Revealing DeepSeek: A more extreme story of Chinese technological idealism

文 | 于丽丽 Wen | Yu Lili

编辑 | 刘旌 Edit | Liu Jing

中国的7家大模型创业公司中，DeepSeek（深度求索）最不声不响，但它又总能以出其不意的方式被人记住。
Of the 7 major model startups in China, DeepSeek is the least silent, but it can always be remembered in an unexpected way.

一年前，这种出其不意源自它背后的量化私募巨头幻方，是大厂外唯一一家储备万张A100芯片的公司，一年后，则来自它才是引发中国大模型价格战的源头。
A year ago, this kind of quantitative private equity giant fantasies that did not mean to derive behind it was the only company outside the large factory that reserved 10,000 A100 chips. One year later, it came from it to trigger the source of China's big model price war.

在被AI连续轰炸的5月，DeepSeek一跃成名。起因是他们发布的一款名为DeepSeek V2的开源模型，提供了一种史无前例的性价比：推理成本被降到每百万token仅 1块钱，约等于Llama3 70B的七分之一，GPT-4 Turbo的七十分之一。
In May, which was bombarded by AI, Deepseek became famous. The reason is that they released an open source model called DeepSeek V2, which provides an unprecedented cost-effectiveness: the reasoning cost of reasoning is reduced to only 1 yuan per million token, which is about one-seventh of LLAMA3 70B, GPT-4 4 Turbo's seventy -tenth.

DeepSeek被迅速冠以“AI界拼多多”之称的同时，字节、腾讯、百度、阿里等大厂也按耐不住，纷纷降价。中国大模型价格战由此一触即发。
At the same time that DEEPSEEK was quickly crowned as "AI Fighting Duoduo", the large manufacturers such as bytes, Tencent, Baidu, Ali and other large manufacturers were also unbearable, and the prices were reduced. The price war in China is from this.

弥漫的硝烟其实掩盖了一个事实：与很多大厂烧钱补贴不同，DeepSeek是有利润的。
The diffuse smoke actually covered a fact: Unlike many large manufacturers burning money subsidies, Deepseek is profitable.

这背后，是DeepSeek对模型架构进行了全方位创新。它提出的一种崭新的MLA（一种新的多头潜在注意力机制）架构，把显存占用降到了过去最常用的MHA架构的5%-13%，同时，它独创的DeepSeekMoESparse结构，也把计算量降到极致，所有这些最终促成了成本的下降。
Behind this is that Deepseek has innovated a full range of model architecture. It proposes a brand new MLA ( a new poly head potential attention mechanism ) architecture, which has reduced the memory of 5%-13%of the most commonly used MHA architecture in the past. The quantity dropped to the extreme, all of which eventually contributed to the decline in cost.

在硅谷，DeepSeek被称作“来自东方的神秘力量”。SemiAnalysis首席分析师认为，DeepSeek V2论文“可能是今年最好的一篇”。OpenAI前员工Andrew Carr认为论文“充满惊人智慧”，并将其训练设置应用于自己的模型。而OpenAI前政策主管、Anthropic联合创始人Jack Clark认为，DeepSeek“雇佣了一批高深莫测的奇才”，还认为中国制造的大模型，“将和无人机、电动汽车一样，成为不容忽视的力量。”
In Silicon Valley, Deepseek is called "mysterious power from the East". The chief analyst of Semianalysis believes that the DeepSeek V2 paper "may be the best article this year." Former OpenAI employee Andrew Carr believes that the paper is "full of amazing wisdom" and applies its training settings to its own model. Jack Clark, the former OPENAI policy director and co -founder of Anthropic, believes that DeepSeek "hired a group of unpredictable wizards" and also believed that the large model made in China, "will be like drones and electric vehicles, which will become unavoidable. strength."

在基本由硅谷牵动故事进展的AI浪潮里，这是罕有的情形。多位行业人士告诉我们，这种强烈的反响源自架构层面的创新，是国产大模型公司乃至全球开源基座大模型都很罕见的尝试。一位AI研究者表示，Attention架构提出多年来，几乎未被成功改过，更遑论大规模验证。“这甚至是一个做决策时就会被掐断的念头，因为大部分人都缺乏信心。”
This is a rare situation in the wave of AI that basically touched the story by Silicon Valley. A number of industry people told us that this strong response originated from the architecture level innovation, which is a rare attempt to be a rare attempt to make large domestic model companies and even global open source base models. A AI researcher said that the Attention architecture has been proposed for many years, and it has almost been successfully changed, let alone large -scale verification. "This is even a idea of being cut off when making decisions, because most people lack confidence."

而另一方面，国产大模型之前很少涉足架构层面的创新，也是因为很少有人主动去击破那样一种成见：美国更擅长从0-1的技术创新，而中国更擅长从1-10的应用创新。何况这种行为非常不划算——新一代模型，过几个月自然有人做出来，中国公司只要跟随、做好应用即可。对模型结构进行创新，意味着没有路径可依，要经历很多失败，时间、经济成本都耗费巨大。
On the other hand, the domestic big model rarely involved the innovation at the architecture level, because few people took the initiative to break such a prejudice: the United States is better at technological innovation from 0-1, and China is better at 1-10 Application innovation. Besides, this behavior is very uncomfortable -a new generation model, naturally someone will do it in a few months. As long as Chinese companies follow and do well. Innovating the model structure means that there is no path to rely on, and a lot of failure is to go through a lot of failures. Time and economic costs are huge.

DeepSeek显然是逆行者。在一片认为大模型技术必然趋同，follow是更聪明捷径的喧哗声中，DeepSeek看重“弯路”中积累的价值，并认为中国的大模型创业者除应用创新外，也可以加入到全球技术创新的洪流中。
Deepseek is obviously retrograde. In a piece of big model technology that is inevitable, Follow is the noise of more smart shortcuts. DeepSeek values the value accumulated in the "detours", and believes that in addition to application innovation, Chinese big model entrepreneurs can also join global technological innovation. In the torrent.

DeepSeek的很多抉择都与众不同。截至目前，7家中国大模型创业公司中，它是唯一一家放弃“既要又要”路线，至今专注在研究和技术，未做toC应用的公司，也是唯一一家未全面考虑商业化，坚定选择开源路线甚至都没融过资的公司。这些使得它经常被遗忘在牌桌之外，但在另一端，它又经常在社区被用户“自来水”式传播。
Many of DeepSeek's choices are different. As of now, among the seven major Chinese model startups, it is the only company that has given up the "both must and also" route and is focusing on research and technology. The open source route has not even finished the company. These are often forgotten from the table, but at the other end, it is often spread by users by users in the community.

DeepSeek究竟是如何炼成的？我们为此访谈了甚少露面的DeepSeek创始人梁文锋。
How is DeepSeek made? We interviewed Liang Wenfeng, the founder of Deepseek, who rarely appeared.

这位从幻方时代，就在幕后潜心研究技术的80后创始人，在DeepSeek时代，依旧延续着他的低调作风，和所有研究员一样，每天“看论文，写代码，参与小组讨论”。
This era of post -80s, who has been studying technology behind the scenes, still continues his low -key style in the DEEPSEEK era. Like all researchers, every day, "look at the dissertation, write code, and participate in group discussions."

和很多量化基金创始人都有过海外对冲基金履历，多出身物理、数学等专业不同的是，梁文锋一直是本土背景，早年就读的也是浙江大学电子工程系人工智能方向。
Different from the founders of many quantitative funds have the overseas hedge fund resumes. Different from the majors of physics and mathematics, Liang Wenfeng has always been a local background. In his early years, he also studied artificial intelligence in the Department of Electronic Engineering of Zhejiang University.

多位行业人士和DeepSeek研究员告诉我们，梁文锋是当下中国AI界非常罕见的“兼具强大的infra工程能力和模型研究能力，又能调动资源”、“既可以从高处做精准判断，又可以在细节上强过一线研究员”的人，他拥有“令人恐怖的学习能力”，同时又“完全不像一个老板，而更像一个极客”。
Several industry insiders and deepseek researcher told us that Liang Wenfeng is a very rare "strong Infra engineering ability and model research ability in the Chinese AI industry, but also can mobilize resources." Those who are more than a front -line researcher in details ", he has" terrifying learning ability ", and at the same time," is not like a boss at all, but more like a geek. "

这是一次尤为难得的访谈。访谈里，这位技术理想主义者，提供了目前中国科技界特别稀缺的一种声音：他是少有的把“是非观”置于“利害观”之前，并提醒我们看到时代惯性，把“原创式创新”提上日程的人。
This is a particularly rare interview. In the interview, this technical idealist provides a very scarce voice in the Chinese scientific and technological community: he is rare to put the "right or wrong view" before the "concept of interest", and remind us to see the inertia of the times. "Original Innovation" on the agenda.

一年前，DeepSeek刚下场时，我们初次访谈了梁文锋：《疯狂的幻方：一家隐形AI巨头的大模型之路》。如果说当时那句「务必要疯狂地怀抱雄心，且还要疯狂地真诚」还是一句美丽的口号，一年过去，它已经在成为一种行动。
One year ago, when DeepSeek first ended, we first interviewed Liang Wenfeng: "Crazy Fantasy Fang: The Road to a Big Model of an Invisible AI Giant". If the phrase "must be embraced madly, and to be madly sincere" is still a beautiful slogan, one year has passed, it is already becoming a action.

以下为对话部分： The following is the dialogue part:

价格战第一枪是怎么打响的？ How did the first shot of the price war began?

「暗涌」：DeepSeek V2模型发布后，迅速引发一场血雨腥风的大模型价格战，有人说你们是行业的一条鲶鱼。
"Dark Surge": After the release of the Deepseek V2 model, it quickly triggered a big model price war with a bloody storm. Some people said that you are a catfish in the industry.

梁文锋：我们不是有意成为一条鲶鱼，只是不小心成了一条鲶鱼。
Liang Wenfeng : We don't intend to be a catfish, but we accidentally become a catfish.

「暗涌」：这个结果让你们意外吗？ "Dark": Is this result surprised you?

梁文锋：非常意外。没想到价格让大家这么敏感。我们只是按照自己的步调来做事，然后核算成本定价。我们的原则是不贴钱，也不赚取暴利。这个价格也是在成本之上稍微有点利润。
Liang Wenfeng : Very unexpected. I did not expect the price to be so sensitive. We just do things according to our own pace, and then calculate the cost. Our principles are not money or profit. This price is also a little profitable on the cost.

「暗涌」：5天后智谱AI就跟进了，之后是字节、阿里、百度、腾讯等大厂。
"Dark Surge": After 5 days, the wisdom spectrum AI followed up, and then large factories such as bytes, Ali, Baidu, Tencent and other large manufacturers.

梁文锋：智谱AI降的是一个入门级产品，和我们同级别的模型仍然收费很贵。字节是真正第一个跟进的。旗舰模型降到和我们一样的价格，然后触发了其它大厂纷纷降价。因为大厂的模型成本比我们高很多，所以我们没想到会有人亏钱做这件事，最后就变成了互联网时代的烧钱补贴的逻辑。
Liang Wenfeng : The wisdom spectrum AI reduces an entry -level product, and the model of our same level is still expensive. The byte is the first to follow up. The flagship model dropped to the same price as us, and then triggered other large manufacturers to reduce prices. Because the cost of the model of the big factory is much higher than us, we did not expect that someone would lose money to do this, and finally became the logic of burning subsidies in the Internet era.

「暗涌」：外部看来，降价很像在抢用户，互联网时代的价格战通常如此。
"Dark surge": It seems that the price reduction is very similar to being robbing users. The price war in the Internet era is usually the case.

梁文锋：抢用户并不是我们的主要目的。我们降价一方面是因为我们在探索下一代模型的结构中，成本先降下来了，另一方面也觉得无论API,还是AI,都应该是普惠的、人人可以用得起的东西。
Liang Wenfeng : Raising users is not our main purpose. On the one hand, our price cut is because in the structure of the next generation of models, the cost drops first, and on the other hand, we also feel that both the API or AI should be inclusive and everyone can use things.

「暗涌」：在这之前，大部分中国公司都会直接copy这一代的 Llama结构去做应用，为什么你们会从模型结构切入？
"Dark Surge": Before that, most Chinese companies will directly Copy's LLAMA structure to apply. Why do you cut in from the model structure?

梁文锋：如果目标是做应用，那沿用 Llama结构，短平快上产品也是合理选择。但我们目的地是AGI，这意味着我们需要研究新的模型结构，在有限资源下，实现更强的模型能力。这是scale up到更大模型所需要做的基础研究之一。除了模型结构，我们还做了大量其他的研究，包括怎么构造数据，如何让模型更像人类等，这都体现在我们发布的模型里。另外，Llama的结构，在训练效率和推理成本上，和国外先进水平估计也已有两代差距。
Liang Wenfeng : If the goal is to apply it, the LLAMA structure is used, and the product is also a reasonable choice. But our destination is AGI, which means that we need to study new model structures and achieve stronger model capabilities under limited resources. This is one of the basic research that Scale UP needs to do a larger model. In addition to the model structure, we have also done a lot of other studies, including how to construct data and how to make the model more like humans, which are reflected in the model we posted. In addition, the structure of LLAMA, in terms of training efficiency and reasoning costs, has two generations gap with advanced foreign levels.

「暗涌」：这种代差主要来自哪里？ "Dark Surge": Where does this difference come from?

梁文锋：首先训练效率有差距。我们估计，国内最好的水平和国外最好的相比，模型结构和训练动力学上可能有一倍的差距，光这一点我们要消耗两倍的算力才能达到同样效果。另外数据效率上可能也有一倍差距，也就是我们要消耗两倍的训练数据和算力，才能达到同样的效果。合起来就要多消耗4倍算力。我们要做的，正是不停地去缩小这些差距。
梁文锋：首先训练效率有差距。 We estimate that compared with the best level in China and the best abroad, there may be double the model structure and training dynamics. We have to consume twice the computing power to achieve the same effect. In addition, there may be double gap in data efficiency, that is, we have to consume twice the training data and computing power to achieve the same effect. It takes 4 times more computing power to close. What we have to do is constantly narrowing these gaps.

「暗涌」：大部分中国公司都选择既要模型又要应用，为什么DeepSeek目前选择只做研究探索？
"Dark Yong": Most Chinese companies choose to use both models and application. Why is DeepSeek choose to only do research and exploration?

梁文锋：因为我们觉得现在最重要的是参与到全球创新的浪潮里去。过去很多年，中国公司习惯了别人做技术创新，我们拿过来做应用变现，但这并非是一种理所当然。这一波浪潮里，我们的出发点，就不是趁机赚一笔，而是走到技术的前沿，去推动整个生态发展。
Liang Wenfeng : Because we feel that the most important thing now is to participate in the wave of global innovation. In the past many years, Chinese companies have been accustomed to making technological innovation. We have taken it for application monetization, but this is not a matter of course. In this wave of waves, our starting point is not to take the opportunity to make a fortune, but to the forefront of technology to promote the entire ecological development.

「暗涌」：互联网和移动互联网时代留给大部分人的惯性认知是，美国擅长搞技术创新，中国更擅长做应用。
"Dark Surge": The inertia cognition left by most people in the Internet and mobile Internet era is that the United States is good at engaging in technological innovation and China is better at applying.

梁文锋：我们认为随着经济发展，中国也要逐步成为贡献者，而不是一直搭便车。过去三十多年IT浪潮里，我们基本没有参与到真正的技术创新里。我们已经习惯摩尔定律从天而降，躺在家里18个月就会出来更好的硬件和软件。Scaling Law也在被如此对待。
Liang Wenfeng : We believe that with the development of the economy, China must gradually become contributors, rather than always taking stools. In the IT wave in the past 30 years, we have basically not participated in real technological innovation. We are accustomed to falling from the sky, and we will come out for better hardware and software when we lie at home for 18 months. Scaling Law is also treated like this.

但其实，这是西方主导的技术社区一代代孜孜不倦创造出来的，只因为之前我们没有参与这个过程，以至于忽视了它的存在。
But in fact, this was created by the Western -led technological community generation, because we did not participate in this process before, so that we ignored its existence.

真正的差距不是一年或两年，而是原创和模仿之差 The real gap is not one year or two years, but the difference between original and imitation

「暗涌」：为什么DeepSeek V2会让硅谷的很多人惊讶？
"Dark Surge": Why does DeepSeek V2 surprise many people in Silicon Valley?

梁文锋：在美国每天发生的大量创新里，这是非常普通的一个。他们之所以惊讶，是因为这是一个中国公司，在以创新贡献者的身份，加入到他们游戏里去。毕竟大部分中国公司习惯follow，而不是创新。
Liang Wenfeng : This is a very ordinary one in the large number of innovations in the United States every day. The reason why they were surprised was because it was a Chinese company, joining the game as an innovative contributor to their games. After all, most Chinese companies are used to Follow, not innovation.

「暗涌」：但这种选择放在中国语境里，也过于奢侈。大模型是一个重投入游戏，不是所有公司都有资本只去研究创新，而不是先考虑商业化。
"Dark Surging": But this choice is too luxurious in the context of China. Large models are a heavy -duty game. Not all companies have capital only to research innovation, rather than considering commercialization first.

梁文锋：创新的成本肯定不低，过去那种拿来主义的惯性也和过去的国情有关。但现在，你看无论中国的经济体量，还是字节、腾讯这些大厂的利润，放在全球都不低。我们创新缺的肯定不是资本，而是缺乏信心以及不知道怎么组织高密度的人才实现有效的创新。
Liang Wenfeng : The cost of innovation is definitely not low. The inertia of the past doctrine is also related to the past national conditions. But now, you can see that regardless of China's economy, or the profits of large factories such as bytes and Tencent, it is not low in the world. We must not be capital, but lack confidence and do not know how to organize high -density talents to achieve effective innovation.

「暗涌」：为什么中国公司——包括不缺钱的大厂，这么容易把快速商业化当第一要义？
"Dark Surge": Why does a Chinese company -including a large factory that is not short of money, so it is easy to take fast commercialization as the first priority?

梁文锋：过去三十年，我们都只强调赚钱，对创新是忽视的。创新不完全是商业驱动的，还需要好奇心和创造欲。我们只是被过去那种惯性束缚了，但它也是阶段性的。
Liang Wenfeng : In the past three decades, we have all emphasized to make money and ignore innovation. Innovation is not entirely commercially driven, but also needs curiosity and creativity. We are just bound by the inertia of the past, but it is also staged.

「暗涌」：但你们究竟是一个商业组织，而非一个公益科研机构，选择创新，又通过开源分享出去，那要在哪里形成护城河？像5月这次MLA架构的创新，也会很快被其他家copy吧？
"Dark": But you are a commercial organization, not a public welfare scientific research institution, choose innovation, and share it through open source. Where can you form a moat? Like the innovation of the MLA architecture in May, will it be Copy soon?

梁文锋：在颠覆性的技术面前，闭源形成的护城河是短暂的。即使OpenAI闭源，也无法阻止被别人赶超。所以我们把价值沉淀在团队上，我们的同事在这个过程中得到成长，积累很多know-how,形成可以创新的组织和文化，就是我们的护城河。
Liang Wenfeng : In the face of disruptive technology, the moat formed by the closed source is short. Even if the OpenAI is closed, it cannot be stopped by others. Therefore, we have precipitated value on the team. Our colleagues have grown in the process, accumulating a lot of Know-How, to form an innovative organization and culture, which is our moat.

开源，发论文，其实并没有失去什么。对于技术人员来说，被follow是很有成就感的事。其实，开源更像一个文化行为，而非商业行为。给予其实是一种额外的荣誉。一个公司这么做也会有文化的吸引力。
Open source, papers, actually did not lose anything. For technicians, it is very accomplished by Follow. In fact, open source is more like a cultural behavior, not a business behavior. Giving is actually an additional honor. A company does this will also be attractive.

「暗涌」：你怎么看类似朱啸虎的这种市场信仰派观点？
"Dark": What do you think of the market beliefs like Zhu Xiaohu?

梁文锋：朱啸虎是自洽的，但他的打法更适合快速赚钱的公司，而你看美国最赚钱的公司，都是厚积薄发的高科技公司。
Liang Wenfeng : Zhu Xiaohu is self -consistent, but his play is more suitable for fast -making companies, and you see that the most profitable companies in the United States are high -tech companies.

「暗涌」：但做大模型，单纯的技术领先也很难形成绝对优势，你们赌的那个更大的东西是什么？
"Dark Yong": But when making a big model, it is difficult to form an absolute advantage with simple technical leadership. What is the bigger thing you bet?

梁文锋：我们看到的是中国AI不可能永远处在跟随的位置。我们经常说中国AI和美国有一两年差距，但真实的gap是原创和模仿之差。如果这个不改变，中国永远只能是追随者，所以有些探索也是逃不掉的。
Liang Wenfeng : What we see is that China AI cannot always follow. We often say that Chinese AI and the United States have a gap between one or two years, but the real Gap is the difference between original and imitation. If this does not change, China can only be followers, so some explorations cannot escape.

英伟达的领先，不只是一个公司的努力，而是整个西方技术社区和产业共同努力的结果。他们能看到下一代的技术趋势，手里有路线图。中国AI的发展，同样需要这样的生态。很多国产芯片发展不起来，也是因为缺乏配套的技术社区，只有第二手消息，所以中国必然需要有人站到技术的前沿。
Nvidia's lead is not just the efforts of a company, but the result of the joint efforts of the entire Western technology community and industry. They can see the technical trend of the next generation and have a roadmap in their hands. The development of Chinese AI also needs such an ecology. Many domestic chips cannot develop, and because of lack of supporting technology communities, there are only second -hand news, so China must need to stand at the forefront of technology.

一群做“高深莫测”事的年轻人 A group of young people who do "unfathomable" things

「暗涌」：OpenAI前政策主管、Anthropic联合创始人Jack Clark认为DeepSeek雇佣了“一批高深莫测的奇才”，做出DeepSeek v2的是怎样一群人？
"Undercurrent": Jack Clark, former policy director of OpenAI and co-founder of Anthropic, believes that DeepSeek hired "a group of unpredictable wizards". What kind of people made DeepSeek v2?

梁文锋：并没有什么高深莫测的奇才，都是一些Top高校的应届毕业生、没毕业的博四、博五实习生，还有一些毕业才几年的年轻人。
Liang Wenfeng : There are no mysterious geniuses. They are all recent graduates from top universities, interns with Ph.D. 4 and Ph. 5 who have not graduated, and some young people who have graduated only a few years ago.

「暗涌」：很多大模型公司都执着地去海外挖人，很多人觉得这个领域前50名的顶尖人才可能都不在中国的公司，你们的人都来自哪里？
"Undercurrent": Many large model companies are persistent in poaching people overseas. Many people think that the top 50 talents in this field may not be in Chinese companies. Where do your people come from?

梁文锋：V2模型没有海外回来的人，都是本土的。前50名顶尖人才可能不在中国，但也许我们能自己打造这样的人。
Liang Wenfeng : There are no people who came back from overseas in the V2 model, they are all local. The top 50 talents may not be in China, but maybe we can build such people ourselves.

「暗涌」：这次MLA创新是如何发生的？听说idea最早来自一个年轻研究员的个人兴趣？
"Undercurrent": How did this MLA innovation happen? I heard that the idea first came from the personal interest of a young researcher?

梁文锋：在总结出Attention架构的一些主流变迁规律后，他突发奇想去设计一个替代方案。不过从想法到落地，中间是一个漫长的过程。我们为此组了一个team，花了几个月时间才跑通。
Liang Wenfeng : After summarizing some mainstream changes in the Attention architecture, he suddenly wanted to design an alternative. However, it is a long process from idea to implementation. We formed a team for this and it took us several months to get through it.

「暗涌」：这种发散性灵感的诞生和你们完全创新型组织的架构很有关系。幻方时代，你们就很少自上而下地指派目标或任务。但AGI这种充满不确定性的前沿探索，是否多了管理动作？
"Undercurrent": The birth of this divergent inspiration is closely related to the structure of your completely innovative organization. In the Magic Square era, you rarely assign goals or tasks from top to bottom. But does AGI, a frontier exploration full of uncertainty, require more management actions?

梁文锋：DeepSeek也全是自下而上。而且我们一般不前置分工，而是自然分工。每个人有自己独特的成长经历，都是自带想法的，不需要push他。探索过程中，他遇到问题，自己就会拉人讨论。不过当一个idea显示出潜力，我们也会自上而下地去调配资源。
Liang Wenfeng : DeepSeek is also all bottom-up. Moreover, we generally do not pre-position division of labor, but natural division of labor. Everyone has their own unique growth experience and comes with their own ideas, so there is no need to push them. During the exploration process, when he encounters problems, he will invite others to discuss them. But when an idea shows potential, we will allocate resources from top to bottom.

「暗涌」：听说DeepSeek对于卡和人的调集非常灵活。
"Undercurrent": I heard that DeepSeek is very flexible in mobilizing cards and people.

梁文锋：我们每个人对于卡和人的调动是不设上限的。如果有想法，每个人随时可以调用训练集群的卡无需审批。同时因为不存在层级和跨部门，也可以灵活调用所有人，只要对方也有兴趣。
Liang Wenfeng : There is no upper limit for each of us to transfer cards and people. If you have an idea, everyone can call the card of the training cluster at any time without approval. At the same time, because there are no hierarchies or cross-departments, everyone can be flexibly called as long as the other party is also interested.

「暗涌」：一种松散的管理方式也取决于你们筛选到了一批强热爱驱动的人。听说你们很擅长从细节招人，可以让一些非传统评价指标里优秀的人被选出来。
"Undercurrent": A loose management method also depends on you selecting a group of people who are driven by strong love. I heard that you are very good at recruiting people based on details, and can select some outstanding people based on non-traditional evaluation indicators.

梁文锋：我们选人的标准一直都是热爱和好奇心，所以很多人会有一些奇特的经历，很有意思。很多人对做研究的渴望，远超对钱的在意。
Liang Wenfeng : The criteria for choosing people have always been love and curiosity, so many people will have some strange experiences, which are very interesting. Many people's desire for research far exceeds their care of money.

「暗涌」: transformer诞生在谷歌的AI Lab，ChatGPT诞生在OpenAI,你觉得大公司的AILab 和一个创业公司对于创新产生的价值有什么不同？
"Undercurrent": Transformer was born in Google's AI Lab, and ChatGPT was born in OpenAI. What do you think is the difference in the value of innovation between a large company's AILab and a startup company?

梁文锋：不管是Google实验室，还是OpenAI，甚至中国大厂的AI Lab，都很有价值的。最后是OpenAI做出来，也有历史的偶然性。
Liang Wenfeng : Whether it is Google Labs, OpenAI, or even the AI Labs of major Chinese companies, they are all valuable. In the end, OpenAI made it, and it was also a historical accident.

「暗涌」：创新很大程度也是一种偶然吗？我看你们办公区中间那排会议室左右两侧都设置了可以随意推开的门。你们同事说，这就是给偶然留出空隙。transfomer诞生中就发生过那种偶然经过的人听到后加入，最终把它变成一个通用框架的故事。
"Undercurrent": Is innovation largely an accident? I see that the row of conference rooms in the middle of your office area has doors on the left and right that can be pushed open at will. Your colleagues said that this is to leave room for chance. In the birth of transformer, there was a story where people passing by by chance heard about it and joined in, eventually turning it into a universal framework.

梁文锋：我觉得创新首先是一个信念问题。为什么硅谷那么有创新精神？首先是敢。Chatgpt出来时，整个国内对做前沿创新都缺乏信心，从投资人到大厂，都觉得差距太大了，还是做应用吧。但创新首先需要自信。这种信心通常在年轻人身上更明显。
Liang Wenfeng : I think innovation is first of all a matter of belief. Why is Silicon Valley so innovative? The first is to dare. When Chatgpt came out, the entire country lacked confidence in cutting-edge innovation. From investors to large manufacturers, everyone felt that the gap was too big, so they should just make applications. But innovation first requires confidence. This confidence is usually more pronounced in younger people.

「暗涌」：但你们不参与融资，很少对外发声，社会声量上肯定不如那些融资活跃的公司，怎么确保DeepSeek就是做大模型的人的首选？
"Undercurrent": But you don't participate in financing, rarely speak out to the outside world, and your social voice is definitely not as good as those companies that are active in financing. How can you ensure that DeepSeek is the first choice for people who want to build large models?

梁文锋：因为我们在做最难的事。对顶级人才吸引最大的，肯定是去解决世界上最难的问题。其实，顶尖人才在中国是被低估的。因为整个社会层面的硬核创新太少了，使得他们没有机会被识别出来。我们在做最难的事，对他们就是有吸引力的。
Liang Wenfeng : Because we are doing the most difficult thing. What attracts top talents the most is definitely solving the world’s most difficult problems. In fact, top talents are underestimated in China. Because there are too few hard-core innovations at the entire social level, they have no chance to be identified. We are doing the most difficult thing, which is attractive to them.

「暗涌」：前一段OpenAI的发布并没有等来GPT5,很多人觉得这是技术曲线明显在放缓，也很多人开始质疑Scaling Law，你们怎么看？
"Undercurrent": The release of OpenAI some time ago did not wait for GPT5. Many people think that the technology curve is obviously slowing down, and many people are beginning to question the Scaling Law. What do you think?

梁文锋：我们偏乐观，整个行业看起来都符合预期。OpenAI也不是神，不可能一直冲在前面。
Liang Wenfeng : We are optimistic, and the entire industry seems to meet expectations. Openai is not a god, it is impossible to rush ahead.

「暗涌」：你觉得AGI还要多久实现，发布DeepSeek V2前，你们发布过代码生成和数学的模型，也从dense模型切换到了MOE,所以你们的AGI路线图有哪些坐标？
"Undercurrent": How long do you think it will take for AGI to be realized? Before releasing DeepSeek V2, you released code generation and mathematical models, and also switched from dense models to MOE. So what are the coordinates of your AGI roadmap?

梁文锋：可能是2年、5年或者10年，总之会在我们有生之年实现。至于路线图，即使在我们公司内部，也没有统一意见。但我们确实押注了三个方向。一是数学和代码，二是多模态，三是自然语言本身。数学和代码是AGI天然的试验场，有点像围棋，是一个封闭的、可验证的系统，有可能通过自我学习就能实现很高的智能。另一方面，可能多模态、参与到人类的真实世界里学习，对AGI也是必要的。我们对一切可能性都保持开放。
Liang Wenfeng : It may be 2 years, 5 years or 10 years. In short, it will be realized in our lifetime. As for the roadmap, even within our company, there is no consensus. But we did bet in three directions. One is mathematics and code, the second is multimodality, and the third is natural language itself. Mathematics and code are the natural testing ground for AGI. It is a bit like Go. It is a closed and verifiable system, and it is possible to achieve high intelligence through self-learning. On the other hand, multi-modal learning that involves humans in the real world may also be necessary for AGI. We are open to all possibilities.

「暗涌」：你觉得大模型终局是什么样态？ "Undercurrent": What do you think the ending of the big model will be like?

梁文锋：会有专门公司提供基础模型和基础服务,会有很长链条的专业分工。更多人在之上去满足整个社会多样化的需求。
Liang Wenfeng : There will be specialized companies providing basic models and basic services, and there will be a long chain of professional division of labor. More people can meet the diverse needs of society as a whole.

所有的套路都是上一代的产物 All routines are products of the previous generation

「暗涌」：过去这一年，中国的大模型创业还是有很多变化的，比如去年开头还很活跃的王慧文中场退出了，后来加入的公司也开始呈现出差异化。
"Undercurrent": In the past year, there have been many changes in China's large model entrepreneurship. For example, Wang Huiwen, who was active at the beginning of last year, withdrew from the company mid-term, and the companies he joined later began to show differentiation.

梁文锋：王慧文自己承担了所有的损失，让其他人全身而退。他做了一个对自己最不利，但对大家都好的选择，所以他做人是很厚道的，这点我很佩服。
Liang Wenfeng : Wang Huiwen took all the losses and let others escape unscathed. He made a choice that was most detrimental to himself but best for everyone, so he is a very kind person, which I admire very much.

「暗涌」：现在你的精力最多放在哪里？ "Undercurrent": Where do you focus most of your energy now?

梁文锋：主要的精力在研究下一代的大模型。还有很多未解决的问题。
Liang Wenfeng : The main focus is on researching the next generation of large models. There are still many unanswered questions.

「暗涌」:其他几家大模型创业公司都是坚持既要又要，毕竟技术不会带来永久领先，抓住时间窗口把技术优势落到产品也很重要，DeepSeek敢于专注在模型研究上是因为模型能力还不够吗？
"Undercurrent": Several other large model startups insist on having both. After all, technology will not bring permanent leadership. It is also important to seize the time window to put the technical advantages into products. DeepSeek dares to focus on model research. Is it because the model capability is not enough?

梁文锋：所有的套路都是上一代的产物，未来不一定成立。拿互联网的商业逻辑去讨论未来AI的盈利模式，就像马化腾创业时，你去讨论通用电气和可口可乐一样。很可能是一种刻舟求剑。
Liang Wenfeng : All routines are products of the previous generation and may not be valid in the future. Use the business logic of the Internet to discuss the future profit model of AI, just like when Ma Huateng started his business, you discussed General Electric and Coca-Cola. It is probably a kind of carving a boat to seek a sword.

「暗涌」：过去幻方就有很强的技术和创新基因，成长也比较顺利，这是你偏乐观的原因吗？
"Undercurrent": In the past, Huanfang had strong technology and innovation genes, and its growth was relatively smooth. Is this why you are optimistic?

梁文锋：幻方某种程度上增强了我们对技术驱动型创新的信心，但也不都是坦途。我们经历了一个漫长的积累过程。外部看到的是幻方2015年后的部分，但其实我们做了16年。
Liang Wenfeng : Magic Square has enhanced our confidence in technology-driven innovation to some extent, but it is not always a smooth road. We have gone through a long accumulation process. What we see from the outside is the part of Magic Square after 2015, but in fact we have been doing it for 16 years.

「暗涌」：回到关于原创式创新的话题。现在经济开始进入下行，资本也进入冷周期，所以它对原创式创新是否会带来更多抑制？
"Dark": Back to the topic about original innovation. Now that the economy has begun to fall, capital also enters the cold cycle, so will it bring more suppression of original innovation?

梁文锋：我倒觉得未必。中国产业结构的调整，会更依赖硬核技术的创新。当很多人发现过去赚快钱很可能来自时代运气，就会更愿意俯身去做真正的创新。
Liang Wenfeng : I don't think it is necessary. The adjustment of China's industrial structure will rely more on the innovation of hardcore technology. When many people find that in the past, they are likely to come from the times, and they will be more willing to lean down to do real innovation.

「暗涌」：所以你对这件事也是乐观的？ "Dark Surging": So are you optimistic about this?

梁文锋：我是八十年代在广东一个五线城市长大的。我的父亲是小学老师，九十年代，广东赚钱机会很多，当时有不少家长到我家里来，基本就是家长觉得读书没用。但现在回去看，观念都变了。因为钱不好赚了，连开出租车的机会可能都没了。一代人的时间就变了。
Liang Wenfeng : I grew up in a fifth -tier city in Guangdong in the 1980s. My father was a primary school teacher. In the 1990s, there were many opportunities to make money in Guangdong. At that time, many parents came to my house. Basically, parents felt that reading was useless. But now when I go back, my concept has changed. Because the money is not easy to make, even the chance of driving a taxi may be gone. The time of a generation has changed.

以后硬核创新会越来越多。现在可能还不容易被理解，是因为整个社会群体需要被事实教育。当这个社会让硬核创新的人功成名就，群体性想法就会改变。我们只是还需要一堆事实和一个过程。
There will be more and more hard-core innovations in the future. It may not be easy to understand now because the entire social group needs to be educated on the facts. When this society allows hard-core innovative people to become successful, group thinking will change. We just need a bunch of facts and a process.

财税茶座（C&S Tea House）cscz.org，跟踪贸易战

搜索此博客