Yuwuwei CEO Huaiting Zhang: Using Generative AI to Achieve "Teaching Students in Accordance with Their Aptitude, and Education Without Discrimination"

Zhang Huating, CEO of AI for Dance, is perhaps China's best-positioned entrepreneur to tackle the "AI + Education" proposition — as the former head of Baidu's Fengchao ad system and co-founder of GSX/TAL's predecessor. We are honored to have backed AI for Dance from the angel round onward, walking this journey with Huating.

Since ancient times, the ultimate goal of education can be summarized as "teaching students in accordance with their aptitude, and providing education for all without discrimination" — enabling every student to receive the most personalized education possible. As AI capabilities improve rapidly and costs drop just as fast, edtech innovators like AI for Dance are leading us swiftly toward this ultimate goal.

— Yusen Dai, Managing Partner of ZhenFund

The 2025 World Artificial Intelligence Conference (WAIC) concluded successfully on July 28 at the Blue Hall of Shanghai World Expo Center.

At the forum, Zhang Huating, founder and CEO of AI for Dance, delivered a speech titled Reflections and Practices on AI Application Entrepreneurship.

In his speech, Zhang stated that the entrepreneurial opportunity in AI applications lies in using generative AI to transform service industries into manufacturing industries, breaking the "impossible triangle" of large-scale personalization, high quality, and low cost.

He argued that the lack of explosive commercial adoption of AI applications stems from persistent issues with large language models: hallucinations, insufficient reasoning accuracy, and unstable outputs. This demands that AI application teams understand both the business domain and AI technology, finding the balance between model uncertainty and business fault tolerance. Teams should first establish a closed business loop, using operations to drive gradual AI capability deployment, and in the process discover the data flywheel suited to their specific scenario.

In the intelligent era, he emphasized, cross-disciplinary talent density and a culture of pragmatic innovation are central to organizational building, while human-AI collaboration will become the foundation of enterprise operations.

This article is republished from IPO Zaozhidao, edited and organized by ZhenFund. Below is the full text of Zhang Huating's speech:

Transforming Services into Manufacturing Through Generative AI

Today I'd like to share what our two-year-old startup has learned from its entrepreneurial practice in AI applications over the past two years.

Over a decade ago, my team and I at a major internet company first applied deep machine learning algorithms to build what was likely China's first large-scale advertising recommendation system. It performed well and gave us our initial understanding of AI.

Later, I started my first company in education with like-minded partners, and we were fortunate enough to list on the NYSE, which deepened my understanding of the education sector. When generative AI emerged, we saw it opening greater possibilities for "technology for good" and equitable education. So in 2023, my partners and I launched our second venture.

We believe that what matters most in education is having good teachers, and achieving education for all. Whether male or female, urban or rural, rich or poor, child or adult — everyone should have access to a great AI teacher for lifelong companionship. One that can provide personalized instruction, knowledge transfer, and Q&A based on each individual's interests, stage, efficiency, potential, state, and personality. Yet in reality, quality educational resources remain scarce and costly.

Based on our understanding of technology and education, the emergence of generative AI led us to conclude that the marginal cost of personalized educational resources will only decrease. Theoretically, it should approach the cost of real-time inference alone — potentially a 90% reduction from current levels.

As technology evolves, the magnitude of cost reduction will only grow. We believe such AI teachers should be available anytime, anywhere, and as intelligent systems strengthen, their knowledge bases will continuously expand to provide the most suitable personalized guidance across all domains for every individual.

I don't plan to discuss products or algorithms in detail today — there have been plenty of demonstrations these past few days that you've likely seen. Instead, I'd like to offer a different perspective, speaking as a serial entrepreneurial team about some of our observations from a business standpoint.

We believe one of the larger opportunities in AI applications is transforming services into manufacturing. Many service industries are labor-intensive and face an "impossible triangle": providing high-quality service at low cost while achieving large-scale coverage.

This is essentially a paradox. Take doctors, for example — we've all had the experience of waiting two hours at a hospital for a ten-minute consultation. Or the doctor orders tests, and you wait in another queue, sometimes getting in, sometimes not, requiring another appointment. For most individuals, quality service is both hard to access and extremely expensive. Generative AI shows us the possibility of delivering personalized services at scale, achieving both quantity and quality.

In the virtual digital world, we often hear the term "thousand people, thousand faces" — "thousand people" represents scale, "thousand faces" represents personalization. Recommendation systems have already solved the coexistence of scale and personalization in content distribution, but not yet in service industries, constrained by the limitations of recommendation system capabilities.

Using generative AI to transform labor-intensive industries primarily means replacing labor costs with computing costs. The trend is clear: computing costs will continue to fall while labor costs rise. Second, labor-intensive enterprises face extreme complexity in hiring, deploying, and developing talent. The loss of excellent people keeps management costs high and full standardization out of reach. Generative AI, however, may enable standardized services. Once services are transformed into manufacturing, perhaps each of us will eventually have our own AI teacher, AI lawyer, and AI family doctor.

For AI applications, why haven't we seen large-scale adoption and explosive growth yet? As a comparison, let's recall what enabled the mobile application explosion over a decade ago.

At that time, 4G networks were largely in place, smartphones were ubiquitous, and the hardware foundation was mature. Phones had positioning, cameras, and payment capabilities, providing the infrastructure for mobile applications. Amap, DiDi, and Meituan relied on positioning; Kuaishou and Xiaohongshu used cameras to record life; online education company GSX used audio-video livestreaming to make learning accessible anytime, anywhere. With payment capabilities, monetization could follow — otherwise, massive opportunities would be lost. Once this infrastructure was in place, mobile internet companies could focus on application-layer innovation without needing to build underlying systems.

Today, we find that models still hallucinate frequently, reasoning capabilities remain insufficiently accurate, and outputs are unstable even with identical context. On the multimodal front, real-time digital human interaction, facial stability under heavy occlusion, real-time generated expressions and vocal tone, and interaction latency all remain relatively weak.

I believe many companies haven't yet encountered this situation: hundreds, thousands, or even tens of thousands of concurrent inference requests at the same moment. This requires optimizing substantial underlying architectural capabilities — a fairly high bar for a startup team today.

On one hand, you need to judge AI's development trajectory and iteration speed. Over the past year, its evolution has already surpassed Moore's Law.

On the other hand, you must clarify current model capabilities and application boundaries. For example, text-to-image may be directly applicable, but text-to-video may not yet meet expectations for short dramas. In applications, you must also balance model uncertainty against business fault tolerance — this is crucial.

With recommendation systems, you can swipe away a short video you don't like, or skip an ad. But if an AI doctor is going to perform surgery on you, can you allow it to make mistakes? So model uncertainty and business fault tolerance are intimately connected. Finding the right balance between when to use model capabilities and when to use system capabilities is the "real problem."

For the path of AI application entrepreneurship, our understanding is this: first establish a business closed loop, validate the effectiveness of the application scenario through that loop, then gradually use models to assist or replace certain links in the loop, ultimately achieving AI-driven transformation of the business. This may be a relatively pragmatic, incremental path.

In this process, the core question is whether all data in the closed loop can reach the cloud. Can the system capture all interaction data and static features, and consolidate them into a high-quality feature set? Only then can you use this effective data to train models and drive the ultimate transformation of AI applications.

This thinking draws from our summarized experience of internet entrepreneurship history. Many disruptive technologies emerge from the pressure of mature businesses. Alibaba Cloud and Amazon Web Services both originated from their e-commerce businesses facing concentrated, explosive usage pressure, which drove the birth of cloud services. Cloud services first supported internal operations, then capability overflowed to serve external customers.

The Finiteness of Carbon-Based Life, the Infiniteness of Silicon-Based Life

For AI applications, there's also a fundamental question: use AI to empower, or use AI to replace? We believe both will exist, manifesting in different businesses or processes.

If AI is empowering, it likely turns people into "Iron Man" — the path being intelligent assistance with human decision-making. Clearly, human ceiling is higher, because today's AI hasn't reached top human levels compared to the most reliable people. But with AI assistance, achieving partial standardization can likely narrow the variance in human performance.

Since final decisions still rest with humans, you can imagine that even if someone makes one decision per second, their daily ceiling is 86,400. Constrained by this, business growth can only be linear. Of course, with AI assistance, costs will decrease somewhat, but ultimately team organizational capability still rests on management and systems.

If a job can be fully replaced by AI — that is, unmanned — the likely path is intelligent system-driven. But since today's AI isn't 100% accurate, human fallback is still needed.

Clearly, in the long term, unconstrained by human limitations and scalable through computing power, exponential growth becomes possible. From current AI levels, the ceiling certainly won't match humans, but the average will be significantly raised, variance can theoretically approach zero, and costs will drop by orders of magnitude. Under this system, organizational capability need only rest on fully intelligent systems — higher efficiency, lower cost, faster iteration.

People often ask: do AI applications actually have data flywheels? Recently, both Google and OpenAI announced that their models now perform very well on difficult problems in Olympic mathematics competitions. For tasks with deterministic answers, today's model capabilities far exceed most humans. In such cases, information gained from human interaction is no longer sufficient to improve its intelligence.

For example, if you ask a model: "Can I still buy NVIDIA stock?" If information is sufficiently complete, there theoretically exists a standard answer — it no longer needs to obtain better solutions through human interaction. Or consider a task with complete constraints: "Book me a second-class high-speed rail seat from Beijing to Shanghai departing at 7 a.m." The agent simply executes — no other choices, no iterative optimization data flywheel.

So what constitutes a data flywheel? Suppose you request: "Order me a good lunch delivery." This task requires the agent to know who you are, likely your usual lunchtime, and to calculate delivery timing. What cuisine do you prefer? What's your taste? Where are you located? Do you prefer budget options or prioritize quality of life? What's your price range? And shouldn't recent orders avoid repetition?

This information heavily depends on personalized interaction records accumulated through continuous use — this forms a data flywheel.

More complex still: "I want to improve my English ability." Which aspect of English ability — listening, speaking, reading, or writing? How should it be improved? What's your current level? What are your personal learning habits and efficiency? These are all static features and dynamic behaviors accumulated through continuous user interaction. Combined with long-term and short-term context to generate corresponding personalized interactions — this is how a true data flywheel gradually forms.

AI Products Are Born as Global Enterprises

Today, what should an AI application organization look like?

First, we believe talent remains paramount. Talent density must exceed business complexity. Especially today, you need both industry domain experts and AI technical talent. Truly integrating these two types is extremely difficult. Our company has experienced situations where AI talent said something wasn't "AI enough," that we weren't using models for everything; while domain talent said current models simply couldn't do it and we needed to stick with existing methods. How to integrate these two types of talent and create synergy is critical.

Second, you need a culture of "pragmatism + innovation." As mentioned earlier, establish the business closed loop first, then use AI to upgrade or transform. You need solid business capability to pragmatically create commercial value, while constantly tracking global technology development and understanding how to apply AI capabilities to your business.

Third, we can already see silicon-based life becoming a necessary member of organizations. In code development, many processes require AI tools like Cursor; in sales, numerous specific tasks are handled by AI. So human-AI collaboration will become the foundational operating paradigm for intelligent-era companies.

Employees with many years of experience may struggle to adapt — what then? On one hand, we give opportunities and time to drive change; on the other, we urge employees: look from the future to the present, change your mindset or be changed.

That's my sharing. To summarize what we follow in building AI applications, in sixteen characters: "Business-driven, intelligence-powered, human-AI collaborative, pragmatically innovative." Our company advocates delayed gratification, guiding everyone to neither overestimate short-term gains nor underestimate long-term accumulation.