殷阿笛：英伟达的技术正在推动AI和生成式AI创新所需芯片和处理器的发展。作为硬件专家，你对AI和生成式AI有什么看法？你认为它会彻底改变我们工作和娱乐的方式吗？

ADI IGNATIUS: So Jensen, NVIDIA is producing technology that is quite literally driving the future, and I’m talking about the chips and processors that are driving innovation in AI and generative AI. I’d love your view as a sort of hardware guy. What’s your thought on AI, on gen AI? Do you think it will revolutionize how we work and play, or is it simply too soon to tell?

黄仁勋：生成式AI的核心是“一个软件理解数据含义的能力”。当然，你可以理解单词、字母、短语、句子和段落的含义，并找出其中的关系和使用模式。从大量的例子中，它可以找出并学习数据的表示方法，甚至理解其含义。

JENSEN HUANG: Well, at the core of generative AI is the ability for software to understand the meaning of data. You can understand the meaning of words of course, letters and words and sentences and paragraphs, and find the relationships and patterns. And from a lot of examples, it figures out – learns – the representation of the data and even understand its meaning.

正因如此，它也能理解信息的结构和含义。这适用于英语、像素、三维物体、蛋白质物体和化学物质。现在，我们已经能够使用这种深度学习方法来学习大量不同类型数据的表达方式，甚至是任何具有结构的数据。因为在物理世界中，在我们生活的世界中，有很多东西都是有结构的。

So it understands the structure and understand how it’s constructed and the meaning of the information. This applies to English, it applies to pixels, 3D objects, proteins, chemicals. We’ve now used this deep learning method to learn the representation of a whole lot of different types of data – any data with structure. And so there are a lot of things in the physical world, in the world that we live in that has structure.

一旦我们能理解数据的含义，并能将一种模态与另一种模态联系起来，就可以进行“翻译”。例如，我们能将图像中代表“猫”的像素与“猫”这个词联系起来。而后，我们可以将英文翻译成中文，也可以将中文翻译成法文。当我们从英文翻译成像素，就成为了生成式AI；而当我们从像素翻译成语言，就成为了字幕。我们也可以将文本翻译成文本摘要等。

Once we can understand the meaning of the data, and we can associate one modality to another modality, for example, we can associate the pixels of a cat in an image with the word cat. Once you could do this for one modality to another modality, then you can translate things. And so you could translate from English to Chinese, Chinese to French. You can also of course translate from English to pixels, which is generative AI or from pixels to language, which is captioning. Or you could go from text to text summarization, so on and so forth.

这种学习了数据表示的基本模型，功能强大得令人难以置信。现在AI行业和计算机行业正在做的，就是将这些生成方法和翻译方法融合到各种有趣的应用中。我认为，它让人兴奋的核心在于学习表征、学习意义以及从一种模式转换到另一种模式的能力，这实在强大得令人难以置信。

This basic model, the model that has learned representation of data is incredibly powerful. And so now what the AI industry, what the computer industry is doing is mixing all of these type of generative methods and translation methods into all kinds of interesting applications. So I think at the core of this excitement is the fundamental ability to learn representation, learn meaning, and to go from one modality to another and that’s just incredibly powerful.

殷阿笛：你是否有应用英伟达技术的实例？这些技术有哪些令人惊叹的应用？

ADI IGNATIUS: I’d love – do you have examples maybe using NVIDIA technology of applications that you’re amazed by that you’re seeing out there already?

黄仁勋：我们正在使用AI来设计芯片。我们用AI找到将晶体管连接到电路中的最佳位置，帮助我们优化设计，而这在以前是不可能实现的。我们用AI排查故障，找到哪行代码出现了问题从而导致后续错误。从设计芯片、优化芯片到了解故障所在，我们都在使用生成式AI来辅助。当然，目前最强大的事情之一，就是生成式AI可以预测下一个单词，或者预测下一个字母、下一个句子、下一个段落，甚至预测下一个像素，或者视频的下一帧。当然，这些预测也可能会产生错误的幻觉。真正有效的方法之一是向AI提供大量的上下文。在引入模型、向它提问之前，或者在要求它做某事之前，人们要做大量的提示，为模型设置背景和前提条件。

JENSEN HUANG: Well, we’re using AI to design our chips, to find optimal placements, to help us connect transistors into circuits and help us optimize designs at a scale that isn’t really possible before. We’re starting to use AI to help us determine from a failure where the bug likely started from, which line of code was responsible for the bug. So we’re using generative AI to help us from everything from designing the chip, optimizing the chip, to understanding where the failures are. One of the most powerful things that’s happening of course, is that generative AI predicts the next word or predict the next letter, next word, next sentence, next paragraph, predicts the next pixel, it predicts the next frame of a video. These predictions could hallucinate.And one of the really effective ways is one of them, of course, is providing a lot of context, a lot of prompts that people do to set the background, precondition the model before you prompt it, before you ask it a question, or before you ask it to do something.

其中一项非常有用的功能就是增强检索。这意味着，无论数据库是结构化的还是非结构化的，是表格形式的还是非表格形式的，都已经被向量化了。向量化或嵌入数据库需要你学习数据之间的所有关系，就像我之前说的学习数据的表示一样。通过学习表示法和学习数据库的向量化，你就能理解它的含义。而现在，真正令人兴奋的是，你可以将数据库向量化后连接到一个大型语言模型，而这个大型语言模型基本上可以让你与数据对话。

One of the things that’s really useful is retrieval augmentation. What that means is a database, whether it’s structured or unstructured, tabular or otherwise that has been vectorized – vectorizing or embedding a database, requires you to learn the relationship between all the data among each other, just like you’re learning the representation of the data that I was talking about before. And by learning the representation and learning the – vectorizing the database, you could understand its meaning. And so what’s really exciting now is that you could vectorize the database, you can connect it to a large language model and that large language model essentially allows you to talk to your data.

每家公司都有大量的数据，但其中大部分数据都处于休眠状态，很难使用。有时你必须大量查询数据，才能使用。但现在，你可以拥有一个真正理解含义的数据库。你可以在数据库中进行语义搜索。它能带出你想要查询的信息，然后你可以利用这种嵌入，并通过引入的模型对数据的提示来增强这种嵌入。这样，给AI提示的上下文和背景信息就丰富多了。而现在，由于大型语言模型能够理解你的意思，与应用程序交互就变得相当普遍和轻松了。

Every company has a ton of data in their company. Most of that data is now dormant. It’s hard to use. Sometimes you have to do a lot of queries on it. But now you could have a database that actually understands meaning. And so you could do a semantic search into the database. It brings back out the information that you meant to query, and you take that embedding and you augment that with the prompt that you’re prompting the model with. And so now the context, the background information with this prompt is much richer And so the thing that is really quite generalized now is the ease of being able to interact with applications because of large language models that understands your meaning.

其次，你现在可以将所有的数据库向量化，来增强提示和查询功能。AI在公司内部的应用真的令人惊奇。当然，最有趣的事情之一是，如果你是一家拥有大型客户服务部门的公司，客户服务人员正在输入他们与客户的互动——无论是问题、投诉还是帮助客户的方式——所有这些都会以某种方式记录在一个非常大的数据库中。如果将这个数据库向量化，你就可以直接与数据库对话，那岂不是很神奇？我们可以问数据库：客户最不满意的是什么？我们能看到哪些趋势？如果想改善公司的客户服务，我们能开始做的两三件事是什么？你可以像与人交谈一样与你的数据库交谈，这真的很强大。

And then secondarily, all of these databases that you can now vectorize, which could then be used to augment the prompts and queries. And so the applications across your company is really quite amazing. One of the fun things, of course, if you’re a company with a large customer service department and customer service agents are inputting their interactions with customers, whether it’s problems or complaints or whatever it is, help that you’re routing the customer through, all of that is somehow captured in a very large database. But wouldn’t it be amazing if now that database was vectorized and you can just talk to the database. What are the things that people are most upset about? What are some trends that we’re starting to see? If we want to improve the customer service of our company, what are the top two or three things that you can do? You could talk to your database like you’re talking to a person, and that’s really powerful.

殷阿笛：市场对新产品、新服务有不断变化的需求，你如何跟上需求？你又该怎样打造一支畅想未来的管理团队，以跟上变化的步伐？

ADI IGNATIUS: And for starters, how do you keep up with constantly shifting demand for new products, for new services? And how do you build a management team that can imagine the future and keep pace with so much change?

黄仁勋：你得学习。首先，我们周围的人都是计算机科学艺术、基础和基本部分的专家，他们可能是编译器专家，可能是芯片设计专家，可能是处理器设计专家，也可能是互连专家，等等。你的公司里到处都是专家，他们都是各自领域的佼佼者。这时，公司需要尽可能保持警惕。随着公司规模越来越大，让信息在组织内自由流动而不被困在孤岛中是非常重要的。尤其是当行业发展如此之快时，你需要警惕来自某处的微弱信号。这可能是一篇突然出现的研究论文，它与数字生物学有关，也可能与机器人技术有关。这些关于机器人的信息，可以以一种泛化的方式呈现，从而对我们如何训练聊天机器人非常有用。

JENSEN HUANG: Well, you got to learn. First of all, we’re surrounded by people who are expert in the art, in the fundamental and the foundational parts of computer science and they could be compiler experts or they could be chip design experts, or they could be processor design experts or interconnect experts or so on and so forth. And so you’ve got experts all over your company and they’re the best in the field of what they do. Meanwhile, you want your company to be as alert as possible. And so the ability for information to be moving freely in your organization without being trapped in silos is something that’s really important as the company gets larger and larger. Especially when the industry’s moving so fast, you want to be alert of very weak signals that could be coming from somewhere. And so it could be a research paper that’s coming out of left field, it could be related to digital biology, it could be related to something in robotics. And somehow that robotics information, it could be generalized in such a way that it could be useful for how we train a chatbot.

也许发明强化学习的最初目的，是为了表达和动画化简单的图形和手指关节等，从而可以玩游戏。但事实证明，强化学习中人类的反馈对于微调大型语言模型非常有用。因此，你要能看到相邻领域甚至无关领域的发展，并以某种方式将这种发展提炼回你的第一原则，然后推导到其他领域的应用中。

Maybe reinforcement learning was something that was really invented for articulating and animating stick figures and articulating fingers and things like that and playing a game. But it turns out that reinforcement learning human feedback could be quite useful in fine-tuning a large language model. And so you want to be able to see developments in adjacent fields or even unrelated fields and be able to somehow distill it back down to first principles and then extrapolate it to useful applications in other domains.

从拥有非常好的基础、公司的专业知识和对微弱信号保持警惕的能力开始，当做出观察时，你就能够回到第一原则，回到为什么这很重要、原因是什么、与该现象相关的主导动力是什么等问题，然后再加以归纳和总结。

That fundamental ability of starting from really good foundation, expertise in the company, the ability to stay alert of weak signals, and then being able to go back to first principles when you make some observation, go back to first principles on why is that important, what is the reason, what are the governing dynamics associated with that phenomenon? And then be able to generalize them.

殷阿笛：你认为AI永远需要人类的监督吗？

ADI IGNATIUS: Do you think that AI will always need human supervision?

黄仁勋：在科技回路中，人类的存在是非常重要的。众所周知，机器人技术中，科技可以有不同级别的自主性，直到达到高度的安全水平，并在各个方面都有高度安全性。例如，在自动驾驶汽车的情况下，如果你希望，家人在后座睡觉时，自动驾驶汽车仍然可以完全自主地行驶，那么你需要确保传感器系统、计算系统、车辆控制等各个方面都具备多样性和冗余性。举例来说，你可能会坚持确保你同时拥有摄像头、雷达和激光雷达等多个传感器，甚至还可能有冗余的转向控制和冗余的制动动力等。这些措施确保了无论在何种条件下，影响功能的关键组件都具有多样性和冗余性，从而为自主性提供了保障。

JENSEN HUANG: A human in the loop is pretty important. As you know in robotics, you could have different levels of autonomy and most of the levels of autonomy until you have a great deal of nines of safety and safety in all kinds of different ways. For example, in the case of a self-driving car, if you want the self-driving car to be able to take your family while they’re sleeping in the backseat fully autonomously, you want to be fairly assured that the sensor system, the computing system, the vehicle control, all of it has diversity and redundancy. And so for example, you might insist on making sure that you have cameras as well as radars, as well as LIDARs. You might even have redundant steering control and redundant braking dynamics and so on and so forth. And these things give you assurance that irrespective of the conditions, that the critical components that affects the functionality has diversity and redundancy.

多样性和冗余性是智能的重要组成部分，自主智能在很多方面都是如此。即使在公司里，我们也有多样性和冗余性。我们经常谈论公司的多样性，因为它能让公司更具弹性。我们有做同一件事的多种方法，我们试图从不同的角度分析同一个问题。也许销售部会得到市场部的补充，而市场部又会得到其他部门的补充……对待工作的不同方式和角度让公司有活力和弹性。而大多数自主系统，无论是机器人汽车还是大型组织，都希望在设计时能够实现多样性和冗余性。

Diversity and redundancy is an important part of intelligence, autonomous intelligence in a lot of ways. Even in companies, we have diversity and redundancy. Notice, we talk a lot about diversity in companies because it allows the company to be more resilient. We have redundant ways of doing the same thing. We’re trying to analyze the same problem from different perspectives. Maybe the sales organization is augmented by the marketing organization, which is augmented by so on and so forth, and different types of ways to ensure the vibrancy and the resilience of the company. And most autonomous systems, whether it’s as a robotic car or it’s a large organization, wants to be designed in such a way that you have diversity and redundancy.

在此基础上，你仍需要人类出现在工作回路中。有时，你会认为，自动驾驶汽车内部的回路不需要人类的介入，你可以在车外设置人类介入，以监控其效果，就像空中交通管制一样。但其实，飞机都有两名飞行员，飞行员之间也会互相监督。这就好比，我们有冗余的自动驾驶系统，但也有空中交通管制。

And then above that, you want human in the loop. Sometimes you get to a level where you don’t believe you need human in loop inside the car. Maybe you have human in loop outside the car and you’re monitoring the effectiveness. This is no different than air traffic control for example:  you have two pilots. You also have redundant autopilot systems, but you also have air traffic control. And pilots keep an eye out for each other.

自主系统有很多不同的设计方法，以实现最大的安全性和可靠性。生成式AI也必须采用同样的思路。我们必须发明安全技术，运用工程学方法，设计安全的系统，验证系统的安全性。最后，在相当长的一段时间内，我们一定会让人类参与进科技回路中。比如，当你让系统学习下一批数据、并将学习成果发布到全世界之前，你需要人类对结果进行验证、测试和红队测试，而不是让系统自己学习、更新。我们有许多不同的方法来确保安全性，而人类介入是最佳方式之一。

And so there’s a lot of different ways that autonomous systems are designed for maximum safety and reliability. The same line of thinking will have to be done for generative AI. We have to invent the technologies for safety. We have to apply engineering methods that allows us to design systems that are safe, validation systems that ensure safety. And then lastly, for quite a long time, we will likely… we will not likely, we will surely have human in loop. For example, while you’re learning the next batch from the next batch of data before you release it out to the world, you have humans validate it and test it and red team it instead of just have it be continuously learning and continuously updating all by itself. So there are a lot of different methods that we ensure safety and human in the loop is one of the best ways.

殷阿笛：有人引用英伟达一位副总裁的话说：“在英伟达工作，每天都有一种存在感。我总是感觉，公司明天就会破产。”你们是否试图灌输一种偏执感来刺激创新？

ADI IGNATIUS: One of your vice presidents was quoted as saying, “Working at NVIDIA feels existential every day. We feel like the company is going to go bankrupt tomorrow.” So I’m wondering, is that deliberate? Do you try to instill a sense of paranoia to try to spur innovation?

黄仁勋：你不必灌输它。如果你认为自己没有危险，那可能是因为你把头埋在了沙子里。没有哪家公司的生存是有保障的。我们的优势在于，公司从无到有。我们并没有夸张表达曾经几次险些倒闭的经历，所以我们不必假装公司总是处于危险之中。公司就是总处于危险之中的，我们能感觉到。

JENSEN HUANG: You don’t have to instill it. If you don’t think you’re in peril, it’s probably because you have your head in the sand. There are no companies that are assured survival. We have the benefit of having built the company from the ground up and having not exaggerated circumstances of nearly going out of business, actual experiences of nearly going out of business a handful of times, we don’t have to pretend that the company’s always peril. The company’s always in peril and we feel it.

当我和公司不需要保证会做得很好时，我们才反而能做到最好。我认为，公司生活在渴望和绝望之间，这比总是乐观或悲观要好得多。我认为，谨慎乐观的态度和现实的认识——即公司随时可能处于危险之中——能让我们必须保持专注和警觉，尽最大努力工作，不断赢得经营权。这些感受和理智对公司来说是非常有益和健康的。

I don’t need and the company doesn’t need assurance that we will do well in order to do our best work. And so I think the company living somewhere between aspiration and desperation, living somewhere within that spectrum is a lot better than either always optimistic or always pessimistic. And so I think a guarded level of optimism and realistic understanding that the company could be in peril at any given time, and we have to stay focused and alert and do our best work and constantly earn the right to be in business. Those feelings and that sensibility I think is really good and healthy for companies to have.

殷阿笛：你的管理风格有些与众不同。你的组织结构非常扁平，很少召开一对一的会议。请谈谈你的管理理念。

ADI IGNATIUS: From what I’ve read, your management style is somewhat unusual that you maintain a very flat organizational structure, that you have very few one-on-one meetings. Talk about your theory of management.

黄仁勋：这种组织架构的设计是为了让公司尽可能高效地开展工作。对于一个像我们这样技术发展迅速的公司来说，一件非常重要的事情是，我们发明了一种叫作加速计算的技术。虽然摩尔定律增加性能的速度不再像以前那么快，但在过去的日子里，摩尔定律每5年将性能提高10倍，每10年提高100倍。

JENSEN HUANG: Well, an architecture is designed for the company to do its work, to operate as efficiently as possible. And one of the things that is really important for a company and technology that’s moving as quickly as it is, we invented this technology called accelerated computing. And whereas Moore’s Law in the past, it is no longer nearly this fast, but Moore’s Law in the older days would increase performance by a factor of 10 every five years and a factor of 100 every 10 years.

但就加速计算而言，我们每10年的提升速度在1000到10万倍之间。因此，当发展速度如此之快时，你要确保信息在公司内尽可能快地传播。你希望公司尽可能保持一致。

But in the case of accelerated computing, we’re moving somewhere between 1000 to 100,000 times every 10 years.And so when you’re moving that fast, you want to make sure that that information is flowing through the company as quickly as possible. You want the company to be as aligned as possible.

此外，我们公司生产的产品非常复杂。人们认为，我们的GPU像是专为游戏玩家设计的插卡。的确如此，我们怀着无比自豪和喜悦的心情制造这些插卡。但我们为AI制造的GPU重达70磅，拥有3.5万个零件，其中8个来自台积电（TSMC）。它们功耗为1万安培。显然，它们太重了，以至于我们需要使用机器人来制造这些GPU。而且它们的计算机非常先进，以至于我们需要用一台超级计算机来测试另一台超级计算机。

And also our company builds really complicated things. People think our GPUs are like the add-in cards that we designed for gamers, and they are, we build those with great pride and great joy. But the GPUs that we build for AI weighs 70 pounds. They have 35,000 parts. Out of those 35,000 parts, eight of them come from TSMC. They consume 10,000 amps. Obviously they’re so heavy, you use robots to manufacture them. And they’re so… the computers are so advanced that it takes a supercomputer to test another supercomputer.

我们正在构建的这些GPU非常复杂，需要大量的软件才能运行。我们设计基础芯片以创建系统、软件、网络和所有硅光子技术，并将其连接到大型数据中心，从而帮助客户公司建立和运营。对于这样一家公司来说，我们的信息必须首先在自己公司内部快速传播，真正做到无障碍、无边界。

And so these GPUs that we’re building are insanely complex and it requires a lot of software to run it. And for a company like us that designs the fundamental chips to create the systems, create the software and creates all the networking and all the silicon photonics and connect it into a large data center and help companies stand it up and operate it, our company information has to really move quickly through the company and really, really have no barriers and no boundaries.

英伟达就是这样设计的一家公司。一方面，这能让我们在非常复杂的计算技术堆栈中工作，另一方面，这能让我们构建非常复杂的东西，并能非常自如地以光速前进。

And so we architected a company that allows us to be able to work across this really complicated stack of computing technology on the one hand, build incredibly complicated things on the other hand, and be very comfortable moving at light speed.

如果你想要一家这样的公司，那么你最不想要的就是按等级传递信息。大多数组织其实最高层只有两三个人。从顶层开始传递消息，在消息传达到真正处理基础问题的基层人员之前，会经过7、8、9、10层的管理层。我不希望出现这种情况。因此，最好的办法就是设计一个从高层开始就尽可能扁平的公司。这样，我们的公司能减少三到四层的管理机构。

If you want a company like that, then the last thing you want is for it to be information to pass along hierarchically. And if you look at most organizations, there are only two or three people at the top. And if that starts that way, then the number of layers of management or layers of managers before you get to somebody who’s actually dealing with ground truth could be 7, 8, 9, 10 layers. And I didn’t want that to happen. And so the best way to do that is to put the most… at the highest level to be as flat as possible up there. And so we probably reduced three or four layers of management out of our company just by doing so.

殷阿笛：你对量子计算有什么看法？你在多大程度上投资了量子计算？

ADI IGNATIUS: What can you say about quantum computing? To what extent are you investing in it and what can you say about that?

黄仁勋：要想制造出下一代史上最快的计算机，你需要目前世界上最快的计算机来帮助你设计它、模拟它、为它创造新的算法。而世界上最快的计算机就是英伟达公司的加速计算系统。因此，我们不制造量子计算机，但我们与量子计算生态系统中的几乎每个人都有合作——从进行基本量子计算设计的科研人员，到算法开发人员，再到研究经典量子计算机架构的架构师等。

JENSEN HUANG: Well, in order to build the world’s next fast computer, you need the world’s fastest computer to help you design it, simulate it, create new algorithms for it. And the world’s fastest computer is NVIDIA’s accelerated computing systems. And so we don’t build quantum computers, but we work with just about everybody in the quantum computing ecosystem from people who are doing the basic quantum computing design to algorithm developers, to architects who are working on classical quantum computers architectures. And so we work with just about everybody out there.

量子计算机可能还要再过十年、甚至二十年才能真正实现工业化。当它出现时，它很可能是一台与英伟达加速计算系统相连的加速器。量子计算机的大量计算将在经典计算机上完成，包括预处理、后处理等等。因此，我期待着量子计算机能够发挥作用。当它出现时，经典计算机，尤其是加速计算和AI，将是其中非常重要的一部分。

The quantum computer won’t come along in a real industrial way for probably another decade and maybe two. And when it gets here, it’ll likely be an accelerator that’s connected to the NVIDIA accelerated computing systems. And some of the computation will be done. Quantum computers, a lot of the computation will be done on classical computers doing pre-processing, post-processing, and so on and so forth. And so I’m looking forward to when quantum computers are going to be useful and effective, and when it comes, classical computers and particularly the accelerated computing and AI will be a very big part of it.

关键词：英伟达