Summary

朱广翔’s core view is that the second half of Web Coding will shift from IDEs to No Code, and No Code will also replace Workflow. At least 200 AI Coding products appeared on Product Hunt in 2025; the first half was dominated by IDEs, while No Code overtook them in the second half and continued to widen the gap. The underlying market is “30M programmers” versus “8B potential creators”: “Code is cheap, show me the idea; idea is cheap, show me the app.”
Miaoda does not use its own ARR as the primary metric; it first validates whether users can make money, then captures value from application-layer revenue. A delivery company with a 12-person engineering team and 17 registered companies compressed project cycles from 6–12 months to roughly 1 week, delivering an elderly-care system and enterprise office platform in just over 1 month for RMB400K and RMB300K, respectively. 朱广翔 wants to create 10K “super-individuals” earning more than RMB100K next year, generating more than RMB1B in application-layer revenue; under his model, users return 10% to the platform, giving Miaoda roughly RMB100M, of which about RMB10M would cover underlying hardware and compute costs.
Miaoda is using an AI-native backend, a product-manager agent and multi-model routing to turn a “one-line demo” into a commercial application that can be deployed, distributed and operated. A single Query can generate the frontend, database and linked logic at once; the number of databases created in 1 week has exceeded the cumulative output of a traditional To B database team over 7 years. Application generation is broken into more than 100 tasks, with the model and Agent strategy selected for each cell according to its Benchmark. “The model is the ingredient, the Benchmark is the recipe, and the routing architecture is the cookware.”
The biggest product risk is not competition but misjudging the model’s extension line, only to see the frontier capability absorbed by the foundation model ahead of schedule. 朱广翔 argues that products and models should “maintain a 15-degree angle”: capture the capability dividend through multi-model routing while retaining areas such as voice, backend systems and the path from product design through deployment and distribution that remain difficult to internalize for now. He estimates that internalizing a compiler could take 3–5 years and an OS 5–10 years, while repeatedly stressing that these are probability judgments; higher hallucination rates could also prevent that end state from ever arriving.
The global “Four Kings” have differentiated by position in the value chain, while domestic IDE, general-purpose Agent and low-code players are converging on No Code. Lovable is strongest in lightweight development, templates and distribution; Replit focuses on heavier enterprise deployment and backend systems; Bolt.new is strongest in IDE capabilities and multilingual development; v0 pushes frontend fidelity to the limit. Domestic adjacent entry points include Trae Solo, CodeBuddy, Qoder, Manus 1.5, Coze Space and Coze Coding: “There are few direct competitors, but a great many indirect ones.”
Coding capability is shifting from a scarce moat to a general-purpose tool, with true product advantage moving toward industry knowledge and scenario understanding. A Sinopec engineer who could not code used Miaoda to build mine-design software now deployed at Daqing, Qinghai and Changqing oilfields and used for graduation projects at a petroleum university; software previously purchased by his company for RMB1.4M proved unusable because its programmers did not understand the business. That is why 朱广翔 says that if he could redo university, he would learn computer science earlier but choose a vertical discipline such as law or finance.
Long-term capital value still depends on technology, data and distribution closing the loop simultaneously; product leadership cannot substitute for operations. Baidu bet on “technology from 2 years in the future” and tolerated Miaoda’s early losses, but the team hired its first operations person only 6 months in and still has no standalone operating budget—an error 朱广翔 considers real. After Koji ruled out investing in Miaoda users as the official answer, he settled on Google: it combines technical DNA, data and global distribution.

Deep dive

1. 朱广翔 Turned 20 Years of Coding into the Starting Point for Self-Elimination

Born in 1993, 朱广翔 started coding in middle school, went on to earn a PhD from Tsinghua, and led 8 teams during 4 years at Baidu. He eventually became the only person in his cohort who was no longer a programmer. He accepted the joke that he had “built a wheel that ran over the past 20 years of his life,” and even felt he had “given it up too late.”
After AlphaGo defeated 李世石, he became the first person in his school’s 10-year history to change advisers, moving from computational biology into reinforcement learning. Before the switch, he had already raised the required impact factor for graduation from 5 to 20, and his protein 3D-modeling work had been selected as one of China’s Top 10 annual advances. When Koji asked why he abandoned such a promising path, his answer was that life sciences might still be 50–100 years from industrial results.
朱广翔 thinks on the scale of an entire life: “The longer I spend on the old timeline, the longer the detour I take.” He uses his first C compiler, written in assembly language, as the analogy—machine language, assembly, high-level languages and finally natural language expression. The history of computing is “a process of repeatedly using a wheel to run over yourself.” Today, roughly half of Miaoda’s requirements are generated by Miaoda itself and about 80% of its code is written by AI; the engineers are undergoing the same revolution.

2. Users Believed in “Building an App with One Sentence” Before the Team Did

Go back to the second half of 2024: DeepSeek did not yet exist, the public was unfamiliar with CoT and Thinking, and tool use was limited to shallow implementations such as Function Call. Coding models were broadly unreliable, and product thinking still centered on Chatbots. To build No Code websites, mini-programs and applications, 朱广翔 admits: “At the beginning, I didn’t believe it.”
Early demos relied on a half-engineering, half-model approach—“when intelligence wasn’t enough, humans filled the gap.” But a roadshow changed his mind: people who had been looking down at their phones immediately looked up when they saw the generation process, then began imagining their own academic homepages and business use cases. “My users believed before I did.”
The real turning point came through an accidental search. A website he examined carefully had in fact been generated by Miaoda, yet he still could not detect any AI traces when he reviewed it. Its creator was a doctor in his 50s who had built a hospital website used over the long term; it ranked first in Baidu search with an official logo. The shock felt like something out of The Truman Show, proving for the first time that this could be a serious, enterprise-grade business system rather than an “AI toy.”

3. User ARR Forms Miaoda’s Inverted-Pyramid Business Model

朱广翔 used the industry framework Robin discussed at the Baidu World Congress to explain the shift. In the old model, applications, model platforms and compute hardware formed an upright pyramid: GPUs made the most money, models came next, and applications generally lost money. A healthy structure should invert that pyramid, with applications creating the most value and returning part of their revenue to tools, models and hardware.
A super-individual using the online name 皇阿玛 built 3 projects with Miaoda: a short-drama platform generating RMB120K, a doors-and-windows website generating RMB30K, and an automotive paint-preview tool. The doors-and-windows project commanded far more than a standard website fee because consumers could upload photos of their homes and preview the installation effect, directly increasing purchase intent.
Another company had a 12-person engineering team and had registered 17 companies; traditional projects took 6–12 months. After switching to 4 project managers working with Miaoda, delivery took roughly 1 week. In just over 1 month, they completed an elderly-care management system and an internal enterprise office platform, generating RMB400K and RMB300K, respectively.
“As long as our users can make money, we will make money sooner or later.” The target for next year is 10K super-individuals earning more than RMB100K, implying more than RMB1B in application revenue. Under his model, users return 10% to the platform, producing roughly RMB100M in platform revenue, with about RMB10M then used to pay for underlying hardware and compute. 朱广翔 admits that, for now, the company relies mainly on participants in the “Dream-Building Plan” to report their results, leaving a large amount of underwater revenue unmeasured.

4. Products Must Maintain a 15-Degree Angle to Models, Not Bet on Permanent Non-Substitution

朱广翔 believes Miaoda’s biggest failure mode is “misjudging the model’s extension line”: the team may think it is building capabilities outside the model’s frontier, only to see the foundation model internalize them rapidly. The product cannot sit orthogonal to the model and miss the gains from iteration; it needs to “maintain a 15-degree angle.”
That angle currently includes multi-model routing, voice, backend systems and end-to-end engineering from product design through deployment and distribution. He believes Manus chose its angle well. But when Koji asked whether the angle would still be absorbed over the medium to long term, he did not evade the question: “Over a very, very long horizon, it will be absorbed.”
Unlike the brain-in-a-vat analogy, which assumes the environment can never be internalized, he uses model-free and model-based reinforcement learning to explain the opposing view. Asking for directions at every intersection means interacting directly with the environment; looking at a map means modeling the environment first. The environment can be abstracted as an MDP from the current state and action to the next state and reward. If a neural network learns that transition, the environment has entered the model.
He once said, “Cursor will definitely be killed by Claude Code, because the IDE is not fundamental; the compiler is.” Further out, the compiler—and even the operating system inside a virtual machine—is also code. He estimates 3–5 years to internalize a compiler and 5–10 years to internalize an OS, but preserves a clear uncertainty: hallucinations produced by heavyweight code generation could make that end state impossible.

5. The Four Kings Each Hold a Slice of the Value Chain as Domestic Players Close In from 3 Directions

Among the overseas “Four Kings,” Lovable is lighter in development and deployment, with stronger templates, tutorials and external distribution; its internal creed is “everyone is a CMO.” Replit is more professional, with greater flexibility and completeness in database configuration and enterprise deployment.
Bolt.new inherits cloud IDE capabilities and strengthens the development environment, multilingual support and related workflows for professional developers. v0 is frontend-centric, pushing high-fidelity reproduction of design drafts and interactive visual polish to the limit. The 4 companies are not the same product; each emphasizes a different center of gravity across development, backend, deployment and design.
The first domestic route comes from IDEs: Trae launched Solo, while Tencent’s CodeBuddy and Alibaba’s Qoder have also begun supporting one-sentence application creation. The second comes from general-purpose Agents: Manus 1.5 added Web Coding, while Coze Space and similar products followed with webpage generation.
The third is the migration of existing low-code, drag-and-drop and Workflow platforms toward Web Coding, with Coze’s shift to “Coze Coding” as the clearest signal. Add direct players such as Meituan No Code, Mashangfei and Xiangzhi, and 朱广翔’s conclusion is: “There are few directly equivalent products, but a great many indirect competitors.”

6. Miaoda Turns User Behavior into a Continuously Growing Data Flywheel

“Miaoda is alive.” Post hoc signals—including likes, dislikes, whether an application is published, and repeated revisions without launch—are fed back into the model. Tool calls, scaffolding code, task decomposition and multi-Agent orchestration strategies are also adjusted dynamically based on real usage outcomes rather than fixed at a single product release.
The plugin system fills capabilities the model cannot complete independently, including Baidu Maps, web search and external services. 朱广翔 stresses that novice users do not need one-shot generation; they need someone to handle development, deployment, launch and operations in advance. Plugins are therefore part of end-to-end delivery, not an auxiliary feature.
The scenario layer extends into distribution. On the private-traffic side, Miaoda trains dedicated mini-program languages, dependency packages and Agent workflows; on the public-traffic side, it connects Baidu, Bing, Google and other search entrances so applications can be published to the public internet with one click. Miaoda wants to control the value chain from “written” all the way to “seen and used.”

7. The Key to an AI-Native Backend Is Not Scale but Granularity and Elasticity

朱广翔 compares the frontend and backend to “a face and a brain”: the face has to look good, while the brain handles databases, authentication, payments and complex business logic. He calls Miaoda’s backend capability “in a class of its own,” and relays that Supabase once viewed its partners Lovable, Bolt.new and Miaoda’s overseas version MeDo as representatives of Europe, the United States and Asia, respectively.
The previous generation of databases was built around massive storage: large in size and few in number. AI applications may require one database per application, making databases smaller and more numerous, with capacity scaling up or down as applications demand. One intuitive figure: Miaoda creates more databases in 1 week than a traditional To B database-services team created cumulatively in 7 years.
The user is also changing from engineer to AI. Previously, developers used SQL and code to perform CRUD; the new database must be rebuilt around MCP, Agents, context and automated operations. Miaoda therefore lets one Query generate the frontend, database and linked logic simultaneously. In its early integration with Supabase, Lovable required multiple rounds of confirmation, a jump to create the project, copying a Token and then configuring it.
Koji’s inference was that this may not be beyond competitors’ capabilities, but instead reflect different product choices around exposing code and databases. 朱广翔 clarified that Miaoda allows users to view code but prohibits direct edits that could destroy an application; users can select a code fragment and request changes in natural language. Database schemas, rows and columns, and import/export remain manageable like Excel, while Lovable Cloud’s embedded approach still requires additional configuration.

8. Product-Manager Agents and Benchmarks Together Define an Application’s Taste

Most Web Coding products receive a Prompt and immediately write code. Miaoda first outputs requirements and product documentation for the user to confirm or revise. 朱广翔’s distinction is: “A one-line statement is not actually a requirement; it is a concept, an inspiration, an Idea.” The distance from concept to specifications executable by a development Agent remains “a hundred thousand and eight thousand miles.”
The not-yet-public “Miaoda Bench” evaluates complete application generation. Unlike SWE-bench, which primarily measures software issue resolution, it reflects a different substitution target: Coding models replace programmers, while Miaoda aims to replace an entire product-and-engineering team. Every link therefore needs to be measured end to end.
He cites 姚顺宇’s view that “in the second half of AI, evaluation matters more than training,” adding: “Evaluation in the lab is not real evaluation; evaluation in user scenarios is real evaluation.” Publishing, forwarding, dwell time, clicks and sustained usage all signal taste and usability. In response to Koji’s concern that early users may have biased tastes, the team added expert evaluation from designers, product managers and operators to reduce Bias.
Application generation is split into more than 100 tasks. For each task, the Benchmark selects the best model and Agent logic available, then switches automatically. 朱广翔 explains the moat through a restaurant analogy: “The Benchmark is the recipe, the model is the ingredient,” while multi-model and multi-agent routing is the cookware that turns those ingredients into the target dish.

9. Baidu Enables Long-Term Bets but Brings Alignment Costs and Brand Friction

Miaoda was born from Baidu’s approach of “using technology from 2 years in the future to guide today’s product”: even if capability was insufficient at the time, the team assumed it would become usable. The deeper first principle was that beyond 30M programmers, 8B people have ideas and business scenarios; turning them into creators represents far more incremental value than improving programmer productivity.
朱广翔 says bluntly, “Without Baidu, there would be no Miaoda.” Technology-driven idealism kept the team alive after it burned through substantial money. The friction is the same as at other large companies: more people, more opinions and more time spent on alignment. In response to comments such as “I would use it if it weren’t made by Baidu” and accusations that Miaoda “started early and finished late,” he says Miaoda is “catching the main market, not arriving late.”
The team culture tries to minimize organizational authority. When a rotating graduate recruit left, he said his greatest strength was “not listening to the boss”; engineers also challenge 朱广翔 directly. When arguments cannot converge, the only “God metric” is the user: return to WeChat groups and communities and let real users decide.

10. Cloud Virtual Machines Trade Isolation and Concurrency for a Safety Boundary for Novices

Discussing Claude Code, which had dominated that week’s conversation, 朱广翔 says it is local and belongs on the same track as IDEs. If a novice directly operates a computer while giving a local Agent unlimited access to files and tasks, it could crash the machine—hence the suggestion on the show to buy another Mac mini.
Miaoda and Manus choose cloud virtual machines, potentially opening a fresh environment for every application. Even if a virtual machine is broken, the user can start over. The cloud also enables horizontal concurrency: “I can open 1,000 Agents, 1,000 virtual machines and do 1,000 jobs.”

11. Claude Skills Standardizes Context Secrets, Sending Product Moats Back to Zero

朱广翔 believes the essence of Claude Skills is not mysterious: it is dynamic loading and progressive disclosure. Only the tools, MCPs, code and context required for a task are loaded, avoiding an overly long window or attention scattered by irrelevant information.
In Miaoda’s early days, user sessions would fail after roughly 50 rounds of edits, fundamentally because of context management. Around midyear last year, the team built a manual-like mechanism: AI first determines which resources to load, then inserts them into the system prompt. After launch, users could continue editing and the success rate per round improved because the model was “more focused.”
He jokes that Skill “gave away our secret,” while recognizing its larger contribution: turning internal industry Tricks into a common solution. Once the secret is standardized, everyone returns to the same starting line, and competition shifts back to how products organize and deliver these capabilities.
Miaoda plans to expose more of the implicit mechanism to users and divide plugins into 3 categories: API plugins connect external services; Prompt plugins carry complex workflows such as Claude Skills and Code Skills; Code plugins use Python and other code to process files, characters and formats. 朱广翔 says Miaoda has placed the plugin system in the top-level directory rather than hiding it as an attachment.

12. The Scarce Resource in No Code Is Not Programmers but People Who Understand Industries

From Day 1, the team viewed itself as heretical. 朱广翔 counts more than 200 serious AI Coding products launched on Product Hunt in 2025: IDEs dominated the first half, No Code’s share exceeded IDEs in the second half, and the gap continued to widen. The counter-consensus was becoming consensus.
The most persuasive example is a Sinopec engineer who could not program. The mine-design software he built with Miaoda is already used at the Daqing, Qinghai and Changqing oilfields and by students at a petroleum university for graduation projects. His company had previously spent RMB1.4M on software built by programmers, but it was unusable because they did not understand the physical logic and presentation requirements of mine design.
If he could redo university, 朱广翔 would make computer science part of the general education core before studying a vertical discipline such as law or finance. For his children, he would use Coding as a tool for developing structured thinking and Web Coding as the means of implementing ideas. Just as he learned Pascal to train his thinking and used C in real work, C may eventually move into a “museum”; when people actually need to build, they may use Web Coding and AI.

13. The Bitter Lesson Led Miaoda to Choose No Code over the More Saleable Workflow

朱广翔 believes the team’s most important correct decision in 2025 was choosing No Code over low-code: “No Code will inevitably replace Workflow.” Workflow was more mature and easier to monetize at the time, but it hard-codes human experience into nodes; the more structured it becomes, the weaker its flexibility and generalization.
The decision was based on Sutton’s 2019 essay The Bitter Lesson: the methods that win over the long term make full use of compute rather than stacking human experience. The implementation path is Learning plus Search. 朱广翔 maps the former to the Model and the latter to the Agent—one performs computation inside the model, while the other searches outside the model through interaction with the environment.
The market remains asymmetric: 30M programmers versus 8B people. Non-programmers also spend years working in vertical industries and therefore understand requirements, scenarios and users better. Baidu’s Hackathon showed the same migration: the standard product-design-engineering team became a setup in which product and operations people could win without programmers, with outcomes determined by ideas rather than code.
He compares the IDE to a “three-legged race for 2 people” between human and AI, in which both must understand each other’s state. No Code is a relay in which “AI runs 99 steps and the human runs 1.” IDE users may shift to No Code out of laziness, just as filters simplified Photoshop and high-level languages ran over assembly. “The appeal of a lower barrier is far greater than people imagine.”

14. The Real Mistake Was Starting Operations Too Late, Not Giving Models Freedom Too Early

朱广翔 initially attributed the biggest mistake of 2025 to “giving the model too much freedom.” The model was like an arrogant young person who did not know its own limits: it would force through tasks it could not do, hallucinating without realizing it. But because it improved month by month, that free-form architecture may simply have been an early decision that turned out to be right rather than a directional error.
When Koji kept pressing for the real mistake, he gave a clear answer: operations started too late. The team hired its first operations person only 6 months after formation and still has no independent operating budget, relying on internal Baidu traffic. Users repeatedly asked, “If the product is this good, why don’t you promote it?” Technical leadership had not automatically become market awareness.
He explicitly rejects the popular view that China has no To C AI PMF. In application development, China has mobile payments, social platforms, information feeds, a large population and many scenarios; the key is to grow the market together with users. The team has also expanded overseas, but its domestic path remains focused on growing through the real needs of 10K super-individuals.

15. The Final Capital-Allocation Answer Still Comes Down to Users, Data and Distribution

Miaoda itself has a physical “crossroads”: product, operations, engineering and strategy sit at the 4 corners, with every member accountable only for Miaoda’s OKR and decisions made in the same work area. 朱广翔 avoids the noisiest central desks but walks to the crossroads to speak directly with each function.
Asked how he would invest $3M if required to back 3 entrepreneurs, he first chose Miaoda’s users because “they can turn it into $300M.” He added that the Dream-Building Plan had effectively already invested about $3M in early entrepreneurial efforts.
After Koji ruled out that official answer, 朱广翔 ultimately chose Google. Google has technical DNA and vast amounts of data; if search data could all be used to support models, other companies would struggle to match it. It also controls global distribution. He does not trade stocks, but said that if he did, he would buy it.

a16z, Anish Acharya: Is SaaS Dead? Do Margins Still Matter? Why We Are Not in an AI Bubble?Rules based investing with Methodical Investment's David Kaiser