Vibe Coding, Act Two: The Big Four, and Those Who Want to Win | A Conversation with Guangxiang Zhu, Baidu Miaoda Product GM

He built a wheel, and it rolled right over his own 20-year self.

👦🏻 Podcast interview: Koji

🥷 Edited by: Crossing

🧑‍🎨 Layout: NCon

🚥 In 2026, as Claude Opus 4.5 advances, unlocking the global viral phenomenon of OpenClaw, the Vibe Coding track enters its second half, splitting into two camps: one side continues to center on code and IDEs, betting on极致提效 for programmers; the other begins to center on "intent," trying to make software the natural result of expression.

This week's guest on Crossing is Guangxiang Zhu, General Manager of Baidu's Miaoda product. He's a Tsinghua PhD in reinforcement learning, wrote code for 20 years, yet at Baidu built a product that "doesn't let anyone write code." In his colleague's words: he built a wheel, and it rolled over his own past 20 years.

In the second half of 2024, when everyone was racing to build IDEs and chase Cursor, Miaoda chose a slightly "heretical" path — No Code. Guangxiang says that at the time, "basically nobody believed in it."

In this episode, we discussed:

➤ The route debate: Why are IDEs like a "three-legged race," while No Code is like a "human-machine relay"? Why does Guangxiang believe Cursor and its ilk might eventually be swallowed by large models?

➤ Real-world cases: A 50-year-old doctor building a hospital website with Miaoda, a super-individual earning six figures annually with Miaoda, a 12-person startup delivering two projects in a month for 700,000 RMB — they're not programmers, but they understand the business better than programmers

➤ The counterintuitive business model: Miaoda doesn't look at its own ARR, only its users' ARR. The "inverted pyramid" theory — as long as users can make money, the platform will eventually make money

➤ Claude Skill: "Claude Skill is just us open-sourcing last year's trick" — what is the essence of context management?

➤ The 15-degree angle theory: How can AI products find breathing room between "being eaten by the model" and "riding the model's dividends"?

➤ As a Baidu "young general" born in '93, how does Guangxiang honestly assess his experience building products at Baidu? — Where's the friction, where's the boost? Will the pattern of "waking up early but arriving late" continue?

A thread running through the entire conversation: history has always rolled over itself with wheels. Assembly rolled over machine language, high-level languages rolled over assembly, and natural language is now rolling over high-level languages. If Guangxiang could do college over, he says he might not study computer science — "Kindergarten-level Miaoda is enough; college should be spent on a vertical specialty where you truly understand the business."

Listen on WeChat:

Listen on Xiaoyuzhou:

🎬 Our video podcast is also available on @Koji's WeChat Channels, Xiaohongshu[1], Bilibili[2], Youtube[3], and other platforms.

As the full interview is quite long (13,838 Chinese characters), here's a table of contents:

🟢 Lightning Round: Age, alma mater, MBTI and zodiac sign, one-sentence intro to Miaoda, revenue and profit, pre-startup experience

🟢 From "Nobody Believed" to "Eyes Lighting Up"

In the second half of 2024, there was no DeepSeek, no CoT, coding models were all subpar — betting on No Code at that moment, basically nobody believed in it.

🟢 "The Inverted Pyramid"

🟢 The Four Heavenly Kings of AI Coding and the Cross-Border Invaders

🟢 Miaoda's Moat

🟢 MiaodaBench: Why Evaluation Matters More Than Training

🟢 Claude Skills

🟢 The Wheel Rolling Over One's Own 20 Years

Started coding in middle school, wrote code for 20 years, after PhD stopped being a programmer to become a product manager, then built a product that lets everyone stop writing code. A colleague joked: you built a wheel, and it rolled over your own past 20 years.

🟢 Why No Code Will Definitely Win

In the second half of 2024, everyone was racing to build IDEs and chase Cursor; Miaoda's choice of No Code was seen as "non-mainstream." A year later, everyone's cramming into No Code.

🟢 What Was Done Right, What Was Done Wrong

🟢 How Not to Get Eaten by the Model?

If Miaoda fails, what's the biggest risk? — Misjudging the model's extension line, thinking what you're doing is off that line, but then getting internalized by the model.

🟢 A Young General's Inside View of Baidu

A colleague said, "There's not a normal person on our team."

🟢 If You Had $3 Million to Invest

Lightning Round

👦🏻 Koji

This week's guest on Crossing is Guangxiang Zhu, General Manager of Baidu's Miaoda product. We'll start with a lightning round to help everyone quickly get to know Guangxiang. Your age?

🧑🏻‍💻 Guangxiang

Born in '93, now 22, wait no. 32, turning 33 soon. My memory might still be stuck at 22.

👦🏻 Koji

Same here, dreaming I'm still 22. Alma mater?

🧑🏻‍💻 Guangxiang

Tsinghua University.

👦🏻 Koji

MBTI and zodiac sign?

🧑🏻‍💻 Guangxiang

ENFP, the happy puppy; Pisces.

👦🏻 Koji

ENFP is the most enviable MBTI.

🧑🏻‍💻 Guangxiang

An MBTI not suited for tech, but suited for product — more on the emotional side.

👦🏻 Koji

Very suited for product. Introduce Miaoda in one sentence.

🧑🏻‍💻 Guangxiang

Build an app in one sentence. And this app isn't an AI toy — it's something that can actually be deployed commercially, so Miaoda is the platform with the most commercial applications in China.

👦🏻 Koji

Can you share the company's current revenue and profit situation?

🧑🏻‍💻 Guangxiang

We don't look at our own ARR, only our users' ARR.

👦🏻 Koji

We'll dive into that later. What were you doing before Miaoda?

🧑🏻‍💻 Guangxiang

Odd jobs everywhere, worked on many products. At Baidu, Miaoda is actually the eighth team I've led.

The Wheel Rolling Over One's Own 20 Years

👦🏻 Koji

I understand you started coding in middle school, wrote code for 20 years, all the way through your PhD. Then a few years into work, you stopped writing code, became a product manager, and built a product that lets others stop writing code too — Miaoda.

A colleague joked that you built a wheel and rolled it over your own past. How did that make you feel?

🧑🏻‍💻 Guangxiang

Getting run over by a wheel. I'm indeed a pretty rebellious person — I'm the first person in ten years since my college (Institute for Interdisciplinary Information Sciences, Tsinghua University) was founded to switch advisors.

👦🏻 Koji

The first?

🧑🏻‍💻 Guangxiang

Yes, at the time I switched because I saw AlphaGo beat Lee Sedol and decided to pivot to reinforcement learning.

👦🏻 Koji

How hard is it to switch directions at IIIS?

🧑🏻‍💻 Guangxiang

Before me, nobody had done it since the college was founded — there wasn't even a process for it.

👦🏻 Koji

Why was it so hard?

🧑🏻‍💻 Guangxiang

Maybe everyone was a good student, a well-behaved kid, not the type to do something that seems deviant to ordinary people.

👦🏻 Koji

Was it because switching advisors feels like a kind of "betrayal"?

🧑🏻‍💻 Guangxiang

It carries a lot of uncertainty. You've been working on a research direction for a long time; switching means starting from zero, and graduation might get delayed.

👦🏻 Koji

But weren't you worried about not graduating?

🧑🏻‍💻 Guangxiang

If you're on a detour, the longer you stay on the old direction, the longer the detour becomes. If you zoom out and look at life, graduating a year earlier or later doesn't really matter.

👦🏻 Koji

What "detour" were you on at first?

🧑🏻‍💻 Guangxiang

I started with machine learning, which wasn't itself a detour. But when looking for application directions, I chose computational biology.

Actually, before switching directions, my paper's impact factor was already at 20 (the graduation requirement was 5), so I had already met the requirements.

I was mainly working on 3D protein modeling, and it even made China's top ten annual research advances.

👦🏻 Koji

So basically a precursor to AlphaFold?

🧑🏻‍💻 Guangxiang

You could say that. AlphaFold came later with larger models.

👦🏻 Koji

Why switch directions when it had so much potential?

🧑🏻‍💻 Guangxiang

Because life sciences are still somewhat distant from industry — more research-oriented, maybe 50 to 100 years away from significant real-world results.

For example, I could see the structure of a particular protein clearly, but turning that into an actual drug to cure disease was still a long way off.

Reinforcement learning was different. I could already see robots surpassing humans in intelligence, making decisions. So I switched then, becoming one of Tsinghua's earliest reinforcement learning researchers.

👦🏻 Koji

At life's crossing, you always make decisive choices, even at a cost?

🧑🏻‍💻 Guangxiang

Yes. After graduating, I was also the only one in my class who didn't become a programmer — I went straight into product management. Twenty years of coding, basically wasted.

At Baidu, I switched teams 8 times in 4 years, essentially because I was always watching the broader industry trends. Though now I've been run over by the very wheel I helped create, abandoning 20 years of foundation, I actually think I gave it up too late. Those 20 years feel like a detour.

History has always been about running over itself with its own wheels.

Do you know what the first C compiler was written in?

👦🏻 Koji

Assembly language.

🧑🏻‍💻 Guangxiang

Right, it was written in assembly. So assembly language created a wheel that ran itself over.

Once you had a C compiler, who still used assembly?

👦🏻 Koji

Learning assembly was an absolute nightmare.

🧑🏻‍💻 Guangxiang

Yes. That's how history iterates. The entire history of computing is a history of languages progressively running over themselves.

From machine language of 0s and 1s that only machines could read, to assembly language that told machines what to do in something closer to human terms, to high-level languages (like C, Java, Python) that used something even closer to natural human language to describe "what I want to do."

So what's the limit of this curve? It's no longer using "human-like" language, but directly using human language to say what I want. That's Vibe Coding today. That's Miaoda.

👦🏻 Koji

You just said you felt you "gave up too late" on coding. But interestingly, there are still programmers on the Miaoda team today. What do you feel when you see them?

🧑🏻‍💻 Guangxiang

They're running over themselves too.

Right now, half of Miaoda's own requirements are generated using Miaoda itself, and 80% of the code is AI-written. They're also using new tools to revolutionize themselves.

From "Nobody Believed" to "Eyes Lighting Up"

👦🏻 Koji

When you first started Miaoda, how did you understand what you were building? Did you have conviction from the start that a no-code application building platform was an inevitable future?

🧑🏻‍💻 Guangxiang

At first I didn't believe it. Rewind to the second half of 2024 — there was no DeepSeek yet, and nobody knew what CoT was, what thinking was, what tool calling was. Though there were some shallow implementations of tool calling, like through function call.

At that time, everyone's coding models were inadequate.

On the product side, Manus hadn't launched yet, and people's understanding of AI still stopped at chatbots. At that point, if you wanted to build a no-code product for creating applications, mini-programs, and web pages — basically nobody believed in it.

👦🏻 Koji

Looking back now, the second half of 2024 already feels distant, like prehistoric times.

When you said that, it suddenly reminded me — we once recorded a podcast episode where a guest said there was one major ChatGPT release you absolutely had to pay attention to: function call.

I remember it very clearly. He spoke about it with such passion, and I was thinking, is it really that important?

🧑🏻‍💻 Guangxiang

Looking back now, it was very important. Time moves so fast — 2025 feels like so much has happened.

👦🏻 Koji

So you didn't believe it at first. What made you gradually start believing?

🧑🏻‍💻 Guangxiang

I really didn't believe it at the time. Our internal demo was still somewhat customized, only targeting a few specific scenarios. The large models' capabilities just weren't there. "When intelligence isn't enough, use human effort to compensate" — that's artificial intelligence for you. So we used a lot of semi-engineering, semi-model approaches to cobble things together.

But as we kept working, I went from disbelief to belief.

The first turning point was when I went out to present this product, I discovered my users believed before I did.

I'd open with "make an app in one sentence," and the audience wouldn't know what I was doing — thought it was some pyramid scheme, completely baffled. But as I talked and demoed, after people saw it, I genuinely witnessed their aha moment. People who had been looking at their phones suddenly looked up, eyes wide open, visibly lighting up.

Many said that watching my demo made them think of so many scenarios of their own that could be built much more easily this way. And I suddenly remembered that I myself had had many such scenarios that Miaoda could address. Like when I was doing my PhD, my advisor required each of us to make an academic homepage — that's a personal-level scenario.

The second moment that shifted my thinking: one day I was searching on Baidu, found a website, and later a classmate told me — that site was built with Miaoda.

I was stunned. It suddenly felt like The Truman Show — I'd been earnestly browsing content, and suddenly someone tells me it's all AI-generated.

👦🏻 Koji

You had no idea it was made with Miaoda?

🧑🏻‍💻 Guangxiang

Right, I didn't even know it was AI-generated, let alone that it was Miaoda. I went back and studied it carefully — still couldn't tell. Later I reached out to the site's creator and found out he was a doctor in his 50s.

👦🏻 Koji

Wow, what kind of site did he make?

🧑🏻‍💻 Guangxiang

The entire hospital's official website. And it was actually live and in use, ranking first in Baidu search results, with an official logo.

This was the first case I saw with my own eyes — a real application in a business scenario, actually solving a real problem.

👦🏻 Koji

And this wasn't a user case you deliberately collected — you discovered it by chance?

🧑🏻‍💻 Guangxiang

Yes, completely accidental.

👦🏻 Koji

So the impact was quite strong.

👦🏻‍💻 Guangxiang

Right, these turning points showed me that the applications we were building could serve larger scenarios — serious, enterprise-level scenarios.

"The Inverted Pyramid"

👦🏻 Koji

You mentioned earlier that you don't look at your own ARR, but at users' ARR. Can you expand on that?

🧑🏻‍💻 Guangxiang

The reason we can afford not to look at our own ARR is that we promised users and management a metric that I believe is more aligned with first principles: Miaoda users' ARR.

What does this mean? Robin Li said something at this year's Baidu World Conference: our industrial structure is shifting from a pyramid to an inverted pyramid.

The old pyramid had three layers: applications at the top, models and platforms in the middle, and compute hardware at the bottom.

Think back — who made the most money in the entire AI circle? It was Jensen Huang's GPUs, the thickest bottom layer. Above that were the models, making the next most. Further up, applications basically lost money.

But a healthy structure should be an inverted pyramid: the application layer makes the most money, and after capturing value, shares some revenue with the middle tools and model layer; the tools and model layer, because they use hardware, then shares some with the bottom hardware layer.

That's the sustainable, healthy industrial model.

What Miaoda considers is an inverted pyramid industrial structure. We want to support more people in building applications with real industry value and making money from them.

There's a popular term lately called OPC (One Person Company). There are already many "one-person companies" making over 100,000 yuan through Miaoda.

👦🏻 Koji

Who's the user who's made the most money through Miaoda?

🧑🏻‍💻 Guangxiang

We haven't specifically tracked this, but from what I've casually learned, one individual made over 100,000 yuan, and one small company made 700,000. I haven't tracked beyond that.

👦🏻 Koji

But how do you track users' ARR?

🧑🏻‍💻 Guangxiang

We created a "Dream Building Program" to support users with commercial ambitions — they reach out to us proactively. But this is just the tip of the iceberg; there's definitely much more below the surface that we haven't captured.

👦🏻 Koji

Like the individual who made over 100,000 and the small company that made 700,000 — what were they each doing?

🧑🏻‍💻 Guangxiang

The super-individual who made over 100,000, online name "Huang Ama," did three projects. The first was a manga-drama platform, earning 120,000. The second was a website for a window and door store, sold for 30,000.

You might ask, how does a website sell for 30,000? Usually 3,000 would be about right. But he discovered a core pain point: users want to know what windows and doors would look like installed in their own homes.

So his website added a feature where users upload a photo of their home, select a window or door style, and see the installed effect online. This boosted purchase intent, and the store owner was so satisfied he paid 30,000.

The third project was a car painting tool that lets you preview effects before painting — also an AI tool.

Miaoda still doesn't support users writing code. He built everything with Miaoda, published many applications to the Miaoda gallery, and got his first customer through a message someone left online.

👦🏻 Koji

What about the small company that made 700,000? What was their background?

🧑🏻‍💻 Guangxiang

They were previously a traditional project delivery company — a 12-person R&D team, registered 17 companies, specializing in government and industry projects.

After adopting Miaoda, they replaced four project managers. Previously, project cycles were six months to a year. With Miaoda, they basically deliver a project in about a week. They'd only been using it for a bit over a month and had already delivered two projects, earning 400,000 and 300,000 respectively. One was a management system for a nursing home, the other an internal office platform for an enterprise.

So to bring it back — whether we're talking about super-individuals or one-person companies, our goal for next year is to incubate ten thousand of them. If each person earns a hundred thousand or so, ten thousand of them means over a billion in output. If the application layer generates over a billion in output, and they give us 10%, that doesn't seem unreasonable, right? Then we'd make a hundred million. We'd put ten million of that back into underlying hardware and IaaS services. That creates a very healthy inverted pyramid structure.

We believe that as long as our users can make money, we'll make money eventually, and the entire industry can develop sustainably. So our north star is our users' ARR.

How Do You Avoid Getting Eaten by the Model?

👦🏻 Koji

If Miaoda fails one day, what do you think would be the biggest reason?

🧑🏻‍💻 Guangxiang

Probably that we misjudged the model's trajectory. We thought what we were doing wasn't on the model's extension line, but accidentally ended up right on it, and finally got internalized by the model. That's probably the biggest risk.

I think the best posture for all AI products is to maintain a 15-degree angle with the model. The product direction absolutely can't be orthogonal to the model — then you wouldn't be able to capture any of the model's dividends.

👦🏻 Koji

What's your 15-degree angle?

🧑🏻‍💻 Guangxiang

We use a multi-model routing architecture to access the best model capabilities in the world.

At the same time, a large part of our work is something models can't internalize in the short term — like the cloud capabilities, backend capabilities, and the end-to-end flow from product design to deployment and distribution that we mentioned earlier. That's where our angle lies.

👦🏻 Koji

But looking at the medium to long term, do you think Miaoda, or Cursor, or Bolt.new, will ultimately get eaten by large models?

🧑🏻‍💻 Guangxiang

If you stretch the timeline to infinity, yes. I once read an interview with Peak Ji where he used "brain in a vat" as a metaphor, arguing that the environment can never be internalized.

I have a slightly different view. I believe that from an endgame perspective, the environment can be internalized too.

I used to work in reinforcement learning. The theories that large models are playing with now — Agent, Multi-agent, tool use, COT — the reinforcement learning field explored all of this long ago. Reinforcement learning has two directions: model-free and model-based.

Model-free doesn't internalize the environment; it learns through interaction with the external environment.

Model-based builds an internal model of the environment and simulates within the "brain in a vat."

For example, getting from Baidu to Tsinghua University by asking people along the way — that's model-free. But if I first look at a map, plan the route, and then walk, the map is a model of the real environment. That's internalizing the environment. The environment is essentially an MDP (Markov Decision Process), and this process itself can be learned by a model. Once it's learned, the environment is internalized.

Specifically for AI Coding products, what's their environment? For products like Cursor, the environment is the IDE, and the core of the IDE is the compiler. The compiler itself is code responsible for lexical analysis, syntactic analysis, and code transformation. If a large model learns the compiler's code, it can run code itself without needing an external compiler. Then products like Cursor might disappear.

Going further, virtual machines like Manus and what we use — their essence is the operating system. The OS is also code. If that gets learned by the model too, then we'll be internalized as well.

👦🏻 Koji

That's a pretty extreme thought experiment?

🧑🏻‍💻 Guangxiang

Yes, an extreme hypothesis.

👦🏻 Koji

Under this extreme hypothesis, what would Miaoda ultimately become?

🧑🏻‍💻 Guangxiang

At that point, the model is the product. But this cycle would be very long — internalizing a compiler might take three to five years, internalizing an operating system might take five to ten years.

👦🏻 Koji

Do you think this will definitely happen?

🧑🏻‍💻 Guangxiang

Maybe not. It's a probability question. Because when code generation becomes very heavyweight, hallucination and uncertainty increase, which could become an obstacle. So while it hasn't happened, or won't happen in the short term, maintaining a 15-degree angle with the model is the optimal state — allowing rapid iteration while capturing the benefits.

👦🏻 Koji

Is there any product you think perfectly leverages this 15-degree angle?

🧑🏻‍💻 Guangxiang

Manus. They've positioned that angle very well. We deeply admire Manus's PM. We joke internally that when hiring PMs, hire someone like Manus's; when hiring engineers, hire someone like Shunyu Yao.

The Four Heavenly Kings of AI Coding and the Cross-Border Invaders

👦🏻 Koji

AI Coding is now a battleground everyone wants to conquer, and a scenario with proven massive commercial value. Looking globally, how would you categorize the players in this space?

🧑🏻‍💻 Guangxiang

Internally we call the foreign players the "Four Heavenly Kings": Lovable, Replit, Bolt.new, and V0. Each has its own strengths.

Lovable's development and deployment process is relatively lightweight, with lots of templates and tutorials, and they've done very well on external distribution. They have a principle of "everyone is a CMO" — everyone needs to be able to build a good product and tell a good product story.

👦🏻 Koji

Sounds like Andy Lau.

🧑🏻‍💻 Guangxiang

Yes. Replit is more professional. Its deployment functionality is more flexible and comprehensive, more convenient in enterprise scenarios, supports various database configurations — they've gone very heavy on backend deployment.

The third, Bolt.new, started as a cloud IDE company. Their IDE capabilities are stronger, more suitable for professional developers. They provide more powerful capabilities in the development environment, like supporting multiple languages — they've gone very heavy on the development process.

The fourth, V0, has taken frontend interaction design to the extreme. It has many high-fidelity, high-restoration capabilities that let designers' original designs be perfectly reflected in applications. It's a frontend-centric product.

Overall, Lovable is relatively balanced; the other three each have their own long suits.

👦🏻 Koji

Domestically, besides Miaoda, which competitors do you pay attention to?

🧑🏻‍💻 Guangxiang

Domestically, there are relatively few products focused on no-code app generation that directly compete with us — like Meituan's "NoCode," "Mashangfei," and "Xiangzhi."

But we've noticed a phenomenon: since the second half of last year, many players who previously didn't do no-code have flooded in, even cross-border ones — like chatbot companies, companion chat companies — they're also penetrating our space.

We can categorize them. The first type is penetrating from the IDE side, like Trae. It used to be an IDE, then released a mode called "solo" that can build an app with one sentence. Tencent's CodeBuddy and Alibaba's Qoder are similar.

The second type is general-purpose agent builders. Their previous tasks were also no-code — making PPTs, doing deep research, setting scheduled tasks — and now they can also build apps without code. The typical example is Manus 1.5, which added Vibe Coding capabilities to make web pages. There are also many domestic products benchmarking Manus, like Coze Space.

The third type is low-code platforms, like drag-and-drop and workflow tools, also migrating toward Vibe Coding. This year, Coze announced it was rebranding as "Coze Programming."

So while we have few direct competitors domestically, we have many indirect ones, because everyone has realized this might be the future.

Miaoda's Moat

👦🏻 Koji

In such intense competition, what's your differentiation and competitive advantage?

🧑🏻‍💻 Guangxiang

First, Miaoda is alive — it iterates itself.

This manifests in two ways:

On one hand, we've built a model and agent self-growth, self-evolution architecture based on a data flywheel. Every time a user uses Miaoda — thumbs down, thumbs up, whether the app goes live, or whether they repeatedly modify something — these signals feed back to the model so it generates better next time.

Beyond the model, our agent strategies are also dynamic: how it calls tools, generates code, schedules tasks — all of this adjusts based on user behavior.

On the other hand, we provide a very comprehensive plugin system to extend Miaoda's capabilities. Need map services? Call Baidu Maps. Need web search? Call Baidu Search.

In short, Miaoda targets novice users who need end-to-end service. From generation to deployment to operations, we need to prepare everything for them in advance.

Second, many models are competing on frontend — making things look very pretty. But if data wants to be stored and used long-term, you still need a powerful backend.

👦🏻 Koji

This backend includes databases, authentication systems, payment systems, and so on?

🧑🏻‍💻 Guangxiang

Right — databases, payments, authentication, and backend logic.

You can think of the frontend as the face — it needs to look good. The backend is the brain — it needs to execute complex logic and store data. In backend capabilities, Miaoda is in a class of its own.

The backend itself is also a niche track. The global leader is called Supabase.

In the second half of last year, Supabase mentioned three leading AI builder partners in their blog: Europe's Lovable, America's Bolt.new, and Asia's MeDo (Miaoda's international version). That's official recognition of our progress.

Why is the backend so important? The AI-facing backend is completely different from traditional backends.

First, the scale and flexibility are different. Traditional databases are very large and heavy; their main job is storage. Databases in this era are very small and flexible, but numerous in quantity, because every app might come with its own database, requiring extreme elasticity for scaling up and down. You could say Miaoda creates more databases in a week than a traditional ToB database team accumulates in seven years. This requires us to have very strong, self-developed cloud capabilities.

Second, the previous generation of databases was built for engineers to read, managed through SQL and code. This generation of databases is built for AI to read, requiring transformation of various logic and optimizations oriented toward Agents.

Third, by mid-last year, we had already achieved generating an application with a single query that encompasses both frontend interaction and backend storage. At that time, Lovable needed n queries to accomplish this. It had to first generate the app, then the AI would ask whether to attach a database, the user would confirm, then jump to Supabase for a series of configurations, then paste the token back — a very cumbersome process.

Because we natively integrate frontend and backend, the experience is better for novice users. Later, Lovable also launched Lovable Cloud, embedding Supabase to reduce jumping around, but the experience remains fragmented, with additional configuration still needed inside.

👦🏻 Koji

I understand this might also be a product choice. Miaoda has a somewhat偏执 choice of not letting users see code, modify code, or directly edit database table structures. Lovable opens all of these up. This seems like a difference in product philosophy?

🧑🏻‍💻 Guangxiang

We don't let users directly edit code, but they can view it. We offer an alternative modification method: users can screenshot a section of code and tell Miaoda how they want it changed — Miaoda can analyze and execute that.

We only prohibit direct input behaviors that could cause damage.

As for databases, Miaoda supports users in viewing table structures, managing data, and uploading or downloading.

👦🏻 Koji

Can they modify database table structures?

🧑🏻‍💻 Guangxiang

Yes, they can edit rows and columns just like in Excel.

Our principle is: we open up capabilities comparable to what white-collar workers use in office software; we don't open up coding abilities that require specialized training to understand.

So writing code is off-limits, but database editing and management is allowed — we just make it more native, with much of the logic encapsulated inside the Agent.

👦🏻 Koji

Beyond these, are there other differentiators? For instance, I noticed Miaoda can build mini-programs, which overseas competitors obviously can't do.

🧑🏻‍💻 Guangxiang

Right. We've done specialized training and Agent logic for mini-programs' unique languages, dependency packages, and environments — that's a private-domain scenario.

In the public domain, users need traffic, and Baidu itself is in the business of website distribution, so we've integrated with Baidu Search, Bing, Google, and other search engines.

Applications built on Miaoda can be published to the public web with one click, reaching more people. We've gone one level deeper on scenario-specificity, offering one-stop service from development, deployment, distribution, to management.

Miaoda Bench: Why Evaluation Matters More Than Training

👦🏻 Koji

Any other highlights?

🧑🏻‍💻 Guangxiang

We also have a unique "Product Manager Agent" workflow.

Most Vibe Coding products take a query and output code directly. Ours takes a query, first generates a requirements document for the user to confirm and modify, aligns on it, and only then writes and executes code.

The "one sentence" in "build an app in one sentence" is the creative idea — it's still miles away from a product requirement that a development agent can actually understand. So we inserted a very professional Product Manager Agent in between.

Additionally, we've built an internal benchmark called "Miaoda Bench," and we'll be publishing the related paper soon.

👦🏻 Koji

What does this benchmark evaluate?

🧑🏻‍💻 Guangxiang

It evaluates application generation. Previously there were benchmarks for code generation, like the well-known SWE-bench, which mainly evaluates based on fixing software issues.

Code models that top these leaderboards replace a single programmer; Miaoda replaces an entire product-engineering team, needing to run through all stages and deliver applications end-to-end.

In this novice-facing scenario, each stage needs its own benchmark for measurement.

To quote Shunyu Yao again: "In the second half of AI, evaluation matters more than training." A product's benchmark determines its taste. We have benchmarks at every stage, and this constitutes Miaoda's moat and taste.

👦🏻 Koji

I understand benchmarks can relatively easily evaluate right versus wrong, but how do you evaluate subjective things like aesthetics?

🧑🏻‍💻 Guangxiang

Let me add to Shunyu's quote: Evaluation in the lab isn't real evaluation — evaluation in user scenarios is.

People often look at various model leaderboards, but many know that those rankings are just for show — the actual ranking in real usage scenarios is completely different.

Whether something looks good or works well ultimately depends on users "voting with their feet." We can analyze users' posterior behaviors: if they think something looks good, they'll immediately publish it, share it with friends.

The interactions, clicks, and dwell time that others leave behind when using it — these are all signals that it looks good.

👦🏻 Koji

But if you rely only on early users' evaluation, won't you be limited by their aesthetic level?

🧑🏻‍💻 Guangxiang

People generally agree on what constitutes usability. But aesthetics do vary from person to person. To reduce this bias, we combine evaluations from multiple channels:

Part comes from users, part from our internal expert team composed of designers.

Another part comes from our own product and operations teams.

👦🏻 Koji

Beyond the points above, anything else to add on differentiation?

🧑🏻‍💻 Guangxiang

One last thing: we have a multi-agent, multi-model routing architecture.

The application generation scenario seems small, but is actually enormous, involving over 100 tasks. This architecture automatically evaluates based on our benchmarks, and at each task node, selects the best model and best Agent logic to execute.

If benchmarks are the recipe and models are the ingredients, then this routing architecture is our cookware — it cooks the ingredients into exactly the dish users most want to eat. This architecture is also one of our core competitive advantages.

👦🏻 Koji

That's an interesting metaphor. The founder of OiiOii also said he's like running a restaurant.

🧑🏻‍💻 Guangxiang

Yes, a restaurant for the human heart.

👦🏻 Koji

At the very beginning of our podcast recording, you mentioned that you didn't actually believe in this at first — it seemed like a company arrangement. Can you talk about how this company strategy was formed?

🧑🏻‍💻 Guangxiang

I think Baidu has always used technology two years out to guide current product development. At the time the technology wasn't there, but they believed it would be in two years.

Also, there are only 30 million programmers globally, but 8 billion people total. Those 8 billion people all have ideas, business scenarios, and needs. If they could become creators, that creative output would far exceed that of 30 million programmers.

And for them it's going from zero to one, from impossible to possible — so the market space is even larger. It's a bet on the future.

A Young General's Inside View of Baidu

👦🏻 Koji

Building Miaoda inside Baidu, what's the biggest help and the biggest obstacle you've felt?

🧑🏻‍💻 Guangxiang

The biggest help is: without Baidu, there would be no Miaoda.

The decision to build no-code application generation aligned well with Baidu's technical DNA — using technology two years out to guide current product development. Baidu is relatively idealistic, less concerned with winning or losing every single battle. We've helped our boss lose a lot of money, but we're still alive because the company has conviction in technology.

The obstacles are the same as at any big company: lots of people means lots of ideas, and you need to spend enormous amounts of time "aligning."

👦🏻 Koji

I saw a comment on Bilibili saying, "If this weren't made by Baidu, I'd try it." How do you feel seeing comments like that?

🧑🏻‍💻 Guangxiang

It might be because our product's vibe and our team's style are somewhat different from traditional products. Many people say there's not a single normal person on our team.

👦🏻 Koji

Not a single normal person?

🧑🏻‍💻 Guangxiang

Right. Starting with myself — at a team dinner recently, a campus hire I'd been mentoring for half a year was leaving, and he told me he thought my greatest strength was not listening to the boss.

Then I thought about our technical engineers — they don't listen to me either, they just directly curse me out. Many of our developers are pretty fierce, they like to challenge things. But fundamentally it's because on our team, the only god metric is the user. When we're arguing to the point of no return, we look at WeChat groups, look at the community, and let users decide.

Our team has many other "abnormal" aspects. For example, our operations team includes former digital nomads who wandered everywhere before putting down roots at Miaoda; others are so deeply embedded in user groups that I can't tell if they're Baidu employees or users.

He shares with us how a user from Hainan sent him a coconut engraved with "Miaoda," how a teacher mailed him a collection of apps made by his entire class, how a guy who made over 400,000 yuan with Miaoda invited him over for New Year's Eve dinner.

Our product team is also a bunch of kids with wild imaginations — we're constantly roasting each other.

👦🏻 Koji

There's a saying online that "Baidu wakes up early but arrives late." Are you worried Miaoda will be the same?

🧑🏻‍💻 Guangxiang

At present, we're definitely catching the "big fair," not the "late fair." I'm very confident — this is absolutely still an "early fair."

Claude Skills

👦🏻 Koji

The week we're recording this podcast, Clawdbot is going viral at home and abroad. What was your reaction seeing it?

🧑🏻‍💻 Guangxiang

It made me think of Peak Ji — he previously talked about a philosophy that the cloud has unique advantages. Clawdbot is local, and should be in the same track as IDEs.

Miaoda chose the cloud because we serve novice users. If we let them operate their local computers directly with unlimited file and Bash permissions, they might crash their computers.

👦🏻 Koji

So a lot of people need to buy a new Mac mini, and don't dare mess with their main machine?

🧑🏻‍💻 Guangxiang

Right. We're similar to Manus — we have a set of virtual machines in the cloud where you can install whatever you want.

And these VMs are application-oriented: you spin up a new VM for each app you develop. Even if a VM gets broken, you just start fresh with a new app — no security concerns.

It's also highly efficient: I can spin up 1,000 Agents simultaneously, doing 1,000 jobs across 1,000 virtual machines.

👦🏻 Koji

Everyone's been discussing Claude Skills lately. What's your take?

🧑🏻‍💻 Guangxiang

Let me start with a hot take: Claude Skill is just making public some tricks we were doing last year. Of course, it's not just us — many peers probably did similar work.

The essence of Skill is dynamic capability loading, with context management at its core. If you dump all context onto the model at once, you either exceed the window limit or the model's attention scatters. This dynamic loading, progressive disclosure approach lets you "click wherever you need," very flexibly.

At the end of last year, Miaoda did a major iteration to solve this problem. Early users probably have painful memories — after 50 rounds of app modifications, Miaoda would say "sorry, I can't modify this anymore." That was a context management problem.

So we restructured it, introducing a Skill-like instruction manual mechanism. Before each AI task execution, it first reads the manual to understand what tools, code, and context to load, then executes.

👦🏻 Koji

This process is invisible to users?

🧑🏻‍💻 Guangxiang

Right. What users directly feel is: after this feature launched, apps could be modified for unlimited rounds, and each round's success rate was higher — the model felt more focused.

It's fundamentally the same principle as Claude's Skill. So internally we joked that Skill made our secret public — now everyone knows it.

Of course, what's impressive about Skill is that it standardized this logic and principle into a universal industry solution, benefiting everyone. We're all back on the same starting line.

👦🏻 Koji

There are lots of Skill aggregator sites online now. Will Miaoda consider letting users import Skills, or opening a Skill marketplace?

🧑🏻‍💻 Guangxiang

That was always in our roadmap. Skill operates on two levels:

First, we use it implicitly on the backend to manage context. Second, we open this capability to users so they can extend functionality further.

We're already working on this. In Miaoda, we call it "plugins." Of all the Vibe Coding tools out there, only Miaoda treats the plugin system as a top-level feature.

We have three types of plugins planned:

The first is API plugins, connecting applications through API services.

The second is prompt-based plugins, where you write an extra-long prompt to describe a complex workflow — this resembles the Claude Skills circulating in the market.

The third is code plugins, where a traditional compiler executes a snippet of code. For example, you might write a Python function for format conversion, then define it as a Skill plugin.

Why No-Code Will Inevitably Win

👦🏻 Koji

Is there any "anti-consensus" belief you hold today — something you're convinced of that others haven't realized or don't agree with?

🧑🏻‍💻 Guangxiang

I think what I'm doing right now is the anti-consensus. No-Code is heresy. Miaoda is heresy. From day one, we've been seen as outliers. Back then, everyone was racing to build IDEs because Cursor had validated PMF with programmers. Nobody thought everyday users could one day build apps themselves. But we stuck with it.

Looking at it today, we seem to have chosen right. Look at the overseas market in 2025 — by our count, over 200 AI Coding products appeared on Product Hunt. The first half of the year was dominated by IDEs, but from the second half onward, No-Code products far outnumbered IDEs.

By audience size, we made the right bet. Though we started as heretics, now even players from unrelated tracks like Manus and Coze are pouring into No-Code. It's slowly becoming consensus.

👦🏻 Koji

What advice do you have for people still studying computer science today, or engineers just starting their careers?

🧑🏻‍💻 Guangxiang

If I could do college over, I probably wouldn't choose computer science. I'd learn computers, Vibe Coding, and Miaoda in kindergarten, then study something more vertical in college — like law or finance.

Right now I really envy people who truly understand business and scenarios. Their ceiling is higher than people who just write code.

We have a user who's an engineer at Sinopec. He can't code, but he used Miaoda to build a mine design software. It's now actually deployed at Daqing, Qinghai, and Changqing oilfields. He even lectures at petroleum universities, where students use his software for their graduation projects.

Before he built it, his company spent 1.4 million RMB on software developed by programmers. But because those programmers didn't understand the business, didn't grasp the physics of mine design or the right interaction patterns, nobody used that software.

So vertical knowledge is key to building good software. If I had treated computing as a general education subject early on, then gone deep into a vertical industry in college, I'd probably have a stronger moat.

👦🏻 Koji

If you had kids, would you have them learn Vibe Coding from a young age, or start with basic programming fundamentals?

🧑🏻‍💻 Guangxiang

I'd have them learn Coding as a tool to develop thinking, and Vibe Coding as a tool to realize ideas.

Just like when I learned Pascal in middle school — that language was already half-obsolete, but it was similar to C. Learning it was just for programming thinking; for actual utility you needed C.

In the future, C might become a museum piece too, used only for teaching. For actual utility, it'll probably be Vibe Coding, using AI to boost efficiency.

What We Did Right, What We Did Wrong

👦🏻 Koji

Looking back at 2025, what did you and the Miaoda team do right, and what did you get wrong?

🧑🏻‍💻 Guangxiang

The single best decision was choosing No-Code over low-code. In other words, we bet that Vibe Coding would inevitably replace workflow.

I'm rebellious, but I also follow industry trends. I first think about the big picture, about what the endgame looks like. Based on the tech capabilities at the time, building workflow was clearly the fastest path to market, the easiest way to earn ARR.

But I come from a reinforcement learning background, and my ENFP personality is probably somewhat idealistic, so I thought about more fundamental questions. I thought of Sutton, the father of reinforcement learning, and his famous 2019 blog post, The Bitter Lesson.

He summarized that decades of AI development show that methods maximizing compute always win in the end, while methods overly reliant on human prior knowledge grow complex and lose generalization. How to best use compute? He proposed two paths: learning and search. This maps exactly to today's Model and Agent.

This theory guided us. Why firmly choose No-Code? Because No-Code is a better way to leverage compute, with less reliance on human experience.

Workflow is fundamentally unprincipled — it's like freezing human experience into fixed nodes, forcibly structuring things, which makes it inflexible.

So we saw that this year, Coze rebranded as "Coze Programming" — though I think they may already be a bit late.

👦🏻 Koji

That choice must have been difficult at the time. Did you struggle with it?

🧑🏻‍💻 Guangxiang

The technology wasn't mature then. All the observable phenomena and user needs actually argued against this judgment. But we returned to first principles.

First, market space: there are only 30 million programmers globally, but 8 billion people. Many of those 8 billion non-programmers are actually closer to real user scenarios. They understand needs better. If they build applications, their ceiling might be higher than programmers'.

We also noticed a trend at Baidu's annual Hackathon, where I've been a judge for years. The earliest teams always had "product + design + dev." After large models arrived, the first year still had programmers. By last year, many teams had no programmers at all — just product and operations people going direct, and they still won. And what determined victory wasn't who coded better, but who had better ideas.

👦🏻 Koji

That's fascinating. It used to be "Talk is cheap, show me the code." Now it's "Code is cheap, show me the idea."

🧑🏻‍💻 Guangxiang

You can even go straight to "Show me the App" — skip the PPT, use Vibe Coding to put the application on the table.

👦🏻 Koji

Right, today nobody has any excuse to pitch with just an idea.

🧑🏻‍💻 Guangxiang

Exactly. And people using No-Code definitely won't use an IDE, but people using an IDE might use No-Code. Take a typical programmer like me — if there's a lazier tool, why wouldn't I use it?

I compare using an IDE to a three-legged race: me and AI are tied together, needing to understand each other, coordinate — exhausting. A No-Code product like Miaoda is more like a relay race: AI runs 99 steps, the human just runs the final step.

And the human doesn't need to know how those 99 steps were run — just take the baton.

👦🏻 Koji

A 99+1 relay, with the human only sprinting the finish.

🧑🏻‍💻 Guangxiang

Or the human runs the first leg, tells it "I want to build this," then waits for it to reach the finish line.

So the experience is definitely better.

👦🏻 Koji

That's the history of computing — from complex Photoshop to one-tap filters on Meitu, constantly simplifying.

🧑🏻‍💻 Guangxiang

The appeal of low barriers far exceeds what people can imagine. Humans are fundamentally lazy. That's why high-level languages completely crushed assembly.

I believe natural language tools will eventually crush high-level languages too.

👦🏻 Koji

So what did you get wrong?

🧑🏻‍💻 Guangxiang

I think we gave the model too much freedom. Our mechanism was designed to unlock the model's potential, but the model is like an arrogant young person who can't recognize its own boundaries.

It constantly forces itself to do things it doesn't understand or can't do, gets it wrong, hallucinates — and doesn't even realize it.

👦🏻 Koji

Like the Jeff Chang lyrics: "How could I bear to blame you for your mistake — I gave you too much freedom."

🧑🏻‍💻 Guangxiang

Yes. But it's also a young person with growth potential. Its boundaries expand every month. So designing a free architecture is also betting that it will break through those boundaries itself.

In other words, this might not necessarily be wrong — just too far ahead of its time.

👦🏻 Koji

Was there anything that was genuinely wrong?

🧑🏻‍💻 Guangxiang

Yes — our operations started way too late. Users often ask me, your product works great, why don't you promote it?

Our team had no operations for half a year. Our first operations hire joined after six months. To this day we have no dedicated operations budget — we just piggyback on internal traffic. We should have been much more aggressive here.

👦🏻 Koji

How do you view the current wave of ToC AI apps going global?

🧑🏻‍💻 Guangxiang

We've gone global too. Many people say there's no PMF domestically, but I don't agree.

I think China is in a league of its own when it comes to application development. China's app market is the most vibrant in the world — the most advanced mobile payments, the most advanced social platforms, including our information feed products that move faster than overseas. And with so many people, there are more scenarios, more needs, a market of massive output value.

So what's the only thing I need to do? We grow alongside our users, we expand the pie together. It's like an inverted triangle — when users start making money, they share it with us. Once users have ARR, we'll have ARR sooner or later.

And with a market this massive in China, there's definitely huge ARR potential.

Next year, we'll grow together with our users and create 10,000 super individuals. Whatever these 10,000 super individuals need, that's what our product becomes.

Koji

I heard your team also has a "Crossing" internally?

Guangxiang

Yes, quite a coincidence. When we first built the team, we wanted to foster a startup atmosphere. Our product, engineering, and operations form a closed loop, with all roles working in the same space — that's rare in a big company.

When we arranged the desks, I drew a crossing to make communication more efficient. The four corners are product, operations, engineering, and strategy. Every decision, big or small, gets made at that crossing.

Koji

Is your desk in the middle of the crossing?

Guangxiang

I deliberately avoid the middle — too noisy. But I walk over to the crossing often to talk with everyone.

Koji

What a serendipity.

If You Had $3 Million to Invest

Koji

One last classic question. If you were given $3 million today and had to invest it as angel funding in three people, who would you choose?

Guangxiang

I'd definitely invest in Miaoda's users — because they can turn it into $300 million. And we're actually doing this; our "Dream Builder Program" provides support for users with entrepreneurial ideas and potential.

Koji

That's too polished of an answer. What if you can't invest in Miaoda users? Say, choosing from your friends — who do you think is most likely to succeed as a founder?

Guangxiang

That one I really need to think about... I feel like anyone who could actually succeed wouldn't let me invest.

Koji

Regardless of whether they'd let you.

Guangxiang

If anyone's fair game, then I'd definitely invest in Gemini.

Koji

You could just buy the stock directly.

Guangxiang

I don't trade stocks. If I did, I'd definitely buy it.

Koji

Why are you so bullish on it?

Guangxiang

First, Google has technology in its DNA. Second, it has the largest volume of data — something OpenAI and Claude simply can't match. Finally, it controls the world's biggest distribution channel.

So in terms of positioning, capability, and accumulated advantage, it's top-tier.

🚥

References

[1] Xiaohongshu: https://www.xiaohongshu.com/user/profile/548251dce779893bcf3f77bc

[2] Bilibili: https://space.bilibili.com/505301413

[3] Youtube: https://www.youtube.com/@kojiyang