DeepSeek boss Liang Wenfeng’s internal speech transcript article addressed to investors has been going viral online all day, and in the afternoon it was gradually taken down on domestic public platforms.

I saw it in a friend group’s document, but I still can’t help but read it with great interest!

In the 42-page text transcript, he discussed AGI, open source, commercialization, compute, and domestic chips—he also talked about organizational management and talent.

There’s a lot of information, but I think what he keeps emphasizing is really only one thing: how to improve the probability of ultimately turning AGI into reality.

For example, what left a strong impression on me is that DeepSeek didn’t choose to launch everywhere and attack constantly—it kept doing subtraction.

No rush to grab users, no chasing the next super App;

For the sake of neither revenue nor anything else, they don’t try to eat everything on the consumer side and business side;

Even though video generation, 3D, and world models are hot, they don’t jump on that trend;

The model could be closed-source to earn more, yet they still choose to open source;

The pricing could clearly be set higher, but they only earn the profit they consider reasonable.

This isn’t because he doesn’t want to make money.

On the contrary, it’s because he knows which money can be earned and which money earned now will, in fact, slow him down.

1️⃣ Restraint is a strategy!

There’s a saying that left a deep impression on me: restraint is a strategy.

Many people understand business competition as grabbing users, grabbing distribution channels, grabbing revenue—anything that can be obtained, you take first.

But Liang Wenfeng’s understanding is completely the opposite.

This AI thing is big enough that no single company can truly monopolize it. If the coming “watermelon” is big enough, then there’s no need to stop for the sake of the small “sesame” in front of you.

Consumer-side users can be pursued, and business-side revenue can be earned too, but all of these are just byproducts on the road to AGI. They shouldn’t become the company’s goal in reverse.

This is actually the most philosophical part of the entire exchange:

The more you want, the harder it may be to get it in the end; the more you’re willing to take less proactively, the more easily you can reach the end.

Open source follows the same logic.

Many people think open source means giving up your moat, but Liang Wenfeng believes that as long as DeepSeek maintains advantages in cost and efficiency, open source won’t harm business—on the contrary, it can attract talent, build an ecosystem, and reduce enemies.

He even judged that if a company wants to take too much of the benefits in the AI era, in the end it will definitely be beaten by another person who is willing to take less.

Take too little and the company can’t survive;

Take too much, and you’ll be challenged by new competitors.

In the end, what can exist long-term must be the one that earns “reasonable profit.”

So his restraint isn’t “peaceful” in a soft sense, and it’s not a lack of desire.

It’s a very strong sense of purpose:

You can take less profit, skip hot businesses, and not fight to grab short-term users—but the things you absolutely cannot lose are only two: the AGI main line, and the stability of the core team.

That’s where restraint becomes truly sharp.

He doesn’t have no desire—he just concentrates all desire into one thing.

2️⃣ Technical route

Technically, Liang Wenfeng’s route was also very clear:

After language models comes the chain of thought; after the chain of thought comes Agent; and after Agent, what truly needs solving is “continuous learning.”

Now AI capabilities are already strong, but it still needs humans to provide full context. It can’t, like a new employee, enter a company, learn for two months, understand the people here, relationships, rules, and work habits, and then grow continuously.

Once AI has the ability for continuous learning, it can start helping humans research the next generation of AI, forming a loop of “AI accelerating AI.” Only after that comes self-iteration and embodied intelligence.

Of course, he also admitted very honestly:

Continuous learning has not truly been solved yet; the whole world is still exploring.

That kind of candor is also hard to come by.

On one hand, believing that AGI will definitely happen; on the other, admitting that the specific path is still full of unknowns. Not for fundraising—he didn’t force investors into a fixed timeline as if everything were certain.

3️⃣ The China–US gap

Regarding the AI gap between China and the US, his judgment was also direct:

The biggest difference isn’t talent—it’s compute and resources.

As long as there remains an order-of-magnitude gap in compute, China can’t comprehensively surpass; it can only catch up in some directions by higher efficiency, lower cost, and more clearly defined trade-offs.

He also believes that the real gap in future large models won’t be some mysterious technology that can never be replicated. It will be three things:

Cost, time, and user experience.

Who can build it earlier, who can provide it at lower cost, and whose product makes people use it more comfortably—whoever does those things better can stay.

4️⃣ Long-termism

My biggest takeaway from this conversation isn’t which AI company we should buy, but a renewed understanding of the four words “long-termism.”

Real long-termism isn’t something you just claim you see far.

It’s whether, when short-term interests are truly laid in front of you, you have the ability to refuse.

Investment is the same.

Every hot topic wants in, every market move wants to make money from, every rise is afraid of missing out—until in the end your capital, attention, and judgment are all scattered.

A person’s real ability circle may not only be knowing what you’re good at; you also need to know:

Which money doesn’t originally belong to you.

But Liang Wenfeng’s whole model also can’t be simply romanticized.

DeepSeek dares to practice restraint because it has technical efficiency, a team, funding, and it believes the future market will be big enough. Opening source, not setting KPIs, and letting employees explore freely are not things that any company can copy and succeed just by doing the same.

Liang Wenfeng also acknowledges that as the company grows in scale, some departments still need clearer organizational structure; DeepSeek also must rely on APIs and commercial revenue to survive.

So what’s truly worth learning isn’t the superficial “not making money,” “not working overtime,” and “not setting KPIs.”

What’s worth learning is first thinking clearly:

What are the things you absolutely cannot lose, and what are the things you can proactively give up.

For DeepSeek, what they cannot lose is AGI and the core team.

For ordinary people, it might be health, family, cash flow, judgment, and the direction you truly want to commit to long-term.

Today everyone is discussing how to have more, but Liang Wenfeng spent nearly four hours explaining why he can do without so much.

I think this may be the most powerful part of the entire exchange:

It’s not that you’re great because you seize every opportunity—it's that, faced with countless opportunities, you still know where you truly want to go.

Many people think the hardest part of long-termism is persistence.

Actually, the harder part might always be giving up.