“So our philosophy here is to integrate the globally optimal models and give users the best results.” — Hang Yu, Head of Product at Qoder, Alibaba

This is the first episode in a series of founder and builder dispatches, featuring interviews with the people creating the future. If you are a founder, builder, or investor in this space and would like to share your story, please reach out.

Today, I am joined by two guests from the Qoder team at Alibaba: Hang Yu, Head of Product, and Christian Hu, Head of Global Marketing and Operations. The Qoder team launched just over two months ago, joining the likes of Cursor, Warp, and Copilot to make coding more agentic, so today we get to learn from them directly about their unique positioning being part of the Alibaba ecosystem.

Hang discusses the thinking behind designing Qoder, how it differentiates itself from peers currently available on the market, the future of agentic work, his fears and excitement about the pursuit of AGI, and finally, challenges the notion that the future of AI may not be based on Transformers.

Christian walks us through Qoder’s business positioning, global ambitions, how it fits into the Alibaba ecosystem, and the reasons for routing between models, beyond just Qwen.

In today’s world, there’s no shortage of information. Knowledge is abundant, perspectives are everywhere. But true insight doesn’t come from access alone—it comes from differentiated understanding. It’s the ability to piece together scattered signals, cut through the noise and clutter, and form a clear, original perspective on a situation, a trend, a business, or a person. That’s what makes understanding powerful.

Every episode, I bring in a guest with a unique point of view on a critical matter, phenomenon, or business trend—someone who can help us see things differently.

For more information on the podcast series, see here.

Topics we covered:

Product

* Introduction to Qoder and AI Coding Agents

* The Transition from Copilot to Agentic AI

* The Future of Developer Productivity with AI

* Addressing Developer Bottlenecks

* Multi-Model Strategy and (Qwen) Integration

* Differentiated Views on AGI and AI’s Future

Business

* Understanding Qoder’s Positioning in the Market

* The Competitive Landscape of Coding Tools

* Qoder’s Role in Alibaba’s AI Strategy

* International Ambitions and Challenges

Transcript (AI-generated)

A. Hang Yu, Head of Product at Qoder

Grace Shao

And again, I just want to say thank you so much for joining me today, Hang and Christian. So today, the first half of our conversation will really focus on the product design of Qoder and the transition that we’re seeing from Copilot to Agentic. And then we’ll move into the second half of the conversation, which will really focus on the business strategy of Qoder, international expansion goals, objectives, and then how it really fits into the bigger Alibaba AI plan and the bigger AI playbook. So with that, I just want to bring in Hang. Hang, it’s lovely to meet you and thank you so much for joining us today. Let’s start with the very, very basics. What is Qoder in plain language and what does it actually do for developers day to day?

Hang

Hey, great. Thanks for having me. So yeah, so in one sentence, Qoder is an AI coding assistant that helps developers maintain and improve existing software system, not just build some new stuff from scratch. And I think that distinction is actually really important. So when we look at what developers actually do every day, we found that 95 % of professional developers spend their time maintaining what we call real software.

So commercially valuable, long-lived systems that are not building new projects every day from zero. So that’s why we design code around that reality. We are optimizing for the messy, important work of understanding existing code, making targeted improvements, and maintaining production systems.

Grace Shao

That’s really interesting. you know, there’s for me again, not a technical person. The hype we hear about is all this vibe coding. And then there’s even these marketing phrases calling it like ⁓ AI coding, build your app from a single prompt. That’s not what’s really happening, right? How does it actually really work?

Hang

Yeah, exactly. actually, that’s the flashy use case everyone talks about, like building one app from a single prompt. But if you’re working at a real company maintaining a five-year-old code base with 500,000 lines of code and 20 different developers who have touched it over time, you need help understanding what’s actually here. So you need help making careful modifications without breaking things.

That’s where code delivers value. So in terms of how we think about the AI developer relationship, we see it evolving through three stages. So the first is assistive programming. In this stage, AI helps the developer while human leads. So like code completion, fixing syntax errors. And the second stage is collaborative programming, like co-pilot. Like the human and AI works together like pair programming.

And the final stage is autonomous programming. In this stage, AI takes on complete tasks independently. So the developer can delegate work, and the AI runs in the background and comes back with results. That’s what our Qoder Quest mode does.

Grace Shao

So the really unique bit of Qoder Quest is the autonomous piece, right? And I really want to dig into that a bit more, a bit later. So right now I’m curious. So when you say the user experience is intuitive even for non-technical users, what design choices really led to that? Because honestly, a lot of developer tools are pretty intimidating, especially for people like myself who’ve never done any coding. But what I’m hearing on the street is people are going out of their way, even as non-technical people, building their own apps now with the help of AI. How are they able to do that?

Hang

Yeah, great question. So we have this philosophy. So don’t make users think about things they shouldn’t have to think about. So for example, if you look at some products, they have like 40 different AI models in a dropdown menu. Honestly, that creates a lot of cognitive load. So developers end up becoming model select or instead of focusing on building their own product. So our philosophy here is integrate the globally optimal models and give users the best results. So we will auto select the right model based on the task. So we believe model selection will be better than human selection. And the same thing with context management. The users shouldn’t have to manually figure out, OK, which files to include, what tokens to optimize. Our context engineering handles that automatically. So the goal here is to remove the cognitive overhead. And let developers focus on what they are trying to build, not on configuring the AI tool.

Grace Shao

I have a really dumb question, but is there a latency then between me prompting like, can you help me build this versus the machine telling me which model is optimized? Is there like a latency? It’s automatic.

Hang

No, no, there would be no latency. Yeah, it’s all automatically. The user will not feel it.

Grace Shao

that’s amazing. Okay, let’s talk about kind of the hype right now, the copilot versus the agentic transition right now in the coding space. Many tools are being called assistants or some are called copilots, You know, the cursors of the world. And cursors essentially been leading this. So where does coders autonomous capability really truly defer or is unique or different? And what’s really the big breakthrough we’ve seen here?

Hang

Yeah, this is the defining question, So Cursor has ⁓ perfected the co-pilot approach, real-time help while you’re coding. Their tab completion, tab, tab, tab, is industry-leading after two years of their custom model training. But I’ll say this. So Cursor’s tab completion compatibility is catching up fast. We have made significant progress in recent months and rapidly closing the gap.

But here’s where we see the real future, the autonomous programming. So you delegate a complete task, implement this feature, fixing this bug, et cetera, whatever. And then the AI works on it in the background. You don’t have to watch every line of code being written. You just come back and review the result. now sophisticated autonomous coding and production skill is still involved.

Hang

We think it’s about two to three years out, maybe 2027 or 2028, before it’s really mature. But that creates an opportunity window.

Grace Shao

That’s actually quite soon. So what does that really mean practically?

Hang

Yeah,so it means With autonomous coding you can actually Delicate your work and make and the the AI agentic will can work it background in the cloud so Like you can close your laptop and the AI keeps working

go to a meeting, go home, do whatever you want, the coder is still running. So in other words, you can spin up 10 parallel sessions working on different tasks, and it doesn’t slow down your own machine. As one colleague said, he said, I’m managing 10 agents. My productivity went up 10 times, and it didn’t mess up my work-life balance. So yeah.

Hang

So this cloud execution model is pretty similar to what cursor recently launched with their cloud agents feature. So both approaches let you handle your tasks to agent running remotely. The key advantage here is you are not tied to your machine. You can dedicate the work and then close your laptop and then come back to complete the result. And then when the

Hang

Yeah, and then when the agent finishes, we just need to review the results.

Grace Shao

So that’s actually my question. Like when you talk about reviewing the results, is it very obvious to kind of find the issues only in the result or do you have to go back to the process? Like how do you audit the whole process actually?

Hang

Yeah, so the agents will present to you summary of, this is what I’ve done. This is the result of the question, or this is the feature. You can see it through the browser, or the agents will submit a PR to your GitHub to your routing workflow. And you can easily check, OK, is it done good, or it needs to be modified again.

Grace Shao

So does that mean that junior developers in a way will be replaced then? Because essentially you only really need people who understand a higher level of the code and the execution can be actually outsourced so easily.

Hang

Hmm. Good question. So simple, agentic task works well today. You can already dedicate things like write a unit task for this function or add log into this module and come back get some good results.

So sophisticated, agentic coding at production scale, where AI can take a high level of business requirements and then autonomously design, implement, test, and deploy a complex feature across multiple systems, that’s still two to three years far away. So what this means for software development is a fundamental shift in how developers spend their time. ⁓

Hang

So right now, developers spend maybe, I think, 30 % of the time on creative problem solving and 30 % of the time on mechanical work, like writing border plates, fixing syntax errors, ⁓ updating documentation, writing tests, et cetera. But as agentic tools mature, that flips. So AI handles the mechanical work. Developers ⁓ can focus on what AI can do yet, like understanding what actually needs to be built.

Hang

making architectural decisions that require domain knowledge and validating that the code actually does what it should be.

Grace Shao

Mm-hmm. So in the end, it’s still like, there are still aspects of humanity that can’t be replaced by machines yet. Let’s get a bit more practical. I want to understand how precisely Qoder can actually turn a prompt or product spec into working code? What types of tasks or repos does it handle best today, right now?

Hang

So Qoder handles best what we call a bounded, well-defined task today. For example, implement authentication for our user login system or add relimits to our API endpoints or generate a dashboard for monitoring system health metrics. So these are tasks where requirements can be clearly specified, the scope is contained, and the success criteria are measurable.

Grace Shao

Mm-hmm.

Hang

But what’s harder today are tasks that require deep domain knowledge or ambiguous requirements like improved user engagement, that’s too vague, or like a required entire authentication system, that’s too large, that’s too large and risky for autonomous execution.

Grace Shao

So it seems like a lot of the tasks it can do is still pretty much something that’s very easily verifiable. It’s a bit more like a black and white kind of answer kind of task, but not so much things with nuance, right? I want to understand better. So, thinking about the developer workflow, and correct me if I’m wrong, there’s the planning stage, the code writing stage, the running tests, the debugging, version control, and deployment, right? Where does Qoder natively sit the strongest, like in the most helpful, and where do you hand off to other tools?

Hang

Great question. Let me break it down. So the first is plan and design stage, So Qoder is strong here through the spec generation. We can help the developer translate business requirements into technical specs. And the second stage is writing code. This is our core strings. The Qoder can write code across multiple files, handle complex logic, and generate boilerplate. And for the run and test stage,

Qoder can generate unit tests, integration tests, and run them either locally or in cloud sandbox. That’s building feature. And for debugging stage, Qoder can diagnose its errors from test results and fix them autonomously. But for production debugging with live user data, you still need the traditional tools. That needs to be taken care of. And as for version control or call out, we integrate with Git, GitHub, and GitLab. So a coder can create branches, commit changes, and create pull requests. But the actual code review and collaboration discussion happens in your existing tools, like GitHub, GitLab, whatever you use. But yeah, in the final stage, the deployment, we hand off here. So deployment involves your CI-CD pipelines, like infrastructure, monitoring system. So Qoder creates the code and the tests, but you own your deployment process.

Grace Shao

I see. I kind of want to take a step back and understand another big kind of hovering question a lot of people have right now, which is if we’re going to have agentic tools really implemented the work process, how do we understand developer productivity for the future? Will developer productivity still be measured the same way that it’s currently being measured?

Hang

Yeah, I think the fundamental shift here is this. So AI changes what developers spend their time on, not just how fast they work. So again, right now, developers spend maybe 30 % of the time on mechanical work, like writing boilerplate, debugging syntax errors, searching documentation, setting up environments. And only 30 % of the time goes to creative high-value stuff, like understanding what needs to be built. Like understanding what needs to be built, making architectural decisions, and validating that solution actually solves the problem. But with AI, agent AI flips that ratio. So AI handles the mechanical work. Developers focus on the part that actually requires human judgment. So when we measure productivity, we are not counting lines of code or tickets closed.

So we are looking at, can developers spend more time on higher value work? Can they ⁓ ship features faster without burning out? So what we are seeing is promising. 99 % of our paid users actively use agent model. Nearly 99 % of our paid users actively use agent mode. So it’s become core to their workflow.

80% renewal rate tells us that people see the real value here. And enterprise reports two to three times improvements in the deployment frequency. So the real metric here is simpler. Developers tell us that they are less frustrated. They are not stuck debugging environments or writing repetitive code. They are solving interesting problems. And that’s the productivity gain that matters.

Grace Shao

I see, I see. That’s really interesting because I think as a writer, when people really initially rejected AI, the idea was also that there was a lot of mistakes, AI slop, hallucination, whatnot. But then what people started realizing is that you can actually use it as a productivity tool. And like to your point, it doesn’t change how you think as a writer building out the framework, using your critical thinking and really still rely on your own creativity. But you know, you what you’re outsourcing is actually just like the execution that was a lot of the grunt work frankly. Right. Interesting, interesting.

Hang

Yeah, yeah, exactly. Yeah, you free up your hands and you, yeah, yeah, you free up your hands and focus on your head.

Grace Shao

Okay, so I wanted to ask another question on productivity and bottlenecks. So what is, I guess, one of the top bottlenecks or what are a few that you see currently in this space for developer right now? And I guess can we actually separate the separately answer this question, the first half being what are professional developers bottlenecks and what are kind of the new age, like casual developers bottlenecks and how are you helping these two differently?

Hang

Good one. Let me hit the main ones. So usually the bottlenecks, like the first big bottleneck is about the environment setup, So this is a huge time thing. The coders, and then the coder can help you set up the environment automatically, whether you are running it locally or in the cloud, pre-configured environments, automatic dependency resolution.

So you don’t have to spend two hours debugging your Python version complex or just trying to build your dev environments. And then for the flaky tests, coder can also detect flaky tests by running them multiple times and identifying inconsistency. So you can also suggest fix based on the failure patterns and the test outcomes. And then for...

For another bottleneck here is the legacy code. This is where the Ripple Wiki comes in. So remember how I said the documentation is always out of date for developers? Ripple Wiki uses... So Ripple Wiki delivers value here. So Ripple Wiki features here can use AI to generate up-to-date documentation from the code itself, plug the Git history. So it’s not just...

Here’s what the function does, documentation. But here’s a business logic and why it was architecture this way and what changed and why documentation. So the documentation itself stays fresh. When code changes, the documentation regenerates automatically. So we have measured about five times speed improvements, so from 60 minutes down to 12 minutes for team documentation.

And the last but not least bottleneck here is the context limit. So this is a technical challenge. Models have tokens limits, So we have our context engineering figures out what’s actually relevant to the task. So we don’t just dump the entire code base, which will crush your token limits. We intelligently select what the task needs and what the AI needs. and we ⁓ gain a better solution from the ⁓ context we select.

Grace Shao

That’s really interesting. So I think I want to talk about the models and orchestration here. So Qoder is a multi-model platform, right? You guys use Qwen, your in-house built models, but you also use other models like you mentioned earlier, use whatever is like kind of state of art model for the right task, right? How do you route between the models, thinking about latency, cost, evaluations, languages? Do you usually like have a preference for Qwen or the third party models when you’re assigning these models to tasks?

Hang

OK, so what we are trying to do here is integrate the globally optimal models. ⁓ So what we are trying to do here is to integrate the globally optimal models and give users the best results. We are not limited to just Alibaba’s model. So if one frontier model is better for a specific task, we use that. So like if GPT Excel somewher, we use GPT. If Qwen is the right fit, we use Qwen. So now, why does this matter? So first, it best has time. I’ll be direct. So Qwen model isn’t the best model in the world for every coding task yet, but it’s improving fast. So by using the best global models today, can serve users well while our own models can catch up. And if Qwen becomes the best model globally in your year, is our goal, then naturally we’ll use it more. And secondly, yes, sir. Yeah.

Grace Shao

But actually, just want to jump in on that. But actually, Curser is even using Qwen isn’t that fascinating that we just found out recently.

Hang

Yeah, so the cursor’s composter is... They didn’t officially admit it, it’s... So the community thinks they’re fine-tuned and post-trained based on Qwen or some Chinese open source model. Yeah. So first is the best time. And second, the cost optimization without compromising quality is our second goal. So not every task needs the most expensive frontier model. So simple completion is a smaller, faster model. But for a complex reasoning, we use the frontier model. And last but not least is about the reliability. So you don’t want your entire product to stop working because one API provider has an outage. Here we saw OpenAI has multi-day outage in 2023. we want to our reliability to our customer. Yeah, that’s why we use multi-modal strategy. Yeah.

Grace Shao

I see, I see. I was actually going to ask you on that. What’s the thinking behind designing the product using a multi-model design versus only proprietary Qwen? I guess you already answered that partially. ⁓ I also wanted to kind of double click on the fact that Alibaba is throwing 53 billion US dollars into AI infrastructure right now. know, isn’t it, is there not some pressure coming from up top to really hone in on Qwen or, know, like, I guess the question’s more about - Is Qoder really developing based on what’s best for the user and use whichever model that’s best for them? Or is it more focused on being part of an ecosystem, another tool out of the Alibaba set of tools that they provide? Does that make sense?

Hang

Yeah, great question. So it comes down to serving users best today while building towards to the future. look, Qwen models costs us like 1 fifth or 1 sixth of what Frontier model API costs. So that’s an 80 % cost reduction, right? That’s a structural advantage. But Qwen isn’t as good as the top Frontier models on context reasoning tasks.

That’s why we use a multi-modal setup at the balance cost, capability, and reliability. But the key in size is that this isn’t a permanent state. Qwen improves, as coding capability, sorry, as Qwen.

So the key inside is that this isn’t a permanent state. As Qwen improves, as its coding capabilities get stronger, the balance will shift. So we are not philosophically opposed to vertical integration. And we are actually pragmatically choosing what serves users best today while building towards our ownership tomorrow. So for enterprise customers in China, components actually require a domestic model anyway. But for international customers, they often trust well-known frontier models or GPT models. So multi-models board let us serve both.

Grace Shao

I think that’s a really interesting business strategy because end of the day, to your point, it’s a very pragmatic approach to actually serve your customers best and to serve your customerbest, you actually get more business. That’s just how it actually the virtuous cycle works, right? Instead of building up these guardrails and the paywalls. I want to kind of pivot to strategy soon, but before we do that, I really want to ask you a few other questions just on more, a more broad general question. One is, you’ve been working in this space for a long time. There’s a lot of hype around Agentic tools, not just a Agentic tools in coding, right? How do you view a Agentic AI in the next 12 to 18 months? And how do you think it will actually affect what even the general mass think of AI use AI?

Hang

I think agents will swallow the whole market. ⁓ People will use more more agents in their daily workflow. So when you’re trying to use a chatbot, you need to copy paste a lot of data, your contacts, your own data, your domain knowledge.

To chatbot and then generates some specific task results. But by using agent, agentic AI, you don’t need to do that. The agentic AI will live in your workflow. It knows and it contains your domain knowledge. And it knows your context by nature. So this means agentic AI knows ⁓ what you’re trying to do, what you did.

What you are trying to do and what you want. That’s actually a big difference between the big difference between the LLM and the agent AI. So I think, yeah, yeah.

Grace Shao

I see. I think that that leads me to the next question, is like, so for teams who are customizing tools or agents, like you just said, they would have domain knowledge. They already know your work processing really well. And how does it actually work? What’s the plug-in API story with Qoder in this sense again?

Hang

Yeah, so this is about the domain knowledge question and the accessibility work. Let me break down how the domain knowledge and the accessibility works. So first, we have MCP supports. We support the modal context protocol, which is becoming the industrial standard for connecting AI agents to external tools and data source. So this means the coder can ⁓

integrated with thousands of tools in the ecosystem, database, API, version control, and project management through a standardized interface. And then we also support subagents and skills. So teams can create their own specialized subagents. Think of this as task-specific AI workers with their own contacts and capabilities. So ⁓ you want a code review subagents turned into a

So you want a code review, sub-agents turn to your team standard. You can define it. You need a security scanning sub-agents for your compilers requirements, build it by your own. So these are the version controlled and shareable across the whole team, and can also run in parallel. And as for the domain knowledge, you can inject your own documentation, design system guidelines, and your coding standards directly into coders context.

And when Qoder generates code, it will follow your rules automatically. So now, ⁓ here’s the key part. How you actually use all this. So you can think of Qoder like a Stripe for payments or Tailor for communications. You just embed it into your existing development workflow. You don’t need to replace your IDE. You integrate Qoder into VS Code, into JetBrains, into Temrano, into your CI-CD pipelines, whatever and wherever you already work. And there’s actually a precedent here. So Alibaba’s Model Studio platform has enabled over 800,000 custom agents across different domains. And we are also trying to bring the same extensibility here to coding workflows.

Grace Shao

That’s really really interesting. Thank you so much for sharing all that and I do apologize if any of the bit that I didn’t like sound that smart because it’s so technical but I really appreciate you breaking it down for me and really explaining it to me in very simple language. I have one last question for you which is something I always ask all my guests. ⁓ What is one differentiated view you hold? And this doesn’t have to be about work. It could be. It could be about your space could be about agenda coding. It could also be about anything else in the world. It could be about your experience in Silicon Valley versus China. Just a differentiated opinion review of something that you think is not that mainstream or it be a bit against consensus.

Hang

So we talk about something related to agentic coding. actually, I don’t think the transformer-based LLMs in the future. Because I actually do agree with the point that LLMs is just a compressor of the tags. They don’t really understand what they are talking about.

But I think as the develops, as the technology goes, one day we will have the truly AGI. But it’s not based on Transformer.

Grace Shao

What does AGI mean then?

Hang

AI means the AI can really understand what it’s talking, what it’s doing, and really have emotions. Yeah.

Grace Shao

Fully human-like, emotionally, intellectually functioning being.

Hang

Yeah, like human are…From this perspective, they have emotions, have logic, they understand the work, the 3D work, and they have logic of, okay, why I’m doing this, and what I’m going to do in the next.

Grace Shao

Is that scary to think? Because in many ways, they would be more powerful than us, right? They can speak every single language on Earth. They will know more knowledge than us as an individual. But can we though? What if, you know, what if there’s charge and they can charge themselves these days, these humanoid robotics?

Hang

we can cut off the power. You mean to take off and charge by themselves?

Grace Shao

Yeah, like, do you ever worry about this AGI achievement going rogue?

Hang

To be honest, I’m both a bit excited about this and a bit scary.

Grace Shao

Yeah, I think especially for the next generations, right? Like what does it mean for them? Yeah. Thank you so much for your time. I really appreciate it.

B. Christian Hu, Head of Global Marketing and Operations at Qoder

Grace Shao

Christian, thank you so much for joining us. I just spoke to your colleague, Yu Hang. he was super helpful in explaining to me the technicalities of Qoder and the design of the product. My interest right now is really shifting towards the strategy and business side of Qoder. I’m really glad that you can join us today. So why don’t we start with the beginning? Why build Qoder inside of Alibaba? What was the thinking behind that? And what unique advantages does that really give you?

Given that you have someone who just threw $53 billion into AI infrastructure that’s backing you, right? In terms of distribution, your infrastructure, your data, your model access, how does that really advantage you in many ways?

Christian

Okay. Thank you. Thank you, Grace. Thanks for having me here. You know, Alibaba has a very big plan and has very big ambition for AI ⁓ and his position in the future in the AI industry. And as you know, Alibaba has a full stack. We call the full step strategy from the cloud to model and to application. So that’s the full stack strategy. And for, ⁓ you know, a home computer, you know, we don’t build Qoder in a vacuum. Yeah, we build Qoder from the real context because we can find, we have got so many, maybe thousands of engineers from inside Alibaba. They are facing very real software problems in real software development. They have many issues in interface and how to fix the issues. But for the existing AI coding tools, they are maybe reactive. cannot...resolve some existing problems, maybe some complex and sophisticated problems in the real software development process. So they are trying to refer to a new product. we build Qoder just from inside. We have so many inside, the mind from the inside engineers. also, as you know, be part of Alibaba give us, mean, have so many advantage because first of all, we own models.

Because you know, you know, we have early access to Qwen and that’s how I’d propose large-scale models We don’t just use Qwen we also reinvent and feed the Qwen with our real data from our AI coding Qoder base. So we just really reinvent the Qwen and we are coordinated with the Qwen team to improve the performance of Qwen models and

Secondly, we own the cloud. I mean, the bottom layer of Alibaba’s AI strategy. that’s the cloud. Every digital tools and maybe even every AI coding tools should be based on large cloud infrastructures, consume very large computing powers. So cloud is also a very important factor in AI coding platforms.

The third, think we, I think the most, last but not least, we also have the advantage of Alibaba has so many enterprise customers and so many business, you know, we got some business lines maybe from the consumer to the from the SMB to large enterprises, from the consumer to industries. So we got so many data across, as real data for us to, how to do involve the, the area AI to, you know, to ⁓ upgrade our coding platforms from the real data. that’s the, so, ⁓ so back to the, the origins of the Qoder So we just not, you know, we’re just not create another AI coding platform just for the tool, which we want to build a Qoder as a new platform for a Agentic platform for the real software. We want to build the Qoder for real software, not just for fun, not just for the fun making, but for the real software.

Grace Shao

Thank you, that’s really helpful for everyone to understand. I think saying we have a model is the most humble thing anyone can refer to Qwen as, because you guys have one of the best leading open source, open weight models in the world right now. Actually, so on that right now, you know, we’ve been hearing a lot of news about people adopting Quinn globally, and it’s really being used not just in China right now. ⁓ Who are the people using Qoder actually?

Are they mostly Chinese developers or are you guys actually expanding globally? That’s, I guess, the first half of that. And the other half is, are these mostly professionals or are they like students or are they enterprises? Like, how do we understand the demographic here?

Christian

Yeah, good question. We are building Qoder for global developers from day one. Yeah, so that’s not just for Chinese developers, but for global ones. And for now, ⁓ actually for now, our users are mainly come from the, we call them individual developers, not just the enterprise users because know we the enterprise editions on it our way we are just we want to start with the individual you just at the first and and the funding you know geographically from China to the overseas market so we want to some difference between you know preference different preference for different you’re just in different regions yeah from in China they would prefer to the more, more integrated coding system existing with the existing customers. It may be fully integrated with existing systems. And in the Western countries, maybe in some in US or maybe some in ⁓ Malaysia, the users will prefer to more flexible workways. So they’d have a different preference for different coding languages. And for now, we are trying to solve the real software development problems. the most of the users come from the professional users. Because they have a real problem to resolve in the real software development process. Because they doing their work. Because they doing their work and they want to increase their productivity levels to solve more problems. So the main users come from the professional users.

But we also find the increasing adoption of the individuals and some new learners, maybe some product designers, maybe some UI, maybe some UX designers, they want to deploy the AI coding tools to to expand some ⁓ new interface or maybe some new mini apps to increase their productivity is, yeah, so that’s the trend. we believe, uh, Qoder is created to solve the real problem real software development problems. But we also find that the, our agentic, uh, models and the questing models can allow, uh, more users, just like the new learners and new indies, new learners to, uh, new individuals to use our software to create something new, something more powerful. Yeah.

Grace Shao

So my understanding is that Qoder itself came out of the desire to help professional developers. But Qoder Quest, that mode you can go into, can actually help people with less technical backgrounds to be able to play around with it and potentially still build their own thing.

Christian

Yes, Quest Mode is maybe the key to unlock more space for web coding. What is web coding? Web coding is for non-professionals to use the Qoder tools to create something new. I think the Quest Mode may be the new way to unlock more web coding. But our Quest Mode is different from the existing agent model of other competitors or maybe other coding platforms because our Quest Mode is always is really is also a Agentic is a Agentic about also ⁓ we called a delegated and you can control the workflow and control the result on the question model. So that’s that’s for real software. Yeah, so the Quest Mode of coding is also is for real software.

Grace Shao

I see. I think I have a question for you. I asked Yu Hang earlier as well. But basically, my question was that, you know, there are a lot of tools out there already, like the co pilots Curser co whisper Warp you know, there’s there’s a whole array of them. How do I understand where Qoder sits so Yu Hang explained it to me in the technical sense of where it sets between co pilot and a agentic. Can you explain to me where Qoder sits in terms of like the business positioning.

Christian

Okay, we have a slogan for a Qoder We Qoder a agentic platform, a agentic coding platform for real software. So we got two keywords, Agentic and the real software. I just explained what is the software. I can explain more about the software because many developers, mean, no matter professional developers or non-professional developers, they found something critical in their developing process because they need to know the existing Qoder base. They need to understand what the existing Qoder means, what the existing documents mean for the codes. So they need to understand so many documents, so many files to understand what the coding process will be like.

So we created a Ripple Wiki and a non-context memory function to understand the codes, understand the documents, understand the behaviors of the developers. So for the new entrants, for the new developers, they know how to get the real software down. So that’s the issue in real software development context.

Yeah, that maybe seems different from what we call the rubber coating. So that’s for real software. And for Agentic, for most coding platforms, we call it your prompt and the Agent, the React. for Qoder we call it your delegate, then Qoder delivers. So then Qoder that deliver the real software and real results and the real apps for your delegations. So that’s the difference from the existing, I mean, maybe some most of the competitors to agentic platforms. So we are just not want to respond because of the delegation that can, you know, to free you from the desk. mean, for agent, for most agents, you also need to speak to the agent.

You need to communicate with the agent while you are sitting alongside the desk. But the dedication can free you. You just don’t need to communicate with the agent. don’t need to interact with it for line by line, word by word. You just need to monitor the per science and maybe some need to check with the workflow so that the Qoder can deliver the result. That’s the difference between way from the other competitors.

Grace Shao

Actually, let’s take a step back. I’ve noticed that when I was doing some research on you guys that, you’re, it’s not just Alibaba that is creating these coding tools. You know, we have even just in China, we have ByteDance creating something similar. Obviously, Microsoft has things like has Copilot. So then we are looking at obviously then the startups and the Frontier Labs all kind of swarming into coding tools. Why is that? What is really the reason for focusing on coding right now as the next kind of use case?

Christian

Yes, actually, coding, as we just talked, coding may be acceptable for every engineer, maybe a non-engineer, maybe you’re just a digital learner, maybe just an analyst for the AI. So coding is the most certain way for token consuming, for the ⁓ infrastructure consuming, the tool to the cloud infrastructure. Maybe the coding is the most important way. And for different players, they are playing different games. For giants, you just mentioned the Bydance, maybe Microsoft, maybe AWS, they have very large cloud infrastructure. They just to integrate the new AI application to the infrastructure. And for SRAPAC, maybe some other front-end labs, are... ⁓

They are trying to find the new path to the developers, maybe some to the application levels. for Cursor, they are just a new service. They are deploying the large-scale models to make it accessible to the new developers. Cursor has...

I think for new startups, have their weakness, maybe they have their cost structure weakness. mean, they are the users of the large-scale algorithm. They can’t dominate the large-scale algorithm model and maybe they will be at risk to disengage with the large-scale algorithm model owners. So that’s the cost structure that is at risk for them.

And for Qoder know, as I just mentioned, the Qoder is born within Alibaba. Alibaba has full stack from ⁓ cloud infrastructure to model to application. So I think that’s a great advantage. As you know, ⁓ as you just mentioned, Qwen is very popular around the world. some, you know, I just kind of found the news that Airbnb, Airbnb,

Grace Shao

Yeah, Brian Chesky

Christian

Yeah, yes, I said adaptedQwen open source model to the real invent their own infrastructure That’s very popular and for it for cursor. Yeah, it’s just anything for a cursor itself They just came up with a new composer composer model because it’s a small small model for coding but We have a speculation that the this small model comes from the Qwen or maybe some cheap. Both of them are from China. So we will find that the full integration and full connection with the model and the client structure will benefit a for Qoder in the future.

Grace Shao

Actually, on the point of how Qoder sits in Alibaba, can we kind of zoom in on that? Can you help us understand the big picture? Where does Qoder sit in the Alibaba Grand AI strategy across the stack?

Christian

Yeah, I think Alibaba AI strategy is a very big picture and also very long roadmap for what we call the super artificial intelligence. We call it super artificial AI. Yes, wait, wait, wait, ASI, that’s the roadmap. so from the map, so when we look at the map, I think it will form the bottom to the top level.

I just mentioned is the cloud infrastructure. Yeah. The middle level, mean, the Qwen, maybe some other large Niagara models in the middle layer. And the Qoder, I think it sits on the application level alongside with some other applications. Maybe, you know, Dink Talk, maybe Quark, maybe some other applications on the application levels. think that’s the full, we call, so we call the full stack strategy for Alibaba in the near future. Yeah.

Grace Shao

I understand you lead Qoder, international operations and marketing, right? And that’s a pretty big title you got here. What’s your GTM focus right now? Where are you kind of focusing on essentially selling your product to? Like which markets, through what channels?

Christian

Yes.

Grace Shao

How is this kind of working out for you?

Christian

Yeah, so I think it’s a tough task. I know because we are new, we’re just a baby. We just launched our products about two months, no longer than three months. We’re just a new one in the market. our ambition is here. We want to ⁓ become the top global coding platform for global users.

So for me, think the go-to-market strategy for ⁓ us to go global, think we need more partnerships, we need more integration with the local communities. So for me, that’s why I’ve followed to different regions to meet up with local community developers, maybe the leaders of local communities. We want to talk about it.

What kind of preference for them, what kind of products they want to prefer in their context, in their context, maybe within the enterprise, maybe within the individual developing context. And also we are aiming to the global products, we are not going to separate.

China with the global markets. We want to offer two different products. We just offering one product, one platform, one user interface for all the users around the global. So that’s the challenge. But we believe we can do that because we are trying to interact directly or maybe talk directly with ⁓ the developers from the different corners of the world.

Grace Shao

Yeah, that’s definitely quite different from how Chinese companies used to sell their products abroad, right? Because it used to always be a one app location or one interface, domestic one interface for the globe, for the rest of the world. So actually on that, think my final question for you really is just how are you navigating the current climate? Obviously it’s not.

Probably not easy given just current situation with geopolitics, with the competition, with everything, right? How are you navigating differences between China and the global markets in terms of adoption, compliance, data residency requirements, and even developer culture? Because I know you were just in Singapore, like you mentioned, you’re flying around a lot, you’re meeting people from different parts of the world. How are you making it all work?

Christian

Mm-hmm. Mm-hmm. Yes.

Quite different. I think it’s quite different. I have seen huge difference in the developers in China from the developers in Southeast Asia. Because I just flew to Singapore and met up with some local developers. The difference comes from different angles.

User interface and the language and maybe some other features of the software. So the difference, is huge difference between the different rappers geographically and demographically. Yeah, so all the differences are very, very huge. And we also find some, because we are trying to offer our enterprise edition in the near future. So we also started a very

We have made a study for the enterprise levels in different markets. We find quite different stories because in China, the developers in enterprise, maybe in the small companies, in large small companies, they would prefer the company to purchase the software. They prefer to use the existing software. The truth is using software can be embedded into the existing OA system or existing software ecosystem. So they prefer the existing, they prefer the integrations. But in other countries, maybe in other markets, they prefer more flexible and more athletic applications. They don’t care about the integration, they don’t care about the...

I mean the sophisticated integrals, maybe sophisticated workflow and controlled by the company owners. So that’s the different culture. So you can find it’s very common for individual users to buy our software, buy our products in other countries, mean outside China. Yeah, so that’s the huge, that’s the huge difference.

Grace Shao

That’s really interesting. I would have thought that integration part is something that everyone kind of wanted because it makes your workflow so much more seamless, right? And that was the selling point for all of Microsoft’s tools essentially, right? In the software era. Interesting. Yeah.

Christian

Yes. But it doesn’t mean that we don’t need to make an integration or some ecosystem with the... You just mentioned Microsoft, mean, the OA or maybe some Workday, some other software platforms. We also want to explore more connection with existing software.

Grace Shao

Thank you so much, Christian. I really appreciate your time. Thank you.

Christian

Thank you. Thank you so much.

AI Proem is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

Get full access to AI Proem at aiproem.substack.com/subscribe