After get off work, on the international route, let Xiaolu take you around the world. Did you know that there are rumors in the market that OpenAI is going to start making its own mobile phones? At the same time, the Chinese company DeepSeek, which initially claimed to be competing with OpenAI, recently launched its new model V4. It not only emphasizes high performance and low cost, but more importantly, it has started to abandon Nvidia chips and switch to Huawei's own chips. Is the US-China tech war entering a new phase? Will this become Huang Renxun's nightmare? Today, I'd like to introduce our guest on the show. He is a guest who is very good at converting complex and difficult-to- understand technology into language that you and I can understand. I'd like to introduce you to Qu Bo, CEO of Knowledge Technology. Hello Qu Bo, welcome to the show. Hello Xiaolu, hello everyone. What we're going to talk about at the beginning is, of course, this new trend of AI. But first, I'd like to ask Qu Bo to estimate how many people out of 1,000 in Taiwan are already using AI chatbots, whether it's OpenAI's ChatGPT or Gemini, how many people are actually using AI agents to help their daily lives? I think it should be no less than 60%, 70%, or even 80%. The reason is that even when we search on Google, it brings up an AI summary. Yes, that's already artificial intelligence; we can chat with it. If we include that, I think it's over 80%. But if we only talk about chatting with chatbots, in other words, excluding search engines, I think it's at least 60%. So, how many out of a thousand people do you think will start building their own AI agent, whether using Claude or other tools? The current percentage is relatively low, including myself. The reason is simple: artificial intelligence still has a serious problem—it can spout nonsense. Let's go back and explain what Open Claw, or artificial intelligence agent, is. Generally, when we talk about artificial intelligence, it's a question-and-answer process, so I can control the entire question-and-answer process. But this concept of an artificial intelligence agent is different. It goes a step further: I tell the computer the problem I want to solve, I tell the artificial intelligence, and the artificial intelligence will help me figure out how to solve it, plan the whole process, and even execute it automatically. Of course, because the process is quite complex, there are many risks involved. Given that AI can sometimes spout nonsense, entrusting such a long process to it carries practical risks. For example, what if this computer could control all the data on your computer, even automatically cleaning your hard drive? Wouldn't you be afraid? Your computer might malfunction at any time and delete data it shouldn't have. This is a possibility. So, I think that until chatbots, or rather, current AI language models, can solve the problem of spouting nonsense, is it safe to use such AI agents? That's the question. Another issue is that the installation process for current AI agents is still somewhat complicated. Unlike older apps where you just click "OK" to install, this is different. For example, the recent trend of "raising lobsters" in mainland China, due to its complex installation process, has created a new profession: helping people raise and install lobsters. Why? Because the process is still somewhat complex and not intuitive. But personally, I think this problem is actually simpler because technology is constantly advancing. I believe that in the short term... It's possible that someone will release a feature within six months to a year that does n't require special installation, and it might really be OK, just installed. This kind of AI agent is possible, but I think solving the problem of it being completely nonsensical won't be easy. It might take two or three years, and it will continue to improve and evolve. I do have some friends who are hesitant to enter the AI ​​agent market because the installation process is too cumbersome. I've also heard the Google CEO say that all software needs to be completely rebuilt in this new AI era because the logic is different now. However, Qu Bo also mentioned that the AI we use daily, whether it's Chat-GPT or various AI tools like Gemini Grok, is still improving. Qu Bo, have you seen the recent hairstyle analyses we did before and after filming, suggesting suitable hairstyles and colors? This is a new feature ChatGPT launched because its image generation and text generation previously produced garbled characters for Chinese, and it 's upgrading various languages besides English, including Korean. But speaking of something newer, which Qu Bo just told me... The company that developed ChatGPT is called OpenAI. They want to move from software to hardware. What are they planning to do? Recently, the news has been that they want to make a smartphone. Let me clarify a bit: they could theoretically make a smartphone without it. Any smartphone can install the ChatGPT application directly. So why do they want to make their own smartphone? From their perspective, I think it's because, as everyone knows, while we can install the ChatGPT application, not everyone will. Furthermore, this application is at the top layer of our software architecture, and its control over the phone's hardware is actually quite limited. More importantly, the computing power of our current smartphones isn't as strong as that of cloud servers. The computing power of terminal devices is relatively limited, so the AI ​​accelerators within them may not be able to optimize for OpenAI's algorithms. Therefore, I think from their perspective, they've probably considered a few things. First, they hope to create a native OpenAI (ChatGPT) smartphone. What does "native" mean? It means that as soon as the phone is turned on, ChatGPT pops up as your intelligent assistant. It can handle everything for you. My favorite example of this is J.A.R.V.I.S. from the Iron Man movie. I think anyone who's seen the movie knows what I'm talking about. The protagonist, Tony, always asks J.A.R.V.I.S. whenever he encounters a problem. He often tells J.A.R.V.I.S., "Do this, do that." Similarly, Xiaolu, if you buy a phone today, and as soon as you turn it on, J.A.R.V.I.S. pops up: "Hello Xiaolu, how can I help you?" You can then tell it, "I want to go abroad to Japan for three days, please help me plan my itinerary." It will automatically search all the information for you, arrange a great itinerary, and even discuss it with you. It might even chat with you when you're in a bad mood, asking, "What's wrong? Why are you in a bad mood?" I think this is what artificial intelligence agents are like—being natively integrated into the phone. This is probably what it's thinking. Qu Bo, I heard you. Isn't it possible that from hardware to software, we'll see a completely different situation because of this native ChatGPT or similar chatbot integrated into the phone... Google is already working on it. Google phones can easily use the voice recognition and query functions from Gemini. Is OpenAI going to create a completely different software and hardware? Personally, I think that since it's only just started and there's news about it, the details are still unclear. But from their perspective, it's obvious. Look at everyone's phones; they all use Android. So, wouldn't it be easy for Google to natively integrate their Gemini into our phones? You're essentially forced into it; you don't need to install an app. You just buy the phone, turn it on, and it's there. The problem is with OpenAI. OpenAI is an app, so you have to install it on your Android phone to use it. So, they want to bridge that gap, hoping to have native OpenAI or ChatGPT on smartphones. Of course, you've got to the point: will they choose an operating system? It's possible. It's uncertain right now, but I think given OpenAI's current influence, it's very likely. They have their own operating system, which I think will likely be Linux-based, possibly similar to Android. They could even directly use Android, since Android is open source. But the key point is... Our software architecture includes the so-called mobile phone chip, which is the hardware; above that is the firmware, which is the software that drives the hardware; above that is the operating system; and above that are the applications. For OpenAI, it is currently only the top layer. It doesn't have enough control over the underlying layers, so it may want to go down and first control the operating system, or even the related firmware. Its software can control all the functions of the phone, right? I even think it's possible that it will want chip manufacturers like Qualcomm and MediaTek to optimize and accelerate its chips based on its OpenAI algorithms. That would make the phone very unique. Because OpenAI itself has billions of users, do you think it's possible that it would ask Qualcomm or MediaTek to repair the chip to meet the needs of major customers? So I think we are just opening up its future possibilities. But in any case, OpenAI seems very ambitious about entering the mobile phone market. Okay, now that we've finished talking about OpenAI, let's talk about DeepSeek, a Chinese technology research firm, which has always been hailed as the most capable of competing with OpenAI in the context of the current US-China tech war. The company that's competing with us has recently been rumored to be using Huawei's Ascend chip instead of NVIDIA's in its newly launched model. This has raised concerns that Jensen Huang's (Huang Renxun's) dominant position in the chip market, supplying chips to various service providers, might be undermined. Could this be another turning point in the US-China trade war? Let's review the technical background. Artificial intelligence has two stages: training and inference. This is the first key point. In the training stage, AI needs to input data into the computer, identify rules, and build a model. So there are three steps. The model is essentially the rules governing the data. After building the model, the second stage, inference, begins. Inference involves using these rules to input the question into the model, and the model calculates the probability of a correct answer based on the rules. This is inference. For example, the chatbot we just discussed is a large-scale language model. Everyone has heard of large-scale language models. In short, it's the rules governing human speech. In other words, scientists input human speech into computers. Find the rules and build a model. This is the rule of human speech. Next, I will use this model to create a chatbot. When you have a question, you ask it. After inputting your question, it calculates the possible permutations and combinations based on the rules of human speech and tells you the answer. In the past, Nvidia's strength was its graphics processing unit (GPU). The characteristic of a GPU is that it can be used for training and inference. It is basically a master. A general-purpose GPU is indeed a master. But there is a problem: you have to tell it what to do. So you have to write a program to tell it what to do. Nvidia chips require this. This is the characteristic of Nvidia chips. The program itself will consume computing resources. So this kind of general-purpose processor, in our professional terms, has a high degree of software programmability. But conversely, it has some disadvantages. For example, its chip is larger, has more transistors, and consumes more power. It will consume more computing power during operation. In the training phase, this is indeed necessary because the training algorithm will be modified. So software programmability is very important. But now the whole The artificial intelligence industry has entered its next phase starting this year. My favorite analogy is that the previous 10 years, from 2015 to 2025, were primarily a training-focused market. From 2025 (2026) to the next 10 years, 2035, the market will gradually move into the inference phase. This means the models are nearly fully trained, and manufacturers are starting to focus on monetization. Monetization is about inference, not training. Because I can only charge for using my model for inference, but training requires investment and significant investment. Now, as we enter the second phase, can we use graphics processing units (GPUs) for inference computation? Theoretically, yes, but the efficiency is poor. This is because inference usually involves repetitive computations since the model is already defined. Repetitive computations, when written as a program and executed by a GPU, suffer from poor performance. Therefore, the solution is to use so-called artificial intelligence accelerators, or what we commonly call ASICs (Application-Specific Integrated Circuits). Some people also call these... A Neural Processing Unit (NPU), whatever you call it, is an accelerator for artificial intelligence. Using accelerators for inference calculations reduces the computational power required for software control. Furthermore, accelerators directly utilize transistors for computation, resulting in smaller chip sizes, higher performance, lower power consumption, and lower costs. This is changing the entire market. Is this the strength of Huawei's Ascend chip? Okay, next we'll talk about Huawei. From the past to the present, Chinese manufacturers have all wanted to develop graphics processors (GPUs), and some are indeed working on it. However, due to the high barriers to entry, they haven't been able to achieve the advanced GPUs of Nvidia. Mainland China has always lagged behind in this area. So how do they catch up? The solution is simple: use artificial intelligence accelerators. Therefore, the Ascend chip initially had a central processing unit (CPU) and a GPU, but the focus wasn't on the GPU itself, as its computational power wasn't strong. Its focus was on the artificial intelligence accelerator. So, regarding the DeepSeek model you just mentioned, its training and inference, especially inference calculations, performed well using the Ascend chip, particularly with the accelerator. This definitely puts some pressure on Nvidia. So you'll find that Nvidia is also starting to adjust its strategy. At this year's GTC (Global Technology Conference), a global technology conference held annually, Nvidia announced its latest progress. Nvidia has invested in a company called Groq (a chip startup), and Groq has launched a so-called Language Processing Unit (LPU), which is an AI accelerator. So this time, Nvidia is also shifting its focus, emphasizing that its servers are not just graphics processors but also language processors, specifically designed for AI inference and understanding. So I think the overall trend is very clear. As I just mentioned, 2025 to 2026 is a dividing line. Originally, the focus was on training, but now it's shifting to inference. Originally, the focus was on graphics processors, but now it's gradually shifting towards accelerators becoming increasingly important. Mr. Chu, what about the accelerator reserves or productization over the next ten years from 2025 to 2035? Which Taiwanese manufacturers are relatively capable of providing such products? Frankly speaking, these AI accelerators are still mainly designed by European and American manufacturers. Besides Groq designing LPUs and Google designing TPUs, as we just mentioned, other manufacturers like Amazon, Meta, and Microsoft are also designing their own accelerators. The reason is simple: think about it. These cloud service providers—Google, Amazon, Meta, and Microsoft—all have data centers, and they all need a large number of processors. If you buy all these processors from Nvidia, Nvidia will make a fortune, and you cloud service providers will be suckers. Such expensive stuff, and you're still buying so much! So, if you're the owner of a cloud service provider, you'll start wanting to reduce costs. And how do you reduce costs? By developing your own accelerators. Why? Because if it's your own, you can earn the money yourself. You just need to find a foundry to manufacture it for you. And here's an important point to remember: the barrier to entry for graphics processing units (GPUs) is higher, while the barrier to entry for accelerators is lower. The reason is simple: GPUs need to be able to execute programs in any situation, so relatively speaking, the complexity of the underlying chip is higher, and therefore, it's more expensive. That's why it's almost impossible for a company to design a new GPU to compete with Nvidia. But what's the shortcut? It 's to make accelerators. It's a bit like overtaking on a curve, or rather, it depends on how you apply it. Inference calculations are suitable for accelerators, which align with this generation. So, as I just mentioned, every cloud service provider can make accelerators. But what supply chain does Taiwan have? After you make an accelerator, do you need to make it into a printed circuit board? Do you need to make it into a server? Do you need cooling? Do you need a chassis, cabinet, or rack? More importantly, do you need a chip foundry? Yes, and that's where it is. TSMC. So I think TSMC's recent strong stock price is reasonable. TSMC's stock price has been very strong these past few days. What I'm curious about is where Intel is now. Okay, in terms of process technology, TSMC is currently leading, while Intel is slightly behind. However, I think Intel's biggest lag is in yield. Because if your chip yield isn't high enough, customers won't dare entrust such important projects to you. Okay, in terms of process technology, the difference between the two isn't significant. But the key point now is that whether you use an accelerator or a graphics processor, what you need is an advanced process technology. So there's no difference in process technology; customers naturally want the most advanced. So TSMC is still the priority. Then why has Intel's stock price been so strong recently? There are several reasons. First, Intel's manufacturing processes are gradually improving. Second, frankly speaking, TSMC, with its high yield and many customers, does n't have much capacity, so its prices are firm, and it's even been considering raising them. So when Musk wanted to launch his own TeraFab project, who did he first approach? TSMC? No, he approached Intel first. The reason is simple: it's not that he did n't want to use TSMC; he certainly considered it. But could TSMC help him? That's uncertain right now because TSMC is too busy. Conversely, Intel urgently needed an important customer, so they wanted a quick agreement. Therefore, Intel first connected with TeraFab. You might wonder why they wanted to use a slightly less advanced process. The reason is simple: if you can't get the top supplier, you have to go for the second best. That's reasonable. And no customer wants to be tied to a single supplier; they always want a second source, a different option. So, from this perspective, I think Intel has an opportunity. At least Intel naturally took over the customers that TSMC couldn't handle, and the same goes for Samsung. That's right. Another point I want to make about Intel is that when the market shifted from training to inference computing, the role of the central processing unit (CPU) became more important. Why is that? Currently, our AI servers are configured with a CPU paired with a GPU or accelerator. Remember, the GPU or accelerator is only needed for AI algorithms. However, when users input their needs, we still need to manage program scheduling, and that's what the CPU does. So think about it this way: when the market is primarily focused on inference computing, aren't there hundreds of millions of people worldwide connecting to servers to make requests, chat, and have you perform calculations for them? Then the CPU has to be responsible for managing these scheduling tasks, allocating resources to the GPU. This increases the importance of the CPU, which is Intel's strength. That's why recently Intel CPUs have started to be in short supply, and the entire market is reassessing Intel. Based on this analysis by Qu Bo, I knew that Intel's new CEO, Chen Liwu, would be coming to Taiwan for Computex, which will be held in Taipei this June. Many people, including those in the market, are eagerly anticipating his appearance and hoping it will bring some positive news. There's a lot of interest in Intel, which is different from the previous phase. However, from the beginning of the program until now, we've focused more on the US. Now, I want to ask Mr. Qu, looking back at China, with its own Huawei Ascend chips and DeepSeek, I believe they will also want to make their own hardware phones. What I'm curious about is, within the AI ​​system, in the context of hardware AI phones, I'm wondering if China will become a self-contained ecosystem in the entire AI industry chain, no longer needing to rely on the US or Taiwan. This is definitely the direction he's striving for, and it's very clear. He just said it's not possible to achieve this yet. I think the most important thing is the chip supply chain. I know mainland China is working very hard on localization, creating a complete vertically integrated system. But at this stage, even SMIC's wafer foundry still needs to import many high-end equipment. So frankly, if all the equipment needs to be domestically produced, I think it's a long road, maybe 5 or even 10 years, and some equipment might not even be produced domestically. For example, regarding extreme ultraviolet (EUV) lithography equipment, a mainland Chinese netizen bet me that we'd be able to mass-produce it by 2026. It's already 2026 now. He told me two years ago that we'd have mass production within two years, but that's definitely impossible. It's not that simple. It's possible to work towards that goal gradually. Another point is advanced packaging equipment, which has a high barrier to entry, but also a higher probability of domestic production. Okay, that's semiconductors. Once we have semiconductor chips, then there's system integration—printed circuit boards, a whole series of systems. I think these are all things that mainland China can easily handle; they're the most fundamental part. As for software, that's a strength of the US, but mainland China's software is also very strong. So, currently, I think mainland China can indeed produce their entire system. It's just that the US knows this, and it might not be easy to compete. So they've cut off our supply chain by controlling some advanced process-related equipment. If you can't buy it, you can't make the most advanced chips. That's why they need accelerators to compensate. Of course, China is also working very hard, but I think there will still be some gap. The US is probably only a few years ahead. So he can only maintain his lead by leveraging this advantage of a few years to continuously create new things. Speaking of China's strong software, I recently saw a news report that the Gaode Maps app has recently become the number one downloaded app in Taiwan. People think its traffic light countdown timer is particularly accurate, so they want to download Gaode Maps, paying close attention to the difference of one or two seconds. But another side says that while convenience and accuracy are one thing, there are also cybersecurity risks. I usually use Google Maps more. The logic behind Gaode Maps is that it collects a lot of data and map data in Taiwan. I believe it must have some big data. Any user who downloads Gaode Maps will provide some information to it. For example, if you are driving on the road, because the map will definitely have GPS on, if a lot of people in Taiwan have installed Gaode Maps, then it will know which traffic light many cars are waiting at and how long they are waiting. It can use this method to deduce... I have seen some information that our government agencies have issued a statement saying that all government APIs... The application interface ( Gaode Maps) is not open to Gaode Maps for reading. In other words, the so-called traffic light durations it displays are calculated, not provided by the government. So, what kind of cybersecurity issues might arise from this calculated data? I think it's okay, because these are mostly applications related to daily life and consumption, so there shouldn't be too much of a problem. It just depends on whether this map can replace Google Maps. I think it depends on each user's habits. If you really care about traffic light durations, Google Maps does n't have that feature, so you might want to switch to Gaode Maps. But regardless, this example shows that mainland China is actually very advanced in the software industry; we can't underestimate them. Since the government hasn't opened the API to Gaode Maps, are people downloading the Gaode Maps app exposing any cybersecurity risks by giving data to Chinese companies? As I just mentioned, your GPS location—this information will definitely leak out. But recently, some people have been calling for regulation of this kind of thing. Personally, I think it's unnecessary to do this kind of free market competition. This is free market competition. If you want to regulate this, then what about Huawei or VIVO phones? I see many of my friends using those phones... Are you going to regulate Xiaomi phones too? Personally, I think this is a free market. Our general consumption behavior, whether it's going out or taking transportation, doesn't seem to involve any great secrets that require such strict regulation. I personally think it's fine. Only sensitive or special individuals should be careful. Perhaps for important people, knowing exactly where they are at a traffic light and how many seconds they're waiting is crucial—that's possible, right? For ordinary people, I don't think it's necessary. The program has been comparing various AI products from China and the US. We just talked about the chip DeepSeek uses, or how some people in Taiwan are starting to download Gaode Maps. On another note, I observed in the latest episode of The Economist that they specifically introduced Mythos, a product from Anthropic, positioning it like a digital nuclear weapon. To me, it seems like this AI tool can detect system vulnerabilities and mobilize resources to directly patch them. So why is this described as a powerful weapon? We've been saying that artificial intelligence is a set of rules for data, and models are rules for data. Every program we write has certain rules. So, using a model like Mythos... By reviewing your written code, you'll find many vulnerabilities, or bugs. In engineering, we used to find these bugs manually. Frankly, no matter how fast the human eye can see, its speed is limited. But a running program is very fast, isn't that great? So, any program, once reviewed, will likely find a bunch of vulnerabilities. That's why everyone is so afraid. Why? Because if hackers use such tools to analyze the code of their targets, it's very dangerous. They might find vulnerabilities, or even use a certain method to attack the website and find its vulnerabilities, using certain rules to exploit them. That's possible. So, they didn't dare to release this model publicly because making it available to anyone would be too dangerous. But we need to be prepared that Anthropic isn't the only company developing such a model; there will definitely be other similar companies making similar models. So sooner or later, such models will still proliferate. More sophisticated hackers will still be able to create them. So, what we need to do now is for every company to review its own code, of course, using artificial intelligence. Finding and patching vulnerabilities is one thing, but the power of the Mythos model isn't just about finding vulnerabilities; it can also write a program to exploit those vulnerabilities within hours. This is very dangerous. If it can be done in such a short time, many industries, especially those related to finance, government, and the military, will be affected. This is the main reason why the entire market is experiencing such a shock. This product can be described as Ragnarok. Qu Bo, I'm curious about your personal stance. With AI companies or software companies developing such powerful products, should the government play a role in regulating them? I personally think it's necessary. However, this is a double-edged sword. The more the government regulates, the more restrictions there are on the industry's development. Less government regulation makes it easier for industries to develop, but it also makes it easier for problems to arise. So I think there's a balance to be struck. From the government's perspective, it should establish basic regulations to prevent these companies from overstepping boundaries. However, it shouldn't regulate too many details; it should allow the industry to develop and thrive. For example, could we allow a robot to operate without human consent? Regarding the issue of pulling the trigger to kill, I think the government should regulate this. For example, after discussion, if the government decides it's unacceptable, then it should establish regulations. These regulations should be fundamental, prohibiting any manufacturer from developing any weapon, firearm, or robot that can be used to pull the trigger without human decision-making. This is what I meant by a principle-based regulation. As for the details, the government should allow industry to develop freely. However, I think the reactions and understanding of governments around the world are actually very slow. Let's not even focus on Taiwan; even in the United States, it took US senators and representatives a very, very long time to understand how Facebook operates and how it makes money before finally enacting regulations on what Facebook cannot do. In fact, the industry has always been faster than the government, which is normal. Technological development is usually pioneered by a group of intelligent people. Regulations are typically the domain of those with legal or political expertise. Only when scientific breakthroughs and technological advancements have been achieved, and the government or legal system recognizes that something is unacceptable and needs regulation, will the government intervene. Therefore, there will naturally be a time lag, which is normal. But in any case, what needs to be done still needs to be done. When we find that a certain technology might get out of control, then you may still need to do appropriate regulation. That concludes our discussion. Finally, I want to return to a very fundamental issue concerning people's livelihoods: whether everyone can keep their jobs. Because of recent news reports, if you've been following the situation, you'll know that companies like Meta, Oracle, and Amazon are all laying off employees, and the layoff rates are quite high. You'll find that these large companies are investing all their money in the AI ​​arms race, but junior engineers, or those with only one or two years of experience, are all being laid off. I'm curious about this wave of layoffs. Qu Bo, how do you predict it will unfold? At the beginning of this year, a research institution published an article. Its main point was that while everyone believes the future of artificial intelligence is wonderful, it presented a completely opposite picture. One aspect is that many software companies no longer need to write their own programs, so they don't need so many engineers. Even many software companies' clients, who used to pay for these software programs, now only need to employ a few engineers because AI can write programs. You can write the software you need yourself, so why would you spend money to buy software from software companies? That's why you'll find that software-related stocks, especially in the US, have fallen quite sharply in the past three months. This is just the beginning. If you read that article, you'll find it chilling. Why? Think about these unemployed engineers. Their spending starts to decrease, and their reduced spending immediately affects all related industries. They stop eating, stop spending on entertainment, and even their housing will be in trouble. Why? Because banks are currently lending money to these engineers, who are high-income earners, so they are happy to lend to them, right? But when they are unemployed, they will reduce their spending and can't repay their loans. Are they sure they can get back the money that was previously lent out? We don't know. Anyway, people have already started talking about this, and it has indeed impacted the entire industry. As you just mentioned, many software companies have started laying off employees. I recently talked to the general manager of a large domestic software company about this. I think his point is quite reasonable. In the past, the personnel structure of software companies was an equilateral triangle, with many engineers at the lower levels and fewer and fewer at the upper levels. So, you have fewer people managing most of the staff. That's how companies used to operate. But now, because the entry-level work no longer needs human intervention—it can be done by AI—the current situation is to lay off entry-level engineers, while leaving many mid-level engineers. The result? If this continues, it will eventually become an inverted triangle. But the inverted triangle still has a problem: senior engineers always have to start from the bottom. If you cut all the entry-level engineers, there's no one to train them. What happens when these mid-level engineers retire? Who will check if the code is written correctly? The current situation is this: writing the code is done by AI, but checking if the code is correct still relies on experienced engineers. Anthropic's Mythos can do that, theoretically. But even Mythos can make mistakes; it still requires human checking. So, what was our conclusion? We believed that the ideal company, especially a software company, should have a rectangular workforce structure. What does that mean? It means you absolutely cannot eliminate all the entry-level engineers. You need to train people who can become senior engineers. In other words, you might have previously needed 100 entry-level engineers. You need to find at least 50 people now. What's the purpose of these 50 people? It's to gradually train them into engineers with senior experience so they can assist in reviewing these programs in the future. This is a very practical business profit consideration, or is it a vision for talent development? No, it's a business consideration. Because if junior engineers don't even have the opportunity to work, how can they become senior engineers? You must train them yourself. Otherwise, one day you'll find that all the senior engineers have retired, and you won't be able to find senior engineers in the market because everyone has laid off all the junior engineers. So our conclusion is that we must train them ourselves. Every company must train its own junior engineers so that there will be enough senior engineers in the market. But the key is, you originally needed 100 people, now you only have 50. So what will the future world be like? The future world will become a K-shaped world. That is, at this point in time, if you keep up, you will rise to the top, and in the future, you will be a senior engineer, and the company will rely on you even more. But if you don't keep up, you will fall behind and be among the group that is eliminated. Frankly, what should this group do? I don't know. I've thought about it for a long time, but I don't have an answer. So I think I want to give you, the audience, the most important concept, especially if you have children who have already been born or are in junior high or high school or above, you need to start thinking about this: It's crucial to cultivate a love of technology in children from a young age. Why? Because in the future, those who lack a passion for technology won't be competitive. Remember what I said? I'm not saying they have to be engineers. They can be lawyers, accountants, or do any job. But they can't be indifferent to technology. When you're indifferent to technology, no matter what job you do, if a new technology is invented, you're afraid you won't want to use it. Yes, you'll be afraid, "I don't want to use it, it's too difficult, it's not my problem." You absolutely cannot have that kind of thinking. Our generation can barely manage, because we're already of a certain age and have certain positions; the competition isn't in that direction. But your child is different. Your child faces that kind of competition. So you must cultivate their "native" interest in technology from a young age. That's why I founded the Qubo Technology Classroom. Sometimes, children are n't necessarily interested in new technologies. How do you spark their interest? You must find a way to explain things in a way that they can understand. Yes, at least understand, so they don't feel like they can't absorb anything you're saying. You might say it's too difficult for young children, but I believe... Junior high school graduates and those above high school level must understand technology. Therefore, when I designed these teaching materials, I designed them at the junior high school graduation level. Those above junior high school can understand them. Why this? Because everyone attends junior high school; it's compulsory education. So I wanted to start from that basic point and cultivate children's sensitivity and understanding of technology. Sensitivity is very important. Mr. Qu, I'll quickly ask a question related to this. I recently heard Huang Renxun say that he believes the most powerful and popular major in the future is English. He believes that if you master the logic of language, you can communicate with technology, regardless of which technology company's product it is. He believes language proficiency is crucial. What do you think of his statement? He also mentioned wanting to study biotechnology. He keeps changing his mind, but I think one concept that will never change is that besides not being indifferent to technology, there's another even more important point: you must be interdisciplinary. Simply put, you absolutely cannot only know one field. The more fields you know, the stronger your competitiveness. So CEO Huang Renxun wasn't wrong; he said that knowing English is a great skill, but he only said half the story. You can't just know English; you also need to know English... If you have strong grammar, are proficient in artificial intelligence, and can write programs, you'll be very competitive. Therefore, you must have at least two specializations. Thank you, Dr. Qu. Today's topics covered a wide range of issues, from the chips used in DeepSseek to AI ethics, to discussions about Gaode Maps (a Chinese mapping service), where users have their own preferences. The global AI layoffs have also raised questions about the priorities for education in the next generation. Dr. Qu explained all these points in a very clear and concise way. We are currently at a very special juncture, a turning point in our era. Will this K-shaped society move upwards or downwards? We will continue to monitor this for you on our program, and we hope to have the opportunity to invite Dr. Qu back to the program to analyze the latest technological trends. Due to time constraints, our program will end here today. See you again next week at the same time on the air. Bye-bye!