Original · Unique Research · 2025-11-06
Editor’s note: This complete English edition retains the original reporting and Zhang Minsong’s interview in full. Century Tianhong, Zhihong Optimization, Xiaohong Teaching Assistant and Tongbu Xue are English renderings of names in the Chinese source. Career history, company scale, product capabilities and assessments of AI are claims of the original report and interviewee, not independently audited findings.
If you compressed the history of Internet entrepreneurship over the past two decades into a single timeline, you would find an interesting trajectory: from games, video, and antivirus software, to online education and intelligent teaching aids, and then to today’s education-focused large language models and AI teaching assistants. Zhang Minsong’s journey has been almost a microcosm of this trajectory.
Early in his career, he worked as an architect and senior expert at companies such as Shanda, Baofeng Video, and Cheetah Mobile. He personally witnessed products transform from 0 to hundreds of millions of users. He also launched three consecutive ventures in the education sector, reaching the industry’s peak before falling to its bottom, and experiencing the rise and fall of online education—from being eagerly courted by investors to the industry’s subsequent correction. In 2021, he chose to join Century Tianhong, a company that had already spent 30 years deeply engaged in the teaching-aids industry. This listed company, which began with printed teaching materials, is now using large language models to rewrite something that seems more traditional than almost anything else: how teachers teach.
On the surface, this is the story of a technology product professional turning toward the education sector; at its core, it is an example of reflecting on how people can reposition themselves amid a wave of technological change.
But Zhang Minsong was not facing the task of building yet another question bank or resource platform. Instead, he was tackling a seemingly traditional but exceptionally difficult setting: teachers’ lesson preparation, test creation, and grading.
From Teaching-Aid Leader to AI Teaching Assistant: Why Has a Printed-Book Company Set Its Sights on Teachers’ Desks?
The story of Century Tianhong is, in fact, highly traditional: a leading K12 teaching-aids company, planning more than 3,000 book titles a year, reaching more than 30 provinces and municipalities nationwide, and with cumulative circulation exceeding 1 billion copies. It is easy to imagine it as a content factory with an enormous distribution network.
What does traditional lesson preparation look like? It means teachers slowly leafing through books, searching the web, cutting and adapting materials, and reorganizing them in the wilderness of available resources, transforming what others have written into lesson plans and courseware suitable for the class at hand. If they are lucky, some good-quality lesson-plan templates circulate within the teaching group; if they are not, they have to stay up late poring over the materials themselves. A teacher’s expertise in teaching and research, time and energy, and the resources available at the school determine the ceiling for the final courseware and that lesson.
Tools have always existed in this process: question-search websites, teaching-resource platforms, oral-language assessment systems, and so on. But most of them follow the logic of discriminative AI—you provide a question, a voice recording, or a template, and the machine makes a judgment or assembles the pieces. They rarely truly participate in generation and design, let alone understand the real differences among the students in front of a particular teacher.
Xiaohong Teaching Assistant is attempting to rewrite precisely this seemingly fragmented yet crucial gap.
It is not meant to replace teachers’ professional judgment. Rather, it takes away the mechanical, time-consuming tasks of researching materials, assembling structures, and producing first drafts, allowing AI to provide teachers with a foundation for content construction, personalized generation, and integration across areas of knowledge. What teachers should truly do is select, edit, polish, adjust the pacing and the way questions are asked—in other words, turn a usable lesson plan into a class worth remembering.
Generative AI Does Not Bring Faster Answers, but the Ability to Reorganize Knowledge
In Zhang Minsong’s view, generative AI has made two particularly important leaps.
One is the shift from function-driven to cognition-driven systems. In the past, educational AI tools mostly dealt with clearly defined questions: Is this answer correct? Is the pronunciation of this spoken sentence accurate? Does this essay contain grammatical errors in its sentence structure? Once a tool provided a yes-or-no judgment, a score, or a pass-or-fail assessment, its responsibility ended. It never cared about what the student truly failed to understand, nor did it care about how the teacher should explain it more effectively next time.
But as large models began to acquire the ability to understand context, identify intent, and perform logical reasoning, what AI could do shifted from making judgments to participating. It can participate in organizing the structure of a knowledge module, designing tiered teaching plans, and tailoring learning paths to students with different levels of preparation. It is beginning to resemble a colleague who can discuss with teachers why something should be taught in a particular way, rather than merely a tool that automatically strings together questions.
The other leap is the shift from being a tool to becoming an ecosystem hub. Generative AI is no longer an add-on for a single stage; it can connect the entire chain of teaching, learning, practice, assessment, management, and research. You can have it generate content during lesson preparation, design real-time questions in the classroom, automatically adjust the difficulty of exercises after class based on students’ performance, and then consolidate the entire process into data that provides insights for the next round of teaching.
For a long time, the education industry has faced a seemingly impossible triangle of demands that cannot all be met at once: high quality, large scale, and personalization. Once good teachers and good content are scaled up, personalization declines; if you want personalization, it becomes difficult to maintain both scale and consistent standards.
The emergence of generative AI has begun to create cracks in this triangle.
That is because it can both customize at scale and continue learning. For schools in remote areas, it can automatically generate lesson plans and exercises adapted to local learning conditions and textbook editions; for students with special needs, it can design content and pacing suited to their cognitive characteristics. In theory, a single agent can serve several thousand students simultaneously, with each student seeing different questions, prompts, and explanatory pathways. These differences are not created through manual labeling, but through a comprehensive assessment of behavioral data, knowledge structures, and context.
If the digitalization of education in the past was more like moving printed content online, then from Zhang Minsong’s perspective, what generative AI is reconstructing is the way knowledge flows through the education system.
Not AI for AI’s Sake: From the First RAG Lesson-Plan Product to a Rational, Scenario-First Approach
Interestingly, the more clearly people see the upper and lower limits of technology, the less likely they are to harbor romantic illusions about AI.
In 2023, the year of the large-model explosion, Century Tianhong moved very quickly to launch a RAG-based lesson-plan generation product—something with almost no mature precedent in education at the time. As they explored ways to integrate large models with their own content, they also tested and refined the product in real teachers’ workflows: Which kinds of generated content would teachers be willing to accept? Which steps would they rather handle themselves? What kinds of interactions would make teachers feel understood rather than talked down to?
There were, in fact, three clear principles behind this.
First, embrace trends, but do not let trends determine your direction. Technology advances far faster than the education sector accepts change, so following trends is a necessary condition, but not a sufficient one. Excessive excitement produces short-lived demos and long-term products that feel impressive yet prove useless.
Second, put scenarios first—do not build AI for AI’s sake. Users’ needs did not become something else simply because large models appeared. Teachers still need to prepare lessons, write test questions, and grade assignments; the existing tools simply have clear shortcomings in efficiency and quality. AI’s role is not to redefine a whole range of seemingly cool features that no one wants to use every day, but to find the most painful point in the existing workflow, open a small breach, prove its value, and then expand gradually.
Third, always stay anchored to your core foundation. For Century Tianhong, that foundation is books, content, and the understanding and trust built over thirty years of serving teachers. An AI product is not an isolated new business; it must connect with books, educational research, and distribution channels to serve the same group of users. This means it cannot simply admire itself for having elegant technology. It must be measured by highly practical metrics—for example, how many times teachers are willing to open it each day, whether principals are willing to pay for it, and whether teaching-and-research groups are willing to restructure their collaboration around it.
When innovation and implementation are viewed from this perspective, it becomes clear that the truly difficult part has never been the technology, but restraint.
Agentic AI: Not a New Buzzword, but a Product-Building Mindset
When the term Agentic AI began circulating in the industry, much of the discussion focused on how to build an Agent that was smarter and more autonomous. Zhang Minsong’s perspective was somewhat different.
In his view, Agentic is neither an entirely new technological breakthrough nor a marketing concept. Rather, it is a measure of how we think about the degree of AI autonomy—and, even more, a set of engineering principles. High-autonomy Agents and low-autonomy Agents are essentially just tools that complete tasks in different ways, not a matter of superiority or inferiority.
This is especially important in education. You cannot have an Agent autonomously replace a teacher in making every judgment, nor can you completely erase its initiative and treat it merely as an improved search engine. What we really need to do is calibrate, around specific scenarios, how much autonomy is appropriate.
For example, during the test-design stage, an Agent can be given a higher degree of autonomy: based on teaching objectives and student-learning data, it can automatically design an entire test structure, which the teacher then reviews and fine-tunes. But when assigning homework to students, it may need to follow strictly the pace and difficulty boundaries set by the teacher; it is not appropriate for it to improvise creatively and add unnecessary flourishes for the students.
For another example, in a teaching-and-research enablement scenario, a multi-Agent collaborative system can assign different responsibilities to different Agents: one focusing on comparing textbooks, another on data analysis, and another on designing classroom questions, with the teacher serving as the overall director. But to make such a system useful over the long term, its boundaries must be tested repeatedly through engineering experimentation, rather than endlessly piling on the number of models, parameter sizes, and seemingly complex flowcharts.
This is what he means when he says that development in the field of AI has never followed a single path. When we look at products through an Agentic lens, the question should not be how powerful our Agent is, but how powerful—or how limited—it should be in this specific scenario so that it truly makes people more capable.
The real threshold of the individual era: how many intelligent agents you can lead to work together
The theme of our conference is “Pioneering Intelligence | The Individual Era.” The title sounds somewhat grand and somewhat abstract. But placed within Zhang Minsong’s framework, it becomes much more concrete.
The combination of advances in large models’ reasoning capabilities, the maturation of the open-source ecosystem at the tools layer, and the rapid decline in computing costs to an affordable range has brought about a new reality: AI is no longer merely software, but increasingly resembles a virtual employee capable of planning autonomously, retaining memories over time, skillfully using tools, and even understanding how to divide work and collaborate.
Traditional organizations have a monopolistic advantage over resources and capabilities: accomplishing something complex often requires the support of a team, a department, or a company. For an individual to make an impact within such a system, they typically must first submit, then accumulate experience, and then prove themselves before gaining even a little power to mobilize resources.
But when one person can direct a dozen or even dozens of AI agents to work together, this monopoly is torn open. You no longer need a complete content team to continuously produce high-quality text, images, videos, and courses; you no longer need a large operations team to conduct user research, plan events, analyze data, and review campaign performance; you do not even need a large product team to use AI to iterate from prototype to an MVP — a minimum viable product.
This is the real-world version of the super-individual: not some romantic theory of innate talent, but the actual productivity gains achieved by one person + N agents.
Under this structure, the coordinates by which we measure an individual’s value also change. In the past, we assessed someone’s value largely by how much they could do with their own hands: write code, create designs, lead teams, and win clients. In an era of AI deployed at scale, the more important questions become: What kinds of problems can you break down into executable tasks? What kinds of workflows can you design and hand over to AI? How will you continually adjust objectives, evaluate results, and take responsibility throughout the process?
In other words, the core of future productivity is not what you can do, but what you can get a group of agents to accomplish for you.
For teachers, this means you do not need to master every technical detail, but you do need to be willing to let AI into your classroom and lesson-planning desk, learn to hand off repetitive work, and shift your time and energy toward the more valuable end of the spectrum—understanding students, designing classes, and inspiring thought.
For entrepreneurs and managers, this means you need not obsess over how many people to hire or how large a team to build. Instead, you should ask: In this company, what responsibilities should people and agents respectively take on? What mechanisms can we use to enable a small team to achieve what previously required an entire large department?
And for every ordinary individual, what is most worth guarding against in this era is actually two extremes: one is fear—the worry that AI will take away all the jobs, leading you to refuse to learn or experiment; the other is blind faith—placing all your hopes in a single supposedly all-powerful tool and expecting success with one click.
Zhang Minsong has offered a more pragmatic answer through his own path: respect the boundaries of technology, and respect the unique value of human beings; always keep moving forward, but always know why you are moving and whom you are moving for.
When we talk about “Pioneering Intelligence | The Individual Era”, we should not be talking only about the hot trend in a particular industry, nor merely about the opportunities available to a certain type of person. We should be talking about a more fundamental choice: Are you willing to treat AI as an entire invisible team, then begin learning how to lead it—breaking down something you originally could not accomplish into a series of small things that can be done, and completing them one by one?
If you are, then the only thing separating you from the so-called super-individual is today—the day you deliberately begin practicing.
Selected Interview Q&A
Q1: Could you briefly introduce yourself and Century Tianhong?
Zhang Minsong: Early in my career, I worked as an architect and senior expert at companies including Shanda, Baofeng Video, and Cheetah Mobile. I also worked on products with DAU in the tens of millions. I later became a serial entrepreneur focused on the education sector, working in areas including Tongbu Xue, online courses, and online public speaking. I have been the market leader, and I have also fallen into difficult periods. I joined Century Tianhong in 2021. It is a listed company that has specialized in K12 educational aids for 30 years. Its core brand, “Zhihong Optimization,” covers more than 30 provinces and cities nationwide, and its books have a cumulative circulation of more than 1 billion copies. The company is now positioning “AI+ education” as its second growth curve.
Q2: What exactly does Xiaohong Teaching Assistant do? Whose problems does it primarily solve, and what are they?
Zhang Minsong: Xiaohong Teaching Assistant is an AI teaching-assistant tool designed specifically for teachers. It focuses on improving efficiency across the workflow of lesson preparation, test creation, grading, and administrative tasks. Put simply, it helps teachers reduce the time-consuming work of searching for materials, preparing lesson plans, creating tests, and grading assignments, giving them more energy to devote to actual teaching and interaction with students.
Q3: Why do you say lesson preparation is particularly well suited to transformation through AI?
Zhang Minsong: Lesson preparation is both the most labor-intensive part of a teacher’s work and the part that most tests their teaching and research capabilities. In the past, it relied mainly on information retrieval and assembling templates, making true personalization difficult. Generative AI can make up for these shortcomings in knowledge integration, cognitive simulation, and personalized generation, allowing high-quality resources to be “customized at scale,” with teachers making the final decisions and refining the results.
Q4: How do you view the current wave of generative AI? What is its biggest impact on education?
Zhang Minsong: Large models have transformed AI from a “functional tool” into a “cognitive assistant.” It is no longer limited to marking questions and assigning scores; it can also participate in designing knowledge structures and teaching pathways. For education, it may, for the first time, balance quality, scale, and personalization simultaneously, driving the entire industry into the intelligent era of large models.
Q5: How do you avoid “AI for AI’s sake” amid rapid technological evolution?
Zhang Minsong: On the one hand, we need to embrace it proactively: we launched a RAG-based lesson-plan generation product at the beginning of 2023. On the other hand, every innovation must ultimately return to teachers’ real-world contexts—we start by testing it on small applications, verify its value, and then scale it up. Most importantly, our AI products must be integrated with our core businesses of books and teaching aids, serving the same group of teachers rather than becoming an “isolated new toy.”
Q6: How do you understand Agentic AI? What are the main focuses of your practice?
Zhang Minsong: Agentic is more like a way of thinking centered on “AI autonomy + engineering.” Both highly autonomous and minimally autonomous AI can be considered agents; the key is matching them to the right scenarios. In specific educational contexts, we continually fine-tune this “autonomy dial,” using model capabilities, specialized data, and workflow design to find the “just-right” level of intelligence, rather than piling on parameters and concepts indiscriminately.
Q7: From your perspective, how do you understand “Pioneering Intelligence | The Individual Era”?
Zhang Minsong: The threefold breakthroughs in large language models, open-source ecosystems, and computing costs have made AI more like an “invisible team,” enabling individuals to mobilize multiple agents to complete complex tasks together. Future competitiveness will no longer be measured by “how much I can do myself,” but by “what I can accomplish with AI.” The people and organizations that understand technology while also appreciating the value of education and the humanities will go further in this era.