The AI Engineer's Guide: Who Am I, Where Did I Come From, Where Am I Going? | A Conversation with SiliconFlow Founder Jinhui Yuan and Indie Developer idoubi

十字路口Crossing十字路口Crossing··1·1

Ambition, curiosity, and the freedom to roam.

"Crossing" published "The AI Product Manager's Guide: Who I Am, Where I Come From, Where I'm Going" last month, one of the most popular pieces in our history. So many engineers started calling for it in our groups and comments: when are you doing an "AI Engineer's Guide"?

So the Crossing podcast recently invited two outstanding figures in AI — Jinhui Yuan, founder & CEO of SiliconFlow, and independent developer idoubi — to share their unique perspectives on the AI engineer profession.

AI is reshaping industries at an unprecedented pace, and standing at the forefront of this transformation are none other than AI engineers.

In this episode, from underlying technology to application development, from big companies to independent entrepreneurship, Jinhui Yuan and idoubi offer a comprehensive analysis of the present and future of AI engineers, opportunities and challenges, and how to stay competitive in this rapidly evolving field.

Additionally, both guests share their predictions for AI's future development trends. They believe that while AI tool applications will see explosive growth in the short term, AI Native applications will become the mainstream in the long run. Whether you're already in the AI industry or considering transitioning to become an AI engineer, this episode will offer you valuable insights.

Listen on WeChat:

Listen on Xiaoyuzhou:

👨🏻 Jinhui Yuan's "Roller Coaster" and "Crossing"

🚥 Koji

Let's start with Professor Yuan. His career trajectory is quite legendary, especially his experiences over the past year, which can only be described as a roller coaster. Many people are well aware of Yuan's standing in China's large model market. Probably no one's 2023 was more dramatic than his — as he himself summarized, it was "a period of sitting on a roller coaster."

At the beginning of 2023, Yuan's previous company OneFlow was valued at $100 million. After being acquired by Lightyears Away, the new company's valuation soared to over $1 billion within a month. Two months later, that number crashed to zero. Another month after that, Yuan founded SiliconFlow and started over, with an angel round valuation of several hundred million RMB.

👨🏻 Jinhui Yuan

Looking back, I think my highest moments were still in technology. Although conventionally speaking, OneFlow's acquisition might be seen as a high point due to the public attention.

But personally, the most exciting moment was when we built OneFlow and gained traction in the technical community. Especially when we were the earliest in the entire industry to foresee the large model development trend and made technical preparations for it. For a long time, the industry didn't reach consensus, but as models grew larger and attention focused on large model system architecture both domestically and abroad, people discovered that a Chinese startup called OneFlow had already produced results in this area. Even PyTorch referenced OneFlow's work when doing similar designs. When our technology gained recognition, that feeling was truly gratifying. As for the most uncertain moment, it was probably when I left Lightyears Away last year. We still firmly believed that large models represented the best opportunity of our era, but we had excellent conditions in hand that suddenly vanished. We felt lost then, unsure where to go. But this state of uncertainty didn't last too long.

🚥 Koji

How long did the uncertainty last?

👨🏻 Jinhui Yuan

The uncertain period lasted about a month. Around July, word started getting out that our company faced the choice of whether to join Meituan. Many colleagues had already received offers from other companies, and we had to make a decision within limited time — whether to start over. We couldn't delay too long because everyone had offer deadlines, and other options might expire. We had to identify a direction that most people could agree on in a very short time. To put it dramatically, most of our engineers had received offers with 3x, 5x, or even 10x their current salaries. In that situation, it was an enormous test and choice for each of us.

🚥 Koji

So within that month, you decided to start SiliconFlow?

👨🏻 Jinhui Yuan

Within that month, the main decision was: we would continue entrepreneuring. Although during the six months at Lightyears Away, many other companies had accumulated substantial resources while we were back at square one, we recognized that the entire industry was still in a very early stage. More importantly, our team's combat effectiveness and capabilities in the inference direction remained among the best in the entire industry. We possessed what you might call "secret weapons" or "special skills," and facing this massive wave that had just begun, we still had tremendous opportunity. This was our fundamental judgment. What we needed to think about was how to leverage this "diamond drill" in our hands, find the right entry point, ensure we had a seat at the table in this wave, and have greater room to play in the future. This was our overarching premise. After confirming this direction, the remaining questions were finding the specific entry point, completing fundraising, establishing the company, and other practical matters.

🚥 Koji

In my circle of friends, SiliconFlow has gained considerable goodwill and reputation since its launch. If Professor Yuan had to introduce SiliconFlow to all AI application developers in one sentence today, what would you say to recommend it?

👨🏻 Jinhui Yuan

Our goal is to let developers who don't understand AI build AI applications. We hope to handle all the AI-related parts of applications for you.

🚥 Koji

Recently, The Fair used SiliconFlow's API to develop an application called "AI Will."

This application aims to encourage young people to think about death and write a simulated will under AI's guidance and companionship. We hope this approach helps everyone reflect on what matters most in life, thereby making better choices and living each day well.

The developer of this program is Chenran, who previously worked mainly as a product manager. This was his first program from zero to launch. He said development using SiliconFlow's API was extremely fast — even remarking that connecting to DeepSeek's API through SiliconFlow was too fast. Interestingly, to create a slower, calmer, more contemplative atmosphere, we deliberately reduced the AI's response speed rather than having replies and token display happen at API speed. He marveled that SiliconFlow could improve upon DeepSeek's already-fast response by another factor of two.

Independent Developer idoubi's AI Journey

🚥 Koji

Next let's chat with idoubi. Idoubi is regarded as a god-like figure in independent developer circles. I'd like you to briefly introduce what projects you've done as an independent developer, and which one is your favorite and the one you're most proud of.

👦🏻 idoubi

After resigning from Tencent last October, I've been in independent development circles. The first thing I did was a GPTs application navigation site, essentially a third-party app store. Three days before New Year, I made an AI red envelope cover — wrote it in one hour at Starbucks, and it unexpectedly went viral for a while. After that, I developed other projects: an AI wallpaper website in January, an AI music player in February, an open-source AI search engine in March, and AI virtual try-on in April.

From last October to this May, I basically maintained a rhythm of releasing one new project per month. I paused new releases from May to July, mainly because the AI search project required substantial effort. In August I started a new project — an AI landing page generator: one-click generation of web landing pages, expected to release soon.

Among all my projects, three left the deepest impressions:

  1. GPTs navigation site: This was my first independent product using full-stack development. At Tencent I mainly did backend development, responsible for Go language development, KYS, cluster architecture, etc.
  2. AI red envelope cover: This was an unexpectedly successful project. Completed in just one hour at Starbucks, it gained considerable popularity and gave me great confidence. From this project onward, I gradually built recognition on Jike, Twitter, and other platforms.
  3. AI search engine ThinkAny: This is probably the most well-known project. Many people were amazed that one person built an AI search engine. This project also brought me many experiences I never had before.

🚥 Koji

So after you started independent development, you made projects quickly and prolifically, and they seemed like completely impossible tasks, right? One person made an AI search engine. So in your development process, how much did AI help?

👦🏻 idoubi

I found that AI indeed provided substantial help during development. Take implementing a music player as an example — previously, developing a player component similar to NetEase CloudMusic would have taken considerable time without AI assistance. But with AI, simply telling it I needed a playback component with three switching modes, it could be completed in under an hour.

I think AI's greatest advantage is that it has standardized elements — something like a music player component has a standard answer, and with AI assistance, coding and implementation speed can be dramatically improved.

🚥 Koji

My background is in computer science — I completed my undergraduate and graduate studies at Beihang University, then worked as a product manager for ten years without writing code. Recently, the emergence of Claude gave me an idea. I'd been struggling with Crossing's WeChat official account formatting — handing it to others never produced satisfactory results, requiring extensive time to adjust, yet not adjusting felt unpublishable.

One day I had a sudden inspiration and decided to use Claude to help me write a formatter. To my surprise, it was completed in under 10 minutes. Now every article on the Crossing official account has very consistent formatting, and the efficiency is extremely high — as long as I write the format properly in Notion, a single official account article can be formatted in 5 seconds.

That day, at that very moment, I felt like I'd been given wings made of code — all that computer science knowledge I'd accumulated in the past suddenly came alive. I'd never learned the granular craft of writing code, but programmatic thinking had always lived in my mind. As a product manager, what I do best is write requirements and give instructions. Now I could hand those requirements documents straight to AI.

On my first attempt, code flowed across the screen like water. Even when I hit snags configuring my Python development environment, AI guided me through with breathtaking thoroughness.

Categories of AI Engineers

🚥 Koji

Let's talk about what we actually mean when we talk about AI engineers. First off, "AI engineer" doesn't seem to be a precisely defined profession. When we use the term, it usually carries a very broad meaning. So my first question is for Teacher Yuan: in AI companies today, what are the typical categories of AI engineers?

👨🏻 Jinhui Yuan

My understanding is that any engineer working with AI can be called an AI engineer. Of those, I'm more familiar with engineers focused on producing AI technology. As for engineers who use AI, idoubi has more authority to speak. I'll mainly discuss engineers who produce AI technology.

In the production of AI capabilities, there are primarily two types: algorithm engineers and systems engineers.

  • Algorithm engineers focus on improving AI performance — increasing speech recognition accuracy, enhancing image recognition precision, generating clearer and more compliant images, and so on. Their goal is to elevate a model's intelligence and accuracy.
  • Systems engineers (or infrastructure engineers) focus on improving AI efficiency. Even when a model performs well, if the computation is too slow, costs may become prohibitive or the user experience may suffer. Their job is to make AI run faster and at greater scale on underlying compute or chips, without compromising model accuracy.

Systems engineers also optimize the model training process to make it faster, more stable, and more scalable. For instance, they solve problems like how to distribute a model across hundreds or even tens of thousands of GPUs. They're also responsible for production deployment, ensuring inference remains fast under heavy load — say, tens of millions of daily active users.

In short, algorithm engineers are responsible for the quality of AI capabilities, while systems engineers handle efficiency and scale. Under the algorithm engineer umbrella, there are many specialized sub-disciplines we can get into shortly.

Once AI capabilities are produced and deployed to devices or the cloud, new job roles emerge.

Take prompt engineering — we're researching how to design better prompts so large models better understand and satisfy our needs. To address model "hallucinations," we've started using retrieval-augmented generation (RAG), which has spawned new positions. Then there are Agent and Workflow concepts, requiring dedicated engineers to build these application development frameworks. Projects like Dify and Coze are typically developed by people we call Infra engineers.

The scope of Infra engineers keeps expanding. Initially, they might have focused on efficiently implementing deep learning models like Transformer on GPUs. Now their work extends to developing various application frameworks that make model capabilities easier to use. They've built all kinds of Workflows, covering text-to-image, plus tools like ComfyUI and SD Web UI. All of this falls under systems engineering or Infra engineering.

One layer above that is applying these AI capabilities to develop specific applications and products. This is where idoubi is most familiar, so perhaps he can elaborate.

Career Development for AI Engineers

👦🏻 idoubi

In AI application development, I mainly focus on the upper-layer applications. Let me use a concrete example to illustrate how to develop an AI application.

The simplest AI applications, like the "哄哄模拟器" (哄哄 simulator), can be called no-code AI applications. They're usually built on platforms like Coze or Dify, or through agents that embed a prompt to achieve functionality. These applications take user input, engage in multi-turn interaction with the model, and produce results. This is the most basic layer of applications.

To develop an AI application, the first task is designing a quality prompt. Over the past year or so, prompt engineering has developed into a dedicated field, even forming something like a methodology. These methods tell us how to write better prompts — using a "general-specific-general" structure, first defining a role, then guiding the model step by step, even having the model proactively ask questions. There are now dedicated books on how to write prompts.

I believe prompts have become the foundation for developing AI applications.

For any AI Native application, the first step is almost always built on prompts. Whether developing an AI search engine or an AI art creation product like Artifact, as AI applications they all depend on prompts. We must tell the model how to understand, process, and interact with the information we give it, and in what form to present the final result.

Beyond prompt engineering, I've noticed AI application development is rapidly moving toward low-code.

While the low-code concept has existed for years, AI has undoubtedly accelerated this trend. Take Vercel — it recently released v0, a chat tool where users can quickly generate code and render components through simple conversation, similar to Claude's functionality.

I've recently been developing an AI to Landing Page configuration tool. Users simply input what they want to create — say, an AI search engine called ThinkAny — and the system generates a complete webpage with one click. This includes the entire page structure, relevant terminology, content, even SEO optimization, everything. This dramatically reduces the time users need to invest in page building.

So I believe that beyond prompt engineering, low-code development is another important research direction in the upper-layer application space. This trend makes AI application development more efficient and accessible, allowing more people to participate in creating AI applications.

I want to give another recent hot example to illustrate this trend. Not long ago, Wordware launched a roast feature called Roast that went viral and significantly boosted interest in their main product. This feature also falls into the no-code or low-code category. While it may require some coding, its core is still built on prompts. I believe we'll see more and more products like this. Developers will fully leverage various AI-based workflows to make application creation faster and simpler. This stands in sharp contrast to the past approach of investing massive time in writing code and designing pages.

👨🏻 Jinhui Yuan

I want to add an important observation and reflection. Looking back at the mobile internet wave, we can see that the products that ultimately succeeded and had major impact were often created by application-layer engineers or product managers. In that era, success didn't necessarily depend on mastery of underlying technology. Rather, having sharp product sense was the key. You could say it was the golden age of application engineers and product managers. People with these qualities were most likely to develop products with broad impact, like the Super APPs we see today.

Before the AI era, roles like systems engineers or Infra engineers already existed. Back then, they mainly focused on high concurrency, distributed systems, and similar areas. As someone who's always worked in this discipline, I feel this deeply. We engineers tend to only truly come into play after a product has already succeeded — you could say we're doing "icing on the cake" work. Only when a product has shown promise and its DAU is climbing does our specialized expertise become crucial. For example, only with surging user numbers does a product truly need high concurrency and distributed systems to support it.

But in a product's early stages, particularly when it's just breaking out or starting to go viral, what really makes the difference are people with outstanding application development ability and product insight.

They can accurately grasp user needs and develop compelling product features.

Thinking back to that era, we constantly heard about product managers like Allen Zhang and Yiming Zhang becoming legendary figures. As a systems engineer, I have to admit this left a bit of a bad taste in my mouth. I often wondered: when would we engineers focused on underlying systems get to step into the spotlight? When could we found a company and develop a product that impacts all of society? That thought kept circling in my mind.

When AI emerged, I once thought that engineers focused on underlying technology were finally entering their golden age. This field involves many technical challenges, including algorithmic efficiency problems like how to make model training and inference faster. But it turns out these underlying technologies remain at the foundational level — they haven't truly reached the product layer or directly addressed our needs.

Looking back over the past decade, indeed many peers threw themselves into AI one after another. But objectively speaking, the purely technical achievements haven't lived up to our initial expectations.

Even after the rise of large models, despite the high technical barriers, we thought the golden age of technical people had finally arrived. But reality has been different.

Take OpenAI, DeepMind, and others — they have many legendary scientists and top systems engineers. We thought these people had finally stepped to center stage. But in fact, these core technologies remain just one component or module within a product. They need to become a complete product to deliver their ultimate value.

What's more, these technologies are rapidly commoditizing. The rise of open-source models, for instance — many large model training techniques are quickly catching up. The result is that even excellent technology, as a standalone business model, faces challenges.

I find that we seem to have returned to our previous state. When these technologies first appeared, we had high hopes for them. But as they've gradually become commoditized tools — like how we use JavaScript or React today — they've become just one part of product development.

Those colleagues focused on underlying technology, algorithms, and infrastructure now seem to have returned to the "icing on the cake" role. To put it bluntly, their contributions, while important, may no longer be decisive.

To create truly impactful products, I feel the opportunity and imagination space lies more at the application layer. At this level, we may have more chance to create products that transform industries or even society.

🚥 Koji

Your observation reminds me of the "unsung hero" concept, and similar phenomena exist in many fields.

We often remember the names of film directors and stars because they create works that become widely known, sometimes even earning entries on Wikipedia. But beneath these glamorous surfaces are countless unsung contributors — the inventor of Sony's cameras, the developers of film stock. We may not know their names, but they are the true behind-the-scenes heroes who paved the way for the industry. This phenomenon also leads me to reflect on career choices. When selecting a career path, our motivations profoundly shape the final decision. Some may crave a prominent Wikipedia entry, a desire that might steer them toward the spotlight — becoming a film director or a public-facing figure. Others may pursue the continuous challenge of surpassing themselves, of building ever-better technical infrastructure. For them, recognition within professional circles already brings immense satisfaction. From our conversations, I sense that you, Jinhui, have a particularly strong drive for to-C impact. I think this is excellent, because I'm the same kind of person. So that sigh you just let out — it really comes from your own internal priorities, wouldn't you say?

👨🏻 Jinhui Yuan

I believe that ultimately, whether we're doing algorithm research, infrastructure development, or application-level work, we're all pursuing the creation of value and impact.

Objectively speaking, the closer work gets to real-world application scenarios, the more pronounced its leverage and direct impact become, which also makes such work more attractive.

Of course, I also understand that some peers pursue breakthrough innovations in technology, or strive for perfection in a specialized domain, aiming to be the best in the world. This pursuit can bring tremendous satisfaction as well, reflecting different personal value orientations. If we want to see our technology directly help others and create practical value, working on the product side may yield greater fulfillment. In this position, we can more intuitively feel how our work affects users' lives — an experience that is truly exhilarating.

🚥 Koji

So idoubi, do you have deep personal experience with this?

👦🏻 idoubi

I feel that most engineers simply enjoy the joy of creation. Just as Jinhui might take pride in creating a widely popular deep learning framework — as the framework's creator, seeing your work widely adopted and enabling the birth of many excellent products undoubtedly brings immense accomplishment. For me, though I may never develop such an advanced framework, I've found that with AI's help, my ability to create products has dramatically improved. Before, I might only develop five apps in a year; now, with AI technology, by simply adding some prompts or calling APIs, I can complete a product in a week and create dozens of different applications annually. Each product solves specific problems for different groups, and this process of product creation brings me continuous satisfaction. The steady stream of positive feedback keeps me deeply immersed in the joy of building products. I thoroughly enjoy this sustained happiness of continuous creation.

I believe that in the current era, the ways to access this creative joy have become more diverse than in the past.

👨🏻 Jinhui Yuan

I'd like to add something about where the joy of foundational technology comes from. The charm of underlying technology or products lies in their generality — their ability to be applied across multiple scenarios and products. If you can truly achieve this, it does bring enormous satisfaction.

But generally speaking, the foundational technology domain lacks diversity and often follows a "winner-takes-all" dynamic, making competition exceptionally fierce. In this field, diversity typically doesn't coexist for long; ultimately only the best technology emerges victorious. We can see this in several examples:

  1. In algorithms, the Transformer model surpassed previous RNNs and has now become virtually the industry standard.
  2. Among deep learning frameworks, PyTorch has gradually become the mainstream choice.

This competitive environment is brutal — only the technology that ultimately wins and achieves widespread adoption brings true satisfaction. For other participants, their work may simply become stepping stones or intermediate steps in the progress of others.

I think this phenomenon exists in products and applications too. Only by ultimately winning can you experience this sense of accomplishment. In the application layer, I've found its diversity to be remarkably rich. Whether in to-B or to-C scenarios, from productivity tools to entertainment applications, each domain can give birth to unique products. But at the technology level, the achievements that can be preserved long-term are often few — this is the cruel side of this industry. Yet when you create something that the entire industry widely adopts, that sense of accomplishment is unparalleled. For instance, the inventor of search engines must feel incredibly proud, because what they created has become part of daily life for people worldwide. I imagine TikTok's developers feel similarly. Another example is Google's Jeff Dean, regarded as a legend in AI and information retrieval. I'm sure he feels tremendous joy — his development of MapReduce laid the foundation for Google's entire system, and he later went on to create TensorFlow and numerous other important projects.

So these different roles do offer somewhat different kinds of enjoyment. Those closer to the terminal end of value creation — that's the application layer. This, I feel, is where the difference in fulfillment lies.

Changes and Challenges Brought by Large Models

🚥 Ronghui

Following up on what Jinhui mentioned about gaining recognition for foundational work and the changes happening around us, I'd like to ask — as far as you understand, is the situation in Silicon Valley the same as domestically?

👨🏻 Jinhui Yuan

I went to Silicon Valley in March, and recently I've mainly been following the situation there through media. While in Silicon Valley, I discovered an interesting phenomenon chatting with friends: people actually in Silicon Valley often know less about what's happening there than people back in China. The reasons are quite intriguing. Whenever something happens in Silicon Valley, Chinese media has already translated and widely disseminated the information that same evening. In contrast, friends in Silicon Valley might not even know what happened yet. The Chinese community there in particular often learns about local events through WeChat Moments or Chinese media. This is quite fascinating.

I've noticed several eye-catching industry events recently, such as foundational model companies being acquired, and movements at DeepMind and Anthropic AI. These have triggered an obvious trend: more and more people are openly discussing whether large model development is slowing down, and whether existing large model technology can lead to artificial general intelligence (AGI). People now seem more focused on AI applications and return on investment value. This shift in trend is very pronounced. For example, previously only Sequoia Capital investors raised the question of whether AI could create $200 billion in value; now that figure has been raised to $600 billion. This reflects a reality: current investment in AI is enormous, already reaching tens or even hundreds of billions of dollars. For these investments to make sense, they must generate $200 billion or $600 billion in returns. Initially this view may have been a lone voice. But now, increasingly more people are embracing it. This represents a shift in thinking. People are beginning to realize that technological legends and myths must ultimately demonstrate value in practical application. We're entering a new phase, beginning to ask how AI technology can truly create value.

🚥 Ronghui

Could both of you share observations about how the AI engineer role has changed, drawing from your own experiences?

👨🏻 Jinhui Yuan

When I was pursuing my PhD, my research focus was algorithms — I was essentially an algorithm engineer. But after graduating and starting work, I gradually shifted toward engineering, particularly Infra. These two domains differ greatly, and my reasons for moving from algorithms to Infra are quite interesting. I found competition in the algorithm field to be exceptionally fierce. As we often say, foundational technology is "winner-takes-all," but algorithms are even more so. If you can't come up with the most advanced ideas in a given direction, your work is unlikely to be adopted. Whether an algorithm scientist can conceive innovations like ResNet, VGG, RNN, or Transformer often involves considerable randomness. Every year, tens of thousands of papers are published worldwide, yet only a handful of algorithms or ideas survive, gain industry recognition, and achieve widespread use.

The emergence of these breakthrough ideas certainly requires continuous exploration and diligent effort, but I believe there's also an element of inspiration's blessing — akin to that "eureka moment" product managers experience.

Thus in the algorithm field, only a small number of people experience true fulfillment — those who created ResNet or Transformer, for instance. Most people's work, while valuable, rarely reaches this degree of accomplishment. This makes the algorithm field extraordinarily brutal. Meanwhile, innovations in algorithms often depend on clever "insights." Once an idea is proposed, it's like piercing a window paper — everyone can understand it. For example, we all now know concepts like "Attention is all you need." Once such ideas appear, they spread extremely rapidly. If we tried to build a commercial company around such an algorithm, we'd face problems. Because once this innovation is made public, it becomes difficult to maintain as a commercial moat. This is a characteristic of the algorithm field, and an important distinction from Infra.

I found that Infra differs significantly from algorithms. While Infra also requires some flashes of inspiration, it relies more on long-term engineering practice and accumulation. This isn't a domain where simply having an idea or breaking through a concept allows others to immediately replicate it. Rather, it demands engineers' intelligence and sustained commitment, typically measured in years. Just as the Black Myth: Wukong game we discussed today required four years of development, many Infra systems need dozens or even hundreds of people working continuously for several years. Even if others have similar ideas, without hundreds or thousands of person-years of investment, success remains elusive. This was precisely my reason for moving from algorithms to Infra. Though both belong to foundational technology, systems software is more likely to develop into a commercial product or company. Just as past big data systems, database products, and companies like Snowflake, Databricks, and Oracle successfully established commercial enterprises. Another change is that large models have also brought shifts to foundational technology. Compared to past models, the most notable characteristic of large models is their need for intensive computational resources. Previously, algorithm engineers only needed one or two GPUs to conduct experiments or invent new algorithms.

But now, without a large model team and experimental conditions involving thousands of GPUs, it's difficult to carry out such work. This has resulted in far fewer people being able to continue working as algorithm engineers than in the past.

Beyond the resource-intensive nature of these models, I've noticed another important trend: model architectures have become highly converged. In the past, we went through a period of blooming innovation with many competing algorithms and architectures. In the CNN space, for instance, we had ResNet, VGG, Inception, and numerous variants. But now, virtually everyone is using Transformer architectures, and they all look remarkably similar. This presents algorithm engineers with a formidable challenge: how to break through again, or even surpass or overturn Transformer? The task has become extraordinarily difficult. Because enormous resources have already been poured into the Transformer architecture, exploring its performance under various conditions and identifying optimal hyperparameter settings.

New ideas struggle to secure comparable resource support for thorough exploration, making major breakthroughs increasingly unlikely.

As model architectures converge and massive resources have already been committed, many algorithm engineers find it increasingly difficult to produce "masterpieces" or disruptive achievements. Most people can only do incremental improvement work within the Transformer framework. In this environment, our focus has shifted toward data. With fixed architectures, model performance gains come primarily from data optimization. We need to consider how much data to use, what kind of data, and how to configure it. The combination of Chinese, English, code data, and even various proprietary datasets has become the critical secret sauce.

🚥 Ronghui

I've noticed an interesting shift in the Bay Area recently, particularly in how people talk about work. It reminds me of a conversation I had with a friend, and some exchanges I've heard myself. Previously, when people discussed high-earning engineers, they didn't particularly emphasize whether someone was a data scientist. But now, the situation seems to have changed significantly. If two engineers both have high incomes, people will specifically point out that one is a data scientist, as if that title makes a golden halo appear above their head.

👨🏻 Jinhui Yuan

I'd add that this isn't absolute. In large model companies, algorithm engineers are indeed relatively scarce—they even use GPUs per capita as a key metric. For example, a company with 10,000 GPU cards might have only 10 algorithm engineers using those resources, meaning per capita could reach 500 or even 1,000 GPUs.

These companies measure productivity not by the number of algorithm engineers, but by how many top-tier algorithm engineers they have.

They believe each engineer should be able to command massive GPU resources. However, I realize this situation may feel distant for most people, since after all, most of us don't work at these large model companies. At the same time, I've noticed substantial work continues in academia—for instance, researching how to better fine-tune models, how to improve alignment techniques, how to address hallucination issues. There's also the important challenge of solving the "black box" problem. Currently, neural networks are generally considered uninterpretable, but developing a theory to explain why deep learning and large models work so effectively remains an open mystery. These are still problems that require our colleagues in algorithms and theory to tackle.

🚥 Ronghui

Professor Yuan, from when large models emerged with ChatGPT to now, how do you think the evolution of their capabilities has changed requirements for engineers' abilities?

👨🏻 Jinhui Yuan

For algorithm engineers, cross-disciplinary capabilities and interdisciplinary knowledge are indeed crucial. Their innovative ideas often extend beyond traditional computer engineering, potentially drawing from physics, biology, and other fields. At companies like OpenAI, many people with algorithm backgrounds originally came from other disciplines, including a notable number with physics backgrounds. For infrastructure engineers—particularly those supporting large model training—they need deep expertise in distributed systems, large-scale cluster management, and high-performance networking. This demands exceptionally strong hands-on coding ability, the capacity to quickly solve complex low-level system problems, and sharp insight and intuition about these challenges. Overall, the primary criterion for evaluating such engineers is outstanding engineering capability—they must be able to rapidly get up to speed and solve highly challenging foundational problems.

As AI technology has developed, we've seen an interesting trend. Initially, AI-related skills were extremely scarce. But as more people have recognized and adopted it, AI is gradually becoming infrastructure—like water and air, ubiquitous yet barely noticed. This trend is also shaping the direction of industry specialization. In this context, I believe roles like idoubi's as full-stack engineers are becoming increasingly important. These engineers need not only solid technical foundations but also strong product sense and insight. They can combine technical capability with product thinking, penetrating industries and various scenarios to become indispensable products in our work and lives.

Super-Individuals

Since I quit my job to become an independent developer, my way of working has changed dramatically. — idoubi

🚥 Ronghui

I actually wanted to ask idoubi next—when you worked as an engineer at a major tech company versus now, essentially running a one-person company, what major changes have you noticed in how you work?

👦🏻 idoubi

Before university, I had basically never touched a personal computer, let alone studied computer science. Unfamiliar with the field, I chose nuclear engineering at Wuhan University. However, in my sophomore year, I stumbled into IT and was immediately captivated. I started watching videos online and teaching myself programming. The sense of accomplishment from building things gave me tremendous motivation. This continuous self-learning and passion for the industry has stayed with me.

If someone wanted to follow my path now—say, studying an uninteresting major, or a freshman wanting to self-teach programming—I think the barrier to entry is much lower today. Resources are abundant, and with AI-assisted programming tools like Copilot, learning has become more accessible. When you encounter something you don't understand, just paste the question and AI gives you a thorough explanation. So I believe this is an excellent era. Whether learning full-stack development, Python, or other technologies, the barriers are lower and progress is faster. In a short time, you can learn much more.

This is my first point: with modern technology, we can learn more.

For my second point about work changes, let me share my experience. Previously, working at a major company like Tencent, there was typically clear division of labor. As Professor Yuan mentioned, there are horizontal layers from bottom to top. But in practice, each layer also has many vertical divisions. In large tech companies, application-layer specialization is indeed extremely fine-grained.

Take my project building WeChat's overseas payment system—our team structure clearly illustrates this: frontend development, backend development, and database engineering each had dedicated engineers. As a backend engineer, our work focused on specific domains like wallet account balance changes, writing API and RPC interfaces, and deploying systems for frontend consumption. Frontend engineers didn't need to understand backend data design details; they simply followed interface documentation to integrate, ensuring the whole process ran smoothly.

I'd also done architecture design and operations roles. As an operations engineer, I didn't need to care about specific business logic. My responsibilities were setting up Kubernetes clusters, planning microservice decomposition, selecting appropriate IPC frameworks, deciding whether network communication went through internal networks or VPC private networks, and so on. My goal was ensuring each component people developed could run smoothly on this cluster, so I didn't need to focus on its business logic.

To summarize, large companies divide every job function very precisely. You only need to excel in your own specialty. In the promotion system, we had different advancement tracks. For example, even if I completely didn't understand frontend development, couldn't write applications or products, I could still become a backend development expert. Similarly, a frontend expert might know little about backend development.

Since I quit to become an independent developer, my way of working has changed dramatically. As a full-stack developer, I no longer need to coordinate interface integration with others, or negotiate data pipeline design. Documentation and data transmission details are all my responsibility. I just select a full-stack development framework, quickly write interfaces, and debug them myself. This transformation has brought remarkable efficiency gains. Previously, developing a system might take a month with a five-person team. Now, I can complete the same system alone, perhaps in three days, sometimes even one day.

The change in working style manifests in two main ways: First, I no longer need extensive collaboration, which dramatically improves efficiency. Second, I've broken free from dependence on fixed components. At WeChat, we often relied on internally developed components from the company's architecture team—event centers, message queues, databases, and so on. Now, I can use AI to recommend suitable components, then directly integrate based on interface documentation. This approach greatly boosts development efficiency.

Personal development and company development differ enormously in workflow and patterns.

👨🏻 Jinhui Yuan

Full-stack engineers like idoubi are very much needed—many companies desperately need such talent. Because with the emergence of AI, where product opportunities lie remains highly uncertain. It requires many ideas and rapid iteration. If you use a formal, large-team approach with divided responsibilities, the pace is too slow, and we see results too slowly. People sometimes say that even if programmers are all skilled, the bandwidth for collaboration and communication between them remains relatively narrow. So with a full-stack engineer, they're communicating with themselves—the bandwidth in their mind is essentially unlimited. But with two people, communication still happens through language, and language bandwidth is quite narrow.

🚥 Ronghui

As idoubi described, the workflow at major companies used to work like that, but we're seeing more and more super-individuals like you. Do you think any jobs will disappear in the future?

👦🏻 idoubi

I don't think certain roles will disappear entirely — it's more likely their proportion will shrink. Looking back at my time at big companies, a product typically involved multiple roles: designers, product managers, testing engineers, frontend and backend developers. Each had clearly defined responsibilities. Testing engineers, for instance, would validate the complete workflow against interface documentation or PRDs. But at WeChat, I noticed an interesting management philosophy: engineers are responsible for their own code. This meant WeChat effectively had no dedicated testing engineers. Every developer had to test their own code and handle their own releases. Now, as an indie developer, things have shifted again. I don't just lack testing engineers — I may not even have clearly separated roles for product manager, frontend, or backend. I'm essentially combining multiple roles into one person. I think as more people choose to become indie developers or freelancers, the boundaries of many traditional jobs will blur. We may no longer need such granular division of labor, but rather expect everyone to have multifaceted capabilities.

Jinhui Yuan

The rise of indie developers and super-individuals is indeed a significant trend both domestically and internationally. Several key factors drive this: increasingly clear division of labor, and the continuous improvement of developer tools and platforms. Much infrastructure that previously required self-hosting can now be easily obtained through cloud services. For example, we no longer need to set up Kubernetes clusters ourselves since cloud platforms offer managed K8s services. If we need AI capabilities, we can directly call various large model APIs like OpenAI's. Even website hosting, server rental, and CDN services have become extremely convenient. This means developers can focus more energy on core business logic rather than being bogged down by infrastructure. I recently read a blog post by idoubi about migrating his website from Vercel to Cloudflare. At first everyone used Vercel, which already saved people considerable effort. Later he found the costs still too high and wanted more autonomy, using CDN instead, which led him to discover products like Cloudflare. These IT infrastructure, cloud products, and services create excellent conditions for indie developers.

Now, a full-stack engineer can leverage their own capabilities, supplemented by external products and services, to build a complete product. This dramatically lowers the barrier to entrepreneurship and makes the emergence of super-individuals possible.

We've also seen some successful cases recently. Midjourney, for instance, reportedly has only about twenty people yet generates hundreds of millions in annual revenue. Many small companies have achieved solid commercial results in their respective niches.

The Challenges of Entrepreneurship

Ronghui

Something just occurred to me — it seems like you, Teacher Yuan, would also want to recruit someone like idoubi, someone with such comprehensive abilities. But someone with his breadth of skills can just start a one-person company and make money on their own. This is actually a contradiction. How do employers like you resolve this? And idoubi, what are your thoughts?

Jinhui Yuan

I deeply understand the restless mindset of entrepreneurial spirits like idoubi, because I've walked that path myself. I used to constantly think about how to accomplish something, create impact, and kept trying. The process is indeed full of challenges and difficulties, but if you have that inner drive and passion, these obstacles can be overcome. Such people cannot be recruited by startups or big companies alike. Talent can be measured across two dimensions: capability and appetite for risk. Those with both ability and a strong love of adventure are best suited for entrepreneurship; those with ability but risk aversion may fit better in large companies; those in between might consider joining other startups. Of course, there are compromise solutions. Many ByteDance businesses, for example, entered through acquisition. This is one approach: though you join a large company, as an acquired unit you can still operate a new business relatively independently. Douyin itself developed this way after being acquired into ByteDance. So between entrepreneurship and employment, multiple possibilities exist.

idoubi

I have a simple idea: respect your own personality. I'm someone who loves freedom, the type who acts on impulse and does whatever comes to mind. Five years at Tencent was already quite a long career for me. Before that, I left two companies within less than a year each. So I felt I'd reached my professional limit at Tencent, and resigned last year without another job lined up. Mainly because I had too many ideas in my head every day without enough time to realize them, which made going to work painful. Since quitting, I've actually felt quite happy. Though you mentioned individual entrepreneurship might earn lots of money without needing to work for others, I'm skeptical about that. At least so far, I haven't had any standout commercial results. But in terms of product development and building personal influence, these nine-plus months of freelancing have been very rewarding. My biggest feeling is that I can act according to my own ideas every day. When inspiration suddenly strikes, I can quickly implement it, launch it, and have many people see it — that feeling is fantastic. Some people like Teacher Yuan, Monica, and Brother Xiaohong have invited me to join their teams. I deeply respect these entrepreneurs and feel there's much to learn from them. But currently, I still have too many unrealized ideas. I want to make the things I'm truly passionate about even better. Then I'll see what other opportunities arise. If I can make money through my own efforts, that would of course be the ideal direction.

Ronghui

I think as you said, you adjust based on current priorities and what your personality needs most.

Jinhui Yuan

I too started as an engineer unwilling to accept the status quo, feeling I shouldn't be limited to narrow capabilities and influence, and thus embarked on this entrepreneurial path. One major advantage of entrepreneurship is that there's no ceiling — you can fully unleash yourself, and your upper limit of influence and achievement depends entirely on you. By contrast, working at large companies often requires conforming to overall planning and facing various constraints. Entrepreneurship indeed tremendously aids personal growth. Moving from mastering a single skill to developing comprehensive abilities involves many aspects. Now entrepreneurship requires not just R&D and product development, but also handling growth operations, financing, and commercialization — you need business thinking to form a complete closed loop. This is a process of comprehensive growth. Behind growth often lie challenges, and entrepreneurship does involve many difficulties. But if someone truly enjoys the process of realizing their ideas, these challenges seem trivial and can be overcome. Even so, entrepreneurship is indeed a difficult path. I believe every entrepreneur personally experiences the full range of emotions these challenges bring. Problems and difficulties emerge endlessly. But having chosen this path, one must be mentally prepared. After all, while entrepreneurship is full of challenges, it also contains infinite possibilities.

Koji

I'm curious — what are the hardest things for you?

Jinhui Yuan

On the surface, many people would consider fundraising the biggest challenge in entrepreneurship. Indeed, fundraising is a major challenge, but I think this is only the surface. The most difficult thing is actually changing yourself, or recognizing your own limitations. No matter what difficulty you encounter, someone can solve it, but if you cannot change yourself, many problems may remain unsolvable. Take fundraising as an example — it has its particular patterns. Some people sail through it while others struggle. The key is finding the underlying patterns, which may require challenging yourself, reflecting on whether you're too stubborn or self-righteous. We need to consider which capabilities need improvement, let go of ego, recognize where we can improve, and work to fill those gaps. This process may require independent effort or team collaboration. So I believe the biggest challenge is your own growth. If you can't keep pace with developments and adjust and improve in time, this becomes the greatest obstacle on the entrepreneurial path. Entrepreneurship isn't just about business development — it's a continuous process of self-renewal and breakthrough.

Koji

This time you still resolutely started a new venture. And knowing there are thousands of difficulties, where do you think the greatest inner motivation comes from?

Jinhui Yuan

My lifelong pursuit has never been fully realized. From the start of my entrepreneurship, I've focused on solid technical work and R&D, hoping to create something valuable for others, to create value, generate influence, and gain a sense of achievement and reward. Though this wish has never been completely fulfilled, if I gave up now, I would certainly regret it later.

Considering how much I've invested and how long I've persisted, now is actually a better timing. Everyone generally agrees that large-model AI represents a transformation comparable to the mobile internet or even the PC revolution. Such opportunities may only come a few times in a lifetime. We've accumulated substantial groundwork, the team is growing — if we gave up now, we would certainly regret it. So to leave no regrets, we must continue pursuing our unrealized dreams.

From beginning to end, I've wanted to explore my own limits, to see what I can achieve, what kind of enterprise I can create. Early on I focused on what level I could reach in research, how creative I could be — where papers could be published, how many citations they could receive. Later I paid more attention to product and business impact on the company — how much revenue I could generate.

Now I hope that amid still-uncertain industry prospects, through my own research and thinking, I can make some forward-looking judgments. To prepare in advance for important events that may occur in the future, whether at the technology or product level. Ultimately, it's all about generating greater influence.

I've always pursued this goal, hoping to rely on my own and my team's strength to see how much influence we can generate, what kind of enterprise we can build. This is what drives my persistence.

Koji

When did you realize you had such a strong drive for impact?

Jinhui Yuan

This pursuit traces back very far. From childhood to now, the scope and depth of this pursuit have continuously expanded, and I feel it's become more ambitious. From a young age, I pursued so-called "excellence" — striving for good grades, publishing high-quality papers. My horizons were perhaps limited then, focused on problems of not much practical significance. As time passed, my pursuits gradually shifted toward more practically meaningful directions. I began focusing more on how to bring positive momentum to people — how to make people's lives happier, how to improve work efficiency, how to push the world toward better development, and so on.

Koji

idoubi, what you do also demonstrates vigorous creativity. To many people it's unimaginable — how can someone have so many ideas and such strong execution? So if you look within yourself, do you feel this drive comes from the same place as Teacher Yuan's pursuit of influence and excellence, or from a different angle?

idoubi

I think, on one hand, it's similar to Teacher Yuan — I also hope to have greater influence. On the other hand, when you build a product that genuinely solves someone's problem and people really love it, there's an enormous sense of accomplishment. I've always been someone driven by accomplishment. Whether it was teaching myself computer science early on or building products now, it's fundamentally the same motivation. This continuous sense of accomplishment is what propels me forward.

Another aspect is that I'm constantly exploring the boundaries of my own capabilities. I used to focus mainly on backend development, but later I tried frontend development and found it wasn't too difficult for me. With some accumulated experience, picking up new technologies became relatively easy. I typically just need to go through the official documentation once, understand the technical architecture, and I can quickly get started building things.

For example, in March this year, I started working on AI search. At first I thought this was a massive domain — companies like Google and Baidu have been at it for decades, and I didn't think someone like me could easily enter it. But as I dug deeper, I realized it was actually within my capabilities. Because I didn't need to retrace the old path of traditional search engines; I could build on top of existing AI capabilities. I found I could even get the basic functionality working within a day or two.

As my exploration deepened, I gradually became "disenchanted" with this industry. I realized that basically every industry I've looked at doesn't seem to have particularly high technical barriers.

Based on this understanding, I developed more desire to experiment. I feel that for any software I use frequently, as long as it's within my capabilities, I want to build it myself.

Doing this lets me engage with different products, gain different kinds of accomplishment, connect with different communities, and develop deeper understanding of each industry. These elements give me continuous new desire to learn, constantly replenish my knowledge base, and apply what I've learned through actual products across different domains. This is what drives my continuous creation.

Jinhui Yuan

Actually, I'm no longer an engineer myself, because I don't write code anymore. I find this quite regrettable. I especially envy someone in idoubi's position — having an idea or product concept and being able to immediately start implementing it, seeing results quickly. Looking back at when I used to write code, I thoroughly enjoyed that state. Now if I have an idea, I can't execute it as quickly myself. I need to find colleagues to do it, discuss it at length, and they might not execute exactly according to my vision. In this situation, I've become more aware of how enviable an engineer's freedom is. That ability to directly translate ideas into reality is truly precious.

Ronghui

Would you consider using agents like devv?

Jinhui Yuan

Although I haven't personally used these AI programming tools yet, I've heard they can solve some preliminary problems. However, I still believe in the value of human engineers. This reminds me of how AIGC has developed in painting and writing. Some say AI has already surpassed painters and writers in certain aspects. But I still feel that AI cannot surpass the most outstanding human creators. The same logic applies to programming. I believe the capabilities of the best engineers still cannot be fully matched by current AI programming assistants. There's still quite a bit of wisdom here that I think AI cannot yet comprehend.

Ronghui

Some of our listeners are interested in entering this industry, and some are already in it. Teacher Yuan, for those transitioning from being classical engineers to entering AI companies, what areas would you suggest they get into first?

Jinhui Yuan

Using APIs based on large models or generative AI, creating interesting things with prompts or approaches similar to Dify — this is a relatively fast method. There are many open-source projects available now for learning and reference. Downloading open-source code, modifying and customizing it, transforming it into what you want — this approach lets you get hands-on more quickly. Of course, there are also abundant tutorials, textbooks, and video resources available now. For engineers who want to deeply understand the underlying principles and are genuinely curious about AI technology, these resources can help them learn about so-called deep learning, gradient descent, matrix operations, the specific calculation methods of attention mechanisms, and so on — there are formal courses available. For example, Stanford University's CS231 series has trained a generation of people and had profound influence. Online learning resources are now extremely rich, whether on Bilibili or YouTube, there's a lot of high-quality content. But the most direct and effective method is still to start from code, from open-source projects. Through actual hands-on operation and modifying code, you can understand and master these technologies more quickly.

Ronghui

Teacher Yuan, as an employer, when your company is hiring engineers, what do you value most?

Jinhui Yuan

In a startup, I think two key qualities are particularly important: First, self-motivation. Startups typically have small teams with flat organizational structures, requiring employees to independently handle various tasks. People with strong self-motivation have the drive to continuously learn and improve, are full of curiosity, and can proactively solve problems. This trait means they won't easily give up when facing challenges, and they have tremendous growth potential. Second, highlights in past experience. This can manifest in multiple ways: excellent academic performance can indicate someone is disciplined and dedicated; even if grades aren't outstanding, there should be achievements in other areas, such as well-written blogs or well-produced videos. What's important is having a domain you love and excel at.

Koji

Let me ask a harsh question. Because many people, including myself, when we hire new colleagues, we hope they have strong self-motivation. The more startup-oriented you are, the more you want self-motivated people, but the truth is many people in this world simply aren't very self-motivated. Teacher Yuan, what advice would you give these people?

Jinhui Yuan

I believe the root of both self-motivation and interest is actually finding your true passion. When someone is genuinely passionate about something, no matter what obstacles lie ahead, they will strive to overcome them and figure things out. This curiosity and interest that comes from within is the most powerful motivation. Now, most people have moved beyond subsistence concerns, and we have more freedom to follow our hearts.

The philosophy of "Follow my heart" is becoming increasingly important.

If you can develop according to your interests and passions, you can not only solve survival problems but also potentially achieve higher accomplishments in your professional field. So you still need to have your own hobbies and interests, and develop according to your interests and career.

Koji

Actually self-motivation is a result, right? Self-motivation is the result of passion. Paul Graham, founder of Y Combinator, wrote an unprecedentedly long article last year exploring how to do great work. The first point he emphasized was to return to yourself and discover where your passion differs from ordinary people's.

For example, suppose you have an unusual interest in caterpillars. In others' eyes, this might seem somewhat eccentric. But if you follow this interest and research it deeply, you might become an expert in caterpillars. This can not only solve your livelihood but could also become the starting point for achieving great things.

Jinhui Yuan

Although AI can do more and more things, seemingly capable of replacing many human abilities, it cannot replace people's individuality and inner creative pursuits. I believe the future trend should be encouraging people to develop according to their own personality and interests. From both a worldly and personal survival perspective, this is also the more appropriate approach.

Advice for AI Engineers

Ronghui

That question wasn't fully answered — the second half was for idoubi. I just asked Teacher Yuan about work for AI engineers, and I'm thinking many people might want to follow your path, which is more free and freestyle. Of course, it also involves taking many risks. I'd like to ask, for engineers who want to develop in your direction, what kind of engineering background, what kind of advice do you have?

idoubi

As an engineer with experience in AI application development, I'd like to offer some advice for those wanting to enter this industry. Whether you have an engineering background or are starting from zero, developing at the AI application layer is relatively easy. Of course, if you want to do underlying development, the difficulty is much greater.

I believe there are mainly three paths into AI application development:

  1. No-code development: There are now many no-code platforms like Dify and Coze. As long as you have ideas, you can use these platforms' workflows to build an application your friends can use. For example, previous apps like the "coaxing simulator" or roast apps — these applications can quickly gain massive traffic, give you positive feedback, and inspire more passion to create. No-code development lets you quickly enter the AI application industry.
  2. Low-code development: There are many low-code frameworks now, such as Gradio and Streamlit. You just need to know Python, without needing to master frontend development or write lots of HTML code, to create powerful applications. For example, you could develop an app where users upload photos and see themselves wearing a red T-shirt. This not only brings a sense of accomplishment but might also generate some income.
  3. Full-code development: This has higher requirements for fundamentals, but compared to ten years ago, learning programming has become much easier too. There are many open-source courses and resources on GitHub. I myself started a full-stack development course last December, and some product managers who joined my community have also learned to code.

For those starting from zero, I suggest beginning with a complete 0-to-1 project. You don't need to understand all the principles at first — just replicate the project and get it running. When you encounter something you don't understand, ask AI. Typically a product manager with zero foundation can achieve the process of coding and developing their own ideas within one to two months.

For people with some technical background, such as those who previously worked on algorithms or backend, learning will be faster — perhaps able to build things themselves within a week. Because there are so many open-source resources now, plus AI-assisted programming, building an application has become very simple.

Overall, I suggest choosing the no-code, low-code, or full-code path based on your situation. If you can read code, you can learn different types of programming. If you just want to build your own things, there are many low-code or no-code tools available.

The most crucial thing is maintaining continuous passion.

You need to have an idea you want to realize, something that doesn't yet exist in the market. With this motivation, you'll seek out various methods to implement it. So my advice is: first have your own creative idea, then find various tools to realize it.

At the current stage, as long as you have your own ideas and a bit of self-motivation, entering the AI application development industry is actually very easy.

Ronghui

For engineers who already have some experience, how should they refine themselves to improve their competitiveness? That is, people working at the Infra level and application level where you respectively are. Because the environment is quite challenging now — we've also heard the overall market has some fluctuations, and some people are being eliminated.

Jinhui Yuan

When discussing future engineer competitiveness, I believe there are two main development directions:

  1. Specialization: This means going extremely deep in a specific domain to become one of the best in the world. For example, some of my colleagues focus on making algorithmic matrix operations run faster on GPUs or Huawei Ascend chips. This requires intimate knowledge of GPU system architecture at the lowest level — instruction sets, shared memory, bandwidth, memory banks, and countless other details. These engineers need depth that surpasses even the chip vendors' own engineers; their understanding can't be weaker than NVIDIA's. This kind of specialization can go extraordinarily deep, making you a top expert in that niche.

  2. Well-rounded development: The other direction is breadth — the ability to synthesize knowledge across domains and manage complex systems. While AI can handle portions of coding work, I don't think current AI is fully capable at the overall system architecture level. That's why the industry desperately needs full-stack engineers in the vein of what a16z advocates — people who can think about and design systems holistically. This is a capability that AI currently struggles to replace.

idoubi

Building on what Teacher Yuan said, I'd like to add some thoughts on engineer development directions.

Depth is indeed necessary, and it maps to certain specialized domains. Take full-stack development: building a simple product might be easy. For example, Monica's recently launched AI roast feature, or Worldware's roast feature before that — for someone like us, it might take a day or even half a day to build.

But what if the product suddenly blows up? Monica's project, for instance, hit tens of millions of PVs and millions of users within five days. Then the question becomes: Can we handle that traffic? This gets at the depth Teacher Yuan was talking about.

Only experience with high-traffic projects teaches you how to do high concurrency, distributed systems, disaster recovery, multi-cluster load handling, and so on.

On the breadth side, the most important thing is having a global perspective. Many people's understanding of full-stack engineers isn't comprehensive enough. In the past, people thought a full-stack engineer was someone who writes their own APIs, writes their own frontend, handles the integration themselves, or uses Node.js to hack together an entire project. But I think that understanding is insufficient.

To truly extend your breadth, you must have a global perspective. This isn't just about being able to handle frontend and backend yourself. It also includes:

  1. Knowing how to design databases for better scalability.
  2. Understanding k8s cluster deployment.

For example, when I worked on system architecture at WeChat, I always considered deployment for every project. At first, for convenience, I might use cloud hosting or Vercel to launch quickly. But later I realized that while convenient, this approach is expensive, and the architecture absolutely wouldn't survive tens of millions of users hitting it. So I'd consider migrating to a self-built k8s cluster.

When designing code, I'd think about how to deploy quickly with Docker, how to achieve hosting-platform-like operation within a cluster. With a global perspective, you not only know how to build the application, but also how to deploy and run it better, and how to scale quickly and smoothly when traffic increases without needing downtime.

If the initial design is poor — say, all components depend on Vercel and don't run easily in Docker — then once the application grows, you'll spend a lot of time refactoring and migrating, potentially impacting user experience.

So for application-layer development, beyond depth, the most important aspect of breadth is having vision. This vision isn't achieved overnight; it's related to work experience and project history. If you've done frontend, backend, DevOps, and architecture, your understanding of these things will be more thorough.

Ronghui

Do you agree with the view that junior AI engineers will be replaced?

idoubi

When discussing engineer competitiveness, I think the challenge junior engineers face is limited vision and narrow capabilities. For example, only knowing how to call someone else's API, or only knowing how to apply off-the-shelf templates. While this approach can quickly validate ideas, if you want better development opportunities at a company, merely knowing these things isn't enough.

As an employer, if I'm hiring someone, only knowing how to apply templates, call APIs, or write simple pages isn't sufficient. Because future product forms may be more diverse, requiring consideration of scalability and handling larger traffic. A more qualified candidate shouldn't just know how to use these tools; more importantly, they should understand why things are done this way, how to do them more reasonably, or whether there are other ways to do them better and more completely.

I can't say people with these junior-level capabilities will be completely replaced, but they may gradually lose competitive advantage. If you're only confined to one domain, don't understand the principles behind the technology, and only know surface-level usage methods, it may be difficult to stand out in the future job market.

To improve your competitive advantage, I suggest:

  1. Learn more about available technical solutions and tools.
  2. Deeply understand the underlying principles of the technologies you use.

For example, if asked to build a system without using off-the-shelf solutions, you should at least know why you'd do it and how to implement it. People with such capabilities will have more advantages in the workplace.

Jinhui Yuan

Beyond pure technical refinement and expanding one's vision, I'd like to add an important angle: the importance of cross-domain learning or interdisciplinary fusion.

In the current software development environment, we're seeing more and more applications composed of both model-generated and human-written code. Using models directly can accomplish some general capabilities, but in terms of domain-specific knowledge and industry applications, current AI models still have many shortcomings.

So if an engineer has not only solid technical foundations but also deep understanding of a specific industry, they can stand out in competition. For example, I have expertise in engineering but also deep knowledge of the legal industry. This interdisciplinary knowledge enables me to better apply general AI model capabilities and software architecture engineering capabilities to the legal industry, thereby creating better solutions than others.

AI Trend Outlook

Ronghui

Based on your assessment of trends, what do you think will happen next?

Jinhui Yuan

In the AI industry, we can divide technology into two main parts: the application layer and the foundation. For a while now, massive investment resources have concentrated on the foundation — building compute clusters, training models, and so on. However, many analyses point out that this investment model is unsustainable if the application layer cannot flourish.

The future focus should shift to the application layer, paying attention to the value created and the role played by technology in practical applications. I also believe AI technology progress is real and substantial; it will certainly play unexpected roles in many application scenarios.

It's just that this process takes time and requires some preconditions. For example:

  1. Improvement in model capabilities
  2. Reduction in model usage costs
  3. Increased participation

We hope AI models can become like some past technology modules — tools that everyone can use, within reach. It should become technology that almost everyone can participate in, rather than a mysterious domain accessible only to a small group of AI engineers.

When millions or tens of millions of developers can participate, we may see unexpected innovations or chemical reactions.

Because the birth of many products has a certain element of chance, possibly related to someone's unique growth or development experience. Only when this person can access and use this technology can such a chemical reaction occur.

We need to improve infrastructure so more people have the capability and opportunity to participate, to create interesting things based on these technologies. When this happens, the explosion of applications and the emergence of so-called "super apps" will come naturally, I believe. I'm still very optimistic about this overall future development.

idoubi

I think that although some people may view AI as a bubble about to burst, AI may actually just be getting started. In the next 5 to 10 years, AI will likely remain a mainstream trend.

Many people like to compare AI's development to the mobile internet era, but I don't think this comparison is entirely appropriate. The mobile internet mainly created massive new opportunities because it satisfied many previously unmet needs, giving rise to many "super apps." But the AI era doesn't have that many entirely new demands; it's more about improving existing experiences.

I think AI's development is better analogized to the cloud computing era. Before cloud computing emerged, enterprises typically used independent server rooms, deploying Oracle or IBM systems to run their applications. After cloud computing appeared, "moving to the cloud" became a trend, with enterprises beginning to migrate their business to cloud platforms. This trend gradually evolved into the concept of "Cloud Native."

In the Cloud Native era, new projects were deployed and developed on the cloud from the very beginning, adopting microservices architecture. This approach changed traditional development paradigms. By analogy with AI, we may see more "AI Native" applications emerge. These applications use AI to build entire product workflows and development experiences from the ground up.

I believe AI applications will flourish in the future, especially these AI Native applications. All their functions are built on AI foundations, using AI to continuously improve application experiences. We may see many traditional applications reimagined by AI, with significantly enhanced experiences.

This trend may not fully materialize in 1 to 2 years, but in the next 5 to 10 years, we will likely continue developing in this direction.

Ronghui

Then what do you think will happen in the short term, in the next 1 to 2 years?

idoubi

In the short term, I think we're entering a period of explosive growth for AI tool applications. So-called tool applications include web UI generation, poster design, design element generation, and so on. The core of these applications lies in using AI technology to comprehensively improve efficiency across all aspects.

I personally believe this kind of AI-empowered tool application will become an important trend. I remember Kai-Fu Lee previously mentioned that AI application development might go through several stages, starting with tool applications, then entertainment applications, then e-commerce applications, and so on. While I may not remember the exact sequence accurately, this view of staged development is very enlightening.

Based on my personal experience, in the past year or two, AI tools have indeed been the fastest-growing domain.