DeepSeek's Liang Wenfeng on AGI, Restraint, and Why the Company Finally Took Capital

I'm not sure if everyone is aware of the previous DeepSeek financing matter. Let me briefly explain the background:

Previously, DeepSeek had always adhered to the principles of no financing, no IPO, and no commercialization.

But later, perhaps driven by various reasons, DeepSeek ultimately decided to formally enter the capital market.

So in mid-May, they organized this 4-hour online investor exchange meeting. At that time, each institution had two participant slots.

The overall process was Liang Wenfeng speaking first, then investors asking questions, and Liang answering.

Because it seemed to be a closed-door meeting at the time, everyone only had some snippets of golden quotes circulating. Recently, Elsewhere and Tencent Technology compiled a transcript of Liang Wenfeng's remarks at the meeting. Those interested can look for the original text. (I recommend Tencent Technology's version, which is more comprehensive and detailed.)

Here is the address: https://baijiahao.baidu.com/s?id=1871476713948433091&wfr=spider&for=pc

The entire speech was extremely informative. After reading it, I couldn't help but admire their vision and way of doing things.

I also discovered from some small details what might be the reason DeepSeek chose to enter the capital market. Of course, it might not be correct; it's just my speculation.

The article from Tencent Technology divided the entire speech into 11 topics. From these, one might be able to glimpse the path DeepSeek is walking.

I will also offer my humble interpretation. I hope you will forgive any immaturity or misinterpretation.

Vision and Restraint

DeepSeek didn't start out wanting to make a lot of money, but rather wanted to do something beneficial to humanity. (Although I believe a company inherently has a commercial nature.)

With this vision and original intention, DeepSeek grew from a dozen people to where it is now.

It relies on a common goal, not rigid KPI requirements.

Moreover, DeepSeek has maintained a high degree of restraint in its work, only doing things on the main line. The main line DeepSeek understands is AGI.

Their long-term goal is AGI, and they will spare no effort in anything related to AGI.

Restraint is one of the core reasons DeepSeek can firmly move forward. Resources are limited, interests are infinite. Sometimes restraining oneself and doing less can actually increase the probability of success.

It is precisely because of this restraint that DeepSeek is not seen as a huge threat in the eyes of peers, but rather a friendly partner.

This partner only does things on its own path and is willing to use its own strength to help others, without encroaching on others' shares.

Although it may seem like they are taking less, in fact, they get what they originally wanted, without incurring hostility, and even gaining favor.

AGI Roadmap

DeepSeek understands the realization of AGI as a gradual process, a step-by-step process that cannot be achieved overnight.

Last year's step was chain of thought, this year's step is Agent. Through the iteration of continuous learning capabilities, a singularity might be reached next. (Here, I understand it might mean a 'turning point where quantitative change leads to qualitative change,' which could be a process, not an instant.)

This singularity is when the Agent can achieve self-iteration -> continuous learning -> self-iteration, at which point it can enter the next step: embodied intelligence.

DeepSeek also consistently adheres to the principle of restraint on this path, only doing the main line of AGI.

It does not work on things on the auxiliary routes of AI, such as image generation, video generation, 3D generation, etc. These things might each represent a huge business, but DeepSeek chooses not to do them given the current resources and stage.

Maintain a high degree of focus, concentrate on the main line of AGI, and strive to push AI into the next step, general intelligence or embodied intelligence.

The main problem to solve at this stage is continuous learning. Currently, it still relies on training to achieve a stronger model. The goal is self-continuous learning and self-iteration.

Being able to develop its own next version, a stronger model, until entering the embodied intelligence stage.

Team and Talent

Everyone gathers not simply for profit, nor is it simply bringing smart people together.

To get everyone to cooperate and move towards a goal, the core is having one's own vision.

If everyone adheres to a vision and a goal, the team will be stable enough, and AGI can be achieved. (Liang's original words here were: As long as I can maintain the stability of the team, I will definitely succeed, definitely achieve AGI, it's that simple. I personally feel it's very confident and firm.)

At this current stage, money is not the problem, resources are not the problem. The stability of the team is actually the company's core interest.

Liang's original words: This is also a very big challenge we face, or I should say, I think it's the biggest risk. Of course, this risk has been significantly mitigated with our recent financing. Because the options everyone received are quite substantial, the amounts are quite large.

So, I understand that one of the purposes of DeepSeek's financing this time is to maintain team stability. Although everyone came for the same goal,

bread is also necessary. This financing allows everyone to get options, which is part of DeepSeek sharing the benefits.

It's not about using money to bind everyone, but using money so everyone isn't so anxious.

After all, the competition for AI talent is very intense right now. With the continuous escalation of poaching investments from various companies, there might be some instability within the team.

Letting everyone eat their fill and then re-clarifying the goal might not be a bad approach.

Because not everyone in the team is doing this solely for money. They themselves have ideals and want to accomplish this. After solving basic needs, they can better firm up their goals.

DeepSeek also doesn't have internal KPIs to constrain people. They hope people have time to research areas they are interested in.

And because of sufficient restraint and focus, there are few things that need to be done, and no need to work overtime. Liang believes that research requires a relatively relaxed environment, and interest can only develop in a relaxed environment. (Our boss would just think the workload isn't full.)

Computing Power and Resources

Currently, graphics cards are insufficient; the more, the better. But ensuring that graphics cards are bought at a reasonable price is a very arduous task. (US embargo)

Liang believes the gap between us and the US is mainly in resources. All differences ultimately boil down to the difference in computing power resources.

We are about two years behind the US and need to use one-twentieth of the US's computing power to achieve AI. Through our continuous efforts, this gap is constantly narrowing, and the lag time is constantly decreasing.

Currently, the V4 model DeepSeek is training is not necessarily large enough, but it's the largest that can be trained given current resources.

Silicon Valley is saying Scaling has hit a wall, but for us, it's far from over. (Here, I understand Scaling likely refers to the predictable power-law improvement in AI intelligence performance and the emergence of new capabilities as parameters, data, and computation increase. This is a methodology, not a law.)

Domestic Chips and Ecosystem

Nvidia's moat will be dismantled later, because AI itself can drive the development iteration of graphics cards.

Gaming cards will later be separated from computing cards. Future computing cards should all use dedicated computing chips.

The ecosystem for domestic cards will rapidly rise later. This should be attributed to the 'necessity' brought by the US embargo, so domestic computing power will catch up.

Competitive Landscape and Industry Judgment

In the entire industry competition, the role China is most likely to play is still the one with the largest production capacity, the most electricity, and the cheapest products.

Just like how we export our goods now: large quantity, high quality, low price.

OpenAI might have initially really aimed to monopolize the world, but it will continuously encounter challengers.

And we, as a challenger, are willing to take less in the entire profit distribution process than they do. Taking more will definitely encounter greater resistance than taking less.

On the path of commercialization, DeepSeek only takes its due share. Product capability will definitely not be worse than the US's, costs will definitely be lower than the US's, so competitiveness will definitely be stronger.

Model R&D and Technology

Future competition among companies will likely be about cost control.

DeepSeek has also been working on multimodality. The version released after V4 will support native multimodality.

The company internally adheres to the idea of only pushing things to others after finding them useful themselves.

Commercialization and Pricing

Currently, DeepSeek's API pricing is within a reasonable range (I personally don't think it's cheap). In the current context, what limits B-end business should be demand, not the technology itself.

But currently, B-end demand is growing rapidly, though it won't grow to infinity.

So, commercialization mainly comes from B-end revenue, plus C-end users.

Given the scale of B-end revenue demand next year, the company might take a long time to achieve net profit positivity.

The worst-case scenario is just selling APIs; even that is enough to support a listed company.

Open Source Strategy

DeepSeek currently still adheres to the open source strategy, because closed source has no benefits.

Even if it's open source, there is a threshold to making it work, and this threshold is very high.

Open source does not affect commercial revenue. We are committed to the development of the AGI main line; open source itself does not cause us to lose anything.

Data and Post-Training

Data annotation has currently become a large part of the cost. The cost of high-quality data annotation is very high, both in China and the US.

At this stage, we still walk on two legs: self-annotation + external annotation. Currently, self-annotation is the lower-cost part.

Organization and Company Positioning

The company has no imitation target; every step is taken based on its own actual situation.

DeepSeek is clearly going to commercialize, because it is a company after all. But it only earns the money it should earn. There are trade-offs in this process.

It hopes to start from the company and make the AI pie bigger. DeepSeek has enough confidence that it can grow into a trillion-level company within this 'big pie'. (Very confident)

Summary

I highly recommend everyone read the original text deeply. The founder's vision is very grand, and the person is sufficiently confident.

The most crucial point is knowing where one's boundaries are. Having a clear idea of what can be achieved with current resources, and being able to accurately give one's own judgment on the industry's direction.

DeepSeek's initiation of financing might be to stabilize the team internally, but with such a highly confident and capable team in China working on AI, success is definitely achievable.