When Alibaba Entered Global AI Coding: 60 Days on the Battlefield | A Conversation with 叔同, Founder of Qoder
Summary
AI Coding’s capital frenzy has expanded from copilots to IDEs, Agents and “digital employees,” but 叔同 believes these formats will likely converge into one. The host’s market markers were Cursor at roughly $500M in annual revenue and a roughly $10B valuation, alongside a sharply accelerating revenue curve for Anthropic after the launch of Claude Code; the shared underlying demand is “using Coding to deliver a software service, satisfy a business scenario and create commercial value.”
Qoder did not chase the flashiest “generate it from one sentence” use case, but instead bet on the “real software” maintained by more than 95% of professional developers. Software that already creates commercial value and must be iterated for five or ten years cannot be modified recklessly; it requires understanding history and controlling hallucinations. Qoder had 500,000 active developers two months after launch, with monetization “roughly in line with expectations,” and 叔同 sees real software as the high-value ground that later entrants can attack directly.
Model ownership materially affects the cost, margins and valuation of AI Coding products, narrowing the window for application-layer startups. 叔同 believes Cursor “will definitely” build its own models, because products that cannot jointly optimize with foundation models risk becoming high-cost “wrappers”; new founders without either a model or a first-mover category advantage must create a new model rather than fight Big Tech, Claude Code and GPT-5 Codex on the same terrain.
Qoder defines the Agent direction as asynchronous autonomous programming: humans clarify the objective, and AI delivers the whole block. Spec-driven planning expands a one-line request into business, architecture, module and acceptance documents before the Agent works for an extended period; Quest Mode takes the spec and codebase into a cloud sandbox, allowing one person to launch 10 Agents in parallel. 叔同 summarizes the result as “a genuine tenfold productivity increase.”
Qoder’s decision not to let users choose the model is a product bet that assumes risk at the outcome level. Asked what happens if Claude 5.0 is a breakthrough but users cannot select it, 叔同 maintained that “machine selection should outperform human selection, and platform knowledge should outperform individual knowledge,” with the system routing each problem to the right model; whether the “impossible trinity” of performance, efficiency and cost can be overcome depends on context engineering, not on stacking up a model menu.
Repo Wiki is Qoder’s concrete differentiator: it regenerates the codebase and its commit history into a continuously updated document that explains the system. It supports handoffs, onboarding, offboarding and module comprehension; generating a codebase initially took about 60 minutes, while 叔同 says the latest version has cut that time by 5x. Qoder rejects locking users in to drive retention: documents can be shared, exported and owned by users, because “differentiated capability and openness” are the basis of stickiness.
Alibaba’s chips for Qoder are models, engineering infrastructure and internal-startup-style autonomy, with a goal of becoming “top three in the world.” The team did not wait for Alibaba’s own model to become the global leader; it combined overseas models with Qwen and pursued developers first with the best global results. Tongyi Lingma continues to serve China, while Qoder has used a single global version since day one. The brand is framed as “the programmer in your pocket,” tied to neither a specific model nor a platform.
叔同 rejects the linear conclusion that AI will replace programmers; the real divide will be between those who can command Agents and create business value and those who cannot. Jevons’ paradox may cause demand for low-cost software to surge, and the total number of programmers could even rise, but human capability must move upward into demand discovery, architecture and acceptance of results. He sees organizational forms potentially changing within a three-year window; further out, Coding could connect foundation models, robots and the physical world, reorganizing everything from real software to building houses on Mars.
Deep dive
1. AI Coding Is Converging on a Single Production Chain Through Three Entry Points
The host put the market heat into concrete terms: Cursor generates roughly $500M in annual revenue at a valuation of about $10B, while Anthropic’s revenue curve suddenly steepened after the release of Claude Code, at one point even surpassing OpenAI—though its base may also be smaller.
叔同 defines the first category as going from 0 to 1: tools such as Lovable and Bolt.new let generalist developers generate websites or services for a few dollars. The results are dazzling but may have very short lifecycles; once those creations need to be bought and used by others, a platform market emerges naturally.
The second category—Cursor, Claude Code and Qoder—serves roughly 30M professional developers, handling the long-term maintenance journey from 1 to 10 and then to 100. The third, represented by Devin, directly delivers “digital employees” that can take on complete assignments and work for extended periods.
2. Product Formats May Converge, but Model Capability Will Rewrite Profit Distribution
叔同 believes the different paths are “very likely to converge into one”: future products will offer stronger capabilities and more varied interfaces, but the core will remain delivering software services through Coding, satisfying business scenarios and creating commercial value.
Application-layer companies without models face not only high inference costs but also an inability to jointly optimize performance, efficiency and cost end to end. 叔同 said bluntly that these products are easily dismissed as “wrappers,” or even merely “a service that helps the model generate traffic.”
Asked whether Cursor will eventually develop its own model, the host received a deliberately unequivocal answer: “They will definitely provide—or, let’s call it, build—their own model.” Model ownership affects product outcomes, company profits and valuation at the same time.
3. Qoder Chose “Real Software”—Less Flashy, but More Valuable
叔同’s key statistic is that more than 95% of professional developers maintain real software. Real software has already created commercial value and comes with accountability to users and customers.
Once a system may have been running for five or ten years, changing it becomes serious business: AI must understand the accumulated history, cannot “change things at random,” and cannot rely on imagination and hallucinations to produce output. Qoder therefore calls this the “high ground of value” in the information industry.
The host’s follow-up is worth preserving: why had so few people pursued such an obvious need first? 叔同 acknowledged both the difficulty and the fact that “generating a beautiful website from one sentence” delivers more creative gratification, so the first wave naturally picked that low-hanging fruit.
The counterintuitive choice for a late entrant is to skip the “first fruit” and attack high-value maintenance scenarios directly, while covering assistance, collaboration and autonomy simultaneously rather than using the next generation to dismiss the completion and Q&A tools developers still use heavily today.
4. A Single Global Version Must Accommodate Developers at Different Stages of Maturity
Qoder has targeted the global market from day one with a single version. 叔同 observes that Chinese developers still mainly use completion, Ask and “the model answers, I edit,” while overseas users are moving faster toward “I’m barely doing it myself anymore; I let the model rewrite everything.”
Broad coverage therefore does not mean averaging everything down. It means allowing completion, Q&A and autonomous Agent programming to coexist, then supporting different habits through the IDE, CLI, plugins and eventually image and voice input. When users are ready, they can switch directly to a more advanced mode of production.
5. Spec-Driven Development Is the Prerequisite for Long-Running Agents, Not a Return to Documentation for Its Own Sake
叔同 describes the evolution as a shift from prompt engineering to context engineering: to give an Agent a larger assignment, the team must first clarify the requirements, business logic, technical architecture, design specifications, module breakdown, implementation requirements and acceptance criteria.
This is not a process invented from thin air. It maps human-to-human development collaboration onto human-to-model collaboration: a boss states the requirement, the team researches it, writes the PRD, defines upstream and downstream systems and data interfaces, develops, accepts and launches the product, then iterates based on market feedback.
Qoder does not ask users to fall in love with documentation again. The user starts with one sentence—“everyone is the boss”—and a hidden spec Agent expands it into a standard design document. Only after the human fine-tunes and confirms it does the execution Agent begin working asynchronously.
叔同’s conclusion is clear: spec-driven development is the “inevitable choice” for making an Agent work for an extended period. The value of the document is not its form, but its ability to turn a large assignment into an executable, verifiable contract.
6. Rules and Memory Are Becoming the Operating System of AI Teams
The host described persistent requirement documents and mood boards as “the constitution of a website.” 叔同 agreed with the metaphor, and Qoder lets users write into rules “what must not be done and how things should be done,” establishing a team-style set of ground rules first.
The Agent can also infer patterns from what users accept, reject and prefer, then write them into memory. But 叔同 acknowledges that such inferences may overreach, so users must be able to edit and change them: “I just happened to have this situation today.”
The new mode of collaboration also makes people conscious of the cost of compressing context, switching models and restarting sessions. The product has to balance tool convenience and productivity against frequent interruptions, so developers continue to feel that they are steering the Agent.
7. In the AI Era, Flow Means Not Having to Understand the Product Itself—not Never Being Interrupted
JetBrains and VS Code were built around letting developers work continuously inside a single platform. But once one person begins pair-programming with, or even managing, a team of Agents, the questions, rework and new requirements make traditional flow difficult to sustain.
叔同 therefore changes the test: interruption is not necessarily the most important problem; what is truly unacceptable is the cost of understanding. “If you can scan something once and still cannot understand it,” then the communication, behavior or interaction design has failed.
He cites Don’t Make Me Think, which he read 15 years ago: the working method can change, but the interface must remain “especially natural.” A sense of control, rather than absolute continuity, becomes Qoder’s key measure of Agent interaction.
8. Not Offering a Model Menu Is Qoder’s Most Contrarian—and Most Testable—Design Choice
Qoder’s highest principle is to integrate the best models globally, including overseas models and Qwen. But the platform decides which model handles each scenario, and users do not see a Cursor-style list of roughly 40 models.
叔同 believes a menu forces developers to memorize prices, strengths and rankings, then retest whenever models are upgraded. In the end, they have “become a model selector and trainer—or, let’s call it, a tester.”
The host offered a sharp counterexample: if “Claude 5.0” suddenly appears and is widely recognized as a tenfold productivity improvement, users who cannot select it directly may become frustrated and churn. 叔同’s answer was to take responsibility for the outcome on the same task: “Our result will definitely be no worse than theirs, and may even be better.”
He calls the constraint the “impossible trinity” of performance, efficiency and cost. Qoder’s two judgments are that “machine selection should outperform human selection” and “platform knowledge should outperform individual knowledge.” The system can reroute every question; humans will not manually switch models for every question.
9. Public Evaluations Will Turn Product Philosophy into Comparable Metrics
Qoder plans to open a public benchmark focused on real software, allowing products to be compared on durability, resource consumption, time to completion and satisfaction with the result. 叔同 acknowledges that user experience is also part of evaluation, but public data will be more objective.
For Alibaba, Coding is a necessary path for large language models toward AGI and can also improve their Coding capabilities end to end. Qoder’s ambition is “top three in the world,” but 叔同 refuses to rank a product only two months after launch: “This is a marathon. Everyone has really only run the first kilometer.”
Cursor remains the most common comparison today, and 叔同 admits Qoder is still catching up. His confidence comes from jointly optimizing models, context, tools and product interfaces—not from presenting “top three” as an already-achieved market share.
10. Tongyi Lingma Was Not Replaced; It Made Room for a New Global Hand
Qoder was conceived, developed and launched in August 2025. Tongyi Lingma had already been operating in China for more than two years and, according to 叔同, held the No. 1 plugin market share in China. The new project therefore did not start from zero; it chose a different global product path.
The real internal debate concerned the fundamentals: which platform to use for globalization; whether to choose a plugin, IDE, CLI or standalone Agent; and whether to use only Alibaba’s own model or integrate models globally. The final answer was for Lingma to serve China and Qoder to serve the world through a new product format.
叔同 puts it practically: if Alibaba’s model becomes the world’s best within a year, Qoder’s model mix will naturally change. But “we cannot wait until our model becomes No. 1 in the world before serving developers; by then there will be no opportunity.” Winning users first also buys time for the model.
11. Qoder Was Designed as an Alibaba Internal Startup, Not Another Product with a Name Attached
The team was not assembled from scratch; core members already had experience with Lingma and other products. 叔同 describes taking over the project as a “two-way journey,” and says that extending Lingma into a global product was something “we believe no one is better suited to do than us.”
“Qoder” is explained as a homophone of coder, deliberately tied to neither a specific model nor a platform. 叔同’s brand image is: “It is the programmer in your pocket, the programmer in your home and on your team.”
Alibaba’s key support was not a slogan but resources, freedom from legacy targets, team efficiency and openness. 叔同 defines it as “internal innovation, and also internal entrepreneurship,” with the central requirement of shedding some baggage and matching the speed of the market and competitors.
12. The Startup Window Is Narrowing, but the Market Is Far from Zero-Sum
叔同 believes “there will be fewer and fewer opportunities for startups.” Even if AI Coding is understood as model packaging, it requires substantial teams and costs, and application-layer companies do not have the advantage.
Claude Code and GPT-5 Codex have the structural advantage of using their own models and clusters, where the resources are almost sunk costs for the model companies. Selling the same capability to Cursor, by contrast, means “not losing a single penny,” enough to lower costs while preserving performance and efficiency.
Early companies such as Cursor and Lovable still have room because they defined the category and product format, then defended their first-mover advantage through iteration speed. New entrants building the same type of product must face Big Tech, foundation-model companies and established startups simultaneously.
Asked which domestic company was doing well, 叔同 first answered, “We’re doing very well,” and then did not name a second company. But he stressed that global professional developers are at most one-third covered, leaving two-thirds untouched. For now, the market looks more like different philosophies unlocking different users than a price-performance bloodbath.
13. Qoder Refuses to Manufacture Retention Through Asset Lock-In
The host’s concern was direct: with the codebase remaining where it is, a user can use Qoder today and switch to Cursor tomorrow without much difficulty, especially if the latter gains access to a new model. Retention could be fragile.
The team discussed whether Repo Wiki should be viewable only inside Qoder and ultimately concluded that this was “impossible”: the asset must belong to the user and be shareable and exportable. 叔同’s straightforward answer is that “differentiated capability and openness” win recognition together, rather than by locking users in.
14. Repo Wiki Starts with the Codebase, Not Old Documentation, to Understand the System
叔同’s starting point is that “the most real-time, freshest thing” in accumulated work is the code. When a boss changes a feature with one sentence, the team usually does not update the documentation, but the code definitely changes. Over time, the original documentation becomes seriously distorted.
Repo Wiki does not merely read slices of the current code. It analyzes commit records, historical versions and intermediate changes to reconstruct system design, the iteration process, sequencing relationships, business architecture and logic.
Every line of documentation is generated by a large language model, adding no writing burden for programmers. As the codebase continues to receive commits, it can refresh incrementally so the documentation changes with the real implementation.
The document is placed inside real collaboration: handing off modules, employee departures, onboarding and learning systems can all start from Repo Wiki. A ten-person team does not need 10 separate copies; once connected to the Git repository, the document is shared.
15. Repo Wiki’s Moat Is the Complete Workflow, Not a One-Off Summary Button
叔同 acknowledges that “implementing the feature” itself may not be especially difficult. The challenge is understanding history quickly, accurately and efficiently. Qoder therefore customizes models, dedicated Agents, maintenance workflows and usage workflows, then deeply integrates the result into developer collaboration.
Generating a codebase initially took about 60 minutes. 叔同 says the latest version has cut the time by 5x; because the token volume remains large, the multiple-fold efficiency gains from a customized model directly determine whether the feature can be used day to day.
Quality is evaluated first against the previous internal version, then against customized models and leading overseas models. Some of the assessment is manual; some borrows from reward models. Adoption, reading volume and user annotations are used to judge whether the product has genuinely improved.
16. Quest Mode Tries to Break Through the Boundaries of Human Hours and Human Computers
The ceiling for autonomous programming is not just model capability. It is also “human working hours and human computers”: once a laptop is closed or the workday ends, a local Agent stops as well.
Quest Mode sends the spec and codebase to a cloud sandbox, allowing the Agent to continue working while the user is on vacation. One person can launch 10 asynchronous sandboxes at once, which 叔同 describes as “one person managing 10 Agents” and tenfold productivity.
Compared with remotely launching an entire IDE, Qoder’s cloud environment retains only the runtime environment and the programming Agent. It does not migrate a large collection of plugins and tools, making it lighter, faster to start and better suited to highly concurrent tasks.
The design makes life and execution asynchronous: humans arrange, inspect and accept the work, while machines use the time when humans are offline to keep moving, rather than requiring developers to stand watch over the workstation for the Agent.
17. Long Tasks Still Require Supervision Because Agents Can Drift—and “Get Clever”
Qoder creates a to-do list at the start of a task. The Agent must execute every item, cannot skip steps and cannot claim completion without finishing. 叔同 sees this constraint as the baseline for trustworthy delivery by a digital employee.
Users can ask at any time what the Agent is doing, inspect the continuously updated workflow and add instructions midstream. The reason is practical: if a task runs for three or four hours and has already drifted halfway through, continuing to spend compute only compounds the waste.
叔同 acknowledges that Agents may drift, skip tasks or “play some clever little tricks.” The future is therefore not one in which humans exit the process, but one in which humans move up into the role of leader, managing an AI team through systems, questions and acceptance criteria.
18. Coding Will Become the Hands of Foundation Models; Humans Will Organize the Agents
叔同 defines Coding as the “hands and actuator” of large language models, as well as the connector between the digital and physical worlds. It translates natural-language objectives into actions that can call tools and change systems.
Humans remain the managers in this relationship: they set clear requirements and mobilize different Agents to understand and collaborate. The people with an edge in the future will be those who can find business scenarios, manage AI and convert the results into commercial value.
He sees AI-native companies as one possible organizational form: many creative organizations, with much of the work completed through collaboration among AIs. This “could” become the mainstream future, not a fixed timetable.
19. The IDE Is the Largest Common Denominator; the CLI Frees Agents from a Single Workbench
Qoder started with the IDE because it remains the core interface for the broadest group of professional developers. But building on VS Code also naturally excludes habits such as Vim, so an IDE alone cannot fulfill the promise of serving every developer.
The CLI externalizes the Agent previously packaged inside the IDE into command-line, scripting and integration capabilities. It can enter maintenance, pipelines, internal platforms and automated tasks, theoretically covering every developer.
Competition with Claude Code is not one-dimensional. 叔同 says its context construction is “fairly simple and brute-force, but the results are very good.” Qoder may pay more when using external models, so it must be more refined in context assembly, Agent flexibility and scenario design.
Application-layer expansion also includes mobile, Slack, IM and customized platforms, allowing users to schedule and follow up on tasks remotely. 叔同 believes model companies will not pursue these product-form innovations, which is precisely where Qoder can differentiate.
20. Context Engineering Is About Doing “More, Faster, Better and Cheaper” Within Limited Tokens
Stuffing the entire codebase into every request is slow and expensive, so Qoder first determines relevance. Vector retrieval, text retrieval and semantic retrieval each fit different scenarios; no single method covers every task.
Engineering choices also include whether to use an Agent or subagent, which tools to design, how to call them in loops and how to aggregate information. 叔同 emphasizes that “context is absolutely not better when it is longer.” The objective is to assemble enough context, and the right context.
Feedback such as “this isn’t what I wanted” and “this is what I like” is abstracted into rules. Some rules enter the system prompt, while others enter memory. This prevents context from growing too quickly while creating the sense that “the more I use this tool, the more it understands me.”
叔同 confirms that the underlying infrastructure—including memory, sandboxes and to-do lists—is built in-house. In his framing, these are the capabilities that determine whether the performance-efficiency-cost triangle can be pushed open.
21. AI Will Not Simply Eliminate Programmers, but It Will Rapidly Shift the Capability Boundary
叔同 explicitly rejects the idea that “programmers will no longer be needed.” His logic comes from Jevons’ paradox: lower software production costs will unlock massive demand, with generalist developers handling small innovations and professional programmers taking on larger, more complex ones. The number of programmers could even rise.
Individuals will nevertheless diverge sharply. Those who use AI to produce fivefold or tenfold productivity gains will be more competitive; those who cannot may “become an ordinary programmer.” The Qoder team is already “using Qoder to build Qoder” and explicitly tracks how much of its code generation is completed by AI.
Boundaries among frontend, backend, iOS, DBA and other roles are being flattened by models. Once models write code faster than humans, people must move upward into demand discovery, intent recognition, requirement definition, overall design, product sense and acceptance of results.
22. Students Must Embrace AI and Fundamentals Together, Because Acceptance Capability Determines Whether They Can Steer It
叔同’s first piece of advice to freshmen is to use AI extensively and actively explore its limits. His second is to study computer science fundamentals seriously, because the von Neumann architecture, operating systems and underlying structures have not undergone a revolutionary change.
Without understanding computer architecture and operating systems such as Linux, Windows and Mac, one cannot judge whether the model’s delivery is good. “It may bluff you.” Expertise is no longer only for writing code by hand; it is also needed to review, correct and select an architecture.
He uses a sharp observation to explain where scarcity will lie: “In the future, a simple piece of software may be worth just 50 cents.” AI Coding products can build an iOS app but cannot build iOS itself; large language models have not yet created a large language model. Work that is sufficiently complex, rare and dependent on top-level design will still carry the highest value.
For existing engineers, 叔同 is more optimistic. This group has spent the past 20 or 30 years learning rapidly and “doesn’t care at all about learning one more thing.” The bottleneck of projects being repeatedly delayed for lack of development capacity may also be significantly weakened by AI.
23. Double 11 Shaped 叔同’s Core Method: Always Extract the Dividend from Technology
When 叔同 joined Alibaba, Double 11 traffic could reach roughly 100x normal daily traffic. Each team acted like “railway police, each responsible for one section”: spare capacity was enough in ordinary periods, but hundreds or thousands of systems assembled at peak load could easily collapse from a local bottleneck.
In 2013, he and his team pioneered full-link stress testing. 叔同 calls them the first team in the world to propose the concept: keep hundreds or thousands of systems at the same capacity level, block traffic beyond processing capability and ensure that every request admitted can complete. He says it later became standard practice for internet services.
The obstacles were not purely technical. Double 11 could not be postponed, and everyone shared the goal of “zero downtime.” When a transformation could not be pushed through, the core team personally backstopped it. Alibaba had unified its technology stack relatively early, allowing middleware to be changed “across a huge swath at once.” The entire effort took about three months.
The team rehearsed repeatedly using real user journeys. At one point it simulated 500,000 transactions entering simultaneously within one second at midnight, exposing capacity, concurrency, elasticity and mismatch problems one by one. Midnight was still tense; there were no major issues after 2013, but “small issues kept coming one after another.”
24. From Zero Failures to Zero Incremental Cost, Qoder Extends the Same Value System
Once stability was solved, the next target was cost: could the group handle roughly 50x traffic without “spending one extra cent”? Containerization, cloud-native infrastructure, hybrid deployment and cross-business time-sharing allowed resources to scale, lend and return within 10 to 20 minutes.
Once high-traffic capability matured and the demographic dividend peaked, 叔同 turned toward technology standardization, open source and commercialization, serving internal capabilities to internet and industrial customers. He summarizes the path as “technology drives business, then creates value.”
On career choices, he advises: “Don’t pay for things that happened in the past; pay for what is about to happen in the future.” He does not see Qoder as overthrowing his own past, but as exploring the next path for programmers: “Stand on the shoulders of technology, rather than become its slave.”
The satisfying picture three years out is for Qoder to produce the most real software with commercial value. Within the three-year window, organizational forms may also change, with multiple Agents taking on complex, long-term objectives. Large language models’ answers, generation and reasoning are inherently uncertain, while “code can deliver certainty.” Further out, he uses Mars as the metaphor: houses on Mars will not be built by astronauts by hand, but jointly by foundation models, Agents, Coding and robots.