Last month, Elsewhere reported on DeepSeek’s fundraising story. Among all the discussions it sparked, none attracted more attention than the now-legendary four-hour investor meeting with DeepSeek founder Liang Wenfeng.
Over the past month, fragments of Liang’s comments have circulated widely. We have collected and cross-checked a number of them from various sources.
Throughout the conversation, Liang said “no” remarkably often: no to the notion of genius, no to maximizing profits, no to chasing user growth, no to closed-source models, no to 3D generation, video generation, world models, and no to building the next super app.
According to Liang, restraint is a strategy—one that sacrifices certain opportunities in exchange for a higher probability of achieving Artificial General Intelligence (AGI).
Among the limited materials available to us, several themes appeared repeatedly: models, cost, AGI, time, open source, and learning.
Most of the time, Liang spoke cautiously and in plain language. Only when discussing a handful of topics he truly cared about did his tone become noticeably sharper:
“As long as I can maintain the stability of the team, I will achieve AGI. It’s that simple.”
Below are 52 remarks collected from various sources. Some wording may differ slightly from the original while preserving the intended meaning.
DeepSeek Has Only One Main Mission
1. This is not the time to optimize for product revenue.
Products are a necessary step on the road to AGI, but we don’t need to devote excessive effort to either consumer-facing or enterprise-facing products. When you’re operating from a higher technological level, building lower-level technologies becomes a form of dimensional advantage. Products are ultimately byproducts of the pursuit of AGI.
2. Many things are not on our main path, such as 3D generation and video generation. The same applies to world models—they are not strongly related to the upper limits of intelligence itself.
3. Multimodality is extremely important for products and for end users, but it is only a component. It is not the core path, nor is it intelligence itself.
4. There are certainly ways to address hallucinations in large language models, but it is a long-term challenge. Internally, we classify hallucinations primarily as a product problem. We will solve it, but it is not our highest priority.
5. At this stage, Coding Agents are the most important area.
Given China’s market conditions, the most rational approach is to focus entirely on building general-purpose agents. Specialized agents for finance, healthcare, and other industries are lower priority.
6. If the AI era produces many trillion-dollar companies, it would be great if DeepSeek is simply one of them.
First Continuous Learning, Then AI Self-Improvement, Eventually Embodied Intelligence
7. AI does not currently lack taste or intuition. What it lacks is the ability to learn continuously.
8. Humans can continuously learn, but for AI, every task requires providing the entire context again. That’s almost impossible.
As a result, AI still cannot fully replace employees. The next generation of models must possess continuous learning capabilities; otherwise, they are not truly next-generation models.
9. We hope future models can help develop our own technology.
To put it simply, the primary goal of the models we build is not that everyone else finds them useful, but that we ourselves find them useful. That is the fastest route to AGI.
10. At present, nobody in the world has found a truly good solution, because learning itself consists of many interconnected components.
11. DeepSeek’s long-term vision is AGI.
If achieving AGI is like climbing a staircase, last year’s step was Chain-of-Thought (CoT) reasoning. This year’s step is agents. The challenge after agents is continuous learning.
12. Once continuous learning is achieved, we may reach a gradual singularity: models will be able to perform everything humans can do, including developing more advanced AI systems themselves.
At that point, AI research will be accelerated by AI. Only after crossing that threshold does embodied intelligence become possible.
13. The endpoint of intelligence may ultimately be embodiment.
For most people, what they actually need is not a computer—they need labor.
Commercialization Is Still a Long Way Off
14. We only seek reasonable profits. We do not price for maximum profit.
15. When one of our models launched, we initially priced it relatively high because we worried demand would be overwhelming. Later, we cut the price to one-quarter of the original level, and many employees celebrated in the company chat.
That was the purpose of building the model so carefully in the first place: so that as many people as possible could fully use it.
16. Low cost is a result, not the goal itself.
Our architecture has always been designed to drive costs down. We want AI to remain affordable, especially in a world where compute resources are scarce.
There is another reason: lower costs enable larger models. When compute is limited, higher computational efficiency allows you to train bigger models. Large companies can simply add more resources; we prioritize cost efficiency.
17. From the outside, it may seem like we chose a difficult path. In reality, it’s quite easy for us.
Lowering prices is certainly bad news for our competitors—they are not celebrating.
Frankly, selling APIs is not that attractive. I only need a few people to maintain the API. We don’t even have customer support, and we don’t need a sales team. Users come on their own.
18. We have always been commercializing, but commercialization has never been our objective.
The point at which DeepSeek fully shifts toward a commercial-first strategy remains very far away.
19. I don’t even need to worry about securing a position in that future market.
If the commercial opportunity is truly that large, there will always be ways to participate.
DeepSeek is a product of its era—a response to reality, not an imitation of someone else’s model.
Open Source Is the Sweet Spot for a Company of Our Size
20. Restraint is a strategy: giving up certain things in exchange for gaining others.
Open source is one example. Internally, it gives employees a sense of accomplishment and strengthens cohesion. Externally, it benefits society and makes both peers and ordinary users happy.
I have no doubt that AGI will create enormous commercial value. Given that, my first priority is not maximizing our share of the value but increasing the probability of success.
21. Open source is actually beneficial if you want AI to succeed commercially.
That may sound counterintuitive.
Historically, software markets were only worth billions of dollars annually. Open-sourcing software often meant giving away the market.
AI is different. It may ultimately account for ten percent of global GDP.
If we attempt to monopolize that value, history will leave us behind. That is an objective reality and a historical pattern.
22. The models we release as open source are the same models we deploy ourselves.
We will not release an inferior model while secretly running a better one internally.
23. I am not worried about others deploying our models and competing with us.
Not every company has the desire or capability to pursue this goal.
Startups are often too small to undertake such efforts, while large corporations struggle with organizational complexity.
This happens to be the sweet spot for a company of our size.
24. Open source has no impact on our business model, provided that we are content with earning reasonable profits.
If you are aiming for one-hundred-times returns, then yes, open source becomes a problem.
25. We do not want to become an adversary of any internet company, large or small.
Given that premise, we are happy to help anyone—including Alibaba, Zhipu AI, or Moonshot AI—become better.
The Gap Is Not Talent
26. In the future, we want to rewrite the narrative of AI.
With only a fraction of the compute resources, we want to reduce the gap from years to six months, or even three months.
27. We believe in scaling.
Larger models generally produce better results.
The reason we train models of this size is not because we think they are large enough, but because these are the resources available to us.
28. There is almost no difference in talent.
In many cases, it is literally the same group of people.
China does not lack talent. Talent shortages are temporary phenomena. Historically, no category of talent remains scarce forever.
In Foundation Models, Cost Comes First
29. Anthropic surpassing OpenAI is not a permanent development—it is a temporary phase.
Over the long run, OpenAI and Google will likely alternate in leadership.
30. There are too many model companies in China doing the same thing, resulting in fragmented resources.
Eventually, the market will consolidate, but that process takes time.
If every company were satisfied with reasonable profits, the industry would not need so many organizations building foundation models. Two large companies and two smaller ones might be enough.
31. I absolutely do not believe foundation-model companies will capture most of the profits in the AI industry.
32. Ultimately, competition among model providers comes down to three things:
- Cost
- Time
- User experience
Cost is the most important.
Given the same quality of service, what does it cost you to provide it?
Time comes second. Being several months earlier or later can make a huge difference.
User experience creates some stickiness and barriers, but it is not fundamental.
No Interest in Becoming the Next Super App
33. We have no desire to become the next super app.
The next ByteDance? The next Tencent?
We have absolutely no interest in that.
34. We are not competing for those opportunities because there are much bigger opportunities ahead.
The opportunities immediately in front of us may seem large, but compared to what’s beyond them, they are relatively small.
35. Last year everyone was competing for chatbot users and consumer traffic.
This year everyone is competing for enterprise revenue.
We don’t think either is particularly important.
What truly matters internally is the AGI roadmap and the next technical breakthrough.
Interestingly, the things you want most are often the hardest to obtain, while the things you care less about tend to arrive on their own.
36. The popularity we experienced during last year’s Lunar New Year was never part of our plan.
Team Stability Is the Core Priority
37. There is only one thing that cannot be compromised: maintaining team stability.
It is one of our greatest risks.
Fortunately, the recent financing has significantly reduced that risk.
38. Many of our decisions are made with team stability in mind.
We do not want to become rivals of internet companies. We prefer to empower and help them.
We do not want enemies.
A better environment for them ultimately creates a better environment for us.
39. Some people think our organization is top-down. Others think it is bottom-up.
I believe both are true.
The top-down portion is what we call “real work,” and ideally it should occupy less than half of an employee’s time.
The remaining time is bottom-up and self-directed. People can explore whatever they find interesting or important without predefined requirements.
40. We generally do not work excessive hours.
One reason is that research requires a relaxed environment.
The other is our extreme focus.
Many of our products remain unfinished, yet we choose not to address every gap. That restraint is part of our culture.
41. Organizations are dynamic, not static.
As the company grows, adjustments will be necessary.
We are unlikely to become a traditional hierarchy, although some additional structure may emerge.
What will not change is being driven by vision.
Acting with Goodwill Toward the World
42. When we started this company, we were not thinking about making huge amounts of money or going public.
The first few dozen employees never thought that way either. If they had, they probably would not have joined.
We built this company with genuine goodwill toward the world because we believed the work would be useful to humanity.
43. “We must hit KPI X” is not how we operate.
We are a vision-driven organization.
That comes with both strengths and weaknesses.
Over time, we will learn how to maximize the strengths and minimize the weaknesses, but this is part of our identity.
44. The vision is not even formally written down.
It exists in how we work and how we view the world.
Different people within the company may interpret it differently, but we are aligned on the broader direction.
45. About twenty years ago, the management figure I admired most was Jack Welch.
Looking back, many of his ideas may no longer hold true.
But he was right about one thing:
The most important thing in a company is its vision.
A vision is not a slogan on the wall. It is not what you say—it is what you do.
Restraint Increases the Probability of Achieving AGI
46. AGI offers the greatest possible return.
Everything else is secondary.
If we have the capacity, we may do other things. If we don’t, we won’t.
Restraint is part of our vision.
47. AI is simply too large an opportunity.
The value at stake is enormous.
If AGI succeeds, even a small share of the benefits would be tremendous.
The more restrained you are, the more likely you are to achieve it.
48. This feels intuitive to me.
Beyond our vision, we do not possess many extraordinary advantages.
49. When we founded the company two years ago, we had little money, limited compute resources, almost no recognition, and very little influence.
We were simply a group of ordinary people.
The story I prefer is not “a group of geniuses accomplished something extraordinary.”
It is “a group of ordinary people accomplished something extraordinary.”
50. Open source is also an expression of restraint.
Our pricing decisions are certainly not based on maximizing revenue or profit.
In the short term, higher prices might generate more revenue.
Over the long term, however, the answer is far less obvious.
For me, restraint is a strategy.
51. Open source and low pricing give employees a strong sense of accomplishment.
That creates organizational cohesion.
It benefits society, and it makes both peers and ordinary users happy.
From a long-term perspective, this form of restraint increases our chances of achieving AGI.
52. If your vision is centered on taking more for yourself, you’ve already lost.
You will likely face even greater difficulties.
That’s simply how the world works.
Cover image source: Rembrandt van Rijn, “Self-Portrait”, 1633, Isabella Stewart Gardner Museum.
