DeepSeek balances its long-term AGI vision with practical execution.

DeepSeek's answer is not merely gazing at the stars or keeping your feet on the ground, but using real-world problems to support a sufficiently ambitious goal.

Article author and source: 36Kr

DeepSeek's answer is not merely gazing at the stars or keeping your feet on the ground, but using real-world problems to support a sufficiently ambitious goal.

Article by Li Jiaxing and Liu Chenyi

Edited by Zhang Wei

Header image source | DeepSeek

Cover source | DeepSeek

DeepSeek is once again in the spotlight.

This company was once one of the most enigmatic entities in the AI community: deeply rooted in an engineering culture, low-key, with founders who rarely appeared in public and a team that didn’t frequently engage in storytelling or publicity like typical internet companies. Yet over the past year, it has repeatedly become the center of attention in China’s tech circles. When it released its models, it went viral for its low cost and open-source approach; when news of its funding emerged, it sparked discussions over its valuation and investors; and now, with a nearly four-hour recording of Liang Wenheng’s investor meeting leaked, the outside world is once again reevaluating this company.

This time, the outside world sees a more complete, yet more contradictory, DeepSeek.

On one hand, there is a sufficiently distant goal. Liang Wenhong speaks of AGI, of vision, of “goodwill toward the world,” of not pursuing maximum profit, and of how open-sourcing is essential to organizing talent. He says DeepSeek does not want to become the next super app, the next ByteDance or Tencent; the company operates on vision rather than KPIs, researchers have ample time for free exploration, and overtime is uncommon. On the other hand, there is grounded realism. Liang Wenhong also discusses API pricing designed to break even in ten months, how fundraising mitigates team stability risks, the need to turn money into cards as quickly as possible, bottlenecks in domestic computing power capacity, how half of the core researchers are engaged in data labeling, and how the company must strengthen its organizational structure as it grows. He even says that, in the worst-case scenario, selling APIs alone could sustain a publicly listed company.

How can an AI company aim for the stars while staying grounded? This is precisely the answer DeepSeek seeks to provide.

Based on the leaked transcript of the "Liang Wenheng Investor Q&A Session," "Future Humanity Lab" has compiled the following summary.

The long-term goal is AGI, but rather than just shouting about the distant future, we break it down into consecutive steps.

Liang Wenfeng defines his long-term goal as AGI, but he doesn’t treat AGI as an abstract concept; instead, he breaks it down into a continuous roadmap: language models, chain-of-thought, agents, continuous learning, self-improvement, and embodied intelligence. The goal is ambitious, yet each step is grounded in current technical challenges.

This is our speculation: we believe the timeline should proceed as follows—first, learning to learn; then reaching the intelligent singularity capable of self-iteration; and finally, embodied intelligence. Once embodied intelligence is achieved, it will enter the real world, helping with household chores and elder care.

We think this is a fairly ideal roadmap, but everyone has different perspectives—there’s no right or wrong. We simply feel this roadmap is the easiest, because very little new action is required at each step.

The development of AI can be understood as a staircase. Last year’s step was CoT—Chain of Thought—because we discovered that using Chain of Thought enables intelligence to reach a higher level; by thinking through problems itself, AI can achieve a higher ceiling and accomplish more tasks.

It chases the agent but does not treat the agent as the end goal.

Currently, the hottest topics outside are Agent products and their commercial applications, but Liang Wenheng is more concerned about the bottleneck that comes after Agents: continuous learning. He participates in today’s technological wave, but he is not defined by today’s trends as the final outcome.

In this agent, the next bottleneck visible is continuous learning. The next problem to solve is how to achieve continuous learning—this is clear and relatively well-defined.

AI cannot replace your employees. But if AI, with its ability to learn continuously, could spend two months learning at your company just like your employees, then it could replace everyone.

Currently, investors see the most about agents; but for those of us who research, we see more about learning.

Think long-term, but proactively exclude technological directions outside the main roadmap.

Video generation, 3D, world models, multimodal learning, and embodied intelligence are all assigned different priorities. DeepSeek does not deny their value, but rather assesses whether they align with the current main trajectory of intelligence—recognizing the rich possibilities of the future while resisting the urge to be distracted by every trend.

When interacting externally, our stance is: we focus solely on the core path of AGI. That means we concentrate on GPT, CoT, Agent, and similar core technologies. The field of AI is broad, and there are many areas we believe lie outside this core path—such as 3D and video generation—which we feel have little connection to the fundamental pursuit of intelligence, and therefore we will not pursue them.

But commercially, it’s a good business—it’s a good business commercially. But that has nothing to do with intelligence. We won’t do it just because it’s a good business; we’ll only do it if it aligns with our intelligent roadmap.

Multimodal layout is something we've always been working on. It’s important for products, and especially important for consumer-facing products. But when it comes to the upper limits of intelligence, it’s just one component—not the core focus itself.

It allows high-uncertainty research to be like a lottery, but also acknowledges the immense amount of hard work behind technological breakthroughs.

Continuously learning through such problems is unsuitable for purely project-based or KPI-driven models, nor does it necessarily require large volumes of data. Liang Wenhong calls it “playing the lottery”: low barriers to entry, allowing many to engage, but with uncertain breakthroughs. On the other hand, he is also highly pragmatic about data labeling, post-training, hallucinations, and cost efficiency. This is not a research culture driven solely by inspiration, but one that embraces both free exploration and the tedious, labor-intensive work. It grants space for uncertainty in research while accepting the hard, unglamorous effort behind technological breakthroughs.

It’s not a project that requires allocating resources; rather, it’s something that needs many people to think about it. What sets us apart from other companies is that we take the time to discuss it, reflect on it, and treat it as something important.

So we call this “scratch and win.” The barrier to entry is very low—anyone can try their luck—but I don’t know whether what you win depends on talent or something else.

So right now, we’re essentially walking on two legs. It’s not that we can’t label data at all—it’s just that some labeling tasks have low costs while others are expensive, so we’re prioritizing the low-cost ones. You could also say that half of our team is currently engaged in data labeling, and half of our core researchers—the most important people—are also involved in labeling data.

There is a way to address the hallucination issue, but it’s a complex topic. Hallucinations can be seen as a solvable and improvable problem through better post-training; it’s just that not enough effort has been devoted to it. Or, in my view, while hallucinations are a problem, we may ultimately classify them as a product issue. We will address them, but they are not our top priority.

It believes in scaling and acknowledges that hashing power is the real bottleneck.

DeepSeek does not deny scaling, nor does it avoid acknowledging the gap between China and the U.S. Liang Wenhong states the gap plainly: it’s not about talent, but about resources. Therefore, low cost, domestic computing power, and underlying compilers are not mere industry slogans—they are integral parts of the technical strategy. Believe in the principles of large models, but use engineering efficiency to gain iteration space under resource constraints.

Scaling—we believe in scaling. The larger the scale, the better the results, and the more features we can unlock. What’s holding us back from scaling is actually computing power, not a lack of desire—we simply don’t have enough computing power to achieve this scaling.

We haven’t reached this upper limit. We trained such a large model not because I believe this size is sufficient, but because I happened to have that much resource. I determined the model size based on the resources I could accept and train with—not because this model size is enough.

It's not about eliminating organization, but about dynamic, long-term organization.

When exploring technical directions, Liang Wenfeng took an unexpected approach, saying that creating this roadmap didn’t require overtime—making it the easiest path to implement, which then became their technical route. This pursuit of ease and flexibility also extended to how he organized and managed his team: Liang Wenfeng said they had no formal organizational structure, no overtime, no KPIs, and team members spent at least half their time working on their own projects.

They also have half of their time unassigned, allowing them to do whatever they choose. This is a space for research, where they can explore whatever they personally find important, without any predefined requirements.

The founder's authority comes from consensus, not decree.

In Liang Wenheng’s view, it is the vision that brings people together. They are driven by this vision—“they have great goodwill toward the world and want to do something meaningful,” which is why they came. The first dozen or so people never considered making money: “If they had, they wouldn’t have come,” Liang said. His own authority is shaped by consensus; something moves forward only because it is agreed upon.

Let me tell you a story about our DDCP model. At first, we were worried about having too much demand, so we set the price relatively high—and not everyone on the team was happy about it. Later, I lowered the price to a quarter of the original amount, and suddenly everyone was thrilled. I believe this truly reflects our real intentions—the vision I mentioned earlier: we want this product to be useful to people, not just maximize our profits, but ensure it’s affordable for everyone while still earning a fair profit. During that price reduction, many people in the company cheered in our group chat—it was clear everyone was genuinely happy.

For example, if I want to do something, I first look at what our collective consensus is—whether everyone wants to do it. I might have some guidance or preferences, but their influence is limited; guidance is very limited and must always be built on consensus. This decision-making process is essentially a consensus-seeking mechanism. It’s not about me being able to push something through—it can only move forward if there’s consensus, and only then will I proceed.

It remains relaxed but not lax, with its own organizational mechanisms.

The team doesn’t just talk—they act. Liang Wenfeng says that vision isn’t a slogan on the wall, but how you act. He clearly understands what he’s doing and has a well-defined organizational logic: the structure, incentive system, and team goals are all thought through meticulously. Employees spend half their time on their own projects and the other half on company tasks. No overtime is indeed a benefit, but Liang Wenfeng hopes it creates a relaxed environment conducive to research. The team does fewer things because their focus is sharp—they are working on AGI. Looking ahead, as the team grows, Liang openly acknowledges that some departments will need to establish stricter hierarchical structures—changes are already underway.

The second issue is planning for the future organizational structure and team size. Previously, our structure was highly fragmented because we didn’t have a formal structure at all. However, as our team continues to grow, these aspects will certainly need to change. Right now, we can only say that many adjustments will be necessary, but it’s difficult to outline all of them at once. Ultimately, we’ll likely need to establish different departments, some of which will require a more formal, hierarchical structure, while others may continue to maintain a more relaxed, flat structure.

We also don’t usually work overtime. There are two reasons for this. First, research requires a relaxed environment. If you push too hard, you can’t do meaningful research—because true research stems from personal interest and the habit of thinking about these issues in your daily life, which only thrives in a relaxed setting. This is essential to our research culture. Second, we are highly focused. Being focused means we have very few priorities, so there simply isn’t enough work to require overtime.

A stable team is the foundation for realizing ideals.

Team stability is a core priority. In Liang Wenheng’s view, many things matter—money, cards, resources—but none of these pose the greatest risk; the greatest risk is the team falling apart. To prevent the team from collapsing, Liang Wenheng seeks dynamic organizational possibilities within an unstructured environment, using hierarchy, technology, and innovation mechanisms as tools to maintain team balance. Recently, DeepSeek completed part of its funding round, and Liang Wenheng said that with everyone receiving stock options, the risk of “the team falling apart” has been significantly reduced.

As long as I can maintain team stability, I will definitely succeed—I will definitely build AGI. It’s that simple. As long as everyone stays and we can keep going, I will succeed. There’s basically no risk; it’s just a matter of time before we encounter setbacks. If we face setbacks and no one leaves, I can simply continue.

From a team stability perspective, as long as the most important and longest-tenured employees remain, others are unlikely to leave. Even if they have fewer options or lower income, they still won’t leave, because they’re not solely motivated by money—everyone wants to be part of an environment where AGI can be achieved. Therefore, it remains attractive to talent.

Historically, our employee turnover has been relatively low, and compared to our peers, it has always been among the lowest. But it remains our greatest challenge—the only challenge, you could say.

Open source is not about abandoning barriers, but about shifting them from models to cost and engineering efficiency.

When Liang Wenfeng talks about open source, he doesn’t just speak idealistically or purely from a marketing perspective. His view is that AI is large enough that no single company can monopolize it; open source helps the ecosystem adopt the technology without inherently harming business models, because true barriers lie not just in the models themselves, but in deployment costs, engineering efficiency, and organizational capabilities.

Why are we so committed to open source? Because this vision inherently requires it. Without this vision, you cannot bring people together.

Regarding open source, we had a very clear vision from the beginning. First, our vision; second, we believed that open source is beneficial for successfully commercializing AI.

But AI is big enough that, in the end, it might account for as much as ten percent of human society’s GDP, for example—so it’s an enormous number. No one person can monopolize this; you must share it with others, or you simply won’t survive.

If we try to monopolize this benefit, we will surely be abandoned by history. I believe this is primarily an objective law, a historical perspective.

Even if the model is open-sourced and you reveal everything to others, the barrier to entry is still very high. It’s also very difficult for others to actually use it. Not only is it hard to get started, but it’s also extremely challenging—and far from easy—to do so at a low cost.

Just because I open-sourced it doesn’t mean he can easily achieve the same deployment cost as me—there’s still a lot of work involved. Although the underlying principles are well understood, not every company is willing, or has the desire and capability, to organize the manpower needed to reach this goal.

Open source, I believe, has no impact on our business model—as long as we only earn six times the profit and recover our costs within ten months, which corresponds to roughly six times the profit. Under these conditions, open source won’t have any effect.

I’m not worried at all about others deploying our model and competing with us—we actually hope they do.

No enemies, but not without business defenses

DeepSeek does not wish to become a competitor to large or small companies; instead, it is even willing to assist its rivals. This stance stems partly from its vision and partly from reducing external obstacles, allowing the company to remain focused on its core mission. Meanwhile, Liang Wenheng believes that competition in large models is ultimately not mysticism—it’s about cost, time, and experience. Those who better control costs earn more; those who struggle with cost control earn less.

So, much of what we’re doing now is aimed at maintaining team stability. Beyond that, I believe everything else can be forgone or restrained. We’ve always been very restrained, unwilling to become competitors with any large or small internet company.

Under this premise, we are very willing to assist and support anyone—even our competitors, including Alibaba, Zhipu, and Moonshot AI—to do better, because we don’t lose anything; after all, we are open source.

My business interests have not been affected by our goodwill—not at all; in fact, it may have even been beneficial.

The only differences are two things: time and cost. So no one should make excessive profits—I don’t believe excessive profits are possible. Those who manage costs well earn a bit more; those who manage costs poorly earn a bit less—that’s all there is to it.

Don't try to control the entire industry chain—just hold the most critical position.

Liang Wenfeng has repeatedly expressed a clear sense of boundaries: DeepSeek does not aim to do everything, does not seek vertical integration into upstream chip development, and does not intend to capture all opportunities in the application layer. It believes the AI era is large enough to support many companies, and DeepSeek only needs to be one of them—focusing solely on what it does best and what lies at its core.

Another question is whether we will vertically integrate upstream. I hope we won’t. So, the question is: will we vertically integrate into upstream applications? We hope not—we’d prefer others to handle that. I don’t want to take on everything; I just want to take one piece—the part we’re best at, or the part we believe is most core to us.

Will we build our own large-scale clusters in the future? I believe building our own large-scale clusters is definitely necessary—we’ve been doing this all along, and all of our clusters are self-built. However, whether we should develop our own chips in the future depends on how significant the returns will be, and how substantial those returns will be.

We hope to focus on just one area. I believe AI is a huge field and doesn’t require me to limit myself to just one piece. If we concentrate here, I think the business opportunities are already substantial enough. If the AI era gives rise to many companies worth trillions, I believe we will be one of them.

Don’t chase the current opportunity, but don’t refuse to catch it if it comes your way.

Should DeepSeek be commercialized? Yes, but commercialization is no longer the ultimate goal. DeepSeek’s objective is to build AGI; along the way, Liang Wenheng recognized opportunities for commercialization and did not reject them—he seized them decisively. “If it’s something we can easily pick up, we’ll take it,” Liang said. He compared this to the distinction between “picking up sesame seeds” and “grabbing watermelons”:

I think this approach is commercially viable and possible. Last year, we could have spent a large amount of money to compete directly with ByteDance for users—that would have been another strategy. But we chose a much more restrained approach: I’m not going to fight you for this, because there are watermelons ahead, and what’s in front of us might just be sesame seeds.

Where there’s a big business opportunity, you’ll always find a way. As for the smaller opportunities in front of us, we’ll pick them up casually—we don’t stop what we’re doing or treat them as major priorities. So last year’s daily active users on the C-side? I think it was just a minor thing. But we still picked it up, maintaining user engagement at a relatively low cost, because it might prove useful in the future.

Commercialization has always existed, but it is not the primary goal.

In Liang Wenheng’s view, achieving ToC and ToB goals were less prioritized than building AGI. Yet, the more one strives for something, the less it seems to come—DeepSeek internally focused intensely on breaking through its technical roadmap, yet unintentionally captured massive C-end traffic and B-end revenue. Looking back on this journey from last year to this year, he concluded that this was a strategic advantage brought about by dimensional reduction.

There’s a strategic advantage here: we’re thinking about and building AGI. When we then develop applications, whether for consumers or businesses, we don’t need to invest much thought or effort—it takes very little to get it right. I believe that when you operate from a higher technological vantage point to build lower-level technologies, it creates a kind of dimensional advantage. At least in last year’s consumer-facing products, we saw this clearly: we didn’t put much effort into the C-end at all; at one point, we even considered abandoning those users—but they simply wouldn’t leave.

Low prices are not charity—they are reasonable profit.

On pricing, Liang Wenfeng is unequivocal: his goal is to recoup equipment costs within ten months through the API. The current price is neither set to maximize profit nor to be the lowest possible. Liang Wenfeng says prices cannot be lowered indefinitely—without sustainable pricing, the company cannot survive, and ultra-low prices do not generate additional social value. A reasonable price ensures long-term advantages within a healthy profit margin. He believes the AI market is large enough that achieving success with a fair return is sufficient; if one tries to extract excessive profits from the start, they will only invite more competitors.

If my goal is to capture 5% of AI and the entire world’s GDP, then, theoretically, this calculation still holds. Look at OpenAI—they seem to be able to make this calculation work; it’s theoretically sound. But here’s the problem: they could be defeated by someone else who is willing to take only 1%. Because that person says, “I’m doing this incredibly well, but I only need 1% of global GDP,” and that would defeat them. Then, if another person emerges saying, “I only need 0.1%,” they would defeat the previous one.

From a macro perspective, it doesn’t matter where these percentages come from—they’re all the same. Those who take more will be outperformed by those who take less. Even worse, you don’t even need to actually take more; if your vision is to take more, you’ll be defeated by those whose vision is to take less. In reality, no one truly gets paid—it’s all just a vision. If your vision is to take more, you’ve already lost, and you’ll face even greater challenges. That’s just how the world works.

What does your ideal AGI look like?

Feel free to share your thoughts in the comments~