他是地球上最有权势的工程师。这就是为什么他认为人工智能进展顺利。
He’s the Most Powerful Engineer on Earth. Here’s Why He Thinks A.I. Is Going Great.
在最近的白宫科技峰会上，英伟达首席执行官黄仁勋坐在总统特朗普身旁的椅子上——名副其实的得力助手。那天晚些时候，黄仁勋使用了总统偏爱的“超级智能”一词。在今年夏天的恶意人工智能袭击事件之后，黄仁勋已成为这种技术所构成危险的主要怀疑论者。作为他的传记作家，我理解他为什么不把这种风险放在眼里。我也担心他没有看到全局。
At the recent White House tech summit, Jensen Huang, the chief executive of Nvidia, sat in the chair next to President Trump — his literal right-hand man. Later that day, Mr. Huang used the term “super intelligence,” just as the president prefers. In the aftermath of this summer’s rogue artificial intelligence attacks, Mr. Huang has emerged as a leading skeptic of the danger this technology poses. As his biographer, I understand why he dismisses the risk. I also worry that he’s not seeing the whole picture.
当我第一次采访他时，黄仁勋告诉我，每一个问题——无论是商业的、个人的，还是其他任何问题——实际上都是一个工程问题：你研究输入，你研究限制条件，然后你得到想要的输出。在撰写他的传记时，我看着黄仁勋将这种方法付诸实践。他对其业务进行工程设计，以确保英伟达成为机器人和医疗保健等行业的默认平台。他对自己的产品进行工程设计，使其比竞争对手的更好。而在2025年，他开始对美国总统进行工程设计。
When I first interviewed him, Mr. Huang told me that every problem — business, personal, whatever — is actually an engineering problem: You study the inputs, you study the constraints and then you get the outputs that you want. While writing his biography, I watched Mr. Huang put this approach into action. He engineers his business to ensure that Nvidia will be the default platform in industries like robotics and health care. He engineers his products to be better than the competition’s. And in 2025, he began to engineer the president of the United States.
在特朗普先生之前，黄仁勋对政治并没有表现出太多兴趣。现在特朗普先生和黄仁勋经常通电话，并且似乎有着真正的个人默契——特朗普甚至在阿拉斯加接上他，乘坐空军一号前往中国。特朗普先生想要接近黄仁勋，因为这位总统喜欢赢家——而黄仁勋是最大的赢家。自1993年在一家丹尼斯餐厅创立以来，英伟达已经发展成为美国历史上价值最高的公司，即使把通货膨胀计算在内也是如此。
Before Mr. Trump, Mr. Huang didn’t show much interest in politics. Now Mr. Trump and Mr. Huang talk on the phone often and seem to have a genuine personal rapport — Mr. Trump even picked him up in Alaska to go to China on Air Force One. Mr. Trump wants to be close to Mr. Huang because the president likes winners — and Jensen Huang is the biggest winner of all. Since its inception in a Denny’s restaurant in 1993, Nvidia has grown to become the most valuable company in American history, even taking inflation into account.
驱动这一增长的是人工智能背后空前的数据中心热潮，黄仁勋将这称为“新工业革命”。数据中心可以容纳数千个被称为“机架”的高大金属柜，这些机架将数十枚英伟达微芯片连接成一个模块化计算单元。一个新机架的大小约为一台冰箱，售价约为800万美元，据某些估计，其耗电量可能超过70栋单户住宅。到2026年，英伟达通过销售此类及类似设备将净赚约2500亿美元。
Driving the growth is the unprecedented data center boom behind A.I., which Mr. Huang has termed a “new industrial revolution.”Data centers can hold thousands of tall metal cabinets called “racks” that wire together dozens of Nvidia microchips into a modular computing unit. A new rack, about the size of a refrigerator, can cost around $8 million and, by some estimates, can draw more power than 70 single-family homes. In 2026, Nvidia will net around a quarter of a trillion dollars selling this and similar equipment.
数据中心的热潮让黄仁勋成为了人工智能的皇帝——其他大型科技公司都在借钱来获取他的产品。正是在英伟达的核心业务——人工智能芯片领域，尚无竞争对手能够获得显著的市场份额，这证明了黄仁勋的工程才能。工程师们就是更青睐英伟达。这些自称“系统专家”的人负责在电路板层面上实现人工智能。
The data center boom has made Mr. Huang the emperor of A.I. — the other big tech companies are borrowing money to access his products. It is a testament to Mr. Huang’s engineering prowess that no competitors have yet managed to gain significant market share in Nvidia’s core business of A.I. chips. Engineers simply prefer Nvidia. These are the self-described “systems guys” responsible for implementing A.I. at the circuit-board level.
黄仁勋本人就是一名系统专家，这塑造了他对人工智能的看法。到目前为止，我采访过至少一百名图形处理器（即G.P.U.）工程师，但我还没遇到一个真正担心人工智能失控的人。另一方面，我交谈过的大多数软件工程师都极其担忧。比尔·盖茨最近警告称，人工智能可能会杀死10亿人类。
Mr. Huang is himself a systems guy and that informs his view of A.I. At this point, I’ve interviewed at least a hundred graphics processing unit, or G.P.U., engineers, and I have yet to meet one who was seriously concerned about runaway A.I. On the other hand, most of the software engineers I talk with are extremely concerned. Bill Gates recently warned that A.I. could kill one billion humans.
这种情况令人困惑：这里有两组理性的、高智商的技术专家，他们看着相同的数据，却得出完全相反的结论。这让人想起那张病毒式传播的图片“裙子”，当时有些人看到的是蓝色和黑色，有些人看到的是白色和金色。
The situation is baffling: Here are two groups of rational, high-I.Q. technical experts looking at the same data and reaching totally opposite conclusions. It reminds one of the viral image “The Dress,” where some people saw blue and black, and some people saw white and gold.
以下是我认为正在发生的事情：现代人工智能的基础是神经网络，它直接汲取了生物学的灵感。事实上，现代人工智能由用计算机代码模拟的许多亿个的人工神经元组成。因此，对软件专家而言，现代人工智能看起来有点像一团未分化的脑组织，既神秘又从根本上难以控制。
Here is what I think is going on: The basis of modern A.I. is the neural net, which draws direct inspiration from biology. In fact, a modern A.I. consists of many billions of artificial neurons, simulated in computer code. Thus, to the software experts, a modern A.I. looks something like a mass of undifferentiated brain tissue, mysterious and fundamentally difficult to control.
但这并不是硬件专家们的看法。他们在实际神经元——这是可能不可预测的生物细胞——与合成神经元之间做出了明确区分，后者是简单的数学模型，其行为可以实现完美准确的预测。当你从微芯片自下而上地工作时，保持对人工智能的控制是一个工程问题，它受制于与其他任何事物相同的输入/输出限制。对于人工智能末日论的怀疑者来说，安全的人工智能是一个数学方程式，或者用黄先生的话说，只是“软件”。
But that’s not how the hardware guys see it. They draw a clear distinction between actual neurons — which are biological cells that can be unpredictable — and synthetic neurons, which are simple mathematical models whose behavior can be forecast with perfect accuracy. When you work from the microchip up, maintaining control of A.I. is an engineering problem, that is governed by the same input/output constraints as any other. To the A.I. doomer skeptics, safe A.I. is a math equation — or, in Mr. Huang’s words, just “software.”
黄先生在20世纪80年代初开始学习电子工程时，微芯片上有数万个元件，通常是在纸上通过手工设计。如今，芯片可以拥有超过一千亿个元件，并且是在人工智能的协助下，使用专业软件设计的。63岁的黄先生亲自经历了这种呈指数级增长的功能提升。他那一代的工程师构建了各种工具，以确保极度复杂的微芯片依然是人类能够理解和控制的东西。
When Mr. Huang began studying electrical engineering in the early 1980s, microchips had tens of thousands of components on them and were typically designed by hand on paper. Today, they can have more than a hundred billion components and are designed using specialized software, with assistance from A.I. Mr. Huang, at 63, has personally managed this exponential growth in capability. His generation of engineers built the tools to ensure that the hyper-complex microchip remains something that humans can interpret and control.
作为一名硬件专家，黄先生在亚洲度过的时间比大多数西方科技公司首席执行官都要多。他出生于台湾；他的母语是台语，英伟达正在台北建立一个卫星总部。黄先生还对中国大陆的人工智能开发者持支持态度。我认为这也影响了他的思维，因为东亚对人工智能的乐观态度远超美国。当我在亚洲做演讲活动时，我从不会遇到关于人工智能风险的问题。在美国，这是人们最常提及的话题。
As a hardware guy, Mr. Huang spends more time in Asia than most Western tech C.E.O.s. He was born in Taiwan; his native language is Taiwanese, and Nvidia is building a satellite headquarters in Taipei. Mr. Huang is also supportive of mainland Chinese A.I. developers. I believe this also influences his thinking, as East Asia is much more optimistic about A.I. than the United States is. When I do speaking events in Asia, I never get questions about A.I. risk. In the United States, it’s the most common thing people bring up.
因此，黄先生认为自己能够让人工智能神经元变得安全，或许也就不足为奇了。在他看来，如果你能亲眼看到机器所做的每一次计算，那么这台机器就应该完全在你的掌控之中。如果机器表现异常——好吧，那是你的错。你需要的只是更好的工程技术。他知道这是可以做到的。
So perhaps it is unsurprising that Mr. Huang believes he can make the artificial neuron safe. From his point of view, if you can see every calculation the machine does, right in front of you, then the machine should remain entirely within your control. If the machine is misbehaving — well, that’s your fault. All you need is better engineering. He knows it can be done.
因此，黄先生驳回了所有关于放慢人工智能发展步伐的呼吁，以及任何会阻碍该技术发展的监管规定。“世界末日发生的几率是零，”他在接受哥伦比亚广播公司新闻（CBS News）采访时表示。上个月在高盛（Goldman Sachs）的一场会议上，他暗示网络安全担忧是一种营销策略。“还有什么比制造一个问题更好的方法来创造需求呢？”他问道。他还暗示，仅凭利润动机就足以激励各公司确保人工智能的安全。
Thus Mr. Huang has dismissed all calls for any slowdown in A.I., or any regulations that would impede development of the technology. “There is 0 percent chance that’s going to be the end of the world,” he said in an interview with CBS News. At a Goldman Sachs conference last month, he suggested that cybersecurity concerns were a marketing tactic. “What better way to create demand than to create a problem?” he asked. He also has suggested that the profit motive alone will be sufficient to incentivize companies to make A.I. safe.
黄先生成功地让特朗普重复了他的谈话要点。9月，在洛杉矶举行的“全明星峰会”（All-In Summit）上，特朗普给正在台上的黄先生打了电话。他将总统的电话设为免提，然后将麦克风凑近手机，向观众广播了特朗普的说法，即任何关于人工智能对人类生命构成风险的言论都是一场“骗局”。“机器人不会接管世界，”总统在赢得一阵掌声后说道。
Mr. Huang has succeeded in getting Mr. Trump to parrot his talking points. In September, Mr. Trump called Mr. Huang while he was onstage at the All-In Summit in Los Angeles. He put the president on speaker, then held a microphone up to his phone, broadcasting to the audience Mr. Trump’s claims that any talk of risks to human life from A.I. was a “hoax.”“The robots are not going to be taking over the world,” the president said, following a round of applause.
这场带有精心策划的公开宣传特技之感的通话，是对特朗普先生抗拒人工智能监管的公开宣示。但在其他人看到这位总统采取强硬立场的景象时，我看到的是黄仁勋先生在“操纵”总统。其输入是厚颜无耻的公开奉承：在通话中，黄仁勋称呼特朗普为“先生”，赞美特朗普的智商（“您对人工智能了解得这么多”），并煽动特朗普对认可的无底洞般的需求（“您听到了吗？成千上万的人在为您鼓掌”）。
The phone call, which had the feel of an orchestrated publicity stunt, was a public declaration of Mr. Trump’s resistance to A.I. regulation. But where others saw the president taking a forceful stand, I saw Mr. Huang engineering the president. The input was shameless public flattery: In the phone call, Mr. Huang called Mr. Trump “sir,” praised Mr. Trump’s intellect (“You know so much about A.I.,”) and stoked Mr. Trump’s bottomless need for validation. (“Did you hear that? Thousands of people are clapping for you.”)
而输出呢？是对人工智能的微弱监管、自由出口英伟达的一些硬件到中国，以及黄仁勋想要的任何其他东西。
And the output? Little regulation on A.I. and freedom to export some of Nvidia’s hardware to China, and whatever else Mr. Huang wants.
这种情况让我感到担忧。黄仁勋是我见过最有能力的人——但恐怕，人工智能的能力还要更强。我怀疑，今天就已经存在利用人工智能来增强危险病原体并关闭电网的能力。与此同时，解释和控制人工智能的科学落后于使其变得更聪明的科学数年之久。在细颗粒度层面上，我们可以控制人工神经元的行为，但技术人员目前无法解释这些神经元是如何协同产生智能的。这种理解上的差距正在扩大，而我不知道黄仁勋是否能看到这一点。
This situation worries me. Mr. Huang is the most capable individual I have ever met — but, I fear, A.I. is more capable still. I suspect that the capabilities exist, today, for A.I. to enhance dangerous pathogens and to shut down electrical grids. At the same time, the science of interpreting and controlling A.I. is years behind the science of making it smarter. At a granular level, we can control the behavior of synthetic neurons, but technologists can’t currently explain how these neurons coordinate to produce intelligence. That gap in understanding is widening, and I don’t know if Mr. Huang can see it.
未来，即使在最好的情况下，也会极其令人迷茫。但它也许会非常棒。也许我能活到205岁。无论发生什么，它的发生都将是因为一个杰出之人的决定和指引，这个人主导着人工智能的未来以及许多其他事情。工程学是强大的——而今天，黄仁勋是活着的、最强大的工程师。
The future, even in the best-case scenario, will be extremely disorienting. But maybe it will be awesome. Maybe I will live to be 205. Whatever happens, it will happen because of the decisions and guidance of a singular, extraordinary man who directs the future of A.I. and much else besides. Engineering is powerful — and today, Mr. Huang is the most powerful engineer alive.
斯蒂芬·威特是人工智能巨头英伟达发展史《思维机器》一书的作者。
Stephen Witt is the author of “The Thinking Machine,” a history of the A.I. giant Nvidia.