Liang Wenfeng’s Four-Hour Investor Meeting Transcript

Original text ⬇️

Last month, “elsewhere” reported DeepSeek’s fundraising story, and what most people discussed was that rumored four-hour investor meeting.
During this month, Liang Wenfeng’s various quotes have spread through the crypto world; we collected and organized 52 of them from multiple sources.

One main thread: AGI. High-frequency words: models, cost, time, open source.
He said “no” many times: not a genius, not earning unreasonable profits, not pursuing user growth, not keeping it closed-source, not doing 3D/video generation/world models, not building the next super app.
Restraint is a kind of strategy.

DeepSeek has only one main thread.
Right now, building products isn’t about maximizing returns. Products are a byproduct on the road to AGI; there’s no need to spend too much effort on B2C/B2B products.
3D and video generation aren’t on the main line; world models aren’t closely related to intelligence upper limits.
Multimodality is important, but it’s just a component, not the main thread.
Hallucinations are a long-term proposition; internally we attribute them to product issues. They will be solved, but they aren’t the focus.

At this stage, the most important thing is Coding Agent. The most reasonable approach domestically is to focus on building general-purpose Agents; vertical Agents like finance and healthcare have lower priority.
If the AI era brings many companies at the trillion scale, it would be great for DeepSeek to be one of them.

First achieve continuous learning, then AI self-iteration; the endpoint is embodied intelligence.
AI isn’t lacking in taste or intuition; what it lacks is the ability for continuous learning.
Humans can continuously learn, but AI cannot. So the next generation of models must have continuous learning capabilities to truly be called the next generation.
For the models we build, our first goal isn’t for everyone to use them well—it’s for us to use them well. This is the fastest way to achieve AGI.

The whole world hasn’t found a good method yet, because “learning” is made up of many things.
DeepSeek’s long-term vision is AGI. Last year’s step was CoT; this year is Agent. After that, what needs to be solved is continuous learning.
After achieving continuous learning, we may reach a gradual singularity: AI can accelerate AI research. Only after completing this step do we reach embodied intelligence.

The endpoint of intelligence may be embodied intelligence, because human needs aren’t computers, but human labor.
It’s still very far from a complete shift to commercialization.
We only earn reasonable profits, not maximize profit pricing.
When a model is cut to one quarter of its price, many people in our company chat cheer—that’s exactly the purpose of making our models well: to let everyone fully use them.

Low cost is the result; the architecture keeps moving toward lower cost. When compute resources are limited, high efficiency is what allows us to train larger models.
We’ve done all this pretty easily. Price cuts aren’t good news for competitors. Selling APIs isn’t that attractive; a few people maintaining it is enough—there may not even need to be customer support or sales.
We’ve always been commercializing, just not with commercialization as the goal. The time when we fully switch to commercialization should be very far away.

As long as the business opportunity is big enough, there’s always a way. DeepSeek is a product of the era, not the result of imitation.
Open source is our sweet spot as a company of our size.
Restraint is a strategy: give up some things to get more of other things. Open source is about sharing benefits—employees get a sense of achievement, and society benefits as well.

This AI thing is big enough that it may ultimately account for 10% of human GDP. If you try to monopolize the利益, history will abandon you.
The models we open-source are the same as the ones we deploy ourselves—we won’t hold anything back.

We’re not worried about others using our models to compete. A startup is too small to have the power; large companies are hard to organize—this is our sweet spot at our scale.
Open source doesn’t affect the business model, as long as we only earn reasonable profits. If you want to make 100x profits, open sourcing will indeed have an impact.

We don’t want to become rivals with any big companies or small ones; we’re willing to help Alibaba, Zhipu, and Moonshot AI do better.
The gap between China and the US isn’t in talent.

In the future, we’ll use only a fraction of the compute resources to shrink the gap to 6 months, or 3 months.
The China-US gap is mainly in resources. We believe in Scaling, but resources are limited.
There’s almost no gap in talent—it's the same group of people. Talent shortages are just phased.

In competition among model factories, cost comes first.
Anthropic surpassing OpenAI isn’t a long-term thing—it’s a phase.
There are too many companies building models domestically; resources are scattered. In the end, they will definitely converge. If each company only takes reasonable profits, two big companies and two small companies are enough.
I absolutely don’t believe that large model companies can take away most of the profits.

Competition ultimately comes down to three things: cost first, time second, and user experience has stickiness but not fundamentally.
Not interested in becoming the next super app.
We don’t want to be the next ByteDance, or the next Tencent.
We’re not fighting for that, because there are watermelon slices ahead, and sesame seeds behind.
Last year we抢 C-end traffic; this year we抢 B-end revenue. We don’t think that’s important. The thing we want most can’t be obtained—what we don’t care about turns out easier to get.
Last year’s Spring Festival popularity wasn’t in our script.

Keeping team stability is core.
There’s only one thing we can’t make concessions on: we must maintain team stability. That is our biggest risk, and this round of financing has provided a fairly large easing.
We don’t want to create enemies, so our environment will also be better.

Organization is called “doing the right things” from top to bottom; from bottom to top, employees explore on their own—both are right, and each takes half the time.
We don’t really work overtime. Research needs a relaxed environment, and we’re also very focused—if the product isn’t perfect, we don’t补 it. This is also a kind of restraint.
The organization is dynamic; in the future there may be a need for structural adjustments, but we hope the vision-driven direction won’t change.

With goodwill toward the world.
When we started, the first few dozen people never thought about making money or going public; we did this with goodwill toward the world.
We’re a vision-driven organization, not one that relies on KPI.
Vision isn’t something written down; it lives in the methods and attitudes we use when doing things.

Twenty years ago, I admired Jack Welch; now I see that most of his views are already wrong, but one thing is right: the most important thing for a company is vision. Vision isn’t slogans—it’s how you do things.
Restraint allows us to be more likely to achieve AGI.

AGI has the highest payoff. For other things, if we have the energy, we do them; if we don’t, we don’t.
This AI thing is too big and the利益 is too big—if you just split a little, it’s still huge. The more restrained you are, the more likely you are to achieve success.

Besides vision, we don’t have many other advantages—this matches my intuition.
When we were founded two years ago, we had no money, no cards, and no reputation—just a group of ordinary people. The story I like is “a group of ordinary people doing extraordinary things,” not “a group of geniuses.”

Pricing isn’t based on maximizing revenue or profits. In the short term, higher prices mean more revenue, but in the long term it’s not necessarily so. Restraint is a strategy.
Open source and low pricing give employees a sense of achievement and benefit society—this also increases, from a long-term perspective, the probability that we achieve AGI.

If your vision is to take as much as possible, then you’re already the one who loses first. That’s just how this world works.