Interview with Zhu Yu, Chairman of New Oriental Ucoding: How does DeepSeek affect programming education?

DeepSeek is stirring the entire education industry to embrace AI, even more vigorously than during the OpenAI era, allowing everyone to see the dawn of the "iPhone moment" in future education.

Recently, Duozhi interviewed Zhu Yu (pen name "Little Wolf"), Chairman of New Oriental UCoding. He analyzed the advantages of DeepSeek, envisioned the opportunities for educational companies after embracing DeepSeek, and predicted the future development trend of the programming education industry.

Last June, UbCode released the UbCode Model, a vertical model in the field of the National Olympiad in Informatics (NOI), which was developed in cooperation with Xizhi Intelligence and uses an AI teacher's lecture mode. At that time, UbCode Model had already been connected to DeepSeek because the latter had low cost and excellent performance in the field of programming assistance.

During the conversation, Zhu Yu told Duozhi: "Although various educational technology companies are cooperating with DeepSeek, the core competitiveness still lies in the specific applications and optimizations of each company."

When talking about the development of the programming education industry, he mentioned: "I think programming education institutions are divided into two categories, one is interest-based and the other is competition-based. In the future, interest-based programming education institutions will face challenges, and there are already signs of this in the market. This year will be a turning point. Interest-based programming education institutions will transform into competition-based institutions to provide IT education. At the same time, new IT institutions will emerge. "

The following is the full text of Duozhi ’s dialogue with Zhu Yu:

Duozhi: There are many AI programming tools now, and parents are also confused about why their children should learn programming. From an industry perspective, has AI affected the children's programming education industry to some extent?

Zhu Yu: Let me first answer the question “Will AI replace programmers?”

Let me give you an example. Logically, the ChatGPT team should be the one who has mastered advanced AI technology, but why didn't they create an algorithm like DeepSeek earlier?

Previously, ChatGPT's approach could be described as "powerful efforts to achieve miracles". It did not require the study of advanced algorithms, but mainly relied on computing power. This is like using the least squares method to find the best fit between data and the model by minimizing the sum of squared errors. That is, ChatGPT collected all the text in the world, inferred what would appear after the text, and made a fit. All the characters were vectorized, and then there would be predictions. This fitting process requires huge computing power, and is a very rough and violent method, which can be calculated using a graphics card.

Why did DeepSeek bring about revolutionary changes? DeepSeek improved the algorithm and used a new training method. Why didn't the ChatGPT team think of this? This shows that managing and improving AI still depends on human resources. DeepSeek has brought together many outstanding talents, including many outstanding people in academia.

I think that technological change will replace basic jobs, and basic programming languages ​​will become less and less important. However, the requirements for more in-depth algorithms will be higher, and the demand for corresponding engineers will increase.

All the links just mentioned, one is the continuous progress and improvement of AI technology, and AI engineers are needed to provide better algorithms. Secondly, to use it in various fields, the docking engineers must also master the relevant knowledge. Otherwise, for some exclusive models, you will not know how to fine-tune or change the parameters. For the general public, AI tools are so convenient, but the effect of each person's use is still different. People who understand the logic behind AI will use it more smoothly. Without understanding technology and programming, many people are still helpless even with AI tools.

This is equivalent to the first and second industrial revolutions replacing traditional handicrafts, but in fact more people need to master knowledge of physics and chemistry to become workers to manage, operate and improve machines. The same is true now. Basic programming is not important, but people who understand in-depth algorithms are needed to improve AI, apply AI, and combine AI with niche fields.

Therefore, I believe that the current development of AI technology has brought positive impacts to programming education.

Duozhi: What development trend do you think the programming education industry will have this year?

Zhu Yu: I think programming education institutions can be divided into two categories, one is interest-based and the other is competition-based. In the future, interest-based programming education institutions will face challenges, and there are already signs of this in the market.

I think this year is a turning point. Interest-based programming education institutions will transform into competition-based institutions to provide computer science education. At the same time, new computer science institutions will emerge.

Duozhi: Many people have a question: Can talents in information and economics be cultivated?

Zhu Yu: This question is equivalent to "Are students from Tsinghua and Peking University trained?" Some are good at it, some are not; some have hobbies, some don't; some have talent, some don't. Why didn't primitive people know how to program, but there are programmers now? Talent is not enough, but there are certain conditions. Whether people with programming talent can be selected and trained as early as possible, so that they can contribute to society in this field, this is something the entire education system needs to face.

In addition, with a population of 1.4 billion, there are definitely top innovative talents in China. Even if the proportion is small, it is still a very large number. But where is the problem? It may be that talented people are buried because they are not discovered in time, or they are discovered but not cultivated. Without knowledge in this area, only talent, it is impossible to innovate new things out of thin air.

Duozhi: Various educational technology companies have embraced DeepSeek. What impact do you think this will have on the education industry?

Zhu Yu: First of all, we need to understand the difference between DeepSeek and other large models. Why is DeepSeek a revolutionary new technology? The core is that DeepSeek reduces costs, and the service cost is 80% lower than GPT-o1.

With lower costs, applications become more possible. Whether many applications can be realized actually depends on how high the cost is. If the cost of technology is high, it will be difficult to promote.

Previously, ChatGPT and other large models were more about piling up hardware and models themselves, using a simple and crude way to make the models bigger and more effective, which led to an infinite increase in costs. China's DeepSeek used advanced algorithms to reduce costs. DeepSeek gave everyone confidence that while achieving the same effect, costs can be reduced by improving algorithms, and costs are expected to be further reduced in the future. This is the most important change brought by DeepSeek.

Moreover, DeepSeek is an open source large model, which means it can be deployed locally, which is completely different from the previous closed source large models. For example, ChatGPT must be connected to the Internet and needs to call the model's interface. DeepSeek local deployment is very important for robots or some smart hardware. In other words, local deployment means that there is a certain degree of artificial intelligence effect without being connected to the Internet, which brings great benefits to the promotion of hardware. For example, some hardware used by children does not need to be connected to the Internet. What they need is some existing knowledge, such as oral practice, listening practice, learning foreign languages ​​or translation, which are very convenient.

Duozhi: When educational companies generally embrace DeepSeek, where will everyone’s differentiated and essential differences be experienced?

Zhu Yu: Currently, various educational technology companies are embracing DeepSeek, but the core competitiveness still lies in the specific application and optimization of each company.

On the one hand, although each company can use DeepSeek for local deployment, its practical performance is still relatively poor. For example, to achieve educational goals, the requirements are still relatively high, so it is still necessary to use DeepSeek's services. Their service costs are 80% lower than GPT-o1, and the effect is also very good.

On the other hand, if all educational technology companies connect to DeepSeek and use it on their own models, they will need to combine the educational data accumulated by the educational technology companies themselves, strengthen the algorithms, and combine them with DeepSeek's reasoning capabilities to produce better functions in education.

Duozhi: Since DeepSeek is so powerful, why do we still need to use vertical models?

Zhu Yu: It is difficult to teach directly with DeepSeek. There will still be hallucinations, wrong question types, and even wrong answers. After the education company makes fine adjustments, the effect will be better.

There are two ways to reduce hallucinations: one is to use DeepSeek's solution to train and fine-tune the model, which greatly reduces the training cost. The other is to adjust the output link. For example, the analysis of some questions needs to be adjusted and then fed into the model to adjust the output. Take the Ushannon model as an example. It outputs programming programs. We have to ensure that it gives users the correct results. If there is something wrong, make adjustments.

When it comes to the specific business of each company, some corrections and adjustments need to be made. For example, companies that use question-searching software have standard answers to questions, so they can combine it with DeepSeek's technology, that is, check the content of DeepSeek's answers with the content of their own question bank to see if they are correct. If not, they will give it negative feedback to correct and adjust its answers. They also need to make some constraints in the output link to meet the corresponding results and values.

Therefore, when educational technology companies combine with DeepSeek, they need to train and fine-tune the model. At the same time, they need to make adjustments in the output link based on knowledge graphs, result feedback, etc. to minimize hallucinations.

Duozhi: Does that mean it will be difficult for DeepSeek to replace existing learning products?

Zhu Yu: DeepSeek cannot be used directly for classes or lectures. After integrating with DeepSeek, educational technology companies still have a lot of work to do. For example, if the Shannon model is not fine-tuned, the effect will not be good. For example, the program it outputs is in C language, but the functions used in engineering must be excluded to make the output more in line with the programming knowledge of primary and secondary school students. If these are not fine-tuned, it will output the terms used by adults in programming, which is not very suitable for computer science teaching.

Duozhi: In what direction do you think educational technology products will iterate in the future?

Zhu Yu: First, the cost will be further reduced; second, it will develop in the direction of personalization.

Our education industry has four components: classroom content, classroom experience, after-school management, and educational philosophy. In the future, these four parts will continue to be refined and personalized. In terms of course content, for example, there will be stratification and customized content suitable for each child; in terms of classroom experience, for example, there may be AI teachers who can generate teaching styles that suit students' personal preferences; in terms of after-school management, pushing homework, writing homework, and correcting homework will become more personalized; in terms of educational philosophy, volunteer application is part of the educational philosophy. I think volunteer application planners are likely to be replaced by AI in the future.

However, I think that in educational scenarios, teachers still provide emotional value, and they still have an effect on students in terms of emotions, values, and concepts. AI cannot do these things for the time being. Therefore, as teachers or educational technology companies, they still have the responsibility for the final teaching results and teaching effects, and they have advantages in urging children to complete their studies through emotions or emotional values. AI is difficult to replace them for the time being.