黄仁勋斯坦福访谈：英伟达创办史与生成式AI的未来

摘要

2024年3月6日，英伟达创始人兼CEO黄仁勋受邀在斯坦福商学院“从顶峰看未来”系列对谈中，回顾了与Chris Malachowsky、Curtis Priem在Denny’s餐厅立下创业誓言的往事，详述公司早期因Direct3D标准差点破产、靠OpenGL手册“自救”的关键转折，并分享“创造技术+创造市场”的战略信条。他强调必须用第一性原理重塑组织和产品，解释自己为何维持约50位直辖高管，以保持信息不被层级“变形”。在面向未来的讨论中，黄仁勋将生成式AI视为计算范式由“检索”走向“生成”的根本转移，并看好数字生物学、计算机辅助药物设计与机器人操作理解等新领域。

关键信息速览

1. Denny’s创业盟约：三位创始人1993年在餐厅讨论“解决普通电脑无法处理的问题”，即从一开始就锁定3D图形这一尚不存在的消费市场。01:21
2. 靠OpenGL起死回生：早期产品与微软Direct3D不兼容、公司濒临破产，黄仁勋购买OpenGL手册重写管线，才让GPU走上正确标准。11:54
3. 创造市场而非争份额：英伟达始终在技术尚无客户时下注，例如深度学习与自动驾驶，充分体现“技术与市场同步创造”的策略。09:52
4. 第一性原理与扁平组织：坚持分解问题、避免类比思维，并以约50位直接汇报对象保持信息原样传递，减少“信息扭曲”。14:12 / 32:57
5. 生成式AI的时代：将过去“检索式计算”转向“生成式计算”，预测软件写法、芯片架构和网络/存储都将因此重塑。39:39
6. 数字生物学愿景：看好计算机辅助药物设计，把理解“细胞的含义”类比为语言模型理解段落；亦关注类人机器人与操作序列的“token化”。48:26

核心概念及解读

加速计算：英伟达的核心理念，即打造专用计算架构来解决通用CPU无法高效处理的问题，涵盖图形渲染、科学模拟和AI训练等场景

第一性原理：黄仁勋反复强调的思维方式，不照搬既有做法，而是从基本原理出发，结合当下条件重新思考如何设计产品与技术

创造市场与创造技术并行：英伟达的战略特征，不仅开发芯片和软件，还主动培育尚不存在的应用市场，如PC游戏、自动驾驶和深度学习

CUDA与可编程着色器：英伟达从图形专用芯片走向通用GPU计算的关键技术转折点，使GPU能够运行多种科学和AI算法

创业融资与信誉积累：黄仁勋通过在LSI Logic的出色工作记录获得推荐，说明职业声誉对创业融资具有决定性影响

主持人： 非常感谢你的到来，能够在这里真是太荣幸了。让我们从你第一次离开的时候开始聊起。关于你加入LSI Logic的记忆，那是一个当时非常令人兴奋的公司，你在科技界建立了卓越的声誉。然而你决定离开，成为一个创始人。是什么激励了你呢？是Chris和Curtis吗？

黄仁勋： 我是LS Logic的工程师，而Chris和Curtis在Sun公司工作。我当时与一些计算机科学领域最聪明的头脑一起工作，包括Andyto Shim等人。我们当时正在构建工作站和图形工作站等等。Chris和Curtis有一天说他们想要离开Sun公司，他们希望我加入他们，一起弄清楚他们将要离开的方向。我那时有一份很好的工作，但他们坚持要我和他们一起弄清楚如何建立一家公司。我们在Denny’s餐厅见面，那是我的母校附近的一家餐厅，也是我第一份工作——洗碗工的地方。我做得很好，无论如何，我们聚在一起，那是在微处理器革命期间，大约是1993年和1992年。PC革命才刚刚开始，显然，Windows 95这个革命性的Windows版本还没有上市，Pentium处理器甚至还没有宣布。所以，这一切都发生在PC革命之前，很明显微处理器将会非常重要。我们认为，为什么不建立一家公司来解决普通计算机无法解决的问题，即由通用计算驱动的计算机无法解决的问题。这就是公司的任务，直到今天，我们仍然专注于此。如果你看看我们打开的所有问题和市场，比如计算药物设计、天气模拟、材料设计，这些都是我们非常自豪的事情。

机器人技术、自动驾驶汽车、自主软件，我们称之为人工智能。然后我们如此努力地推动技术，最终计算成本降到了几乎为零，这使得一种全新的软件开发方式成为可能，即计算机自己编写软件，也就是我们今天所知的人工智能。这就是我们的旅程。

主持人： 这些应用今天都在我们脑海中，但回到过去，当时的市场甚至还不存在。我是如何说服Don Valentine投资的？

黄仁勋： LSI Logic的CEO说服他最大的投资者Don Valentine与你会面，他显然是Sequoia的创始人。现在我看到很多创始人都急切地向前倾，但你是如何说服硅谷最受追捧的投资者投资于一个由首次创业者组成的团队，他们正在为一个甚至不存在的市场打造一个新产品？

我当时不知道如何写商业计划，所以我去了书店。那时候还有书店，在商业书籍区有一本书，作者是我认识的人Gordon Bell。这本书非常具体，是关于如何写商业计划的，它是为非常小众的市场写的。看起来他就像是为像我这样的人写的，我是其中之一。所以我买了这本书，我应该立刻意识到这是个坏主意，因为Gordon非常非常聪明，聪明的人有很多话要说，他们想要教我如何写商业计划。我拿起这本书，它有450页长，我甚至没有读完，我只翻了几页，我想，等我读完这本书，我的生意就完了，我会没钱了。Lori和我银行里只有大约6个月的存款，我们已经有了Spencer、Madison和一只狗。所以我们五个人必须靠银行里的钱生活，我没有太多时间。所以，我没有写商业计划，而是直接去找了W Coran。有一天他打电话给我说，嘿，你离开了公司，你甚至没有告诉我你在做什么，我想让你回来向我解释一下。所以我回去向他解释了，他听完后说，我不知道你在说什么，这是我听过的最糟糕的电梯推销之一。然后他拿起电话，给Don Valentine打电话，他说，Don，我想让你给这个孩子钱，他是LSI Logic有史以来最好的员工之一。所以我发现的一件事是，你可以进行一次很棒的面试，甚至可以有一次糟糕的面试，但你无法逃避你的过去。所以，拥有一个好的过去很重要。你知道，当我说我是个好洗碗工时，我是认真的。我可能是Denny’s最好的洗碗工。我计划我的工作，我很有组织，我是Misan plus，然后我洗出了盘子上的污垢，然后你知道，他们提升我做巴士男孩，我确信我是Denny’s最好的巴士男孩。我从不空手离开一个站点，我从不空手回来，我非常高效。然后不管怎样，最终我成为了CEO。我仍在努力成为一个好CEO，但说到坏，你是如何决定接下来做什么的？

主持人： 我们需要成为89家获得资金的公司中最好的一家，这些公司都在做同样的事情，然后只剩下6到9个月的运营资金，你意识到最初的愿景就是行不通。在这种情况下，你是如何决定下一步如何拯救公司的，当时所有的牌都对你不利？

黄仁勋： 我们成立了这家公司，叫做加速计算，问题是它用来做什么的，什么是杀手级应用？我们的第一个重大决策就是，第一个杀手级应用将是3D图形，技术将是3D图形，应用将是视频游戏。那时，3D图形是不可能便宜的，它是数百万美元的图像生成器，来自硅谷图形公司，视频游戏市场是零亿美元。所以你有这种令人难以置信的技术，很难商品化和商业化，然后你有一个不存在的市场，这就是我们公司的起点。这个市场不存在，这是我们公司的起点。我仍然记得，当Don在我演讲结束时说，你知道，他对我说的其中一件事，这在当时很有意义，今天仍然很有意义，他说，初创公司不应该投资初创公司，或者初创公司不应该与初创公司合作。他的意思是，为了使NVIDIA成功，我们需要另一家初创公司成功，那家公司就是电子艺界（Electronic Arts）。然后他出门时提醒我，电子艺界的首席技术官只有14岁，需要他的妈妈开车送他上班。然后他提醒我，我就是依靠这样的人，那是谁，然后他说，如果你亏了我的钱，我会杀了你。那是我对第一次会议的记忆，但尽管如此，我们创造了一些东西，我们继续在接下来的几年里创造了PC游戏市场。这花了很长时间才做到，我们今天仍在做这件事。我们意识到，你不仅要创造技术，还要发明一种新的计算机图形学方式，让原本价值一百万美元的东西现在只值三四百美元，可以装进电脑里。你还要创造这个新市场，所以我们必须创造技术，创造市场。这个想法定义了NVIDIA今天，几乎我们所做的一切都是创造技术，创造市场。这就是为什么人们说我们有一个，你知道，人们称之为一个堆栈，一个生态系统。这些词汇，但基本上就是这样，在过去的30年里，NVIDIA意识到我们必须做的是，为了创造条件，让人们可以购买我们的产品，我们必须发明这个新市场。

这就是为什么我们在自动驾驶领域早早就开始了，这也是为什么我们在深度学习领域早早就开始了，这也是为什么我们在包括计算药物设计和发现在内的所有这些不同领域都早早开始了。我们试图在创造市场的同时创造技术，所以这就是，嗯，好吧。然后我们继续前进，然后微软引入了一个叫做Direct3D的标准。这催生了数百家公司，我们发现自己几年后与几乎每个人竞争，而我们发明的技术，3D图形，消费者化的3D，结果与Direct3D不兼容，所以我们开始了这家公司。我们有这个3D图形的东西，我们试图让它变得消费者化。我们发明了所有这些技术，然后不久之后它变得不兼容，呃，所以我们不得不重置公司或者破产，但我们不知道如何按照微软定义的方式来构建它。我记得有一次在周末的会议上，讨论的是我们现在有89个竞争对手。我明白我们的做法是不对的，但我们不知道如何正确地做，幸运的是，还有另一家书店，书店叫做Fry’s Electronics，我不知道它是否还在这里。所以我带着我的女儿开车去了Fry’s。我在那里看到了OpenGL手册，它定义了硅图形公司是如何做计算机图形的。就像68美元一本书，所以我花了几块钱买了三本书，带回办公室，我说，伙计们，我找到了我们的未来。我把它交给了和我一起创办公司的那几个天才，我们像从未有人实现过OpenGL管线一样实现了它。我们建立了世界上从未见过的东西，所以很多教训就在那里，那一刻给了我们公司巨大的信心。这就是为什么，即使在你完全不了解的情况下，你也可以成功地发明未来。我的态度对所有事情都是这样，当有人告诉我关于某件事，我从未听说过，或者我听说过但从不理解它是如何工作的，我的第一反应总是，你知道，有多难呢？可能只是一本教科书的距离，你可能只是一篇档案论文的距离就能弄清楚这一点，所以我现在花了很多时间阅读档案论文，嗯，这是真的，你当然不能学习别人是如何做某件事的，然后以完全相同的方式去做，并希望有不同的结果。但你可以学习某件事是如何完成的，然后回到第一原则，问自己，鉴于今天的条件，鉴于我的动力，鉴于工具和仪器，鉴于事物的变化，我会如何重新做这件事？如果今天我来设计一辆汽车，我会怎么做？如果今天我来建造一台电脑，我会怎么做？如果今天我来写软件，我会怎么做？这有道理吗？

所以我总是回到第一原则。即使在公司内部，我们也会不断重置自己，因为世界已经改变了。我们过去编写软件的方式是整体式的，它是为超级计算机设计的，但现在它是分散的，等等。我们今天如何看待软件？我们今天如何看待计算机？我们总是因为这样的原因回到第一原则，它创造了很多机会。是的，你将技术应用的方式变成了…

主持人： 你是如何决定要进行革命性的转变的？你在IPO时获得了所需的所有动力，然后在成功的过程中，你决定要稍微改变一下Nvidia的创新重点，这都源于你与一位化学教授的电话交谈。你能告诉我们那个电话是怎么回事，你是如何从你所听到的内容中连接到你最终去向的？

黄仁勋: 记住，公司的核心是开创一种新的计算方式。计算机图形学是第一个应用，但我们总是知道会有其他应用。图像处理来了，粒子物理来了，流体力学等等。所以我们让处理器更具可编程性，以便我们可以表达更多的算法。然后有一天，我们发明了可编程着色器，这使得所有形式的成像和计算机图形学都变得可编程，这是一个巨大的突破。我们在那之上发明了，我们尝试寻找更多复杂算法的表达方式，这些算法可以在我们的处理器上计算，这与CPU非常不同。所以我们创建了这个东西，叫做CG，我想那是2003年左右，CG代表GPUs，它比Cuda早了大约三年。写这本教科书救了公司的人，Mark Hilgard，也写了那本书。CG非常酷，我们为此写了教科书，开始教人们如何使用它，我们开发了工具等等。然后许多研究人员发现了它，很多在斯坦福的学生，很多后来成为Nvidia工程师的工程师都在玩弄它。麻省总医院的几位医生拿起它用于CT重建。我飞过去看他们，问他们，你们用这个东西做什么？他们告诉我关于那件事，然后，一位计算量子化学家用它来表达他的算法。所以我意识到，有一些证据表明人们可能想要使用这个，这给了我们越来越多的信心，我们应该去做这件事，这种计算形式可以解决普通计算机无法解决的问题。这加强了我们的信念，并让我们继续前进，每次你听到一些新的东西。

主持人： 你真的很享受那个惊喜，这似乎是你在Nvidia领导过程中的一个主题。你似乎总是提前做出技术转折的赌注，当苹果最终从树上掉下来时，你就穿着黑色皮夹克站在那里等着接住它。你是如何找到那个机会的？

黄仁勋: 你做的事情似乎总是基于核心信念，你知道我们深信我们可以创造一台计算机，解决常规处理无法解决的问题。CPU能做什么是有限制的，通用计算能做什么也是有限制的。然后还有一些有趣的问题我们可以去解决。问题是，这些问题只是有趣的问题，还是它们也可以是有趣的市场？如果它们不是有趣的市场，那就不可持续。Nvidia经历了大约十年的时间，我们在投资这个未来，市场并不存在。那时只有一个市场，就是计算机图形学。对于10到15年的时间，今天推动Nvidia的市场并不存在，所以…你如何继续，呃，与所有围绕你的人，你的公司，你知道，Nvidia的管理团队，所有与你一起创造未来的惊人工程师，你的所有股东，你的董事会，你的所有合作伙伴，你带着所有人一起前进，没有证据，这真的非常具有挑战性。技术可以解决问题，你有一些研究论文，因为有了它而成为可能，这很有趣，但你总是在寻找市场，尽管如此，在市场存在之前，你仍然需要早期的成功指标。你知道我们公司有一句话叫做关键绩效指标（KPIs）。不幸的是，KPIs很难理解，我发现KPIs很难理解。什么是好的KPI？你知道，当我们寻找KPIs时，我们寻找的是毛利率。那不是KPI，那是结果。你在寻找的是早期指标，预示着未来积极的结果。尽可能早地寻找，因为你想早期迹象表明你正在朝着正确的方向前进。所以我们有这个说法叫做EO ifs FS，你知道早期指标（Early Indicators）eFS（早期成功指标）。这有助于人们，因为我一直在使用它，给公司带来希望，嘿，我们解决了这个问题，我们解决了那个问题，我们解决了那个问题。市场并不存在，但有重要的问题，这就是公司的目的，我们想要解决这些问题。我们想要可持续，因此市场最终必须存在，但你想要尽可能早地找到指标，表明你正在做正确的事情。所以，这就是你如何解决投资于非常非常遥远的东西的问题，并且有信念坚持在路上，找到尽可能早的指标，表明你正在做正确的事情。所以，呃，这就是你如何解决这个问题的。

你想要找到的早期指标是什么，早期指标已经被Nvidia的产品团队使用了？各种各样的，嗯，我看到了一篇论文，在我看到这个论文之前，我遇到了一些人，他们需要我的帮助。他们需要我们创建一种特定领域的语言，这样他们的所有算法都可以轻松地表达在我们的处理器上，我们创建了这个东西叫做cdnn，它本质上是神经网络计算的SQL。SQL在存储计算中，这是神经网络计算的SQL。我们为他们创建了一种特定领域的语言，你知道，就像深度学习的OpenGL。他们需要我们这样做，这样他们就可以表达他们的数学，而他们不理解Cuda，但他们理解深度学习，所以我们为他们做了这件事。我们之所以这样做，是因为即使这些研究人员没有钱，这也是我们公司的一项伟大技能，即使财务回报完全不存在，或者非常遥远，只要你相信它，我们就会问自己，这是值得做的工作吗？这是否推进了某个重要领域的科学？注意，这是我从一开始就一直在谈论的事情。我们从工作的重要意义中找到灵感，而不是从市场的规模中找到灵感，因为工作的重要意义是未来市场的早期指标。没有人需要为它写一个商业案例，没有人需要向我展示一个财务预测，唯一的问题是，这是重要的工作吗？如果我们不做，它会在没有我们的情况下发生吗？如果我们不做，而某件事可以在没有我们的情况下发生，那会给我巨大的喜悦。你能想象世界变得更好，而你不需要动一根手指吗？那是终极懒惰的定义。在很多方面，你希望那种习惯，因为在很多方面，你希望公司对其他人总是做的事情保持懒惰，如果有人能做到，让他们去做吧。我们应该选择那些如果我们不做，世界就会崩溃的事情。你必须说服自己，如果我不做这件事，它就不会完成。那是Inc，如果那项工作很难，那项工作有影响力，那项工作重要，那么它就给了你一种目的感，这样说有道理吗？所以，我们公司选择了这些项目，深度学习只是其中之一。第一个成功的迹象是这个，你知道，模糊的猫，Andrew Ng想出了这个，然后Alex Krizhevsky检测到了猫，你知道，不是一直成功，但足够成功，以至于它可能会带我们去某个地方。然后我们推理了深度学习的结构，你知道，我们是计算机科学家，我们理解事物是如何运作的。我们说服了自己，这可能会改变一切。无论如何，这就是一个例子。所以你做出的选择，它们付出了巨大的回报，无论是字面上的还是比喻上的。但你必须引导公司度过一些非常具有挑战性的时刻，比如当它在金融危机中失去了80%的市值，因为华尔街不相信你对ML的赌注。在这样的时候，你如何引导公司，让员工保持对当前任务的动力？

这是我在那个时候的反应，和我今天早些时候问你的这个星期的反应是一样的。这个星期和上个星期或上上个星期没有什么不同。所以，当你的股价下跌80%时，情况有点尴尬，好吧。当你的股价下跌80%时，你不想起床，不想离开房子，这些都是真的，但然后你回到做你的工作。我醒来的时间是一样的，我以同样的方式优先安排我的一天，我回到我的核心，回到你相信的最重要的事情，然后勾选它们。有时候，这很有帮助，你知道，家人爱我，好的，勾选。你知道，一切都对了，所以你就勾选了，然后你回到你的核心，然后回到工作。然后每次对话都回到核心，让公司专注于核心。你相信吗？股价变了，但还有其他东西改变了吗？物理改变了吗？重力改变了吗？我们假设的，我们相信的，导致我们做出决定的所有事情，有任何事情改变了吗？如果那些事情改变了，你必须改变一切。但如果没有任何东西改变，你就什么也不改变，继续前进。

就是这样，你怎么做。在与员工交谈时，他们说你尽量避免公开演讲。他们说过，包括员工在内的领导力，我只是个玩笑，不，领导必须被看到，不幸的是，这是困难的部分。你知道，我我曾是一个电子工程学生，我上学时还很年轻，我16岁就上大学了。所以，我做每件事都很年轻，我有点内向，你知道，我害羞，我不喜欢公开演讲。我很高兴来到这里，我不是说嗯，但这不是我自然而然会做的事情，所以当事情变得困难时，呃，呃，这不容易站在你最关心的人面前。你知道，公司会议上，我们的股票价格刚刚下跌了80%，作为CEO，我最重要的任务是来面对你们，解释这一切，部分原因是你不确定原因，部分原因是你不确定会持续多久，是吧，你只知道这些。这些事情，但你必须面对他们，面对所有这些人，你知道他们在想什么，你知道他们在想什么，但你必须面对他们，你必须做这个艰难的工作。他们可能在想这些事情，但你知道他们在想什么，但你仍然必须面对他们，和他们打交道。你知道他们在想什么，但你必须面对他们，和他们打交道。你知道，你的领导团队中没有一个人在这样的时期离开，事实上，我提醒他们，我只是一个玩笑，我被天才包围着。我被天才包围着，不可思议的Nvidia，众所周知，拥有世界上最优秀的管理团队，这是世界上见过的最深的技术管理团队，我被一群天才包围着，他们只是天才，业务团队、市场团队、销售团队，只是不可思议，工程团队，不可思议，研究团队。

主持人： 不可思议，你的员工说你的领导风格非常投入，你有50个直接下属，你鼓励组织中的所有人向你发送他们心中的前五件事，你不断提醒人们，没有任务是低于你的。你能告诉我们为什么你有意设计了这样一个扁平化的组织结构，以及我们应该如何思考我们未来设计的组织吗？

黄仁勋： 对我来说，没有任务是低于我的，因为记住，我曾经是个洗碗工，我意思是，我曾经清洁过厕所，我清洁过很多厕所，我清洁过的厕所比你们所有人加起来都多。[笑声] 我不知道，我不知道该告诉你什么，你知道，那是生活，所以，你不能给我展示一个任务，那是低于我的。现在我不会做它，不是因为它低于我或者不低我。如果你送我一些东西，你希望我对此提出意见，我可以为你提供服务，我在你身上做出了贡献，我在审查它时与你分享了我是如何思考的，我已经做出了贡献，我已经让你看到了我是如何思考的。我已经让你看到了我是如何推理一个非常模糊的事情的方式。这就是你如何推理一个无法计算的事情的方式，这就是你如何推理一个看起来非常可怕的事情的方式，这就是你如何看起来的方式，你明白吗？所以我总是向人们展示如何推理事物，战略事物，你如何预测一些事情，如何分解问题。你只是一个赋能者，到处都是，所以这就是我如何看待它的，如果你送我东西，你希望我帮助审查它，我会尽我所能，我会向你展示我将如何做它，我在过程中。当然，我从你那里学到了很多东西，你给了我很多信息，我学到了很多东西，所以我觉得我通过这个过程得到了回报，它确实需要很多能量。有时候，因为你要知道，为了给某人增加价值，他们起点非常聪明，我被一群非常聪明的人包围着，你至少要达到他们的水平，你知道，你必须至少达到他们的思考空间，这真的很难，这真的很难。这需要大量的情感和智力能量，所以我感觉在处理这样的事情之后筋疲力尽。我被很多伟大的人包围着，一个CEO应该有最多的直接报告代表，嗯，通过定义，因为CEO所拥有的知识。信息据说如此宝贵，如此机密，你只能和两个人或三个人分享，他们的信息如此宝贵，如此机密，他们只能和几个人分享。嗯，我不相信在一个文化环境中，你所拥有的信息是你拥有权力的原因。我希望我们所有人都能为公司做出贡献，我们在公司中的地位应该与我们通过复杂事物的能力有关。领导其他人实现伟大，激励其他人，支持其他人，这些是管理团队存在的原因，为所有在公司工作的人创造条件，让他们能够做出他们生命中的工作，这就是使命，你知道你可能听说过我非常清楚地说这个。我说过你，我相信我的工作非常简单，就是为你创造条件，让你能够做出你生命中的工作，所以我该怎么办？这个条件应该是什么样子？这个条件应该让你有大量的赋能。你只能在理解情况的情况下被赋能，不是吗？你必须理解你所处的环境，才能产生伟大的想法，所以我必须创造一个环境，让你理解情况，这意味着你必须了解。你必须了解情况，最好的了解方式是信息之间的层级尽可能少，这样才不会有信息的扭曲。这就是为什么我经常在这样的场合中说，首先，这是事实，这些是我们拥有的数据，这是我如何推理的。这些是一些假设，这些是一些未知数，这些是一些已知数，所以你推理它，现在你创造了一个高度赋能的组织，Nvidia的30,000人，我们是最小的大公司，我们是一个小公司。但每个员工都非常赋能，他们每天都在代表我做出明智的决定，原因就是他们了解他们了解我的情况，他们了解我的情况，我非常透明地与人沟通。我相信我可以把信息托付给你，信息有时很难听，情况很复杂，但我相信你能够处理它，你是成年人，你知道，你在这里，你可以处理这个，有时候他们不是真的成年人，他们刚刚毕业。我刚刚毕业时几乎还是个成年人，我很幸运，有人信任我，给了我重要的信息，所以我想这样做，我想创造条件让人们能够做到这一点，我现在想谈谈。

提问： 上周你谈到的生成性AI，你说生成性AI和加速计算已经达到了临界点，随着这项技术变得更加主流，你最兴奋的应用是什么？

黄仁勋： 你必须回到第一原则，问自己什么是生成性AI。发生了什么？我们现在已经有能力让软件理解事物，它们可以理解为什么。 你知道，首先，我们数字化了一切，比如基因测序，你数字化了基因，但那串基因序列意味着什么？我们数字化了氨基酸，但那意味着什么？所以我们现在已经有能力通过大量的学习和数据，以及它们的模式和关系，我们不仅理解了这些事物的含义，我们还可以在它们之间进行翻译，因为我们在同一个上下文中学习了关于这些事情的含义。我们没有分别学习它们，所以我们发现了它们之间的相关性，它们都是相互关联的。所以现在我们不仅理解了每种模态的含义，我们可以在它们之间进行翻译，所以显而易见的事情，你可以将视频字幕翻译成文本，这就是字幕。文本到图像的旅程，文本到文本的聊天GPT，令人惊奇的事情。所以我们现在知道我们理解了含义，我们可以翻译，翻译某件事就是生成信息。突然间，你必须退后一步，问自己，这对我们所做的每一层意味着什么？所以我现在在你面前推理，我在你面前推理，就像我在15年前的一个季度，当我第一次看到AlexNet，大约13、14年前，我如何推理它。我看到的是什么，它有多有趣，它能做什么，非常酷，但最重要的是，它对每一层计算意味着什么，因为你们在计算世界中，所以它意味着我们未来处理信息的方式将会根本不同。这就是Nvidia建造的，你知道芯片和系统，我们未来编写软件的方式将会根本不同，我们将能够编写的软件类型将会不同，处理这些应用程序的方式也会不同。 历史上，我们是基于检索的模型，信息是预先录制的，如果你愿意的话，我们预先编写了文本，预先录制了信息，然后基于一些推荐系统算法检索它。在未来，一些信息的种子将是起点，我们称之为提示，然后我们生成其余的部分。所以，未来的计算将会高度生成性。让我给你举个例子，比如说我们现在正在进行的对话，我传达给你的信息中，非常少的部分是检索的，它叫做智能。所以在未来，我们的计算机将会以这种方式运行，它将会是高度生成性的，而不是高度基于检索的。你回到过去，你问自己，对于企业家来说，你必须问自己，哪些行业将会被颠覆？我们还会以同样的方式考虑网络吗？我们还会以同样的方式考虑存储吗？我们还会像今天这样滥用互联网流量吗？可能不会。注意我们现在正在进行的对话，我每次问题都要上车，我们不必像以前那样滥用信息传输。什么会变得更多，什么会变得更少，你会问自己，整个工业分布将会是什么样子？所以你可以问自己，什么将会被颠覆，什么将会不同，什么将会得到新的开始？所以这个推理从正在发生的事情开始，什么是基于生成性AI的基础？发生了什么？回到所有事物的第一原则。

你开始从基础原则出发，所有的事情都是这样。有一个问题是，什么是Nvidia？Nvidia建立了一个组织，这个组织是为了让我们能够更好地构建我们所构建的东西。所以如果我们都在构建不同的东西，为什么我们要以相同的方式组织？为什么这个组织机器会完全相同，不管你在构建什么，这没有意义。你构建计算机，你以这种方式组织；你提供医疗服务，你也以完全相同的方式组织；这完全没有意义。所以你必须回到基础原则，问自己，输入是什么，输出是什么，这个环境的特性是什么？你知道，这个森林是什么，这个动物必须生活在什么样的环境中？这个特性是什么，它是稳定的，大多数时间你都在试图挤出最后一滴水分，还是它一直在变化，被所有人攻击？

所以你必须理解，作为CEO，你的工作是构建这家公司，这是我的首要工作，创造条件，让你们能够做你们一生的工作。架构必须是正确的，所以你必须回到基础原则，思考这些事情。我很幸运，当我29岁的时候，我有机会退后一步，问自己，你知道，如果我要为未来构建这家公司，它会是什么样子？

你知道，什么是操作系统，我们鼓励什么样的行为，增强什么，我们不增强什么？

主持人： 无论如何，我想节省时间给观众提问，但今年的“从顶峰看未来”的主题是重新定义明天。我们问了我们所有的嘉宾一个问题，Jensen，作为Nvidia的联合创始人和CEO，如果你闭上眼睛，神奇地改变明天的一件事，那会是什么？

黄仁勋： 我们被要求提前考虑这个问题。我会给你一个可怕的答案。我不知道，你看，有很多事情我们无法控制。你的工作是做出独特的贡献，过一个有目的的生活，做一些没有人能做的事情。所以，当你做完后，每个人都会说，世界因为你而变得更好。所以我认为，对我来说，我这样生活，我向前看，然后回头看。所以你们问我的问题，正好是从计算机视觉的角度来看的，正好相反。我从不从我现在的位置向前看，我向前看，然后回头看。这就是为什么，想象一下Nvidia，做出独特的贡献，推动计算的未来，这是人类最重要的工具，现在它不是关于我们自大，但这就是我们擅长的。这是非常非常难做到的，我们相信我们可以做出绝对独特的贡献，这已经花了我们31年的时间来到这里，我们仍然只是开始我们的旅程。所以这是非常非常难做到的。

当我现在回头看，我相信我们会因为做了一件非常非常难做的事情而被记住，不是因为我们出去改变了一切，而是因为我们做了这一件事情，这是非常非常难做的，我们非常擅长做，我们喜欢做，我们做了很长时间。我相信我们会因为这件事情而改变一切，不是因为我们走出去改变了一切，而是因为我们做了这一件事情，这是非常非常难做的，我们非常擅长做，我们喜欢做，我们做了很长时间。

提问： 我是GSP Lead的成员，我将在2023年毕业。所以我的问题是，你如何看待你的公司在未来十年的发展，你认为你的公司将面临什么挑战，你是如何为这些挑战做好准备的？

黄仁勋： 首先，我可以告诉你，当我思考这个问题时，我脑海中闪过的挑战列表是如此之大。我试图弄清楚该选哪一个。嗯，事实的真相是，当你问这个问题时，对我来说，大多数挑战都是技术挑战，因为那是我的早晨。如果你昨天问我，可能是市场创造挑战。有一些市场我非常非常渴望创造，我们已经做了很久了，但我们还是做不到。Nvidia是一家技术平台公司，我们在这里服务于其他公司，这样他们就能通过我们实现我们的希望和梦想。所以，有一些我非常希望发生的事情，我非常希望生物学的世界能够达到40年前芯片设计的世界，计算机辅助设计和电子设计自动化（EDA）这个行业真正使得我们今天的世界成为可能。我相信我们将使得他们的明天成为可能，计算机辅助药物设计，因为我们现在能够代表基因和蛋白质，甚至细胞现在非常接近能够代表和理解一个细胞的含义。一个细胞意味着什么？这有点像我们如何理解一个段落的含义。如果我们能够像理解一个段落那样理解一个细胞，想象一下我们能做什么。所以，所以我很期待这件事发生。我有点兴奋，我知道我们即将在这方面取得突破。例如，类人机器人非常接近，原因是如果你能将语音标记化并理解它，为什么你不能将操作标记化并理解它？所以这些计算机科学技术，一旦你弄清楚了，你就会问自己，如果我能做到这一点，为什么我不能做到那一点？所以我很兴奋这些事情，所以这个挑战是一个快乐的挑战。当然，其他一些挑战当然是工业和地缘政治的，它们是社会和…但你们已经听说过所有这些事情了。这些都是真的，你知道，世界上的社会问题，世界上的地缘政治问题，为什么我们不能相处得好一些？为什么我要说出这些事情，然后放大它们？为什么我们要在世界上评判人们这么多？你知道所有这些事情，你不必让我再说一遍。

提问： 我的名字是Jose，我是2023年的班级成员，来自GSB。我的问题是，你是否担心我们发展AI的速度？你认为我们需要某种形式的监管吗？

黄仁勋： 是的，答案是肯定的和否定的。我们需要，你知道，现代AI最伟大的突破当然是深度学习，它带来了巨大的进步。但另一个令人难以置信的突破是人类一直在实践的东西，我们刚刚为语言模型发明了它，叫做强化学习人类反馈。我每天提供强化学习人类反馈，这就是我的工作。对于他们的父母在房间里，你们每天都在提供强化学习人类反馈。现在我们刚刚弄清楚如何在系统层面上为人工智能做到这一点。生成如何遵守物理定律的标记，现在的事物在空间中漂浮，做着各种事情，它们并不遵守物理定律。这需要技术护栏，需要技术微调，需要技术对齐，需要技术安全。飞机之所以如此安全，是因为所有的自动驾驶系统都被多样性和冗余所包围。所有种类的功能安全和主动安全系统都被发明出来，我需要所有这些尽快被发明出来。你也知道你，网络安全和人工智能之间的界限将变得越来越模糊。我们需要在网络安全领域尽快推进技术，以保护我们免受人工智能的侵害。所以，以很多种方式，我们需要技术更快地发展。好的，监管，有两种类型的监管。有社会监管，我不知道该怎么办，但有产品和服务监管，我知道该怎么办。所以，FAA、FDA、NHTSA，你说出来吧，所有这些都有针对特定用例的产品和服务的监管。嗯，律师资格考试，医生资格考试，你们都有资格认证考试，你们都必须达到一定的标准，你们都必须持续认证，会计师等等，无论是产品还是服务，都有很多监管。请不要添加一个超级监管，横跨所有这些监管。监管会计的人不应该是对医生进行监管的人，你知道我喜欢会计师，但如果我需要进行开心手术，他们能够关闭账本，这很有趣，但这还不够。如果监管会计的人来监管医生，那就有问题了。我希望所有这些已经拥有产品和服务的领域也能在AI的背景下增强他们的监管，但我也遗漏了一个非常重要的问题，那就是AI的社会影响。你如何处理这个问题？我没有很好的答案，但你知道，足够多的人正在讨论这个问题，但这很重要。将所有这些细分成块，这样做是有意义的，这样我们就不会过度关注这一件事，而忽略了我们可以做的许多日常工作。结果人们因为汽车和飞机而死亡，这没有意义。我们应该确保我们做正确的事情。

主持人： 好吧，非常实用。最后，我可以问一个问题吗？我们有一些快速提问要问你，作为“从顶峰看未来”的传统。

黄仁勋： 好的，我试图避开这个。

主持人： 好吧，你的第一个工作是在Denny’s，他们现在有一个专门给你的展位。你在那里工作的美好回忆是什么？

黄仁勋： 我的第二份工作是在AMD，顺便说一下，那里有一个专门给我的展位吗？我爱我的工作，那是一家很棒的公司。

主持人： 如果你必须写一本书，你会给它起什么名字？

黄仁勋： 我不会写一本，你在问我一个假设性的问题，这是不可能的。

主持人： 如果你能分享一条离别的建议，你会对斯坦福广播什么？

黄仁勋： 这不是一个字，但我会每天进行核心信念检查，全力以赴追求它，用你爱的人包围自己，带上他们一起追求正确的道路。这就是Nvidia的故事。

主持人： Jensen，这最后一小时非常愉快，非常感谢你抽出时间。谢谢你的分享，非常感谢。

访谈稿中译稿

Intro 0:00 [音乐] Jensen，非常感谢你的到来 0:06 能够在这里真是太荣幸了，感谢你为了回到斯坦福大学而来到这里。我决定我们从你第一次离开的时候开始聊起 0:12 关于你加入LSI Logic的记忆，那是一个当时非常令人兴奋的公司，你在科技界建立了卓越的声誉 0:18 然而你决定离开，成为一个创始人。是什么激励了你呢？是Chris和Curtis吗？Chris和Curtis…… 0:32 我是LS Logic的工程师，而Chris和Curtis在Sun公司工作。我当时与一些计算机科学领域最聪明的头脑一起工作，包括Andyto Shim等人。我们当时正在构建工作站和图形工作站等等。 0:46 Chris和Curtis有一天说他们想要离开Sun公司，他们希望我加入他们，一起弄清楚他们将要离开的方向。我那时有一份很好的工作，但他们坚持要我和他们一起弄清楚如何建立一家公司。 0:54 我们在Denny’s餐厅见面，那是我的母校附近的一家餐厅，也是我第一份工作——洗碗工的地方。我做得很好，无论如何，我们聚在一起，那是在微处理器革命期间，大约是1993年和1992年。 1:01 PC革命才刚刚开始，显然，Windows 95这个革命性的Windows版本还没有上市，Pentium处理器甚至还没有宣布。所以，这一切都发生在PC革命之前，很明显微处理器将会非常重要。 1:16 我们认为，为什么不建立一家公司来解决普通计算机无法解决的问题，即由通用计算驱动的计算机无法解决的问题。这就是公司的任务，直到今天，我们仍然专注于此。如果你看看我们打开的所有问题和市场，比如计算药物设计、天气模拟、材料设计，这些都是我们非常自豪的事情。 2:07 机器人技术、自动驾驶汽车、自主软件，我们称之为人工智能。然后我们如此努力地推动技术，最终计算成本降到了几乎为零，这使得一种全新的软件开发方式成为可能，即计算机自己编写软件，也就是我们今天所知的人工智能。这就是我们的旅程。

2:30 这些应用今天都在我们脑海中，但回到过去，当时的市场甚至还不存在。我是如何说服Don Valentine投资的？ 3:28 LSI Logic的CEO说服他最大的投资者Don Valentine与你会面，他显然是Sequoia的创始人。现在我看到很多创始人都急切地向前倾，但你是如何说服硅谷最受追捧的投资者投资于一个由首次创业者组成的团队，他们正在为一个甚至不存在的市场打造一个新产品？ 3:48 我当时不知道如何写商业计划，所以我去了书店。那时候还有书店，在商业书籍区有一本书，作者是我认识的人Gordon Bell。这本书非常具体，是关于如何写商业计划的，它是为非常小众的市场写的。 4:24 看起来他就像是为像我这样的人写的，我是其中之一。所以我买了这本书，我应该立刻意识到这是个坏主意，因为Gordon非常非常聪明，聪明的人有很多话要说，他们想要教我如何写商业计划。 4:45 我拿起这本书，它有450页长，我甚至没有读完，我只翻了几页，我想，等我读完这本书，我的生意就完了，我会没钱了。Lori和我银行里只有大约6个月的存款，我们已经有了Spencer、Madison和一只狗。 5:08 所以我们五个人必须靠银行里的钱生活，我没有太多时间。所以，我没有写商业计划，而是直接去找了W Coran。有一天他打电话给我说，嘿，你离开了公司，你甚至没有告诉我你在做什么，我想让你回来向我解释一下。 5:27 所以我回去向他解释了，他听完后说，我不知道你在说什么，这是我听过的最糟糕的电梯推销之一。然后他拿起电话，给Don Valentine打电话，他说，Don，我想让你给这个孩子钱，他是LSI Logic有史以来最好的员工之一。 6:04 所以我发现的一件事是，你可以进行一次很棒的面试，甚至可以有一次糟糕的面试，但你无法逃避你的过去。所以，拥有一个好的过去很重要。你知道，当我说我是个好洗碗工时，我是认真的。我可能是Denny’s最好的洗碗工。 6:25 我计划我的工作，我很有组织，我是Misan plus，然后我洗出了盘子上的污垢，然后你知道，他们提升我做巴士男孩，我确信我是Denny’s最好的巴士男孩。 6:43 我从不空手离开一个站点，我从不空手回来，我非常高效。然后不管怎样，最终我成为了CEO。我仍在努力成为一个好CEO，但说到坏，你是如何决定接下来做什么的？

7:00 我们需要成为89家获得资金的公司中最好的一家，这些公司都在做同样的事情，然后只剩下6到9个月的运营资金，你意识到最初的愿景就是行不通。在这种情况下，你是如何决定下一步如何拯救公司的，当时所有的牌都对你不利？ 7:16 我们成立了这家公司，叫做加速计算，问题是它用来做什么的，什么是杀手级应用？我们的第一个重大决策就是，第一个杀手级应用将是3D图形，技术将是3D图形，应用将是视频游戏。 7:31 那时，3D图形是不可能便宜的，它是数百万美元的图像生成器，来自硅谷图形公司，视频游戏市场是零亿美元。所以你有这种令人难以置信的技术，很难商品化和商业化，然后你有一个不存在的市场，这就是我们公司的起点。 8:11 这个市场不存在，这是我们公司的起点。我仍然记得，当Don在我演讲结束时说，你知道，他对我说的其中一件事，这在当时很有意义，今天仍然很有意义，他说，初创公司不应该投资初创公司，或者初创公司不应该与初创公司合作。 8:41 他的意思是，为了使NVIDIA成功，我们需要另一家初创公司成功，那家公司就是电子艺界（Electronic Arts）。然后他出门时提醒我，电子艺界的首席技术官只有14岁，需要他的妈妈开车送他上班。 9:08 然后他提醒我，我就是依靠这样的人，那是谁，然后他说，如果你亏了我的钱，我会杀了你。那是我对第一次会议的记忆，但尽管如此，我们创造了一些东西，我们继续在接下来的几年里创造了PC游戏市场。 9:24 这花了很长时间才做到，我们今天仍在做这件事。我们意识到，你不仅要创造技术，还要发明一种新的计算机图形学方式，让原本价值一百万美元的东西现在只值三四百美元，可以装进电脑里。 9:45 你还要创造这个新市场，所以我们必须创造技术，创造市场。这个想法定义了NVIDIA今天，几乎我们所做的一切都是创造技术，创造市场。这就是为什么人们说我们有一个，你知道，人们称之为一个堆栈，一个生态系统。这些词汇，但基本上就是这样，在过去的30年里，NVIDIA意识到我们必须做的是，为了创造条件，让人们可以购买我们的产品，我们必须发明这个新市场。

9:56 这就是为什么我们在自动驾驶领域早早就开始了，这也是为什么我们在深度学习领域早早就开始了，这也是为什么我们在包括计算药物设计和发现在内的所有这些不同领域都早早开始了。我们试图在创造市场的同时创造技术，所以这就是，嗯，好吧。然后我们继续前进，然后微软引入了一个叫做Direct3D的标准。 10:05 这催生了数百家公司，我们发现自己几年后与几乎每个人竞争，而我们发明的技术，3D图形，消费者化的3D，结果与Direct3D不兼容，所以我们开始了这家公司。我们有这个3D图形的东西，我们试图让它变得消费者化。 11:01 我们发明了所有这些技术，然后不久之后它变得不兼容，呃，所以我们不得不重置公司或者破产，但我们不知道如何按照微软定义的方式来构建它。我记得有一次在周末的会议上，讨论的是我们现在有89个竞争对手。 11:20 我明白我们的做法是不对的，但我们不知道如何正确地做，幸运的是，还有另一家书店，书店叫做Fry’s Electronics，我不知道它是否还在这里。所以我带着我的女儿开车去了Fry’s。 11:47 我在那里看到了OpenGL手册，它定义了硅图形公司是如何做计算机图形的。就像68美元一本书，所以我花了几块钱买了三本书，带回办公室，我说，伙计们，我找到了我们的未来。 12:18 我把它交给了和我一起创办公司的那几个天才，我们像从未有人实现过OpenGL管线一样实现了它。我们建立了世界上从未见过的东西，所以很多教训就在那里，那一刻给了我们公司巨大的信心。 12:57 这就是为什么，即使在你完全不了解的情况下，你也可以成功地发明未来。我的态度对所有事情都是这样，当有人告诉我关于某件事，我从未听说过，或者我听说过但从不理解它是如何工作的，我的第一反应总是，你知道，有多难呢？ 13:30 可能只是一本教科书的距离，你可能只是一篇档案论文的距离就能弄清楚这一点，所以我现在花了很多时间阅读档案论文，嗯，这是真的，你当然不能学习别人是如何做某件事的，然后以完全相同的方式去做，并希望有不同的结果。 13:56 但你可以学习某件事是如何完成的，然后回到第一原则，问自己，鉴于今天的条件，鉴于我的动力，鉴于工具和仪器，鉴于事物的变化，我会如何重新做这件事？如果今天我来设计一辆汽车，我会怎么做？如果今天我来建造一台电脑，我会怎么做？如果今天我来写软件，我会怎么做？这有道理吗？

14:08 所以我总是回到第一原则。即使在公司内部，我们也会不断重置自己，因为世界已经改变了。我们过去编写软件的方式是整体式的，它是为超级计算机设计的，但现在它是分散的，等等。我们今天如何看待软件？我们今天如何看待计算机？我们总是因为这样的原因回到第一原则，它创造了很多机会。是的，你将技术应用的方式变成了…

14:53 你是如何决定要进行革命性的转变的？你在IPO时获得了所需的所有动力，然后在成功的过程中，你决定要稍微改变一下Nvidia的创新重点，这都源于你与一位化学教授的电话交谈。你能告诉我们那个电话是怎么回事，你是如何从你所听到的内容中连接到你最终去向的？

15:00 记住，公司的核心是开创一种新的计算方式。计算机图形学是第一个应用，但我们总是知道会有其他应用。图像处理来了，粒子物理来了，流体力学等等。所以我们让处理器更具可编程性，以便我们可以表达更多的算法。 15:32 然后有一天，我们发明了可编程着色器，这使得所有形式的成像和计算机图形学都变得可编程，这是一个巨大的突破。我们在那之上发明了，我们尝试寻找更多复杂算法的表达方式，这些算法可以在我们的处理器上计算，这与CPU非常不同。 16:02 所以我们创建了这个东西，叫做CG，我想那是2003年左右，CG代表GPUs，它比Cuda早了大约三年。写这本教科书救了公司的人，Mark Hilgard，也写了那本书。CG非常酷，我们为此写了教科书，开始教人们如何使用它，我们开发了工具等等。 16:36 然后许多研究人员发现了它，很多在斯坦福的学生，很多后来成为Nvidia工程师的工程师都在玩弄它。麻省总医院的几位医生拿起它用于CT重建。 17:04 我飞过去看他们，问他们，你们用这个东西做什么？他们告诉我关于那件事，然后，一位计算量子化学家用它来表达他的算法。 17:29 所以我意识到，有一些证据表明人们可能想要使用这个，这给了我们越来越多的信心，我们应该去做这件事，这种计算形式可以解决普通计算机无法解决的问题。这加强了我们的信念，并让我们继续前进，每次你听到一些新的东西。

17:50 你真的很享受那个惊喜，这似乎是你在Nvidia领导过程中的一个主题。你似乎总是提前做出技术转折的赌注，当苹果最终从树上掉下来时，你就穿着黑色皮夹克站在那里等着接住它。你是如何找到那个机会的？

18:15 你做的事情似乎总是基于核心信念，你知道我们深信我们可以创造一台计算机，解决常规处理无法解决的问题。CPU能做什么是有限制的，通用计算能做什么也是有限制的。然后还有一些有趣的问题我们可以去解决。问题是，这些问题只是有趣的问题，还是它们也可以是有趣的市场？如果它们不是有趣的市场，那就不可持续。 18:54 Nvidia经历了大约十年的时间，我们在投资这个未来，市场并不存在。那时只有一个市场，就是计算机图形学。对于10到15年的时间，今天推动Nvidia的市场并不存在，所以… 19:02 你如何继续，呃，与所有围绕你的人，你的公司，你知道，Nvidia的管理团队，所有与你一起创造未来的惊人工程师，你的所有股东，你的董事会，你的所有合作伙伴，你带着所有人一起前进，没有证据，这真的非常具有挑战性。 19:33 技术可以解决问题，你有一些研究论文，因为有了它而成为可能，这很有趣，但你总是在寻找市场，尽管如此，在市场存在之前，你仍然需要早期的成功指标。你知道我们公司有一句话叫做关键绩效指标（KPIs）。 19:52 不幸的是，KPIs很难理解，我发现KPIs很难理解。什么是好的KPI？你知道，当我们寻找KPIs时，我们寻找的是毛利率。那不是KPI，那是结果。你在寻找的是早期指标，预示着未来积极的结果。尽可能早地寻找，因为你想早期迹象表明你正在朝着正确的方向前进。 20:27 所以我们有这个说法叫做EO ifs FS，你知道早期指标（Early Indicators）eFS（早期成功指标）。这有助于人们，因为我一直在使用它，给公司带来希望，嘿，我们解决了这个问题，我们解决了那个问题，我们解决了那个问题。市场并不存在，但有重要的问题，这就是公司的目的，我们想要解决这些问题。 20:59 我们想要可持续，因此市场最终必须存在，但你想要尽可能早地找到指标，表明你正在做正确的事情。所以，这就是你如何解决投资于非常非常遥远的东西的问题，并且有信念坚持在路上，找到尽可能早的指标，表明你正在做正确的事情。所以，呃，这就是你如何解决这个问题的。

21:04 你想要找到的早期指标是什么，早期指标已经被Nvidia的产品团队使用了？各种各样的，嗯，我看到了一篇论文，在我看到这个论文之前，我遇到了一些人，他们需要我的帮助。 21:49 他们需要我们创建一种特定领域的语言，这样他们的所有算法都可以轻松地表达在我们的处理器上，我们创建了这个东西叫做cdnn，它本质上是神经网络计算的SQL。SQL在存储计算中，这是神经网络计算的SQL。 22:16 我们为他们创建了一种特定领域的语言，你知道，就像深度学习的OpenGL。他们需要我们这样做，这样他们就可以表达他们的数学，而他们不理解Cuda，但他们理解深度学习，所以我们为他们做了这件事。 22:42 我们之所以这样做，是因为即使这些研究人员没有钱，这也是我们公司的一项伟大技能，即使财务回报完全不存在，或者非常遥远，只要你相信它，我们就会问自己，这是值得做的工作吗？ 23:11 这是否推进了某个重要领域的科学？注意，这是我从一开始就一直在谈论的事情。我们从工作的重要意义中找到灵感，而不是从市场的规模中找到灵感，因为工作的重要意义是未来市场的早期指标。 23:31 没有人需要为它写一个商业案例，没有人需要向我展示一个财务预测，唯一的问题是，这是重要的工作吗？如果我们不做，它会在没有我们的情况下发生吗？如果我们不做，而某件事可以在没有我们的情况下发生，那会给我巨大的喜悦。 23:51 你能想象世界变得更好，而你不需要动一根手指吗？那是终极懒惰的定义。在很多方面，你希望那种习惯，因为在很多方面，你希望公司对其他人总是做的事情保持懒惰，如果有人能做到，让他们去做吧。 24:15 我们应该选择那些如果我们不做，世界就会崩溃的事情。你必须说服自己，如果我不做这件事，它就不会完成。那是Inc，如果那项工作很难，那项工作有影响力，那项工作重要，那么它就给了你一种目的感，这样说有道理吗？所以，我们公司选择了这些项目，深度学习只是其中之一。 24:39 第一个成功的迹象是这个，你知道，模糊的猫，Andrew Ng想出了这个，然后Alex Krizhevsky检测到了猫，你知道，不是一直成功，但足够成功，以至于它可能会带我们去某个地方。然后我们推理了深度学习的结构，你知道，我们是计算机科学家，我们理解事物是如何运作的。 25:03 我们说服了自己，这可能会改变一切。无论如何，这就是一个例子。所以你做出的选择，它们付出了巨大的回报，无论是字面上的还是比喻上的。但你必须引导公司度过一些非常具有挑战性的时刻，比如当它在金融危机中失去了80%的市值，因为华尔街不相信你对ML的赌注。 25:27 在这样的时候，你如何引导公司，让员工保持对当前任务的动力？这是我在那个时候的反应，和我今天早些时候问你的这个星期的反应是一样的。这个星期和上个星期或上上个星期没有什么不同。 25:51 这个星期和上个星期或上上个星期没有什么不同。所以，当你的股价下跌80%时，情况有点尴尬，好吧。当你的股价下跌80%时，你不想起床，不想离开房子，这些都是真的，但然后你回到做你的工作。 26:26 我醒来的时间是一样的，我以同样的方式优先安排我的一天，我回到我的核心，回到你相信的最重要的事情，然后勾选它们。有时候，这很有帮助，你知道，家人爱我，好的，勾选。你知道，一切都对了，所以你就勾选了，然后你回到你的核心，然后回到工作。 27:00 然后每次对话都回到核心，让公司专注于核心。你相信它吗？股价变了，但还有其他东西改变了吗？物理改变了吗？重力改变了吗？我们假设的，我们相信的，导致我们做出决定的所有事情，有任何事情改变了吗？如果那些事情改变了，你必须改变一切。但如果没有任何东西改变，你就什么也不改变，继续前进。

27:25 就是这样，你怎么做。在与员工交谈时，他们说你尽量避免公开演讲。他们说过，包括员工在内的领导力，我只是个玩笑，不，领导必须被看到，不幸的是，这是困难的部分。你知道，我我曾是一个电子工程学生，我上学时还很年轻，我16岁就上大学了。 27:52 所以，我做每件事都很年轻，我有点内向，你知道，我害羞，我不喜欢公开演讲。我很高兴来到这里，我不是说嗯，但这不是我自然而然会做的事情，所以当事情变得困难时，呃，呃，这不容易站在你最关心的人面前。 28:26 你知道，公司会议上，我们的股票价格刚刚下跌了80%，作为CEO，我最重要的任务是来面对你们，解释这一切，部分原因是你不确定原因，部分原因是你不确定会持续多久，是吧，你只知道这些。 28:52 这些事情，但你必须面对他们，面对所有这些人，你知道他们在想什么，你知道他们在想什么，但你必须面对他们，你必须做这个艰难的工作。他们可能在想这些事情，但你知道他们在想什么，但你仍然必须面对他们，和他们打交道。 29:14 你知道他们在想什么，但你必须面对他们，和他们打交道。你知道，你的领导团队中没有一个人在这样的时期离开，事实上，我提醒他们，我只是一个玩笑，我被天才包围着。 29:33 我被天才包围着，不可思议的Nvidia，众所周知，拥有世界上最优秀的管理团队，这是世界上见过的最深的技术管理团队，我被一群天才包围着，他们只是天才，业务团队、市场团队、销售团队，只是不可思议，工程团队，不可思议，研究团队。 30:00 不可思议，你的员工说你的领导风格非常投入，你有50个直接下属，你鼓励组织中的所有人向你发送他们心中的前五件事，你不断提醒人们，没有任务是低于你的。 30:21 你能告诉我们为什么你有意设计了这样一个扁平化的组织结构，以及我们应该如何思考我们未来设计的组织吗？对我来说，没有任务是低于我的，因为记住，我曾经是个洗碗工，我意思是，我曾经清洁过厕所，我清洁过很多厕所，我清洁过的厕所比你们所有人加起来都多。 30:58 [笑声] 我不知道，我不知道该告诉你什么，你知道，那是生活，所以，你不能给我展示一个任务，那是低于我的。现在我不会做它，不是因为它低于我或者不低我。 31:09 如果你送我一些东西，你希望我对此提出意见，我可以为你提供服务，我在你身上做出了贡献，我在审查它时与你分享了我是如何思考的，我已经做出了贡献，我已经让你看到了我是如何思考的。 31:25 我已经让你看到了我是如何思考的，通过推理，赋能你，你去哦，我的天哪，这就是你如何推理这件事的方式，它看起来并不像看起来那么复杂，这就是你如何推理一个非常模糊的事情的方式。 31:41 这就是你如何推理一个无法计算的事情的方式，这就是你如何推理一个看起来非常可怕的事情的方式，这就是你如何看起来的方式，你明白吗？所以我总是向人们展示如何推理事物，战略事物，你如何预测一些事情，如何分解问题。 32:00 你只是一个赋能者，到处都是，所以这就是我如何看待它的，如果你送我东西，你希望我帮助审查它，我会尽我所能，我会向你展示我将如何做它，我在过程中。 32:17 当然，我从你那里学到了很多东西，你给了我很多信息，我学到了很多东西，所以我觉得我通过这个过程得到了回报，它确实需要很多能量。 32:28 有时候，因为你要知道，为了给某人增加价值，他们起点非常聪明，我被一群非常聪明的人包围着，你至少要达到他们的水平，你知道，你必须至少达到他们的思考空间，这真的很难，这真的很难。 32:48 这需要大量的情感和智力能量，所以我感觉在处理这样的事情之后筋疲力尽。我被很多伟大的人包围着，一个CEO应该有最多的直接报告代表，嗯，通过定义，因为CEO所拥有的知识。 33:00 信息据说如此宝贵，如此机密，你只能和两个人或三个人分享，他们的信息如此宝贵，如此机密，他们只能和几个人分享。 33:26 嗯，我不相信在一个文化环境中，你所拥有的信息是你拥有权力的原因。我希望我们所有人都能为公司做出贡献，我们在公司中的地位应该与我们通过复杂事物的能力有关。 33:57 领导其他人实现伟大，激励其他人，支持其他人，这些是管理团队存在的原因，为所有在公司工作的人创造条件，让他们能够做出他们生命中的工作，这就是使命，你知道你可能听说过我非常清楚地说这个。 34:30 我说过你，我相信我的工作非常简单，就是为你创造条件，让你能够做出你生命中的工作，所以我该怎么办？这个条件应该是什么样子？这个条件应该让你有大量的赋能。 34:48 你只能在理解情况的情况下被赋能，不是吗？你必须理解你所处的环境，才能产生伟大的想法，所以我必须创造一个环境，让你理解情况，这意味着你必须了解。 35:01 你必须了解情况，最好的了解方式是信息之间的层级尽可能少，这样才不会有信息的扭曲。这就是为什么我经常在这样的场合中说，首先，这是事实，这些是我们拥有的数据，这是我如何推理的。 35:27 这些是一些假设，这些是一些未知数，这些是一些已知数，所以你推理它，现在你创造了一个高度赋能的组织，Nvidia的30,000人，我们是最小的大公司，我们是一个小公司。 35:52 但每个员工都非常赋能，他们每天都在代表我做出明智的决定，原因就是他们了解他们了解我的情况，他们了解我的情况，我非常透明地与人沟通。 36:03 我相信我可以把信息托付给你，信息有时很难听，情况很复杂，但我相信你能够处理它，你是成年人，你知道，你在这里，你可以处理这个，有时候他们不是真的成年人，他们刚刚毕业。 36:22 我刚刚毕业时几乎还是个成年人，我很幸运，有人信任我，给了我重要的信息，所以我想这样做，我想创造条件让人们能够做到这一点，我现在想谈谈。 36:42 上周你谈到的生成性AI，你说生成性AI和加速计算已经达到了临界点，随着这项技术变得更加主流，你最兴奋的应用是什么？

36:59 你必须回到第一原则，问自己什么是生成性AI。发生了什么？我们现在已经有能力让软件理解事物，它们可以理解为什么。你知道，首先，我们数字化了一切，比如基因测序，你数字化了基因，但那串基因序列意味着什么？我们数字化了氨基酸，但那意味着什么？ 37:25 所以我们现在已经有能力通过大量的学习和数据，以及它们的模式和关系，我们不仅理解了这些事物的含义，我们还可以在它们之间进行翻译，因为我们在同一个上下文中学习了关于这些事情的含义。 37:52 我们没有分别学习它们，所以我们发现了它们之间的相关性，它们都是相互关联的。所以现在我们不仅理解了每种模态的含义，我们可以在它们之间进行翻译，所以显而易见的事情，你可以将视频字幕翻译成文本，这就是字幕。 38:18 文本到图像的旅程，文本到文本的聊天GPT，令人惊奇的事情。所以我们现在知道我们理解了含义，我们可以翻译，翻译某件事就是生成信息。 38:36 突然间，你必须退后一步，问自己，这对我们所做的每一层意味着什么？所以我现在在你面前推理，我在你面前推理，就像我在15年前的一个季度，当我第一次看到AlexNet，大约13、14年前，我如何推理它。 38:55 我看到的是什么，它有多有趣，它能做什么，非常酷，但最重要的是，它对每一层计算意味着什么，因为你们在计算世界中，所以它意味着我们未来处理信息的方式将会根本不同。 39:21 这就是Nvidia建造的，你知道芯片和系统，我们未来编写软件的方式将会根本不同，我们将能够编写的软件类型将会不同，处理这些应用程序的方式也会不同。 39:39 历史上，我们是基于检索的模型，信息是预先录制的，如果你愿意的话，我们预先编写了文本，预先录制了信息，然后基于一些推荐系统算法检索它。 40:02 在未来，一些信息的种子将是起点，我们称之为提示，然后我们生成其余的部分。所以，未来的计算将会高度生成性。 40:23 让我给你举个例子，比如说我们现在正在进行的对话，我传达给你的信息中，非常少的部分是检索的，它叫做智能。 40:30 所以在未来，我们的计算机将会以这种方式运行，它将会是高度生成性的，而不是高度基于检索的。你回到过去，你问自己，对于企业家来说，你必须问自己，哪些行业将会被颠覆？ 40:43 我们还会以同样的方式考虑网络吗？我们还会以同样的方式考虑存储吗？我们还会像今天这样滥用互联网流量吗？可能不会。 40:50 注意我们现在正在进行的对话，我每次问题都要上车，我们不必像以前那样滥用信息传输。什么会变得更多，什么会变得更少，你会问自己，整个工业分布将会是什么样子？ 41:16 所以你可以问自己，什么将会被颠覆，什么将会不同，什么将会得到新的开始？所以这个推理从正在发生的事情开始，什么是基于生成性AI的基础？发生了什么？回到所有事物的第一原则。

41:22 你开始从基础原则出发，所有的事情都是这样。有一个问题是，什么是Nvidia？Nvidia建立了一个组织，这个组织是为了让我们能够更好地构建我们所构建的东西。所以如果我们都在构建不同的东西，为什么我们要以相同的方式组织？

42:36 为什么这个组织机器会完全相同，不管你在构建什么，这没有意义。你构建计算机，你以这种方式组织；你提供医疗服务，你也以完全相同的方式组织；这完全没有意义。所以你必须回到基础原则，问自己，输入是什么，输出是什么，这个环境的特性是什么？

43:11 你知道，这个森林是什么，这个动物必须生活在什么样的环境中？这个特性是什么，它是稳定的，大多数时间你都在试图挤出最后一滴水分，还是它一直在变化，被所有人攻击？

43:33 所以你必须理解，作为CEO，你的工作是构建这家公司，这是我的首要工作，创造条件，让你们能够做你们一生的工作。

43:38 架构必须是正确的，所以你必须回到基础原则，思考这些事情。我很幸运，当我29岁的时候，我有机会退后一步，问自己，你知道，如果我要为未来构建这家公司，它会是什么样子？

43:51 你知道，什么是操作系统，我们鼓励什么样的行为，增强什么，我们不增强什么？无论如何，我想节省时间给观众提问，但今年的“从顶峰看未来”的主题是重新定义明天。

44:10 我们问了我们所有的嘉宾一个问题，Jensen，作为Nvidia的联合创始人和CEO，如果你闭上眼睛，神奇地改变明天的一件事，那会是什么？

44:35 我们被要求提前考虑这个问题。我会给你一个可怕的答案。我不知道，你看，有很多事情我们无法控制。你的工作是做出独特的贡献，过一个有目的的生活，做一些没有人能做的事情。 45:03 所以，当你做完后，每个人都会说，世界因为你而变得更好。所以我认为，对我来说，我这样生活，我向前看，然后回头看。所以你们问我的问题，正好是从计算机视觉的角度来看的，正好相反。

45:32 我从不从我现在的位置向前看，我向前看，然后回头看。这就是为什么，想象一下Nvidia，做出独特的贡献，推动计算的未来，这是人类最重要的工具，现在它不是关于我们自大，但这就是我们擅长的。 46:09 这是非常非常难做到的，我们相信我们可以做出绝对独特的贡献，这已经花了我们31年的时间来到这里，我们仍然只是开始我们的旅程。所以这是非常非常难做到的。

46:28 当我现在回头看，我相信我们会因为做了一件非常非常难做的事情而被记住，不是因为我们出去改变了一切，而是因为我们做了这一件事情，这是非常非常难做的，我们非常擅长做，我们喜欢做，我们做了很长时间。

46:45 我相信我们会因为这件事情而改变一切，不是因为我们走出去改变了一切，而是因为我们做了这一件事情，这是非常非常难做的，我们非常擅长做，我们喜欢做，我们做了很长时间。我是GSP的领导，我将在2023年毕业。

46:53 所以我的问题是，你如何看待你的公司在未来十年的发展，你认为你的公司将面临什么挑战，你是如何为这些挑战做好准备的？首先，我可以告诉你，当我思考这个问题时，我脑海中闪过的挑战列表是如此之大。 47:20 我试图弄清楚该选哪一个。嗯，事实的真相是，当你问这个问题时，对我来说，大多数挑战都是技术挑战，因为那是我的早晨。如果你昨天问我，可能是市场创造挑战。 47:44 有一些市场我非常非常渴望创造，我们已经做了很久了，但我们还是做不到。Nvidia是一家技术平台公司，我们在这里服务于其他公司，这样他们就能通过我们实现我们的希望和梦想。 48:01 所以，有一些我非常希望发生的事情，我非常希望生物学的世界能够达到40年前芯片设计的世界，计算机辅助设计和电子设计自动化（EDA）这个行业真正使得我们今天的世界成为可能。 48:26 我相信我们将使得他们的明天成为可能，计算机辅助药物设计，因为我们现在能够代表基因和蛋白质，甚至细胞现在非常接近能够代表和理解一个细胞的含义。 48:47 一个细胞意味着什么？这有点像我们如何理解一个段落的含义。如果我们能够像理解一个段落那样理解一个细胞，想象一下我们能做什么。所以，所以我很期待这件事发生。 48:55 我有点兴奋，我知道我们即将在这方面取得突破。例如，类人机器人非常接近，原因是如果你能将语音标记化并理解它，为什么你不能将操作标记化并理解它？ 49:13 所以这些计算机科学技术，一旦你弄清楚了，你就会问自己，如果我能做到这一点，为什么我不能做到那一点？所以我很兴奋这些事情，所以这个挑战是一个快乐的挑战。 49:25 当然，其他一些挑战当然是工业和地缘政治的，它们是社会和…但你们已经听说过所有这些事情了。 49:38 这些都是真的，你知道，世界上的社会问题，世界上的地缘政治问题，为什么我们不能相处得好一些？为什么我要说出这些事情，然后放大它们？ 50:04 为什么我们要在世界上评判人们这么多？你知道所有这些事情，你不必让我再说一遍。我的名字是Jose，我是2023年的班级成员，来自GSB。我的问题是，你是否担心我们发展AI的速度？ 50:17 你认为我们需要某种形式的监管吗？是的，答案是肯定的和否定的。我们需要，你知道，现代AI最伟大的突破当然是深度学习，它带来了巨大的进步。 50:37 但另一个令人难以置信的突破是人类一直在实践的东西，我们刚刚为语言模型发明了它，叫做强化学习人类反馈。我每天提供强化学习人类反馈，这就是我的工作。 50:57 对于他们的父母在房间里，你们每天都在提供强化学习人类反馈。现在我们刚刚弄清楚如何在系统层面上为人工智能做到这一点。 51:11 生成如何遵守物理定律的标记，现在的事物在空间中漂浮，做着各种事情，它们并不遵守物理定律。 51:24 这需要技术护栏，需要技术微调，需要技术对齐，需要技术安全。飞机之所以如此安全，是因为所有的自动驾驶系统都被多样性和冗余所包围。 51:43 所有种类的功能安全和主动安全系统都被发明出来，我需要所有这些尽快被发明出来。你也知道你，网络安全和人工智能之间的界限将变得越来越模糊。 52:03 我们需要在网络安全领域尽快推进技术，以保护我们免受人工智能的侵害。所以，以很多种方式，我们需要技术更快地发展。好的，监管，有两种类型的监管。 52:27 有社会监管，我不知道该怎么办，但有产品和服务监管，我知道该怎么办。所以，FAA、FDA、NHTSA，你说出来吧，所有这些都有针对特定用例的产品和服务的监管。 52:51 嗯，律师资格考试，医生资格考试，你们都有资格认证考试，你们都必须达到一定的标准，你们都必须持续认证，会计师等等，无论是产品还是服务，都有很多监管。 53:17 请不要添加一个超级监管，横跨所有这些监管。监管会计的人不应该是对医生进行监管的人，你知道我喜欢会计师，但如果我需要进行开心手术，他们能够关闭账本，这很有趣，但这还不够。 53:44 如果监管会计的人来监管医生，那就有问题了。我希望所有这些已经拥有产品和服务的领域也能在AI的背景下增强他们的监管，但我也遗漏了一个非常重要的问题，那就是AI的社会影响。 54:00 你如何处理这个问题？我没有很好的答案，但你知道，足够多的人正在讨论这个问题，但这很重要。将所有这些细分成块，这样做是有意义的，这样我们就不会过度关注这一件事，而忽略了我们可以做的许多日常工作。 54:30 结果人们因为汽车和飞机而死亡，这没有意义。我们应该确保我们做正确的事情。好吧，非常实用。最后，我可以问一个问题吗？我们有一些快速提问要问你，作为“从顶峰看未来”的传统。 54:41 好的，我试图避开这个。好吧，你的第一个工作是在Denny’s，他们现在有一个专门给你的展位。你在那里工作的美好回忆是什么？ 55:01 我的第二份工作是在AMD，顺便说一下，那里有一个专门给我的展位吗？我爱我的工作，那是一家很棒的公司。如果你必须写一本书，你会给它起什么名字？ 55:35 我不会写一本，你在问我一个假设性的问题，这是不可能的。如果你能分享一条离别的建议，你会对斯坦福广播什么？ 56:03 这不是一个字，但我会每天进行核心信念检查，全力以赴追求它，用你爱的人包围自己，带上他们一起追求正确的道路。这就是Nvidia的故事，Jensen，这最后一小时非常愉快，非常感谢你抽出时间。 56:23 谢谢你的分享，非常感谢。[音乐]

访谈原文字稿

Intro 0:00 [Music] Jensen this is such an honor thank you 0:06 for being here I’m delighted to be here thank you in honor of your return to Stanford I decided we’d start talking 0:12 about the time when you first left you joined LSI logic and that was one of the 0:18 most exciting companies at the time you’re building a phenomenal reputation with some of the biggest names in Tech 0:24 and yet you decide to leave to become a Founder what motivated you uh uh Chris and Curtis Chris and 0:32 Curtis uh uh I was an engineer at LS logic and Chris and Curtis were at Sun 0:38 and I was working with with uh some of the brightest Minds in computer science at the time of all time uh including 0:46 andyto shim and others uh building building workstations and Graphics workstations and so on so forth and uh 0:54 Chris and Curtis uh uh said one day that they like to leave some son and they 1:01 like uh me to go figure out what they’re going to go leave four and and um I had a great 1:10 job but they they insisted that I uh figure out you know with them how to how 1:16 to build a company and so so we hung out at Denny when whenever they Dro by and 1:21 and uh uh which was which is by the way my alma marter my my first company uh you know my first job before 1:29 for before CEO was a was a dishwasher and so and and I did that very 1:35 well and and so anyways uh we got together and and we we DEC and it was during the the microprocessor Revolution 1:42 this is 1993 and and 1992 when we were getting together the PC Revolution was 1:48 just getting going you you know that Windows 95 obviously which is the Revolutionary version of Windows uh 1:54 didn’t even come to the market yet and Pentium wasn’t even announced yet and so and this is this is all before the right 2:01 before the PC Revolution and it was it was pretty clear that that uh the microprocessor was going to be very 2:07 important and we we thought you know why don’t we build a company uh to go solve 2:13 problems that a normal computer that is powered by general purpose Computing 2:18 can’t and and so that that became the company’s Mission uh to go to go build a 2:24 computer uh the type of computers and solve problems that normal computers can’t and to this day uh we’re focusing 2:30 on that and if you look at all the the problems that that um and the markets that we opened up as a result uh it’s 2:37 you know things like uh computational drug design um uh weather simulation 2:42 materials design these are all things that we’re really really proud of uh robotics uh self-driving cars uh 2:49 autonomous autonomous uh software we call artificial intelligence and then all you know of course uh we uh we drove 2:57 the the uh U the techn techology so hard that that eventually the computational 3:03 cost uh uh went to approximately zero and then enabled enabled a whole new way 3:09 of developing software where the computer wrote the software itself artificial intelligence as we know it today and so so I that was that was it 3:16 that was the journey yeah thank you all for [Laughter] 3:21 coming well these applications are on all of our minds today but back then the How did you convince Don Valentine to invest 3:28 CEO of LSI logic convinced his biggest investor Don Valentine to meet with you he is 3:34 obviously the founder of seoa yeah now I can see a lot of Founders here edging forward in anticipation but how did you 3:41 convince the most sought-after investor in Silicon Valley to invest in a team of firsttime Founders building a new 3:48 product for a market that doesn’t even exist I I didn’t know how to write a 3:53 business plan and and uh uh so I went to a went 3:58 to a book bookstore and back then there were bookstores and and and um in the 4:05 business book section there was this book and it was written by somebody I knew Gordon Bell and this book I should 4:11 go find it again but it’s a very large book and the book says how to write a business 4:17 plan and and that was you know a highly specific title for a very niche market 4:24 and it seems like he wrote it for like you know 14 people and I was one of them and and so I I bought the book I I 4:31 should have known right away that that it was a bad idea because that you know Gordon is super super smart and super 4:38 smart people have a lot to say and and they wanted you know and I I’m pretty sure Gordon wants to teach me how to 4:45 write a business plan uh completely and so I I picked up this book it’s like 450 4:50 pages long well I never got through it not even close I I flipped through it a 4:56 few pages and I go you know what by the time I’m done reading this this thing I’ll be out of business I’ll be out of 5:02 money and and uh Lori and I only had about 6 months uh in the bank and we had 5:08 already Spencer Madison and and uh and a dog and so the five of us had to live off of you know uh whatever money we had 5:15 in the bank and and so I didn’t have much time uh and so instead of writing the business plan uh I just went to talk 5:22 to to W Coran he turn he called me one day and said hey you know you left the company you didn’t even tell me what you 5:27 were doing I want you to come back and explain it to me and so I went back and I explained it to Wi and wi wi at the 5:33 end of it he he said I have no idea what you said and and um that’s one of the worst 5:42 elevator pitches I’ve ever heard um and then he picked up the phone 5:48 and he called Don Valentine and he he called Don and he says Don I want you to give I’m going to send a kid over I want 5:55 you to give him money he’s one of the best employees l logic ever ever had and um I and and so 6:04 the thing I learned is is uh uh you you can make up a great 6:11 interview you could even have a bad interview but you can’t run away from your past and so have a good past you 6:19 know try to have a good past and and and in a lot of ways I was serious when I said I was a good dishwasher I was 6:25 probably Denny’s best dishwasher um I I planned my work I was 6:31 organized you know I was Misan plus and then I washed The Living Daylights out of the dishes and then and then you know 6:38 they promoted me to bus I was certain I’m the best bus boy Denny’s ever had 6:43 you know I was I never left a station with empty-handed I never came back empty-handed I was very efficient and 6:49 then they and so anyways eventually I became you know a CEO I’m working I’m 6:55 still working on being being a good CEO but you talk about being the bad you How did you decide what to do next 7:00 needed to be the best among 89 other companies that were funded after you to 7:05 build the same thing and then with 6 to9 months of Runway left you realized that 7:11 the initial Vision was just not going to work MH how did you decide what to do 7:16 next to save the company when the cards were so stacked against you well we started uh this company 7:22 called for Accelerated Computing and the question is what is it for what’s the killer app and and uh that was that that 7:31 came our first great decision um and this is what sequa 7:36 funded the first great decision was the first killer app was going to be 3D 7:41 graphics and the the the technology was going to be 3D graphics and the application was going to be video games 7:49 at the Time 3D Graphics was impossible to make cheap it was Million dooll image generators from Silicon 7:57 graphics and the video and so it was a million dollars and and it’s hard to make cheap um and the video game Market 8:04 was0 billion doar so you have this incredible technology that’s hard to uh 8:11 commoditize and commercialize and then you have this Market that doesn’t exist that was that intersection was the 8:17 founding of our company and and I still remember uh when when Don at the end of 8:23 my presentation uh you know Don was still kind of he he said you know know one of 8:29 the things he said to me which made a lot of sense back then makes a lot of sense today he says startups don’t 8:36 invest in startups or startups don’t partner with startups and his point is 8:41 that in order for NVIDIA to succeed we needed another startup to succeed and 8:47 that other startup was Electronic Arts and and then on the way out he he 8:52 reminded me that electronic arts’s CTO is 14 years old and had to be driven 8:59 to work by his mom and he just wanted to remind me that that’s who I’m relying on that that and 9:08 then and uh and then after that he said if you lose my money I’ll kill you and that that was that was kind of my 9:14 memories of that first meeting uh but nonetheless uh we created 9:19 we created something uh we went on uh the next several years to go create the 9:24 market to create the gaming market for PCs and it took a long time to do so 9:30 we’re still doing it today uh we realize that not only do you have to create the technology and uh invent a new way of 9:37 doing computer Graphics so that what was a million dollars is now you know three 400 $500 um that fits in the computer 9:45 and you have to go create this new market so we have to create technology create markets the idea that a company 9:51 would create technology create markets defines Nvidia today almost everything 9:56 we do we create technology we create markets that's that's the reason why people say we have a you know people call it a stack an ecosystem words like 10:04 that um but that's basically it at the core for 30 years what Nvidia realized 10:09 we had to do is in order to uh create the conditions by which somebody could 10:14 buy our products we had to go invent this new market and uh it's the reason why we were early in autonomous driving 10:20 it was the reason why we're early in deep learning it was the reason why we're early and just about all these things including uh computational drug 10:27 disc drug design and and Discovery um all these different areas we're trying to create the market while we're 10:33 creating the technology and so that that's um uh okay and then we got we got 10:38 going and and then and then um Microsoft introduced uh a standard called direct 10:45 3D and that spawned off hundreds of companies and we found ourselves a 10:51 couple years later competing with just about everybody and and the thing that that we invented the company the 10:56 technology we invented uh 3D graphics with the consumerized 3D with turns out 11:01 to be incompatible with direct 3D so we started this company we had this 3D Graphics thing we million-dollar thing 11:07 we're trying to make it consumerized and so we invented all this technology and then shortly after it became 11:13 incompatible and um uh so we had to reset the company uh or go out of business but we didn't know how to we 11:20 didn't know how to build it the way that Microsoft had defined it and um and I 11:26 remember I remember a meeting at at you know on a weekend and the conversation was you know we now have 89 11:34 competitors uh I understand that the way we do it is not not right but we don't 11:40 know how to do it the right way and and um thankfully there was 11:47 another bookstore and um and the bookstore is called fries 11:53 Fries electronics I don't think I don't know if it's still here um and so I had 11:58 I had I had um I I I think I drove madis and my daughter on a weekend to fries 12:04 and and it was sitting right there the openg manual uh which would 12:12 defined uh how silicon Graphics did computer graphics and so it was it was right there it was like $68 a book and 12:18 so I had a couple hundred dollar I bought three books I took it back to the office and I said guys I found it our 12:24 future and I handed out I had three versions of it handed out had a big nice centerfold you know the centerfold is 12:32 the opengl pipeline which is the computer Graphics Pipeline and um uh and 12:37 I handed it to uh the same Geniuses that I founded the company with and we 12:44 implemented the openg pipeline like nobody had ever implemented the opengl 12:49 pipeline and we built something the world never seen and so uh a lot of lessons are right there that moment in 12:57 time for our company uh gave us so much confidence and the reason for that is you can 13:05 succeed in doing something inventing a future even if you were not informed 13:11 about it at all and is kind of the my attitude about everything now you know 13:16 when somebody tells me about something and I’ve never heard of it before or if I’ve heard of it never don’t understand 13:23 how it works at all my first thought is always you know how hard can it be 13:30 and it’s probably just a textbook away you know you’re probably one archive paper away from figuring this out and so 13:38 I spent a lot of time reading archive papers and um and it it’s true it’s true 13:43 you can you can um now of course you can’t learn how somebody else does something and do it exactly the same way 13:49 and hope to have a different outcome but you could learn how something can be done and then go back to First 13:56 principles and ask yourself um giving the conditions today given my motivation 14:02 given the instruments the tools um given you know how things have changed how would I redo 14:08 this how would I reinvent this whole thing how would I design a how would I build a car today would I build it 14:14 incrementally from 1950s and 1900s how would I build a computer today how would I write software today does that make 14:21 sense and so I go back to First principles all the time uh even in the company today and just reset ourselves 14:29 you know because the world has changed and U the way we wrote software in the past was monolithic and it’s designed 14:35 for supercomputers but now it’s disaggregated it’s you know so on so forth and how we think about software 14:41 today how we think about computers today how we think just always cause your company always cause yourself to go back 14:46 to first first principles and it creates lots and lots of opportunities yeah the way you applied this technology turns to How did you decide to Pivot 14:53 be revolutionary you get all the momentum that you need to IPO and then some more because you grow your Revenue 15:00 nine times in the next four years but in the middle of all of this success you decide to Pivot a little bit the focus 15:08 of innovation happening at Nvidia based on a phone call you have with this 15:13 chemistry professor can you tell us about that phone call and how you connected the dots from what you heard 15:19 to where you went uh remember at the core the company was uh pioneering a new way of doing 15:26 Computing computer Graphics was the first application uh but we already always knew that there would be other 15:32 applications and so image processing came particle physics came fluids came so on so forth all kinds of interesting 15:38 things that we wanted to do uh we made the processor more programmable so that 15:43 we could express more algorithms if you will and then one day we invented um uh 15:51 programable shaders which made all forms of Imaging and computer Graphics programmable that was a great 15:57 breakthrough so we invented Ed that on top of that we invented uh we we tried 16:02 to look for ways to express um uh more comp more sophisticated algorithms uh 16:09 that could be computation that could be computed on our processor which is very different than a CPU and so we we 16:15 created this thing called CG this I think it was 2003 or so C for 16:21 gpus it predated Cuda by about three years um the same person who wrote The 16:27 textbook that saved the company Mark Hilgard wrote that textbook and um I and so CG was was 16:36 super cool we wrote textbooks about it we started teaching people how to use it we developed tools and such um and then 16:42 several several researchers discovered it uh many of the researchers here students here at Stanford was using it 16:49 um many of the the engineers that that then became uh engineers at Nvidia were were uh playing with it uh 16:57 uh a doctor a couple of doctors at at Mass General picked it up and used it 17:04 for uh CT reconstruction so I flew out and saw them and said you know what are you guys doing with this thing and uh 17:10 they told me about that and then and then uh a um uh a 17:15 computational a Quantum chemist uh used it to um uh Express his his algorithms 17:23 and so I I realized that that there’s there’s some evidence that people might want to use this 17:29 uh and and it gave it gave us gave us you know incrementally more more confidence that that we ought to go do 17:36 this that that this field this form of computing could solve problems that normal computers really can’t and and um 17:44 reinforced our belief and and kept us going every time you heard something new How do you find the conv 17:50 you really savored that surprise and that seems to be a theme throughout your leadership at Nvidia U it feels like you 17:58 make the these bets so far in advance of Technology inflections that when the 18:03 Apple finally falls from the tree you’re standing right there in your black leather jacket waiting to catch 18:10 it how do you find the conv always seems like a diving catch oh it does seem like 18:15 a diving catch you do things based on core beliefs you know we we uh we we 18:21 deeply believe that that we uh we could create a computer that solves 18:27 problems Norm processing can’t do that there are limits to what a CPU can do there are limits to what general purpose 18:33 Computing can do and then there are interesting problems uh that we can go solve the question the question is 18:39 always are those in interesting problems only or are they can they also be interesting markets because if they’re 18:46 not interesting markets it’s not sustainable and Nvidia went through about a decade where we were investing 18:54 in this future and the markets didn’t exist there was only One Market at the time was computer Graphics uh for 10 15 19:02 years the markets that fuels Nvidia today just didn’t exist and so so how do 19:08 you continue um uh with all of the people around you you know our company and you 19:14 know nvidia’s management team and all of the amazing Engineers that they’re creating this future with me um all of 19:21 your shareholders your board of directors all your partners you’re you’re taking everybody with you and 19:27 there’s no evidence uh of a market that is really really challenging you know 19:33 the fact that the technology can solve problems and the fact that you have research papers that that are used that 19:39 that are made possible because of it are interesting but you’re always looking for that market but nonetheless before a 19:46 market exists you still need early indicators of future success you know we we have this phrase 19:52 in the company is is you know there’s a phrase called key performance indicators 19:58 unfortunately kpis are hard to understand I find kpis hard to understand what’s a good 20:05 kpi you know a lot of people you know when when we look for kpis we go gross 20:10 margins that’s not a kpi that’s a result you know you’re looking for 20:16 something that’s an early indicators of future positive results okay and as 20:22 early as possible and the reason for that is because you want early indic that early sign that you’re going in the 20:27 right direction and so we have this phrase is called EO ifs FS you know early indicators e FS 20:35 early indicators of future success and and um and it helps people uh uh because 20:42 I was using it all the time to give the company hope that hey look we solved 20:48 this problem we solved that problem we solved this problem the markets didn’t exist but there were important problems 20:53 and that’s what the company’s about to solve these problems uh we want to be sustainable 20:59 and therefore the markets have to exist at some point but you you want you want 21:04 to decouple the result from um uh from evidence that you’re doing the right 21:10 thing okay and so so so that’s how you that’s how you kind of solve this problem of investing into something 21:17 that’s very very far away um and having the the conviction uh to stay on the road is to find as early as possible the 21:25 indicators that you’re doing the right things and so uh start with a core belief unless something you know changes 21:32 your mind you continue to believe in it and um look for early indicators of future success what are some of those Early indicators 21:38 early indicators that have been used by product teams at Nvidia uh all kinds 21:44 um uh uh I saw I saw I saw a uh a paper 21:49 uh long before I saw the paper I met some people that needed my help on on um 21:55 uh on this thing called Deep learning at a time I didn’t even know what deep learning Le was and um and they needed 22:00 us to create a domain specific language so that um all of their algorithms could 22:06 be expressed easily on our on our processors and we created this thing called cdnn and it’s essentially the SQL um uh 22:16 SQL is in in storage Computing this is um neuron network computing and uh we 22:22 created a a language if you will domain specific language for that you know kind of like the openg GL of of uh deep 22:29 learning and so we we uh they needed us to do that so that they they could express their mathematics and uh they 22:35 didn’t understand Cuda but they understood their deep learning and so we created this thing in the middle for them uh and the reason why we did it was 22:42 because uh even though there were zero I mean this you know these researchers had no money uh and and this is kind of one 22:50 of the the great skills of our company that that you’re willing to do something even though the financial returns are 22:57 complet completely non-existent or maybe very very far out even if you believed in it uh we we ask ourselves you know is 23:05 this worthy work to do um does this Advance a field of science somewhere that matters notice this is something 23:11 that I I’ve been talking about you know since the very beginning of time uh we ex we we find inspiration uh not from 23:20 the size of a market from but from the importance of the work uh because the importance of the 23:25 work is the early indicators of a future Market and nobody has to write a nobody has to 23:31 do a a um a business case on it nobody has to show me a a pnl uh nobody has to 23:37 show me a financial forecast the only question is is this important work and if we didn’t do it uh would it happen 23:43 without us now if we didn’t do something and something could happen without us it gives me tremendous Joy actually and the 23:51 reason for that is could you imagine the world got better you didn’t have to lift a finger that’s the definition of you 23:58 know of of uh ultimate laziness and and and in a lot of ways in a lot of ways you want that habit and the reason for 24:05 that is this uh you want the company to be lazy about doing things that other people 24:10 always do can do if somebody else can do it let them do it we should go select 24:15 the things that if we didn’t do it the world the world would fall apart you have to convince yourself of that that 24:21 if I don’t do this it won’t get done that is Inc and and if that work is 24:27 hard and that work is impactful and important then it gives you a sense of purpose does that make sense and so our 24:33 company has been selecting these projects deep learning was just one of them and the first indicator of of the 24:39 success of that was this you know fuzzy cat that that Andrew an came up with and 24:45 um then Alex kvki uh detected cats um you know not 24:51 all the time but you know successfully enough that it was you know this might take us somewhere and then we reasoned 24:57 about the structure of deep learning and you know we’re computer scientists and we understand how things work and and so 25:03 we we uh we convinced ourselves this could change everything and and um and 25:09 anyhow that but that’s an that’s an example so these selections that you’ve made they’ve paid huge dividends both Dealing with challenges 25:15 literally and figuratively um but you’ve had to steer the company through some 25:20 very challenging times like when it lost 80% of its market cap amid the financial crisis cuz what Wall Street didn’t 25:27 believe in your bet on ML um in times like these how do you steer the company 25:33 and keep the employees motivated at the task at hand uh it’s this is the my 25:39 reaction during that time is the same reaction I had about this week uh earlier today you asked me about this 25:45 week my pulse was exactly the same this week is no different than last 25:51 week or the week before that um and so the opposite of that you know when you 25:57 drop it 80% um it don’t get me 26:02 wrong when when your share price drops 80% it’s a little embarrassing okay and 26:09 and um you just want to you just want to wear a t-shirt that says wasn’t my 26:15 fault um but even more than that you just you just don’t want to you you don’t want to 26:21 get out of your bed you don’t want to leave the house um all of that is true 26:26 all of that is true um but then you go back to go back to just doing your job I 26:31 woke up at the same time I prioritize my day in the same way uh I go back to what 26:36 do I believe uh you got to gut check always gut check back to the court you know what do you believe uh what are the 26:43 most important things uh and uh just check them off you know sometimes 26:48 sometimes it’s helpful to you know family loves me okay check um you know double you know right and so you just 26:55 got to check it off and and you go back to your core um and then go back to work 27:00 and and then every conversations go back to the core uh keep the company focused back on the core do you believe in it 27:06 did something change the stock price changed but did something else change the physics change the gravity 27:13 change did did all of the things that that that we assumed uh that we believed 27:19 that led to our decision did any of those things change because if those things change you got to change everything but if none of those things 27:25 change you change nothing you keep on going yeah yeah that’s how you do it in speaking with your employees they Speaking with employees 27:32 say that you try to avoid the public in speaking with your employees 27:38 they’ve said that your leadership including the employees I’m just kidding no le lead leaders have to be 27:45 seen unfortunately that’s the hard that’s the hard part you know I I I was I was I was at I was I was an electrical 27:52 engineering student and I was quite Young when I went to school um when I went to went to College I was I was 27:59 still 16 years old and so I was I was young when I did everything and and so I was a bit of an introvert kind of you 28:06 know I’m shy I don’t enjoy public speaking I’m delighted to be here I’m not suggesting um but but it’s it’s not 28:13 something that I do naturally and and um I and so so when when things are 28:20 challenging um uh it’s not easy to be in front of precisely the people that you 28:26 care most about you know and the reason for that is because could you imagine a company 28:32 meeting we just our stock prices dropped by 80% and the most important thing I have to do as the CEO is this to come and 28:40 face you explain it and partly you’re not sure why 28:47 partly you’re not sure how long uh how bad yeah you just don’t know these 28:52 things and and but you still got to explain it face face all these people 28:58 and you know what they’re thinking you know you you know some of them are probably thinking we’re doomed uh some 29:04 people are probably thinking you’re an idiot and some people are probably thinking you know something else and so 29:09 I um there are a lot of things that people are thinking and you know that they’re thinking those things but you 29:14 still have to get in front of them and and and deal you know do the hard work they may be thinking of those things but 29:20 yet not a single person of your leadership team left during times like this and in fact 29:26 unemployable that’s what I keep reminding them I’m just kidding I’m surrounded by Geniuses 29:33 I’m surrounded by Geniuses yeah other Geniuses un un unbelievable uh Nvidia is 29:40 well known to have singularly the best management team on the planet this is the deepest technology management team 29:48 the world’s ever seen I’m surrounded by a whole bunch of them and they’re just genius business teams marketing teams 29:55 sales teams just incredible and engineering teams my research teams 30:00 unbelievable yeah your employees say that your leadership style is very engaged you have 50 direct reports you No task is beneath me 30:08 encourage people across all parts of the organization to send you the top five things on their mind and you constantly 30:15 remind people that no task is beneath you can you tell us why you’ve 30:21 purposefully designed such a flat organization and how should we be thinking about our organizations that we 30:26 designed in the future uh no task is is to me no task is 30:32 beneath me because remember I used to be a dishwasher and I and I mean that I used to clean toilets I mean you know I 30:38 cleaned a lot of toilets I’ve cleaned more toilets than all of you combined and and some of them just can’t 30:46 [Laughter] unsee I don’t know I I don’t know what 30:51 to tell you you know that’s life and and so so uh uh you can’t show me and you 30:58 can’t show me a task that is that’s beneath me um now I’m not doing it I’m 31:03 not doing it uh only because because of uh you know whether it’s beneath me or 31:09 not beneath me U if you send me something and you want my input on it and I can be of service to you and in my 31:17 in my review of IT share with you how I reason through it uh I’ve made a contribution to you I’ve made I’ve made 31:25 it possible for you to see how I reason through something and and by reasoning 31:30 as you know how someone reasons through something empowers you you go oh my gosh 31:35 that’s how you reason through something like this it’s not as complicated as it seems this is how you reason through 31:41 something that’s super ambiguous this is how you reason through something that’s incalculable this is how you reason 31:47 through something that you know seems to be very scary this is how you seem do you understand and so I show people how 31:54 to reason through things all the time strategy things you know how to 32:00 forecast something how to break a problem down uh and you’re just you’re 32:05 empowering people all over the place and so that’s how I see it if you send me something you want me to help review it 32:11 uh I’ll do my best and I’ll show you how I would do it um I in the process of 32:17 doing that of course I learned a lot from you is that right you gave me a seat of a lot of information I learned a 32:23 lot and so I I feel rewarded by the process um it does take a lot of energy 32:28 sometimes because you know you got in order to add value to somebody and they’re incredibly smart as a starting point and I’m surrounded by incredibly 32:34 smart people you have to at least get to their plane you know you have to get into their head space and that’s really 32:41 hard that’s really hard um and that takes just an enormous amount of emotional and intellectual energy and so 32:48 I feel exhausted after after I I work on things like that um I’m surrounded by by 32:54 a lot of great people a CEO should have the most direct report rep s um uh by definition because the people that 33:00 reports to the CEO requires the least amount of management it makes no sense to me that 33:06 CEOs have so few people reporting to them except for one fact that I know to be true the the knowledge the 33:14 information of a CEO is supposedly so so valuable so secretive you can only share 33:20 with two other people or three and their information is so 33:26 invaluable so incredibly secretive that they can only share with a couple 33:31 more well um I don’t believe in in in a culture an 33:37 environment where the information that you possess is the reason why you have 33:44 power I would like us all to to to contribute to the company and our 33:50 position in the company should have something to do with our ability to reason through complicated things lead 33:57 other people to um achieve greatness um Inspire Empower other people um support 34:03 other people those are the reasons why the the management team exists in service of all of the other people that 34:09 work in the company to create the conditions by which all of the all of these amazing people who volunteer to 34:15 come work for you instead of all the other amazing high-tech companies around the world they elected they volunteer to 34:22 work for you and so you should create the conditions by which they could do their life’s work which is Mission you know you probably heard it 34:30 i’ I’ve said that you know pretty clearly and I and I believe that what my 34:35 job is is very simply to create the conditions by which you could do your life’s work and so how do I do that what 34:42 does that condition look like well that condition should um result in great deal of empowerment you should you can only 34:48 be empowered if you understand the circumstance isn’t it right you have to understand the cont you have to understand the context of the situation 34:54 you’re in in order for you to come up with great ideas and so I have to create a circumstance where you understand the context which 35:01 means you have to be informed and the best way to be informed is for there to be as little layers 35:09 of information mutilation right between us and so 35:15 that’s the reason why it’s very often that I’m reasoning through things like 35:20 in an audience like this I say first of all this is the beginning facts these are the data that we have um this is how 35:27 I would reason through it these are some of the assumptions these are some of the unknowns these are some of the 35:33 knowns and so you reason through it and now you’ve created an organization that’s highly empowered nvidia’s 30,000 35:40 people we’re the smallest large company in the world we’re tiny little company 35:45 but every employee is so empowered and they’re making smart decisions on my behalf every single day and the reason 35:52 for that is because you know they understand that they understand my condition 35:57 they understand my condition I’m very transparent with people um and uh and I 36:03 believe that that I can trust you with the information often times the information is hard to hear and uh the 36:09 the situations are complicated uh but I trust that you can handle it you’re you know a lot of people hear me say you 36:16 know these you’re adults here you can handle this sometimes they’re not really adults they just 36:22 graduated I’m just kidding I know that when I first graduated was barely an adult and um I but I was I was fortunate 36:30 that I was trusted with with uh with uh important information so I want to do 36:36 that I want to create the conditions for people to do that I do want to now address the topic What is generative AI 36:42 that is on everybody’s mind AI last week you said that generative Ai and 36:47 accelerated Computing have hit the Tipping Point so as this technology becomes more mainstream what are the 36:53 applications that you personally are most excited about well you have to go back to First 36:59 principles and ask yourself what is generative AI what happened um what happened was we have a we now have the 37:06 ability to have software that can understand something they they can understand why you know what is first of 37:12 all we digitized everything that was you know like for example Gene sequencing you digitized genes but what does it 37:19 mean that sequence of genes what does it mean we’ve digitized amino acids um but 37:25 what does it mean uh and so we now have the ability we dig digitize words we digitize sounds uh we digitize images 37:33 and videos we digitize a lot of things but what does it mean we now have the ability through um a lot of study a lot 37:40 of Da data and from their patterns and relationships we We Now understand what they mean not only do we understand what 37:46 they mean we we can translate between them because we learned about the meaning of these things in the same 37:52 world we didn’t learn about them separately so we we learned about speech and and words and and paragraphs and 37:59 vocabulary in the same context and so we found correlations between them and they’re all you know registered if you 38:05 will registered to each other and so now we uh not only do we understand uh the 38:11 modality the meaning of each modality we can understand how to translate between them and so uh for obvious things you 38:18 could caption video to text that’s captioning uh text to uh images M 38:23 Journey uh text to text chat GPT I amazing things and so so we now we now 38:29 know that uh we understand meaning and we can translate uh the translation of something is generation of information 38:36 and and um uh and all of a sudden you you have to take your you take a step back and ask yourself um uh what is the 38:43 implication in every single layer of everything that we do and so I’m 38:48 exercising in front of you I’m reasoning in front of you uh the same thing I did a quarter uh 15 years ago when I first 38:55 saw um uh alexnet some 13 14 years ago I guess um I how I reasoned through it uh 39:02 what did I see how interesting what can it 39:08 do very cool but then most importantly what does it mean what does it mean what 39:13 does it mean to every single layer of computing because you know we’re in the world of computing and so what it means is that that the way that we um process 39:21 information fundamentally will be different in the future that’s what Nvidia builds you know chips and system 39:27 the way we write software will be fundamentally different in the future the type of software we’ll be able to write write in the future will be 39:33 different new applications and then ALS also the processing of those 39:39 applications will be different what was historically a retrieval based model 39:44 where uh in uh information was pre pre-recorded if you will almost you know 39:50 we wrote the text pre-recorded and we retrieved it based on uh some recommender system algorithm in the 39:57 future uh some seed of information will be will be uh the starting point we call 40:02 them prompts you as you guys know and then we generate the rest of it and so 40:07 the future of computing will be highly generated well let me give you an example of what’s happening for example uh we’re having a 40:14 conversation right now very little of the information I’m trans I’m conveying to you is Retreat most of it is 40:23 generated it’s called intelligence and so in the future we’re going to have a lot more generative our computers will 40:30 will perform in that way it’s going to be highly generative instead of Highly retrieval based you go back and you got 40:36 to ask yourself you know now for for you know entrepreneurs you got to ask yourself uh what industries will be 40:43 disrupted therefore will we think about networking the same way will we think about storage the same way will we think about would we be as abusive of internet 40:50 traffic as we are today probably not notice we’re having a conversation right now and and I to get in my car every 40:57 every question so we don’t have to be as abusive of of transformation information 41:05 transporting as we used to um uh what’s going to be more what’s going to be less uh what kind of applications you know 41:11 etc etc so you can go through the entire industrial spread and ask yourself what’s going to get disrupted what’s 41:16 going to get be different what’s going to get NED you know so on so forth and and that reasoning starts from what is 41:22 happening what is generative AI Foundation Al what is happening go 41:28 back to First principles with all things there was something I was going to tell you about organization you asked the question and I forgot to answer it the 41:34 way you create an organization by the way someday um don’t worry about how 41:39 other companies or charts look you start from first principles remember what an 41:44 organization is designed to do the organizations of the past where there’s a king you know 41:53 CE and then then you have all all these you know the Royal subjects you know the 41:58 Royal Court and then eaff and then you keep working your way down eventually they’re employees well the reason why it 42:05 was designed that way is because they they wanted the employees to have as low information as possible because their 42:10 fundamental purpose of the soldiers is to die in the field of battle to die without asking questions 42:18 you guys know this I don’t I only have 30,000 employees I would like them none of them 42:24 to die I would like them to question everything does that make sense and so the way you 42:31 organize in the past and the way you organize today is very different to Second the question is what is nid what 42:36 does Nvidia build an organization is designed so that we could build what it whatever it is we build 42:43 better and so if we all build different things why why are we organized the same 42:48 way why would why would this organizational Machinery be exactly the 42:54 same irrespective of what you build it doesn’t make make any sense you build computers you organize this way you 42:59 build healthare Services you build exactly the same way it makes no sense whatsoever and so you had to go back to 43:06 First principles just ask yourself what kind of Machinery what what is the input what is the output what are the 43:11 properties of this environment you know what what is the what is the what is the forest that this animal has to live in 43:19 what is this characteristics is it stable most of the time you’re trying to squeeze out the last drop of water or is 43:25 it changing all the time being attacked by everybody and so you got to understand you know you’re the CEO your 43:33 job is to architect this company that’s my first job to create the conditions by which you can do your life’s work and 43:38 the architecture has to be right and so you have to go back to First principles and think about those things and I was 43:44 fortunate that that when I was 29 years old you know I had the benefit of of of taking a step back and asking myself you 43:51 know how would I build this company for the future and what would it look like and you know what’s the operating system which is called culture what do we what 43:57 kind of behavior do we en encourage enhance and what what do we discourage and not enhance you know so on so forth 44:04 and anyways I want to save time for audience questions but um this year’s Redefining tomorrow 44:10 theme for view from the top is redefining tomorrow and one question we’ve asked all of our guests is Jensen 44:16 as the co-founder and CEO of Nvidia if you were to close your eyes and magically change one thing about 44:22 tomorrow what would it be 44:29 were we supposed to think about this in 44:35 advance I I’m going to give you a horrible answer 44:42 um I I don’t know that it’s one thing look there are a lot of things we don’t 44:48 control you know there are a lot of things we don’t control um your job is to make a unique contribution live a 44:56 life of purpose to do something that nobody else in the world would do or can do to make 45:03 a unique contribution so that in the event that after you done 45:09 um everybody says you know the world was better because you were 45:15 here and so I think that that to me um I 45:20 live I live my life kind of like this I go forward in time and I Look Backwards 45:26 so you asked me a question that’s exactly from a from a computer vision pose perspective exactly the opposite of 45:32 how I think I never look forward from where I am I go forward in time and look 45:38 backwards and the reason for that is it’s easier I would look backwards and kind 45:44 of read my history we did this and we did that way and we broke that prom down does that 45:49 make sense and so it’s a little bit like um how you guys solve problems you 45:55 figure figure out what is the end result that you’re looking for and you work backwards to achieve it and so I imagine 46:02 Nvidia uh making a unique contribution to advancing the the future of of uh of computing which is the single most 46:09 important instrument of all Humanity now it’s not about our self self-importance but this is just what 46:16 we’re good at and it’s incredibly hard to do and we believe we can make an absolute unique contribution it’s taken 46:22 US 31 years to be here and we’re still just beginning our journey and so this is insanely hard to 46:28 do and uh uh When I Look Backwards I believe that we made I believe that that 46:34 we’re going to be remembered as a company that kind of changed everything not because we went out and changed 46:39 everything through all the things that we said but because we did this one thing that was insanely hard to do that 46:45 we’re incredibly good at doing that we loved doing we did for a long time I’m part of the GSP lead I graduated in 2023 Challenges 46:53 so my question is how do you see see your company in the next decade as what 46:59 challenges do you see your company would face and how you are positioned for that first of all can I just tell you what 47:04 was going on through my head as you say what challenges the list that flew by my 47:12 head was so so large uh that that I was trying to figure out what to 47:20 select um now the honest truth is is that when you ask that question 47:26 most of the challenges that showed up for me were technical challenges and the reason for that is 47:32 because that was my morning if you were to you know chosen yesterday um it might have been Market 47:39 creation challenges there are some markets that I gosh I just desperately would love to 47:44 create I just can’t we just do it already you know but we can’t do it 47:49 alone Nvidia is a technology platform company we’re here in service of a whole 47:55 bunch of other the companies so that they could realize if you will our hopes 48:01 and dreams through them and and so some of the things that I would love I would love for the world 48:08 of biology to to be at a point where it’s kind of like the world of Chip 48:13 design 40 years ago computer AED and design um Eda that entire industry 48:20 really made possible for us today and I believe we’re going to make possible for them tomorrow 48:26 computer AED drug design because we’re able to now represent genes and proteins 48:32 and even cells now very very close to be able to represent and understand the meaning of a cell a combination of a 48:39 whole bunch of genes um what is a cell mean it’s kind of like what does that paragraph mean well if we could 48:47 understand a a cell like we can understand a paragraph imagine what we could do and so uh so so I’m I’m anxious 48:55 for that to happen you know I’m kind of excited about that uh there’s some that I’m just excited about that I know we 49:02 around the corner on for example uh humanoid robotics very very close around the corner and the reason for that is 49:08 because if you can tokenize and understand speech why can’t you tokenize and understand uh 49:13 manipulation and so so these kind of computer science techniques you once you figure something out you ask yourself 49:19 well if got do that why can’t I do that and so I’m excited about those kind of things um and so that challenge is kind 49:25 of a happy challenge uh some of the some of the other challenges some of the other challenges of course are industrial and 49:33 geopolitical and they’re social and and but you’ve heard all that stuff before 49:38 these are all true you know the social issues in in the world uh the geopolitical issues in the world uh why 49:44 can’t we just get along uh things in the world why do I have to say those kind of things in the world um why do we have to 49:50 say those things and then amplify them in the world uh why do we have to judge people so much in the world uh you you know all those things you guys all know 49:57 that I don’t have to say those things over again my name is Jose I’m a class of the 2023 uh from the GSB my question Regulation 50:04 is uh are you worried at all about the pace at which we’re developing AI um and 50:09 do you believe that any sort of Regulation might be needed thank you uh yeah that’s uh the answer is yes and no 50:17 um we need uh you you know that the the the greatest breakthrough in uh modern 50:23 AI of course deep learning and it enabled great progress but another incredible breakthrough is something 50:30 that that humans know and we practice all the time uh and we just invented it for uh for language models called uh 50:37 grounding reinforcement learning human feedback um I provide reinforcement learning human feedback every day that’s 50:44 my job um and their for their parents in the room uh you’re providing reinforcement learning human feedback 50:50 all the time okay now we just figured out how to do that um at a system systematic level for artificial 50:57 intelligence there are a whole bunch of other technology necessary to uh guardrail uh 51:04 fine-tune ground for example how do I generate um how do I generate uh uh uh 51:11 tokens that obey the laws of physics you know right now things are floating in space and doing things and they don’t 51:18 they don’t obey the laws of physics um how do that requires technology Guard railing requires technology fine-tuning 51:24 requires technology alignment requires technology safety requires technology the reason why planes are so safe is 51:30 because you know all of the autopilot systems are are surrounded by diversity and redundancy and all kinds of 51:36 different functional safety and active safety systems that were invented I need all of that to be 51:43 invented much much faster uh you also know that that the border between 51:49 security and artificial intelligence cyber security and artificial intelligence is going to become blurry and blurry we need technology to advance 51:56 very very quickly in the area of cyber security in in order to protect us from artificial intelligence and so so in a 52:03 lot of ways we need technology to go faster a lot faster okay uh regulation 52:09 there’s two types of Regulation uh there’s social regulation I don’t know what to do about that but there’s 52:15 product and services regulation know exactly what to do about that okay so um the fa the FAA the FDA the uh Nitsa you 52:24 name it all the the fs and all the NS and all the you know fcc’s the they all 52:30 have regulations for products and services that are have particular use cases uh um uh bar exams and doctors and 52:38 you know so on so forth um you all have uh qual qualification exams you all have standards that you have to reach you all 52:44 have to uh continuously be certified uh accountants and so on so forth whether it’s a product or a service there are 52:51 lots and lots of regulations please do not add a super regulation that cuts across of it the 52:57 regulator who is regulating accounting should not be the regulator that regulates a 53:04 doctor you know I love accountants um but I I just you know if I ever need an 53:10 open heart surgery the fact that they can close books is interesting but not sufficient and so and so I I would like 53:17 I would like um all of those all of those fields that already have products and services um to also enhance their 53:24 regulation in context of in the context of AI okay but I left out this one very 53:30 big one which is this the social implication of AI and how do you how do you deal with that I don’t have great 53:36 answers for that um but you know enough people are talking about it but it’s important to subdivide all of this into 53:42 chunks does that make sense so that we don’t we don’t become super hyperfocused on this one thing at the expense of a 53:48 whole bunch of routine things that we could have done and as a result people are getting killed by cars and planes 53:53 and you know it doesn’t make any sense we should make sure that we we do the right things there okay very practical Rapid Fire Questions 53:59 things may I take one more question well we have some rapid fire questions for you as view from the tradition 54:07 okay I was trying to avoid that okay all right far away far away 54:14 well your first job was at Denny’s they now have a booth dedicated to you what was your fondest memory of working my 54:20 second job was AMD by the way is there Booth dedicated to me there 54:26 I’m just kidding um I’m I love my job there I did 54:34 I love there it’s a great company yeah yeah um if there were a worldwide shortage of black leather jackets what 54:41 would we be see you wearing oh no I’ve I’ve got a large reservoir of black 54:47 jackets I’m the I’ll be the only person who is who is not 54:53 concerned um you spoke a lot about textbooks if you had to write one what would it be 55:01 called I wouldn’t write one you’re asking me a hypothetical question that has no possibility of of 55:08 of uh that’s fair and finally if you could share one parting piece of advice to broadcast across Stanford what would 55:15 it be uh it’s not a word but but um I you 55:23 know have a core belief um gut check it every 55:30 day I pursue it with all your 55:35 might pursue it for a very long time surround yourself with people you 55:42 love and take them on that right so that’s the story of Nvidia Jensen this 55:48 last hour has been a treat thank you for spending thank you very 55:54 much [Music] 56:23 than