Keynote by NVIDIA CEO Jensen Huang at 2024 SIEPR Economic Summit
开场与主持人介绍
英文清理版
Welcome back everyone, after the short break. I know that many of you are looking forward to hearing from our next speaker, Jensen Huang. Jensen is at the cutting edge of artificial intelligence and all of the innovation, technology, and human capital that is needed to support it. My good friend and superb colleague John Shoven is going to introduce Jensen. I hope he is here somewhere, so I am just going to keep talking until the two of them have a conversation and then take some of your questions.
John Shoven certainly requires very little introduction to most people in this crowd. As my predecessor as the director of SIEPR, John is the one who started the SIEPR Economic Summit twenty years ago. I would like all of us to give John Shoven a huge round of applause and appreciate the community that he had the foresight to build.
For those of you who have not been touched by John’s research, mentorship, or friendship, here is just a snippet of what you might like to know about him. Along with being the former SIEPR director and a senior fellow emeritus, John is the Charles R. Schwab Professor of Economics. He is also a senior fellow at the Hoover Institution and a research associate of the National Bureau of Economic Research. He specializes in public finance and corporate finance and has published many articles over the years on social security, health economics, corporate personal taxation, mutual funds, pension plans, economic demography, applied general equilibrium economics, and much more.
中文翻译
欢迎大家在短暂休息后回来。想必很多人都很期待接下来这位发言人，Jensen Huang。Jensen 站在人工智能以及支撑它所需要的创新、技术和人才的最前沿。我的好朋友、优秀的同事 John Shoven 将为 Jensen 做介绍。我希望他就在现场，所以我就先继续说几句，等他们两位展开对话之后，再留些时间给大家提问。
John Shoven 对在座的大多数人来说几乎无需介绍。作为我在 SIEPR 的前任主任，John 正是二十年前创办 SIEPR Economic Summit 的人。我想请大家一起为 John Shoven 送上热烈掌声，也向他当年有远见创建的这个学术共同体致意。
如果你还没有接触过 John 的研究、指导或者友谊，这里简单介绍几项他的成就。除了曾任 SIEPR 主任、现在是荣休资深研究员之外，John 还是 Charles R. Schwab 经济学教授，也是 Hoover Institution 的资深研究员和 National Bureau of Economic Research 的研究合作者。他专攻公共财政与公司金融，长期研究社会保障、健康经济学、企业与个人税收、共同基金、养老金、经济人口学、应用一般均衡经济学等多个领域。
Jensen Huang 的美国梦
英文清理版
So I have always thought that the more famous the speaker, the shorter the appropriate introduction. And if I were to follow that rule, I would stop right now and say Jensen Huang. But I am not going to do that.
The Oxford English Dictionary defines the American dream as a situation where everybody has an equal opportunity for success through hard work, dedication, and initiative. I would like to say that Jensen Huang is an example of the American dream. Jensen was born in Taiwan, came to the US at age nine with his brother, not with his parents, went to a rough school in Kentucky, survived that, and when his parents came two years later he moved to Oregon, skipped two grades, graduated from high school, and went to Oregon State as an electrical engineering major.
He was sixteen, looked like he was twelve, and had no chance with the women. But he liked one of them and said, why don’t we work on homework together? He did that over and over again. Six months later he asked her out on a date, and he is still married to her. So that is another American dream.
Now to skip ahead to age thirty, he co-founded Nvidia. He has been the only CEO Nvidia has ever had. It has had ups and downs, more ups and downs, and now it is the fourth largest company in the world and the third largest American company. That sounds to me like the American dream.
He also got a master’s degree from Stanford. Now of course, we were here last week, and Nvidia announced its earnings in a finance crowd. This got more attention than the Super Bowl and occurred a couple of weeks earlier. It was pretty amazing. His company is at the absolute center of the most exciting technological development of the twenty-first century, and he deserves to be congratulated on that.
He has received a lot of awards and a lot of recognition. Last month, he was elected as a member of the National Academy of Engineering. I asked ChatGPT how many CEOs of S&P 500 companies are members of the National Academy of Engineering, and I did not get an absolutely clear answer, but I think it is three. And two are in this room. Aravind? Arvind? of Cadence Design Systems was awarded it last year, so the two of them have that in common.
中文翻译
我一直觉得，发言人越有名，介绍就应该越简短。按这个标准，我本来现在就可以直接说一句 “Jensen Huang” 然后停下来了，不过我不会这么做。
《牛津英语词典》把“美国梦”定义为这样一种状态：每个人都可以通过勤奋、投入和主动性获得平等的成功机会。我想说，Jensen Huang 就是美国梦的典型。他出生于台湾，九岁时和哥哥一起来到美国，不是和父母一起来的。他先在肯塔基州的一所条件很艰苦的学校上学，后来坚持了下来；两年后父母才过来，他又搬到俄勒冈，跳了两级，顺利读完高中，之后进入俄勒冈州立大学，主修电气工程。
那时他十六岁，看起来像十二岁，在女生面前几乎没有机会。不过他看上了其中一位，就说：“我们要不要一起做作业？”他一遍又一遍地这样约她，六个月后终于约到第一次正式约会，而他现在依然和她结婚生活在一起。所以，这也是另一种美国梦。
再快进到三十岁时，他共同创办了 Nvidia。他一直是 Nvidia 唯一的 CEO。公司一路起起伏伏，但最终成长为全球第四大公司，也是美国第三大公司。对我来说，这就是美国梦。
他后来还拿到了 Stanford 的硕士学位。前不久我们就在这里，Nvidia 在金融圈发布财报，热度甚至超过了超级碗，而且发生得还更早一些，真的非常惊人。他的公司正处在二十一世纪最激动人心的技术变革中心，值得我们向他致敬。
他获得过很多奖项和认可。上个月，他当选为美国国家工程院院士。我还问过 ChatGPT，S&P 500 公司 CEO 里有多少人是国家工程院院士，虽然没有得到特别清楚的答案，但我觉得大概只有三位。而且今天这里就有两位。Cadence Design Systems 的 Arvind 也是去年获得的，所以他们两位在这一点上很有共同点。
计算的转折
英文清理版
So in my lifetime, I thought the biggest technical development, the technology breakthrough, was the transistor. It was pretty fundamental. But should I rethink, is AI now the biggest change in technology that has occurred in the last seventy-six years? That is a hint at my age.
Well, first of all, the transistor was obviously a great invention, but what was the greatest capability that it enabled was software. The ability for humans to express our ideas, algorithms, in a repeatable way, computationally repeatable way, was the breakthrough.
What have we done? We have dedicated our company in the last thirty-one years to a new form of computing called accelerated computing. The idea is that general-purpose computing is not ideal for every field of work. We invented a new way of doing computation. It was really, really good at trying to figure out this form of computation called deep learning. It is really good at this thing called AI.
We said, why don’t we use computers to write this software? Because the computational cost is approximately zero. So you might as well let the computer go off and grind on a massive amount of experience. We call data digital experience. Human digital experience is called data.
We thought if we could reduce the marginal cost of computing down to approximately zero, we might use it to do something insanely amazing: large language models to literally extract all of digital human knowledge from the internet and put it into a computer, then let it figure out what the knowledge is. That idea of scraping the entire internet and putting it in one computer sounds crazy, but that is where we are.
中文翻译
在我的一生中，我一直认为最大的技术突破是晶体管。它当然非常基础、非常重要。但我现在是不是该重新思考一下：AI 会不会才是过去七十六年里最大的技术变革？这句话也顺便暴露了我的年龄。
当然，晶体管毫无疑问是伟大的发明，但它真正启发出的最大能力，是软件。人类能够用可重复、可计算的方式表达自己的想法，也就是算法，这才是关键突破。
那我们这些年都做了什么？过去三十一年里，我们把公司投入到一种叫做 accelerated computing 的新型计算方式上。因为通用计算并不适合所有工作领域，所以我们想发明一种新的计算方法。它特别擅长深度学习这种计算形式，也特别擅长我们今天所说的 AI。
我们说，为什么不能让计算机来写这类软件呢？因为此时的计算成本几乎可以忽略不计。既然如此，不如让计算机去处理海量经验。我们把这种经验称为数据，也就是数字化的经验；而人类的数字化经验，本质上就是数据。
我们当时想，如果能把计算的边际成本降到接近零，也许就能做一些非常惊人的事情，比如用大语言模型把互联网中的数字化人类知识全部提取出来，装进一台计算机里，再让它自己去理解这些知识。把整个互联网抓取下来装进一台电脑里，听上去很疯狂，但我们今天就是在这个方向上前进。
从数据到知识
英文清理版
Gene sequencing is digitizing genes, but now with large language models we can go and understand the meaning of that gene. Amino acids, we digitized through mass spec. You know, you ask it, what is the meaning of it, summarize it for me, what is the meaning? This is no different than a hard, huge long page of genes. What is the meaning of that big long page of proteins? What is the meaning of that? We are on the cusp of all this. This is the miracle of what happened.
中文翻译
基因测序其实就是在把基因数字化，但现在有了大语言模型，我们已经可以进一步理解这个基因到底意味着什么。氨基酸也可以通过质谱技术被数字化。你可以问它：“这是什么意思？给我总结一下。”这和一长页复杂的基因信息并没有本质区别。那一大页蛋白质到底意味着什么？我们现在正站在这一切的门槛上，这就是正在发生的奇迹。
数据中心与未来算力
英文清理版
The chip that John just described weighs seventy pounds. It consists of thirty-five thousand parts. Eight of those parts came from TSMC. That one chip replaces a data center of old CPU systems. The savings because we compute so fast are incredible. It computes at data center scale.
What’s going to happen in the next ten years? The computational capability for machine learning and deep learning will increase by another million times. And what happens when you do that? Today we kind of learn and then we apply it, but in the future the computer will watch videos, read text, and continuously improve itself through the training process, the inference process, and the application process. Those will become one.
The marginal cost of transportation has gone to zero. I can fly from here to New York relatively cheaply, though if it had taken a month I probably would never go. This is exactly the same in computing. We are going to take the marginal cost of computing down to approximately zero, and as a result we will do a lot more computation.
中文翻译
John 刚刚描述的那颗芯片重达七十磅，由三万五千个零件组成，其中八个零件来自 TSMC。这样一颗芯片，实际上就替代了一个由旧式 CPU 系统组成的数据中心。因为计算速度大幅提升，带来的节省是惊人的。它已经是在数据中心规模上运算了。
未来十年会发生什么？机器学习和深度学习的算力还会再提升一百万倍。那会带来什么结果？今天我们通常是先学习，再应用；但在未来，计算机会去看视频、读文本，并在训练、推理和应用过程中不断自我改进，这几个过程将逐渐合一。
运输的边际成本已经降到了接近零。我可以相对便宜地从这里飞到纽约，但如果这趟路程要花一个月，我大概就不会去了。计算也是一样，我们会把计算的边际成本降到接近零，而结果就是，人类会去做更多计算。
推理市场
英文清理版
There are recent stories saying that Nvidia will face more competition in the inference market than it has in the training market, but what we are really talking about is a single market. The question is whether there will be a separate training chip market and inference chip market.
Inference is an installed base problem. This is no different than somebody writing an application on an iPhone. The reason they do so is because the iPhone has such a large installed base. If you wrote an application for that phone, it can benefit everybody. In the case of Nvidia, our accelerated computing platform is the only accelerated computing platform that is literally everywhere.
If you write an application for inference and deploy that model on Nvidia architecture, it literally runs everywhere, so you can touch everybody and have greater impact. The problem with inference is really the installed base, and that takes enormous patience.
中文翻译
最近有人说，Nvidia 在推理市场会面临比训练市场更多的竞争，但从本质上说，我们谈的其实还是同一个市场。问题在于，训练芯片市场和推理芯片市场会不会真的分成两个完全独立的市场。
推理本质上是一个安装基数的问题。这和在 iPhone 上开发应用没有区别。之所以大家愿意为 iPhone 写应用，是因为它有巨大的装机量。你只要为这台手机写一个应用，就有机会服务所有人。对 Nvidia 来说，我们的加速计算平台是事实上无处不在的那一个。
如果你写了一个推理应用，并把模型部署到 Nvidia 架构上，它几乎可以在所有地方运行，这样就能真正触达更多人，产生更大的影响。推理的问题核心，其实是安装基数，而这需要极大的耐心。
科学与结构化数据
英文清理版
SQL was born in the nineteen sixties and IBM in the nineteen seventies in storage computing. Structured data is as important as it gets. There are hundreds of zettabytes of data being created every couple of years, but most of it is in structured databases. Wherever we can accelerate that, we can accelerate quantum physics, Schrödinger equations, fluids, particles, and lots and lots of code. What Nvidia is good at is the general field of accelerated computing.
中文翻译
SQL 诞生于二十世纪六十年代，IBM 在七十年代推动了存储和计算的发展。结构化数据的重要性怎么强调都不为过。每过几年，人类都会创造出数百泽字节的数据，但其中大多数其实都在结构化数据库里。只要我们能把这一部分加速，就能加速量子物理、薛定谔方程、流体、粒子，以及大量代码的计算。Nvidia 擅长的，正是更广义的加速计算。
与客户竞争
英文清理版
I want to apologize. I came across as a little competitive. I could have probably done that more artfully. I will next time. But he surprised me with a competitor. I thought I was in an economic forum. I had sent some questions to his team and asked whether he had looked at them. He said no, because he wanted to be spontaneous. Besides, he might start thinking about it.
中文翻译
我想道个歉。我刚才听起来确实有一点竞争性的味道，这件事我本可以说得更委婉一些，下次我会注意。但他给我准备了一个竞争对手的问题，我本来以为自己是在参加一个经济论坛。我之前还把一些问题发给了他的团队，问他有没有看过，他说没有，因为他想保持即兴发挥。再说，他一旦开始想，可能就会开始认真分析了。
让模型理解世界
英文清理版
When we talk about AI, there are certain properties that it must obey in the world. It has to create what is called a world model. We have to understand multimodality. There are all these other modalities: genes, amino acids, proteins, cells, which lead to organs and so on. We would like to have multimodal capabilities. Second is greater and greater reasoning capabilities. A lot of the things we already do, reasoning skills are encoded in common sense.
We already have questions and answers, but today we are mostly doing generative visualization. I am not spending a whole lot of time reasoning about the question. However, there are certain problems, like planning problems, where I am going, that is interesting, let me think about that. I am cycling it in the back, I am coming up with multiple plans, traversing a tree, maybe going through my graph, pruning my tree and saying this does not make sense.
中文翻译
当我们谈 AI 时，它必须遵守现实世界中的一些属性。它需要建立所谓的“世界模型”。我们也必须理解多模态。除了文本之外，还有很多其他模态，比如基因、氨基酸、蛋白质、细胞，进一步又会导向器官等等。所以我们希望模型具备真正的多模态能力。第二点，是越来越强的推理能力。很多我们已经在做的事情，本质上都把推理技能编码在常识里。
现在我们已经有问答能力了，但今天很多时候我们做的还是生成式内容输出。我并不会在每个问题上都花很多时间去推理。不过对于某些问题，比如规划类问题，我会想：这个有意思，让我想一想。然后我在后台不断循环，生成多个方案，像是在遍历一棵树，也可能是在图里搜索，再把不合理的分支剪掉。
未来的知识交互
英文清理版
It would be great just as you can chat with GPT, you can chat with a PDF. You take a PDF file, it does not matter what it is. My favorite is when you take a PDF file of a research paper and load it into ChatGPT, then start talking to it. It is like talking to the researchers: what inspired this research, what problem does it solve, what was the breakthrough, what was the state of the art before then, what were the novel ideas. Just talk to it like a human.
中文翻译
就像你可以和 GPT 聊天一样，未来你也可以和 PDF 聊天。你拿一个 PDF 文件，不管它是什么都可以。我最喜欢的场景，是把一篇研究论文的 PDF 放进 ChatGPT，然后直接和它对话。那感觉就像在和研究者本人交流：这项研究最初的灵感是什么？它解决了什么问题？关键突破在哪里？在它之前的技术水平是什么？有哪些新想法？你就像和一个人对话一样去问它。
计算成本下降带来的结果
英文清理版
That miracle happened about a decade and a half ago. We saw it coming and took the whole company and shaped our computer, which was already driving the marginal cost of computing down to zero, and pushed it into this whole domain. As a result, in the last ten years, we reduced the cost of computing by one million times, the cost of deep learning by one million times.
A lot of people said to me, “But Jensen, if you reduced the cost of computing by a million times, people buy less of it.” And it is exactly the opposite. We saw that if we could reduce the marginal cost of computing down to approximately zero, we might use it to do something insanely amazing.
Large language models can literally extract all of digital human knowledge from the internet and put it into a computer, and let it go figure out what the knowledge is. That idea of scraping the entire internet and putting it in one computer sounds crazy, but that is where we are.
中文翻译
那个奇迹大约发生在十五年前。我们看到了它的到来，于是把整个公司都动员起来，把我们的计算机体系改造成现在这个样子。它本来就在把计算的边际成本往零压，我们又把它推进到这个全新的领域。结果就是，在过去十年里，我们把计算成本降低了一百万倍，也把深度学习的成本降低了一百万倍。
很多人跟我说：“Jensen，如果你把计算成本降了一百万倍，人们会买得更少吧？”事实正好相反。我们发现，如果能把计算的边际成本降到接近零，就可以拿它去做一些极其惊人的事情。
大语言模型可以真的把互联网上所有数字化的人类知识提取出来，放进一台计算机里，然后让它自己去理解这些知识。把整个互联网抓下来，再装进一台计算机里，这想法听起来很疯狂，但我们现在就在做这件事。
H100 与数据中心
英文清理版
The chip that John just described weighs seventy pounds and consists of thirty-five thousand parts. Eight of those parts came from TSMC. We sell the world’s first quarter-million-dollar chip. But the system it replaced, the cables alone, cost more than the chip. That H100 and the cables connecting all those old computers, that is the incredible thing that we did. We reinvented computing, and as a result the marginal cost of computing went to zero.
We took this entire data center and shrunk it into this one chip. This one chip is really, really good at trying to figure out this form of computation called deep learning. It is really good at this thing called AI. The way that this chip works is magnificent. It weighs a lot, with miles and miles of cables, hundreds of miles of cables, and the next ones are coming as liquid cooled. It computes at data center scale.
中文翻译
John 刚刚描述的那颗芯片重达七十磅，由三万五千个零件组成，其中八个零件来自 TSMC。我们卖的是世界上第一颗二十五万美元的芯片。但它所替代的那套系统，单是线缆的成本就比这颗芯片还高。那块 H100，加上连接所有旧计算机的线缆，就是我们做成的惊人之处。我们重新发明了计算，而结果就是计算的边际成本降到了零。
我们把整个数据中心压缩进了这一颗芯片里。它非常擅长处理一种叫做深度学习的计算形式，也非常擅长我们今天所说的 AI。这颗芯片的工作方式非常了不起。它本身很重，系统里有成千上万米、甚至数百英里的线缆，而下一代很快就会采用液冷。它是在数据中心尺度上运算的。
低预期与公司文化
英文清理版
One of my great advantages is that I have very low expectations. Most Stanford graduates have very high expectations, and you deserve to have high expectations because you came from a great school, you were very successful, you were at the top of your class, and you were surrounded by other incredible students.
In our company, I have learned to talk to people all the time. I write no reviews for any of them; I give them constant reviews, and they provide the same to me. My compensation for them is the bottom right corner of Excel, I just drag it down. Many of our executives are paid the same, exactly that: a dollar. I know it is weird, but it works. I do not do one-on-one meetings with them unless they need me. Then I will drop everything for them. I never have meetings with them just alone.
中文翻译
我最大的优势之一，就是我的期待值很低。Stanford 的大多数毕业生往往期望值非常高，而你们也确实有资格拥有高期待，因为你们来自一所很棒的学校，你们非常成功，在班里总是最优秀的一批，而且身边都是同样出色的人。
在我们公司里，我学会了一直和大家沟通。我不会给他们写正式评价，而是持续不断地给反馈，他们也同样持续给我反馈。我给他们的“薪酬”，就在 Excel 右下角，往下一拖就完了。我们很多高管的薪酬都是一样的，准确地说就是一美元。我知道这听起来很奇怪，但它有效。除非他们真的需要我，否则我不会单独和他们开一对一会议；一旦他们需要我，我会立刻放下手头所有事情去帮他们。我也不会和他们只在私下单独开会。
高期待与韧性
英文清理版
You should have very high expectations. You naturally have very high expectations. But people with very high expectations have very low resilience, and unfortunately resilience matters in success. I do not know how to teach it to you except that I hope suffering happens to you. I was fortunate that I grew up with my parents providing a condition for us to be successful on the one hand, but there were plenty of opportunities for setbacks and suffering. To this day, I use the phrase pain and suffering inside our company with great glee, because that is how you build resilience.
中文翻译
你们本来就应该有很高的期待。你们天然就会有很高的期待。但问题在于，期待很高的人，往往韧性很低，而不幸的是，韧性在成功中非常重要。我也不知道该怎么教你们这一点，除非希望你们真的经历一些苦难。我很幸运，成长过程中父母一方面给了我们成为成功者的条件，另一方面人生里也有很多挫折和受苦的机会。直到今天，我还会在公司里带着一点“开心”地说“pain and suffering”这个词，因为韧性就是这样锻炼出来的。
解释 AI 目标
英文清理版
We are training these models to be multimodal, meaning we will learn from sounds, words, and vision, and we will just watch TV and learn, so on and so forth, just like all of us. The reason why that is so important is because we want AI to be grounded, grounded not just by human values. We had large language models before, but it was not until reinforcement learning from human feedback that the AI was grounded to something that we would call human intelligence.
There are certain properties that AI must obey in the world. It has to create what is called a world model. We have to understand multimodality. There are all these other modalities, like genes, amino acids, proteins, and cells, which lead to organs and so on. We would like to have multimodal capabilities. Second is greater and greater reasoning capabilities. A lot of the things we already do, reasoning skills are encoded in common sense.
We already have questions and answers, but today we are mostly doing generative visualization. I am not spending a whole lot of time reasoning about the question. However, there are certain problems, like planning problems, where I am going, “That is interesting, let me think about that.” I am cycling it in the back, coming up with multiple plans, traversing a tree, maybe going through my graph, pruning my tree, and saying this does not make sense.
That long-thinking AI is not good at today. Everything that you prompt into ChatGPT gets a response instantaneously. We would like to prompt something into ChatGPT, give it a mission statement, give it a problem, and for it to think a while. That kind of system, what computer science calls system two thinking, long thinking, planning, reasoning, those kinds of problems, I think we are working on them, and I think you can see some breakthroughs.
And if you ask me, “Jensen, AI is a list of illicit tests,” the answer is that an engineer can only know whether we have a specification and whether we know what the definition of success is. We are not sure yet how to specify all of human intelligence, and therefore it is hard to achieve as an engineer. But we are endeavoring to make it better and better.
中文翻译
我们正在训练这些模型，让它们具备多模态能力，也就是从声音、文字、视觉中学习，就像我们自己会看电视、听声音、读文字并学习一样。之所以这很重要，是因为我们希望 AI 具有“落地性”，不仅仅由人类价值观去约束。我们以前已经有大语言模型了，但直到出现基于人类反馈的强化学习，AI 才真正被“锚定”到我们可以称之为人类智能的东西上。
AI 必须在现实世界中遵守某些属性。它需要建立所谓的世界模型。我们必须理解多模态。除了这些模态之外，还有很多其他模态，比如基因、氨基酸、蛋白质、细胞，进一步又会导向器官等等。所以我们希望模型拥有真正的多模态能力。第二点，是越来越强的推理能力。我们已经在做的很多事情，本质上都把推理技能编码在常识里。
现在我们已经有问答能力了，但今天很多时候我们做的还是生成式可视化。我不会在每个问题上都花太多时间推理。不过对于某些问题，比如规划类问题，我会想：“这个有意思，让我想一想。”然后我在后台不断循环，生成多个方案，像是在遍历一棵树，也可能是在图里搜索，再把不合理的分支剪掉。
这种长思考型 AI 今天还不够好。你现在往 ChatGPT 里提问，它几乎会立刻返回答案。我们希望未来你给它一个任务、一个问题，它能先想一会儿。这类系统，也就是计算机科学里说的 system two thinking，长思考、规划、推理这一类问题，我们正在推进，也已经能看到一些突破。
如果你问我：“Jensen，AI 是一系列非法测试吗？”那我的回答是，工程师只能知道我们有没有规格说明，能不能定义什么叫成功。我们到现在也还不完全知道该如何完整描述人类智能，所以作为工程目标来说这很难。但我们一直在努力，把它做得越来越好。
长思考与规划
英文清理版
Today, when we prompt ChatGPT, we get a response instantaneously. But in the future we would like to prompt something into ChatGPT, give it a mission statement, give it a problem, and let it think for a while. That kind of system, what computer science calls system two thinking, long thinking, planning, reasoning, those are the kinds of problems we are working on. I think you can already see some breakthroughs.
When I am reasoning about a planning problem, I cycle it in the back, come up with multiple plans, traverse a tree, maybe go through my graph, prune the tree, and say this does not make sense. Then I simulate it in my head and maybe do some calculations and so on. That long-thinking AI is not good at today, but that is the direction we want.
中文翻译
今天我们往 ChatGPT 里提问时，它几乎是立刻给出答案。但未来我们希望，你给它一个任务、一个问题，让它先想一会儿。这类系统，也就是计算机科学里说的 system two thinking，长思考、规划、推理这一类问题，正是我们正在做的。我觉得现在已经能看到一些突破了。
如果我在思考一个规划问题，我会先在后台循环，生成多个方案，像是在遍历一棵树，也可能是在图里搜索，再把不合理的分支剪掉。然后我在脑海里模拟，做一些计算等等。今天的长思考型 AI 还不够成熟，但这正是我们想要前进的方向。
生成式未来
英文清理版
What is written by someone, created by someone, is basically pre-recorded. All the words, all the videos, all the sound, everything that we do is retrieval-based. It was pre-recorded. Every modality that you know today is basically pre-recorded. In the future, almost everything will be generated.
When you and I ask about the economy, we probably mean very different things and in very different contexts. In the future, the system will understand that context and generate exactly the right information for you. Most computing will be generative. Today, one hundred percent of content is pre-recorded. In the future, a hundred percent of content will be generative. The question is how that changes the shape of computing.
中文翻译
今天由人写下、由人制作的内容，本质上都可以算是“预先录制”的。无论是文字、视频还是声音，几乎所有我们现在接触到的内容，都是检索式的、预先存在的。你今天看到的各种模态，本质上也都可以看作是预先录制的内容。而在未来，几乎一切都会变成生成式的。
当你和我问“经济”这个词时，我们心里想的东西其实可能完全不同，所处的语境也完全不同。未来系统会理解这种语境，并为你生成刚刚好的信息。大多数计算都会变成生成式。今天，百分之百的内容都是预先录制的；未来，百分之百的内容都会变成生成式。问题在于，这会如何改变计算的形态。
编程的变化
英文清理版
No, not even a little bit. The number of coders in the world will surely continue to be important, and Nvidia needs coders. However, in the future, the way you interact with the computer is not going to be C++. Mostly for some of us that is true, but for you, why program in Python? In the future, you will tell the computer what you want, and the computer will help you. If you do not like the answer the first time, you can fine-tune it and get better and better results.
You can even ask it to write the program altogether to generate that result in the future. My point is that programming has changed in a way that is probably less valuable. On the other hand, because of artificial intelligence, we have closed the technology divide of humanity today.
中文翻译
不，完全不是这样。世界上的程序员数量依然会很重要，Nvidia 也仍然需要程序员。不过在未来，你和计算机交互的方式不会再是 C++ 这种传统方式了。至少对我们中的一部分人来说确实如此，但对你们来说，为什么还要用 Python 去“编程”呢？未来你只要告诉计算机你想要什么，它就会帮你完成。如果第一次结果不满意，你还可以继续调整，让结果越来越好。
你甚至可以让它直接帮你把程序写出来，从而生成最终结果。我的意思是，编程这件事已经发生了变化，传统编程方式的重要性可能没有以前那么高了。另一方面，人工智能正在帮助我们缩小人类之间的技术鸿沟。
Prompt Engineering
英文清理版
It is called prompt engineering, how you interact with people, how you interact with computers, how you make a computer do what you want it to do, how you fine-tune the instructions with that computer. There is an artistry to that. For example, most people are surprised by this, but if you ask Midjourney to generate a picture of a puppy on a surfboard in Hawaii at sunset, it generates one. Then you say, make it cuter, and it comes back cuter. You say, no, cuter than that, and it comes back again. That capability exists in a computer in the future. If you do not like the answer the first time, you can ask it again in a way that gives you better and better results.
We all can program computers. You all know how to prompt the computer to make it do things. Look at YouTube and look at all the people using prompt engineering. Kids are making amazing things and they are not writing programs. They are just talking to ChatGPT. They know that if I tell it to do this, it will do that. It is no different than interacting with people in the future. That is the great contribution the computer science industry has made to the world. We closed the technology divide.
中文翻译
这叫做 prompt engineering，也就是你如何与人交互、如何与计算机交互、如何让计算机做你想让它做的事，以及如何不断细化对它的指令。这件事本身是有技巧的。比如很多人会惊讶，但如果你让 Midjourney 生成一张“海边冲浪板上的小狗、黄昏时分”的图片，它真的会生成。然后你说“再可爱一点”，它就会变得更可爱；你说“再可爱一点”，它还能继续调整。未来的计算机就会具备这种能力。如果第一次答案不满意，你可以换一种方式继续问，得到越来越好的结果。
我们每个人都可以编程。你们其实都知道怎么通过 prompt 让计算机做事情。去看 YouTube 上那些使用 prompt engineering 的人，很多孩子做出了惊人的东西，但他们并没有在写传统程序，他们只是在和 ChatGPT 对话。他们知道，只要我让它做这个，它就会去做那个。未来这和人与人交互没有区别。这就是计算机科学行业为世界做出的巨大贡献，我们缩小了技术鸿沟。
数字主权
英文清理版
We first of all have to understand these policies and stay agile so that we can comply with them. On the one hand, it limits our opportunity in some places, and it opens up opportunities in others. In the last six to nine months, every single country and every single society has awakened to the fact that they have to control their own digital intelligence. India cannot outsource its data so that some country transforms that digital data into India’s intelligence and imports that intelligence back to India. That awakening is sovereign AI.
中文翻译
首先，我们必须理解这些政策，并保持足够灵活，确保能够遵守它们。一方面，这会限制我们在某些地方的机会；另一方面，它也会在其他地方打开新的机会。过去六到九个月里，每一个国家、每一个社会都意识到，他们必须控制自己的数字智能。印度不能把自己的数据外包出去，让别的国家把这些数字数据转化成印度的智能，再把这种智能“进口”回印度。这种觉醒，就是主权 AI，也就是 sovereign AI。
地缘政治与主权 AI
英文清理版
Geopolitics, on the one hand, limited opportunities, but it created enormous opportunities elsewhere. Every country, every society, now realizes that they have to invest in their own sovereign AI. They have to protect their language, protect their culture, and protect their own industries. That awakening happened in the last six to nine months. The first part was that we have to be mindful about safety, and the second part was that we all have to do this.
中文翻译
地缘政治一方面限制了一些机会，另一方面也在别处创造了巨大的机会。现在每一个国家、每一个社会都意识到，他们必须投资自己的主权 AI。他们必须保护自己的语言、保护自己的文化，也要保护自己的产业。这种觉醒大概发生在过去六到九个月里。第一层意思是我们必须重视安全，第二层意思是我们每个人都必须这么做。
最后问答
英文清理版
There are more horizontal solutions today, but we are willing to customize the answers. The reason the bar is relatively high is because each generation of our platform has a GPU, a CPU, a networking processor, a switch, and two types of switches. I build five chips for one generation, and people think it is one chip, but it is really five different chips. Each one of those chips costs hundreds of millions of dollars just to get to launch, what we call tape-out. Then you have to put them into a system, add networking, transceivers, optics, and a mountain of software. It takes a lot of software to run a computer as big as this room.
If the customization is too different, you have to repeat the entire R&D process. However, if the customization leverages everything and adds something to it, then it makes a great deal of sense. Maybe it is a proprietary security system, maybe it is a confidential computing system, maybe it is a new way of doing numerical processing that could be extended. We are very open-minded to that. Our customers know that I am willing to do all that, and they recognize that if you change it too far, you have basically reset nearly a hundred billion dollars of work that it took us to get here. They want to leverage our ecosystem to the extent that it can be done. I am very open to it.
中文翻译
今天市场上更多是横向的解决方案，但我们也愿意为客户定制答案。之所以门槛很高，是因为我们平台每一代产品都包含 GPU、CPU、网络处理器、交换机，以及两类交换芯片。我们一代产品其实要做五颗芯片，但外界常常以为只有一颗，其实是五颗完全不同的芯片。每一颗芯片从设计到流片，也就是 tape-out，成本都要几亿美元。之后你还得把它们组装成一个系统，再加上网络、收发器、光模块，以及大量软件。要运行一台像这个房间这么大的计算机，所需的软件量非常大。
如果定制需求和原有架构差异太大，那就等于要把整个研发流程重新来一遍。但如果这种定制是在充分利用现有能力的基础上，再增加一些新东西，那就非常有意义。也许是一个专有安全系统，也许是一个保密计算系统，也许是一种新的数值计算方式，而且这种方式还可以继续扩展。对于这些方向，我们都非常开放。我们的客户知道我愿意做这些事情，也明白如果改动得太远，基本上就相当于把我们花了将近一千亿美元才建立起来的生态重新推倒重来。所以他们通常希望尽可能利用我们的生态体系。我对此一直很开放。
结语
英文清理版
I think with that we need to wrap up. Thank you so much to John and Jensen.
中文翻译
我想我们可以就此收尾了。非常感谢 John 和 Jensen。