pursuit. If we want long-term stability, we still need to start from the problem, not from the technology itself. Li Mingyang: We used to do mathematical modeling, but now AI has almost completely revolutionized mathematical modeling competitions. Many people think that mathematical modeling competitions are about who uses the most powerful model and who wins the grand prize. But throughout the competition, you'll find that there are thousands of papers written using very powerful AI. While you might get a national third or even second prize with a very powerful model, the papers of those who actually win the national first prize are not written by AI. The difference lies in the modeling used for National Level 2 and 3 exams. Because AI's database is designed for human use, they only use those "popular models." However, for National Level 1, almost every model is purely conceived by the user, unlike AI which learns from those popular human models. They truly start from the problem itself. Li Mingyang: True innovation and breakthroughs, achieving National Level 1, still rely on human breakthroughs from 0 to 1. Only in this way can we achieve long-term stability. Technological waves are like tides, rising and then receding immediately. If you want to truly stay on this beach, you must return to the essence. Tina: That's indeed true. It comes back to what you just said: We cannot let AI do everything for us; instead, we should let it do what it is good at. From Participant to Manager

Tina: You are both a winner of the mathematical modeling competition and the head of the Innovation and Entrepreneurship Training Program. How do you feel about the transition between these two roles?

Li Mingyang: For me, mathematical modeling was about using myself as a tool to make breakthroughs and changes in a certain field, which is how I won awards. The Innovation and Entrepreneurship Training Program is different. It's not just a tool, but about being a manager in a laboratory, coordinating, linking, and regulating, bringing together various departments to break through in one direction. This not only tests individual ability but also the ability to coordinate teamwork—some teammates may be very capable, while others may be lazy or even unwilling to work. At this time, you need to use your coordination skills to mobilize them to work together in one direction. It's more of a transition from participant to manager. Tina: Which role do you prefer? Li Mingyang: In mathematical modeling competitions, you can feel the tension and excitement, the tension and excitement of breaking through alone. But in the National Undergraduate Innovation and Entrepreneurship Competition, you'll find that you need a lot of communication skills to coordinate with classmates. For example, our team might not be able to finish by ourselves, so we need to coordinate with the computer science team and the mathematics and physics team. There are also interdisciplinary issues, and a kind of diplomatic communication—you have to make others understand what you are doing so that they can participate better. In mathematical modeling, you can feel that you can finish it in your mind; but in the whole project, you not only have to solve problems, but also better express the problems so that others can understand your problems so that they can participate better. Tina: This actually develops different abilities, allowing you to review more comprehensively and look at the same problem from different perspectives. Did your parents choose your AI major, or did you choose it yourself? Li Mingyang: I chose it myself. In my first year, I wasn't majoring in AI; I studied Information and Computing Science in the School of Mathematical Sciences. Later, I thought that AI was becoming increasingly popular, and my classmates often talked to me about intelligent agents. I was very interested in AI, so I started working hard in my freshman year, consistently raising my GPA to the top 10 in my grade. In my sophomore year, I switched to the Artificial Intelligence major.

Tina: So you saw the potential of the Artificial Intelligence major, and since you liked it, you gradually switched.

Review: Four Core Elements

Tina: Today's conversation, from a young participant in AI innovation, from the logic of mathematical model training to the understanding of AI's application in industries, and then to thoughts on AI entrepreneurship, provided us with a very unique perspective. Here's a brief recap of a few key points: First, the barriers to entry for AI startups have lowered. What's scarce isn't tools, but our understanding of needs and the drive to get to the front lines. Second, when applying AI to traditional industries, the most difficult thing to grasp is the industry's implicit experience; we need to go deep into the field and get close to the situation. Third, technology always serves commercial value; we must make the greatest improvements within the limited framework. Fourth, entrepreneurs who can weather economic cycles are not those who blindly chase trends, but those who are truly thinking and able to use AI to do what it does best, rather than letting AI completely replace them. This concludes today's 18th episode of the "Aspiring Generation Salon"—featuring our youngest guest's breakthrough in both mathematical modeling and scientific innovation. Thank you again to our guest Mingyang for his insightful sharing, and thank you to all the viewers for joining us. I'm Tina, see you next time in the Progressive Generation Salon. (This article is a transcript of the live broadcast and is for exchange and learning purposes only; some product and model descriptions are subject to change.)