Vibe Coding: The Second Half — The Four Kings and Those Who Want to Win | A Conversation with Zhu Guangxiang, Baidu Miaoda Product GM

🚥 In 2026, as Claude Opus 4.5 advanced and OpenClaw went viral worldwide, the Vibe Coding track also entered its second half, splitting into two camps: one side continued to center on code and IDEs, betting on极致提效 for programmers; the other began to center on "intent," trying to make software the natural result of expression.

This week's Crossing guest 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 that ran over his own past 20 years.

In the second half of 2024, when everyone was racing on IDEs and chasing Cursor, Miaoda chose a slightly "heretical" route — No Code. Guangxiang said at the time, "basically nobody believed it."

In this episode, we discussed:

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

➤ Real-world cases: A 50-year-old doctor building a hospital website with Miaoda, a super-individual earning over 100K RMB annually with Miaoda, a 12-person small company delivering two projects in a month for 700K 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 make money sooner or later

➤ Claude Skill: "Claude Skill is just making public the tricks we used last year" — 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 "enjoying model dividends"?

➤ As one of Baidu's "young backbone generals" born in '93, how does Guangxiang honestly evaluate his experience building products at Baidu? — Where's the friction, where's the boost? Will "rising early but arriving late" continue?

A thread running through the entire conversation: history has always been about wheels running over themselves. Assembly ran over machine language, high-level languages ran over assembly, and natural language is now running over high-level languages. If Guangxiang could redo college, he says he probably wouldn't study computer science — "Kindergarten-level Miaoda is enough; college should be spent on a vertical specialty where you truly understand the business."

🎬 Our video podcast is now live on @Koji Yang Yuancheng's channels on WeChat Video, Xiaohongshu, Bilibili, YouTube, and other platforms.

📒 The transcript will be published on the @CrossingCrossing WeChat official account.

🟢 00:09 Lightning Round: Age, alma mater, MBTI and zodiac sign, one-sentence intro to Miaoda, revenue and profit, pre-startup experience

🟢 01:16 From "Nobody Believed It" to "Eyes Lighting Up"

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

"My users believed before I did"

A 50-year-old doctor built a hospital website with Miaoda — Guangxiang himself didn't recognize it as AI-made

The turning point: "I suddenly had this Truman Show feeling — I'd been earnestly browsing content, then someone told me it was all AI-generated"

🟢 04:38 "The Inverted Pyramid"

The industrial structure is shifting from a正金字塔 to an inverted pyramid — the application layer makes the most money, then反哺 the tool layer and hardware layer

The money-making story of super-individual "Huang Ama"

The comeback of a 12-person R&D team: originally project cycles were half a year to a year, with Miaoda they deliver one project per week, 700K RMB落地 in a month

Next year's goal: cultivate 10,000 super-individuals, users have over a billion RMB in ARR, we take 10% that's 100 million

🟢 08:56 AI Coding's Four Heavenly Kings and Cross-Border Invaders

The four foreign kings each have their strengths: Lovable is lightweight and balanced, Replit is backend-heavy, Bolt.new has strong IDE, v0 is frontend极致

Trae launched Solo mode, Manus 1.5 added Vibe Coding, Coze changed its name to Coze Coding

"We have very few direct competitors domestically, but many indirect ones, because everyone realizes this might be the future"

🟢 11:55 Miaoda's Moat

Miaoda is alive: every user thumbs-up or thumbs-down, whether an app goes live, where they get stuck and revise repeatedly — all feed back into the model, teaching it how to generate better next time

Backend capabilities are in a league of their own: one of three leading AI builders officially recognized by Supabase

Databases created by Miaoda in one week exceed the total accumulated by Baidu's traditional ToB database team over 7 years

"One sentence isn't a requirement, it's an idea — from idea to requirement is a world apart"

🟢 19:34 MiaodaBench: Why Evaluation Matters More Than Training

Shunyu Yao said in AI's second half, evaluation matters more than training. Guangxiang added: evaluation in the lab isn't real evaluation; evaluation in user scenarios is.

App generation has 100+ tasks, we have Benchmarks at every step — this set of Benchmarks is Miaoda's taste

Whether it looks good, whether it's usable — let users vote with their feet: going live, sharing, dwell time, all are signals

Multi-agent multi-model routing architecture: Benchmark is the recipe, models are ingredients, routing architecture is the cookware — we cook ingredients into what users most want to eat

🟢 24:09 Claude Skills

Hot take: The essence of Skill is dynamic loading, is context management — load what's needed, click where needed

Last year Miaoda users would get stuck after 50 rounds of revision, we made transformations using the same principle as Skills, now unlimited rounds of revision are possible

"Claude Skills made our secret public, everyone's back at the same starting line"

Miaoda's plugin system has three categories: API plugins, Prompt plugins (similar to Skills on the market), Code plugins

🟢 27:25 The Wheel Running Over 20 Years

Started coding in middle school, wrote code for 20 years, after PhD didn't become a programmer but a product manager, then built a product that lets everyone write no code. Colleague joked: you built a wheel that ran over your own past 20 years.

"I'm the first person in the ten-year history of Tsinghua's IIIS to switch advisors" — because seeing AlphaGo defeat Lee Sedol, decided to switch to reinforcement learning

The entire history of computing is wheels running over themselves: machine language → assembly language → high-level language → natural language

What was C's compiler written in? First version was assembly — assembly built a wheel that ran itself over

"If I could redo college, I probably wouldn't study computer science. Kindergarten-level Miaoda is enough; college should be spent on a vertical specialty where you truly understand the business"

🟢 34:21 Why No Code Will Definitely Win

In the second half of 2024 everyone was racing on IDEs chasing Cursor, Miaoda's choice of No Code was considered "non-mainstream." A year later, everyone's crowding toward No Code.

Sutton's "bitter lesson": methods that leverage compute best are most efficient; over-reliance on human experience loses generalization capability

Workflow is not essential

IDEs are like three-legged races: AI and human bound together running, each hindering the other, exhausting. No Code is like a relay: AI runs 99 steps, human runs 1 step, just need to take the baton and sprint

"People who use IDEs might use No Code, but people who use No Code will definitely not use IDEs."

🟢 41:30 What Was Done Right, What Was Done Wrong

Done right: chose no-code over low-code, chose 8 billion people over 30 million programmers

Done wrong: gave the model too much freedom — "the model is an arrogant young person, can't recognize its own boundaries, insists on doing what it clearly can't do"

Really done wrong: operations started too late, no operations team for half a year, still no independent operations budget to this day, all蹭 internal traffic

🟢 46:06 How Not to Get Eaten by the Model?

If Miaoda fails, what's the biggest risk? — Misjudging the model's extension line, thinking what we're doing isn't on that line, then getting internalized by the model.

The best approach for all AI products is maintaining a 15-degree angle with the model: enjoy model dividends, but have some portion that won't be internalized in the short term

From a reinforcement learning perspective: the environment can also be internalized — compilers are code, operating systems are code, once the model learns them it doesn't need external environment

"Cursor will definitely be replaced by Claude" — IDE interaction is not essential, compilers are essential, but compilers can be internalized by the model

But this cycle is long: internalizing compilers takes 3-5 years, internalizing OS takes 5-10 years — so maintaining a 15-degree angle is the best state for now

Who handles the 15-degree angle best? Manus.

"We hire PMs like Manus's, we hire engineers like Shunyu Yao's"

🟢 49:38 A Young Baidu General's Internal Perspective

Colleagues say "there's no normal person on our team."

How do you view the boost Baidu gives you?

How do you view the friction Baidu gives you?

How do you view "Baidu rises early but arrives late"?

Miaoda team's Crossing: product, operations, engineering, strategy — four corners, all decisions emerge here

🟢 54:15 If Given $3 Million to Invest

"I'd invest in Google/Gemini" — technical DNA, the largest volume of search data, the world's largest distribution channel

"Baidu can be considered China's Google, so Baidu also has hope — depends on whether they can manage internally well, haha"

Welcome to subscribe to Crossing: 🚦 We follow the new industry changes and entrepreneurial opportunities brought by the new wave of AI technology.

🚦 Crossing is Steve Jobs's metaphor for Apple, describing it as standing at the crossing of technology and liberal arts, where great products are often born. AI is bringing change to all industries; we seek out, interview, and gather a new generation of AI entrepreneurs and active actors in the AI era, and together with them, explore and embrace new changes, new possibilities.

👦🏻 Host Koji: I founded Crossing, launched AI Hacker House — a community space for a new generation of AI entrepreneurs, and serve as Venture Partner at ZhenFund. I believe technology, especially AI, is the greatest value creation opportunity of our generation. Koji on Jike, Koji's website

👧🏻 Host Ronghui: I co-founded Crossing, worked at a USD VC, and spent five years as a Silicon Valley correspondent, following tech development and business stories. Welcome to chat with me, exchange with me. Ronghui on Jike