“As long as I can keep the team stable, I will definitely achieve AGI. It’s that simple.”
TechFlowPost on July 23 published a long reconstruction of remarks attributed to DeepSeek founder Liang Wenfeng during a four-hour investor meeting. The outlet said the session had become one of the most discussed parts of DeepSeek’s fundraising story after elsewhere reported on that process last month.
The article collected 52 statements and said some wording may differ slightly from the original while preserving the intended meaning. Across those remarks, several themes appeared repeatedly: models, cost, AGI, time, and open source. Liang was also described as returning again and again to a series of “nos” — not a genius, not seeking unreasonable profit, not chasing user numbers, not going closed source, not focusing on 3D generation, video generation, or world models, and not trying to build the next super app. In his framing, restraint is a strategy meant to improve the probability of reaching AGI.
DeepSeek has one main line
Liang said this is not the moment to build products for maximum financial return. Product sits on the path to AGI, but DeepSeek does not need to put too much time or energy into either consumer-facing or enterprise-facing products. If a team operates from a higher technical level and works on a lower-level problem, he said, that becomes a form of dimensional advantage. Products are a byproduct of the road to AGI.
He placed several areas outside the company’s main line, including 3D, video generation, and world models. In his view, those directions do not have much to do with the upper bound of intelligence. Multimodality matters for products and for consumer users, he said, but it is a component rather than the main line, and it is not intelligence itself.
On hallucinations in large models, Liang said methods do exist, but the issue is a long-term one. Internally, he said, DeepSeek treats hallucinations as a product problem. It will address them, but they are not the current focus. The priority right now is Coding Agent. Given conditions in China, he said, the most reasonable approach is to go all in on a general-purpose agent, while specialized agents in finance, healthcare, and other sectors rank lower.
He added that if the AI era produces many trillion-dollar companies, it would be enough for DeepSeek to be one of them.
From continuous learning to self-improving AI, then embodied intelligence
Liang argued that AI is not short on taste or intuition. What it lacks is the ability to learn continuously. Humans can keep learning, he said, but AI typically needs all of the context supplied each time for the same task, and that is “almost impossible.” In his telling, that is why AI still cannot replace employees. For that reason, the next generation of models must be able to learn continuously; otherwise they do not deserve to be called next-generation models.
He said DeepSeek wants the next generation of models to help with its own development work first. Put simply, the first goal is not to make a model that everyone else finds easy to use. The first goal is to make a model that DeepSeek itself can use well, because he sees that as the fastest route to AGI. Even so, he said the world still has not found a good solution, because “learning” is made up of many things.
Liang described AGI as DeepSeek’s long-term vision and laid out a staircase analogy for how the company thinks about progress. Last year’s step was CoT, or Chain-of-Thought. This year’s step is Agent. After Agent, the next major problem is continuous learning. Once continuous learning is achieved, he said, a gradual singularity may emerge in which a model can do everything a human can do, including developing more advanced AI models on its own. At that stage, AI could accelerate AI research. Only after that step, in his account, does embodied intelligence come into view.
He added that intelligence may ultimately end in embodiment because, for an ordinary person, the need is not for a computer but for labor.
Commercialization remains distant
Liang repeatedly said DeepSeek aims for “reasonable profit,” not profit-maximizing pricing. He gave an example involving one of the company’s models. The price was initially set relatively high because the team worried demand would be too large, then later cut to one quarter of the original level. He said many people in the company group chat celebrated the reduction because the point of building the model carefully was to let more people use it fully.
Low cost, he said, is a result, but it is also tied to architecture choices that have consistently pushed toward lower cost. DeepSeek wants costs to remain affordable, especially under tight compute conditions. There is another reason as well: lower cost makes it easier to support a larger model. When compute is limited, better computational efficiency allows a team to train larger systems. Big companies can solve problems by adding more resources, he said, but DeepSeek prioritizes cost efficiency.
From the outside, Liang said, DeepSeek may look as if it picked a difficult model, yet internally the company is operating “very easily.” Price cuts are obviously bad news for competitors, he said, and they will not be cheering. He also said selling APIs is not especially attractive in itself. A small number of people can maintain the API, there is no need for customer service, no need for sales, and users will come on their own.
He said DeepSeek has always been commercializing in some sense, just not with commercialization as its central objective. In his words, the point at which DeepSeek fully turns toward commercialization should still be “very far away.”
On the future commercial opportunity, Liang said he does not even need to think about whether he will occupy a certain position when that time comes. If the opportunity is large enough, there will be a way. He described DeepSeek as a product of its era, a response to real conditions rather than an imitation.
Open source as the sweet spot for a company of DeepSeek’s size
Liang defined restraint as a strategy of giving up some things to gain others. Open source falls squarely in that bucket. Internally, he said, it gives employees a sense of accomplishment and strengthens cohesion. Externally, it benefits society, while peers and ordinary users are happy to see it.
He said he has no doubt that AGI will carry enormous commercial value. Even on that assumption, his first priority is not to capture a larger share but to increase the odds of succeeding.
He argued that open source can help make AI commercially viable, even if that sounds counterintuitive. Historically, a software company might have had a market worth only a few billion dollars a year, and open source could wipe that out. AI is different, he said, because the opportunity is large enough that it could eventually account for 10% of global GDP. If one player tries to monopolize that upside, it will be abandoned by history. He called that an objective rule and a matter of historical perspective.
Liang also said DeepSeek releases the same model in open source that it deploys for itself. The company will not publish an inferior version while keeping a better one for internal deployment.
He said he is not worried that others will deploy DeepSeek’s models and compete with it. In his view, not every company has the will or the ability to reach the same goal. Startups may be too small to do it, while large companies are difficult to organize. That, he said, is the sweet point for a company of DeepSeek’s size.
He added that open source does not affect DeepSeek’s business model at all as long as the company only wants a certain level of profit. If a company wants 100x profit, open source would indeed matter. On the same logic, he said DeepSeek does not want to become the enemy of any major or smaller internet company, and is willing to assist others — including Alibaba, Zhipu, and Moonshot AI — in doing better.
The China-U.S. AI gap is about resources, not talent
Liang said the future task is to rewrite the narrative around China and the U.S. in AI by using a fraction of the compute and shrinking the gap to six months or three months.
He said the primary difference between the two countries lies in resources. DeepSeek still believes in scaling and in the idea that larger scale produces better results. The reason the company trains models at such large size, he said, is not because he thinks that scale is sufficient, but because “that is all the resources” available.
On talent, Liang was blunt. There is almost no gap, he said — “it is the same group of people.” China does not lack talent, in his view. Talent shortages are cyclical and temporary. Historically, he said, no category of talent has been in permanent shortage.
In model competition, cost comes first
Liang said Anthropic is ahead of OpenAI right now, but only in a temporary sense rather than a lasting one. Looking ahead, he said OpenAI and Google will most likely alternate in moving up.
In China, he said, too many model companies are doing the same things and resources are spread too thin. The field will eventually consolidate, though the process will take time. If every company only seeks reasonable profit, he said, the market does not need so many teams building large models. Two large companies and two smaller ones might be enough.
He also said he does not believe large model companies will capture most of the profits in the broader AI industry.
As he sees it, competition among model companies will eventually come down to three factors: cost, time, and user experience. Cost ranks first, meaning the cost at which a provider can deliver the same quality of service. Time ranks second because being early by a few months matters. User experience can create some stickiness and barriers, but it is not fundamental.
No ambition to become the next super app
Liang said DeepSeek does not want to become the next super app. It is not trying to turn into the next ByteDance or the next Tencent either.
He explained that by saying there are bigger prizes ahead. What sits in front may be “sesame,” while the “watermelon” comes later. Even if that sesame is already large, he still does not consider it that large. Last year, he said, the market was competing for chatbot traffic and consumer adoption. This year it is competing for B2B revenue. Inside DeepSeek, neither is viewed as the most important issue. The company’s real attention is on the AGI roadmap and the next technical breakthrough.
He also remarked that the things people want most can be the hardest to get, while things they care less about can arrive more easily. Referring to DeepSeek’s popularity around last year’s Lunar New Year period, he said that surge was not part of the company’s script.
Team stability as the non-negotiable condition
Among all the themes in the meeting, Liang treated team stability as the one point on which he cannot compromise. He said it is the company’s biggest risk. That risk, however, was significantly eased by this round of financing.
He said many decisions at DeepSeek are made with team stability in mind. That includes the company’s unwillingness to become a direct adversary of large or small internet companies. Instead, it wants to empower and help them. It does not want enemies. In Liang’s view, that makes DeepSeek’s own environment better as well.
On organization, he said outsiders sometimes describe DeepSeek as top-down and sometimes as bottom-up, and both are correct. The top-down part is called “doing the right things,” and he generally does not want that to consume more than half of an employee’s time. The other half is left open for bottom-up exploration, with no assignment and no prerequisite, so people can investigate whatever they think matters.
He said the company generally does not work much overtime. One reason is that research needs a relatively relaxed environment. Another is that the company is highly focused. Many products remain incomplete, but the team has not rushed to fill every gap, which he again framed as a form of restraint.
Liang said the organization is dynamic rather than fixed. As DeepSeek grows, some changes may come, and some structure may become necessary, but the company will not fully turn into a traditional hierarchy. What will not change, he said, is the role of vision.
Vision, goodwill, and restraint
Liang said the company was not founded with the goal of making as much money as possible or going public through the capital markets. The first several dozen people who joined, he said, were not thinking that way. If they had been, they would not have come.
He said the company is being built with “great goodwill toward the world” and with the belief that the work is useful to humanity. That is why DeepSeek does not run on the logic of “achieve this KPI.” He described it as a vision-driven organization. That has strengths and weaknesses, he said, and the company will try to build on the former while reducing the latter, but the model itself is part of DeepSeek’s character.
In Liang’s account, vision does not even need to be written down. It lives in the way people work and in the attitude they hold toward the world. Individuals inside the company may understand that vision differently, but they are aligned in the broad direction.
He also said that around 20 years ago, the management figure he admired most was former GE chief executive Jack Welch. Looking back now, Liang said much of what Welch said may no longer hold up, but one point remains right: the most important thing in a company is vision. Vision is not a slogan on a wall. It is not about what is said, but about what is done.
That same idea led back to restraint. Liang said AGI has the biggest payoff. Other things can be done if the team has the bandwidth, and left aside if it does not. Restraint, in his words, is part of DeepSeek’s vision. AI is too large an opportunity, and the interests involved are too large as well. If the company can succeed, even a small share of the value will be enormous. The more restrained it is, the more likely it is to get there.
He said that logic fits his intuition, and that beyond vision DeepSeek does not have many other advantages. When the company was founded two years ago, he said, it did not have much money, many cards, much name recognition, or much ability to rally people. It was simply “a group of very ordinary people.” The story he prefers is not that geniuses did something extraordinary, but that ordinary people did something extraordinary.
He placed open source inside that same framework of restraint. Pricing, too, is not set with the sole aim of maximizing revenue or profit. A higher price may produce more revenue in the short term, he said, but the long-term picture is less clear. For him, restraint is a strategy.
He closed that line of thinking by saying open source and low prices give employees a stronger sense of accomplishment, create organizational cohesion, benefit society, and make peers and ordinary people happy. Over the long run, he said, that kind of restraint can raise DeepSeek’s odds of achieving AGI. If a company’s vision is simply to take more for itself, then it has already lost. “That’s how the world works.”