AI 的进展与局限：与李飞飞博士的对话

有一句来自 1970 年代关于 AI 的引述，我认为今天依然适用。它说，即使是最先进的计算机 AI 算法，在房间着火时仍然会下出一手好棋。这句话意在表明，机器被编程来执行任务，但不像人类，我们拥有情境化的态势感知能力，而今天的 AI 还不具备这一点。她是斯坦福大学计算机科学系的首任 Sequoia 教授，也是斯坦福以人为本 AI 研究所的联合主任。她发表了 300 多篇科学文章，曾任谷歌副总裁和谷歌云 AI/ML 首席科学家。以人为本的 AI 是一种开发和运用 AI 的框架，将人类价值和尊严置于中心，这样我们就不会开发出对人类有害的技术。现在，我并不天真，我知道技术是一把双刃剑。我知道我们的文明、我们的物种，始终在光明与黑暗、善与恶的斗争中定义。所以从这个角度来看，我对 AI 的任何希望都不在于 AI 本身，而在于人类。希望在于，当我们创造出类似人类智能的机器时，我们应该……是的，朋友们，欢迎来到节目。我非常兴奋今天能更深入地探讨人工智能领域，以及它未来如何影响我们所有人的生活。今天我们有一位特邀嘉宾，她在 AI 的发展中扮演了重要角色，并且很可能在未来 AI 的使用中发挥更大的作用。李飞飞博士是斯坦福大学计算机科学教授，也是斯坦福以人为本 AI 研究所的联合主任。她的工作专注于推进 AI 研究、教育和政策，以改善人类状况。李博士的新书叫做《我看见的世界》，它将她的个人叙事与 AI 的历史和发展交织在一起。今天我们将讨论她以人为本的 AI 方法，探讨她如何通过计算机视觉为 AI 创造眼睛，并了解未来 AI 这项前景光明但有时令人恐惧的技术会走向何方。李博士，欢迎来到 Young and Profiting 播客。

There's a quote from the 1970s about AI that I think is still true today. It says that the most advanced computer AI algorithm will still play a good chess move when the room is on fire. It's a quote to show that machines are programmed to do tasks, but unlike humans, we have a contextual situational awareness, and that is not what AI is today. The inaugural Sequoia professor in the computer science department at Stanford and co-director of Stanford's Human-Centered AI Institute, she's published more than 300 scientific articles, was vice president at Google and chief scientist of AIML at Google Cloud. Human-centered AI is a framework of developing and using AI that puts human values, human dignity in the center, so that we're not developing technology that's harmful to humans. Now, I'm not naive, I know technology is a double-edged sword. I know that our civilization, our species, is always defined by the struggle of dark and light and by the struggle in good and bad. So from that point of view, any hope I have for AI is not about AI, it's about humans. And the hope is that when we create machines that resemble our intelligence, we should... Yeah, fam, welcome to the show. I'm super pumped to be digging even deeper today into the field of artificial intelligence and how it might impact all of our lives in the years to come. Today we have a special guest who has played a big role in the development of AI and will likely even play a bigger role in how it gets used in the future. Dr. Fei-Fei Li is a professor of computer science at Stanford University, as well as the co-director of the Stanford Institute for Human-Centered AI. Her work focuses on advancing AI research, education, and policy to improve the human condition. Dr. Li's new book is called The Worlds I See, and it weaves together her personal narrative with the history and development of AI. Today we're going to talk about her human-centered approach to AI, we're going to discuss how she's creating eyes for AI with computer vision, and we'll also learn what the future holds for the promising yet sometimes scary technology of AI. Dr. Li, welcome to the Young and Profiting podcast.

谢谢你，Hala。我非常高兴能参加这个节目。

Thank you, Hala. I'm very excited to join this show.

同样。我很荣幸能和你这样的人交谈，鉴于你的所有资历。事实上，《连线》杂志将你列为少数几位科学家之一，也许小到可以围坐在一张厨房餐桌旁，正是这些人推动了 AI 近期的显著进步。所以感觉 AI 每天都在变化，总有新的发展。那么我的第一个问题是：你能带我们回顾一下 AI 的发展吗？比如，它目前能做什么，不能做什么？

Likewise. I'm so honored to talk to somebody like you given all your credentials. In fact, Wired named you one of a tiny group of scientists, perhaps small enough to fit around a kitchen table, who's responsible for AI's recent remarkable advances. So it feels like AI is changing every day, there's new developments all the time. So my first question to you is: can you walk us through the development of AI? Like, what can it currently do now and what can't it do right now?

是的，好问题。确实，即使作为一名 AI 科学家，我也觉得几乎跟不上 AI 的进展。这是一个大约 70 岁的年轻领域，但发展得非常非常快。那么 AI 现在能做什么？首先，它已经无处不在，就在我们身边。AI 的另一个不那么炒作的名字是机器学习。它实际上只是由计算机程序构建的数学模型，程序可以迭代学习，使模型更好地预测或决策数据。所以它本质上是机器学习。例如，如果我们在亚马逊应用上购物，我们得到的推荐就是通过机器学习或 AI。如果你从 A 地到 B 地，规划路径的算法就是机器学习。如果你去 Netflix，有推荐，那是机器学习。如果你看电影，有很多机器学习、计算机视觉、计算机图形学来制作特效和动画，那也是机器学习。所以机器学习和 AI 已经无处不在。它不能做什么？今天没有机器能帮我叠衣服或做煎蛋卷。它无法取代复杂的人类推理。它无法像人类那样创造，结合推理、逻辑，还有美和情感。有一句来自 1970 年代关于 AI 的引述，我认为今天依然适用。它说，即使是最先进的计算机 AI 算法，在房间着火时仍然会下出一手好棋。这句话表明，机器被编程来执行任务，但不像人类，我们对自己的思维、情感以及周围环境有更流畅、更有机、更具情境性的态势感知。而今天的 AI 还不具备这一点。

Yeah, great question. It's true, even as an AI scientist, I feel that I can hardly catch up with the progress of AI right now. So it's a young field of around 70 years old, but it's progressing really, really fast. So what can AI do right now? First of all, it's already everywhere, it's around us. Another name for AI that is a little less of a hype name is machine learning. It's really just mathematical models built by computer programs so that the program can iterate and learn to make the model predict or decide on data better. So it's fundamentally machine learning. For example, if we shop on the Amazon app, the kind of recommendations we get is through machine learning or AI. If you go from place A to place B, the algorithm that maps out the path is machine learning. If you go to Netflix, there is a recommendation, that's machine learning. If you watch a movie, there is a lot of machine learning, computer vision, computer graphics to make special effects and animations, that's machine learning. So machine learning and AI is already everywhere. What can it not do well? No machines today can help me to fold my laundry or cook my omelet. It cannot take away complex human reasoning. It cannot create in the way humans create, in the combination of both reasoning, logic, but also beauty, emotion. There is a quote from the 1970s about AI, and I think that quote is still true today. It says that the most advanced computer AI algorithm will still play a good chess move when the room is on fire. It's a quote to show that machines are programmed to do tasks, but unlike humans, we have a much more fluid, organic, contextual, situational awareness of our own thinking, our own emotion, as well as the surrounding. And that is not what AI is today.

真深刻。我喜欢你说的它有点像机器学习的进化，因为我一直想知道机器学习和 AI 有什么区别？听起来很相似。所以机器学习几乎像是 AI 的基础，AI 的工具。AI 是，你知道，有点像物理学，对吧？牛顿时代的物理学，最重要的工具是微积分，但我们称之为物理学。所以人工智能是一个科学领域，研究和开发技术让机器像人类一样思考，但我们使用的工具，数学计算机科学工具，主要由机器学习主导，尤其是神经网络算法。很好。所以 AI 其实我最近刚接触过，因为两天前我采访了斯蒂芬·沃尔夫勒姆博士。我不知道你是否认识他，Mathematica，是的，他做了沃尔夫勒姆项目和计算机语言 Wolfram。我刚刚采访了他，我们讨论了 ChatGPT 以及它是如何工作的。他向我解释说，当他们开发 ChatGPT 时，令人惊讶的是，他们发现这些简单的规则竟然能创造出所有这些复杂性。他们可以给 ChatGPT 简单的规则，然后它就能像人类一样写作。而事实证明，我们实际上仍然不太理解 AI 是如何学习的，这对我来说太不可思议了。我们创造了一个东西，却不知道它到底是如何工作的？你能详细说明一下吗？

So insightful. And I love that you said that it's sort of like an evolution of machine learning, because I always wondered, like, what's the difference between machine learning and AI? It sounds pretty similar. So machine learning was almost like the basics of AI, the tool of AI. AI is, you know, it's a little bit think about physics, right? Physics in Newtonian time, the most important tool of physics was calculus, and yet we call the field physics. So artificial intelligence is a scientific field that is researching and developing technology to make machines think like humans, but the tools we use, the mathematical computer science tool, is dominated by machine learning, especially neural network algorithms. So good. So AI is actually fresh on my mind because two days ago I interviewed Dr. Stephen Wolfram. I don't know if you know him, Mathematica, yeah, he did the Wolfram project and computer language Wolfram. So I just interviewed him and we talked about ChatGPT and how ChatGPT works. And he was explaining to me that when they were developing ChatGPT, what was surprising is that they found out that these simple rules would create all this complexity. That they could give ChatGPT simple rules and then it could write like a human. And it turns out that we actually still don't really understand how AI learns, which to me is mind-boggling. How did we create something and yet we don't even know how it really works? Can you elaborate on that a bit?

是的，归根结底，有些事情我们理解，有些事情我们不理解。所以并不是完全不懂。它既不是白盒也不是黑盒。我称之为灰盒。根据你对 AI 技术的理解，它要么是深灰色，要么是浅灰色。我们知道的是，背后是神经网络算法，比如 ChatGPT 模型或大型语言模型。当然你听说过 Transformer 模型、序列到序列等名称。归根结底，这些模型接收数据，比如文档数据，并学习单词甚至子词（单词的一部分）之间是如何相互连接的。有模式可循，对吧？如果你看到单词'how'，它往往后面跟着'a'，然后'r'，然后往往跟着'u'。所以'how are you'是一个频繁出现的序列。这个模式就被学习到了。一旦你在一个巨大的神经网络中学到了足够多的模式……

Yeah, it really at the end of the day there are things we understand, there are things we don't. So it's not like completely we don't. So it's neither a white box nor a black box. I would call it a gray box. And depending on your understanding of the AI technology, it's either darker gray or lighter gray. So the things we know is that it is a neural network algorithm that is behind, say, a ChatGPT model or a large language model. Of course you hear the names of Transformer models, sequence to sequence, and all that. At the end of the day, these models take data, like document data, and it learns how the words and sometimes even subwords, parts of words, are connected with each other. There are patterns to see, right? If you see the word 'how', it tends to be followed by 'a' and then 'r', and then it tends to be followed by 'u'. So 'how are you' is a frequently occurring sequence. So that pattern is learned. And once you learn enough in a big huge neural network...

你预测下一个词的能力高得惊人，几乎可以像人类一样对话。而且因为训练数据中包含了大量知识——无论是化学、影评还是地缘政治事实——它都记住了，所以能给出很好的答案。这些是我们知道的：我们知道算法如何工作，知道它需要训练，知道它在学习和预测模式。但我们不知道的是，因为这些模型非常庞大——有数十亿、数千亿的参数——模型内部有这些小节点，每个节点都有相互连接的数学函数。那么，我们如何确切知道这数十亿参数是如何学习模式的？模式存储在哪里？为什么有时它会幻觉出一个模式，而有时给出正确答案？目前还没有精确的数学解释。我们没有一个方程能告诉我们：‘哦，我确切知道为什么此刻 ChatGPT 给出的是“你好吗”而不是“他是谁”。’这就是伟大之处：这些是大模型，其行为无法用数学精确解释。

Your ability to predict the next word when given a word is really amazingly high, to the point that it can converse more or less like a human. And because in the training data there is so much knowledge—whether it's chemistry, movie reviews, or geopolitical facts—it has memorized all of them, so it can give very good answers. So those are the things we know: we know how the algorithm works, we know it needs training, we know it's learning and predicting patterns. What we don't know is that because these models are huge—there are billions and billions, hundreds of billions of parameters—and inside these models there are these little nodes, each one of them has a little mathematical function that connects to each other. So how do we know exactly how these billions and billions of parameters learn the pattern, and where is the pattern stored, and why sometimes it hallucinates a pattern versus giving a correct answer? There is not yet a precise mathematical explanation. We don't know at the level of an equation that can tell us, 'Oh, I know exactly why at this moment ChatGPT gives you the word 'how are you' versus 'how is he'.' So that's where the greatness comes from: these are large models with behaviors that are not precisely explained mathematically.

所以根据我的理解，这些神经网络是为了某种程度上复制人脑的工作方式，对吗？

So from my understanding, these neural networks are made to sort of replicate how the human brain works, basically.

我不会用‘复制’这个词。它们是受人类大脑启发。有相似之处：例如，它们由小的神经元节点组成，按层级连接。但人脑从根本上是以化学电方式工作的。神经元交流的方式非常复杂——有时通过脉冲，脉冲还会释放化学物质。这些功能很微妙。此外，连接方式——大脑一个区域如何与其他区域连接——也与神经网络不同。所以我们是受启发，而非复制。

I would not use the word 'replicate'. They are inspired by the human brain. It has resemblance: for example, they are made of small neuron nodes, they are connected in hierarchies. But human brains fundamentally work in a chemical-electrical way. The way neurons communicate is very complex—sometimes through spikes, and the spike also releases chemicals. There are just these kinds of nuanced functions. Also, the connectivity—how one area of the brain is connected to others—is not the same as neural networks. So we are inspired, but not replicating.

这个区分非常有帮助。是的。我想对所有雇主说几句。我们来谈谈公司文化。在 Yap Media，我们有非常独特的公司文化，我们都追求卓越，甚至自称‘Scrappy Hustlers’。我的团队正在快速扩张，招聘很麻烦，尤其是当你寻找愿意撸起袖子干的 A 级人才时。但幸运的是，在招聘方面，我不再为寻找完美候选人而烦恼，因为我使用了 Indeed，这个终极招聘平台。Indeed 的匹配引擎总能为我提供一批高质量候选人，完全符合我的职位描述。如果你厌倦了在招聘池中挣扎，Indeed 来拯救你。你可以用 Indeed 安排面试、筛选和联系候选人，让整个招聘过程变得轻松。Indeed 不仅帮你更快招人——根据最近一项 Indeed 调查，93%的雇主认为 Indeed 比其他招聘网站提供更高质量的匹配。我招到了一些最好的员工，一些最好的 Scrappy Hustlers。Indeed 每天分析超过 1.4 亿个资格和偏好，它不断从你的招聘偏好中学习，所以你用得越多，它就越能找到你的完美匹配。加入全球超过 350 万家企业的行列，它们已经选择 Indeed 快速招聘优秀人才。我们节目的听众将获得 75 美元的赞助职位积分，让你的职位获得更多曝光，请访问 indeed.com/profiting。现在就访问 indeed.com/profiting，并告诉节目你是从本播客了解到 Indeed 的，以支持我们。Indeed.com/profiting。条款和条件适用。需要招聘？你需要 Indeed。

That's a really helpful distinction right there. Yes. I want to talk to all of my employers out there. Let's talk about company culture for a second. At Yap Media, we have a really unique company culture, and we're all obsessed with excellence, and we even call ourselves Scrappy Hustlers. We're all Scrappy Hustlers at Yap Media, and my team is growing fast, and hiring is a pain in the butt, especially if you're looking for A-players that are going to roll up their sleeves. But luckily, when it comes to hiring, I no longer feel overwhelmed by the search for the perfect candidate because I use Indeed, the ultimate hiring platform. Indeed's matching engine always presents me with a pool of high-quality candidates that match my job description to a T. If you're tired of drowning in your hiring pool, Indeed is here to rescue you. You can use Indeed for scheduling, screening, and messaging your candidates, making the entire hiring process a breeze. And Indeed doesn't just help you hire faster—93% of employers agree that Indeed delivers the highest quality matches compared to other job sites, according to a recent Indeed survey. I've hired some of my best employees, some of my best Scrappy Hustlers, with over 140 million qualifications and preferences analyzed every day. Indeed is constantly learning from your hiring preference, so the more you use Indeed, the better it actually gets at finding your perfect match. Join the ranks of more than 3.5 million businesses worldwide that have already chosen Indeed to hire great talent fast. And listeners of my show will get a $75 sponsored job credit to get your jobs more visibility at indeed.com/profiting. Just go to indeed.com/profiting right now and support our show by saying you heard about Indeed on this podcast. Indeed.com/profiting. Terms and conditions apply. Need to hire? You need Indeed.

那么跟我们讲讲 AI 模型是如何训练的。比如，AI 通常是如何学习的？

So talk to us about how AI models are trained. Like, how does AI learn? Typically?

通常，AI 模型会获得大量数据，其中一些数据带有人类监督的标签。例如，如果我给 AI 模型数百万张图像，有些标注为‘猫’、‘狗’、‘微波炉’、‘椅子’等等，它们会学习将模式与标签关联起来。有时，尤其是在最近的语言领域，我们使用所谓的自监督学习：你给它数百万、数万亿份文档，它不断学习预测下一个音节、下一个词，因为所有训练数据都展示了这些词序列。这时你不需要额外标签，只需提供文档，这就是自监督学习。所以无论是带额外标签的监督学习，还是不带额外标签的自监督学习，都是从数据开始的。数据进入算法，算法必须有一个学习目标。通常在语言模型中，目标是尽可能准确地预测下一个音节，就像训练数据展示的那样。以带猫标签的图像为例，目标是预测一张有猫的图像，并给出正确标签‘猫’，而不是错误标签‘微波炉’。然后，因为有这个目标，如果在训练中犯了错误——比如没有正确预测下一个词，或者把猫标错了——它会返回并根据错误迭代更新参数。它有一些数学规则或学习规则来更新，然后不断重复，直到人类要求停止或它不再更新，无论停止标准是什么。最后你会得到一个巨大的神经网络，已经由海量数据训练而成。在这个神经网络中，所有数学参数都已经学习好了。现在你可以输入一个新句子，它通过这个模型，因为拥有所有已学习的参数，它会根据新句子预测应该说什么，比如‘你好，今天早餐怎么样？’它会预测‘我今天早餐很棒’之类的。这就是它的使用方式。

Typically, an AI model is given a vast amount of data, and then some of the data are labeled with human supervision. For example, if I give AI models millions and millions of images, some are labeled 'cat', 'dog', 'microwave', 'chair', and all that, and they learn to associate the pattern with the labels. Sometimes, especially in the language domain recently, we use what we call self-supervision: you give it millions and millions, trillions of documents, and it just keeps learning to predict the next syllable, the next word, because all the training data shows you all these sequences of words. There you don't have to give additional labels; you just give the documents, and that's called self-supervised learning. So whether it's supervised with additional labels or supervised without additional labels (self-supervised), it starts with data. Now data goes into the algorithm, and the algorithm has to have an objective to learn. Typically in the language model, the objective is to predict the next syllable as accurately as the training data shows you. In the case of images with cat labels, for example, the objective is to predict an image that has a cat with the right label 'cat' instead of the wrong label 'microwave'. Then, because it has this objective, if during training it makes a mistake—if it didn't predict the next word right, or if it labeled the cat wrong—it goes back and iterates and updates its parameters based on the mistake. It has some mathematical rules or learning rules to update, and then it just keeps doing that until humans ask it to stop or it no longer updates, whatever the stop criteria. Then you're left with a ginormous neural network that's already trained by a ginormous amount of data. In that neural network, it has all the mathematical parameters that have already learned. Now you can take this and have a new sentence come in, and it goes through this model. Because it has all the parameters it has learned, it predicts what it should say given the new sentence, like 'Hello, how is your breakfast today?' and it would predict 'I had a great breakfast today' or whatever. So that's how it's going to be used.

这太有趣了。基本上 ChatGPT 就是在根据所有不同模式预测下一个词、再下一个词，试图找出接下来什么合理。这非常清楚。但我不理解的是，像 ChatGPT 这样的东西，它非常擅长写人类语言，但却会犯一些很小、很简单的数学错误，对吧？它怎么可能擅长人类语言，而在数学上却犯愚蠢的错误呢？

So it's so interesting. Like basically ChatGPT is just predicting the next word and the next word and the next word based on all the different patterns and trying to figure out what makes sense to come next. So that's super clear. What I don't understand with something like ChatGPT is that it's so good at writing human language, but it's known to make really small, simple math mistakes, right? How is it possible that it's good at doing human language but then on math, for example, it's known to make stupid mistakes?

这是因为数学，我们人类大脑做数学的方式，与做语言的方式不同。语言有非常清晰的序列模式——下一个词预测对语言很有效，因为语言本质上是序列的。而数学需要精确的逻辑推理和符号操作，这与预测下一个词不同。模型没有内置的数学规则理解，它只是从数据中学习模式。所以如果训练数据包含很多数学问题的例子，它可以模仿模式，但并不真正‘理解’数学。这就是为什么它会犯简单错误——它只是根据模式预测它认为最有可能的下一个词，而不是进行实际计算。

It's because math, the way we do math in the human mind, is different from the way we do language. Language has a very clear sequential pattern—next-token prediction works well for language because language is inherently sequential. Math, on the other hand, requires precise logical reasoning and symbolic manipulation, which is not the same as predicting the next word. The model doesn't have a built-in understanding of mathematical rules; it just learns patterns from data. So if the training data contains many examples of math problems, it can mimic the pattern, but it doesn't truly 'understand' math. That's why it can make simple mistakes—it's just predicting what it thinks is the most likely next token based on patterns, not performing actual calculation.

序列到序列的模式，比如我说单词'how'，你知道通常后面会跟'r'和'u'，但有时也不是，对吧？所以我必须学习这些模式。但如果说'one plus'，通常后面不会跟'five'或'two'，对吧？实际上有一个更深的规则：1+2=3。当然，当它见过足够多这样的例子后，它应该预测出 3。对于今天的语言模型来说，它确实能做到。这个例子太简单了，但关键是数学需要比仅仅遵循统计模式更高层次的推理，而大语言模型基本上遵循统计模式，所以一些数学推理是欠缺的。

Pattern of sequence to sequence, like I say the word 'how' and you know the word 'r' and 'u' typically follow, but sometimes it doesn't, right? So I have to learn these patterns. But if I say the word 'one plus', it's not like 'five' typically follows or 'two' typically follows, right? There is actually a deeper rule: 1+2=3. Of course, when it has seen enough of that, it should predict three. For today's language model, and actually it does. This is too simple an example, but the point is that math takes a higher level of reasoning than just following statistical patterns, and large language models by and large follow statistical patterns, so some of the mathematical reasoning is lacking.

完全说得通。

Totally makes sense.

你有一本新书，叫《我看见的世界》，你说你看到的世界有不同的维度。能跟我们聊聊为什么给这本书取这个名字吗？

So you've got a new book, it's called 'The Worlds I See', and you say that the worlds you see are in different dimensions. So can you talk to us about why you titled the book this way?

是的，这个书名是在我写完书稿后才想到的。我意识到写书的过程其实是在剖析不同的经历。有作为科学家体验的 AI 世界——这本书是一个年轻科学家的成长故事，所以我经历了不同阶段的科学世界。但也有作为移民的世界，我在世界不同地方生活，如何应对和经历这一切。还有一个更微妙但深刻的世界，比如学习做人。我知道这听起来很傻，但尤其是在 AI 科学家的背景下，这非常重要。书的一部分探索了我与年迈父母共同生活、照顾他们的旅程，这段经历如何塑造了我的性格，我们如何互相帮助、互相支持。在书的结尾，这段经历如何让我以不同的视角看待科学，与其他可能没有这种深刻人类经历的科学家相比。所以我确实经历了不同的世界，它们都融入了这本书。

Yeah, this title came about after I finished rewriting the book, and I realized the journey of writing the book is really peeling into different experiences. There is the world of AI that I experience as a scientist. The book is a coming-of-age of a young scientist, so I experience the world of science in different stages. But there is also the world as an immigrant, right? I go through life in different parts of the world and how do I handle or go through that. And then there is a more subtle but profound world, like learning to be a human. I know that sounds silly, but especially in the context of an AI scientist, it's really important. Part of the book is exploring my journey of living with and taking care of aging parents, and how that experience built my own character, how we help each other, support each other, and towards the end of the book, how that experience made me see my science in a different light compared to maybe other scientists who haven't had this very profound human experience. So it really is different worlds that I experience, and it's blended into the book.

我喜欢这个说法。我也喜欢你称它为科学回忆录。你说你参与了 AI 的科学方面，但也参与了 AI 的社会方面。那么社会方面具体指什么？

I love that. And I love how you call it a science memoir. And so you say that you're involved in the science of AI, but you're also involved in the social aspect of AI. So what do you mean by the social aspect exactly?

是的，我最初进入 AI 领域是一段非常个人化的旅程。只是一个热爱冷门领域的年轻科学极客，你知道，一个没人知道的领域，但我私下里非常着迷：我们如何让机器思考，如何让机器看见？我很开心，老实说，如果余生都这样我也很满足。即使世界上没人听说过 AI，我也会快乐地在实验室里做科学家。但真正改变的是 2017、2018 年左右，我感觉作为科学家，我和科技界都醒悟过来，意识到'哇，这项技术已经成熟到影响社会的程度了。'因为它是 AI，它受人类思维启发，受人类行为启发，它在个人层面和社会层面都有太多人类影响。所以作为科学家，我觉得自己被推入了一个从未意识到的更混乱的现实。现在我有了选择。我的许多科学家同行会继续待在实验室，我认为这非常令人钦佩和尊重，他们仍然专注于科学。但我的另一个选择是认识到，作为科学家、教育者、公民，我有社会责任。我的责任更侧重于：我需要教育年轻人。虽然我可以教他们方程和编码等等，但我也想与他们分享这门科学的社会影响，因为这是我的责任。我也有责任与世界沟通，因为即使从几年前开始，现在因为大语言模型更糟了，关于 AI 的公共讨论太多了，其中很多是信息不足的，这很危险，对吧？这不公平，很危险，它往往伤害那些没有权力的人。我有责任去沟通。第三，我也觉得斯坦福，尤其是作为美国的高等学府，有责任帮助让世界变得更好，帮助我们的政策制定者，帮助公民社会，帮助公司，帮助企业家，去教育、告知并提供见解。所有这些都是面对现实世界的混乱，我觉得我不应该回避，我应该承担起这个责任。

Yeah, so I started in AI as a very personal journey. It's just a young science nerd who loves an obscure field, you know, like nobody knows the field, but I'm just fascinated in a private way: how do we make machines think, how do we make machines see? And I was happy, and I would have been content with that through the rest of my life, honestly. Even if nobody in the world had heard of AI, I would be happily in my lab being a scientist. But what really changed is around 2017, 2018, I felt like me as a scientist and the tech world woke up and realized, 'Oh wow, this technology has come to a maturation point that is impacting society.' And because it's AI, it's inspired by human thinking, it's inspired by human behavior, it has so much human implication at the individual level as well as the society level. So as a scientist, I feel I was thrust into a messier reality that I never really realized. Now I have a choice. A lot of my fellow scientists would just continue to stay in the lab, which I think is very admirable and respected, and just still focus on the science. But my other choice is to recognize as a scientist, as an educator, as a citizen, I have social responsibility. My responsibility is more focused on: well, I need to educate young people. While I can teach them equations and coding and all that, I also want to share with them what the social implications are of this science, because it's my responsibility. I also have a responsibility to communicate with the world, because even starting quite a few years ago, now it's even worse because of the large language model, there's just so much public discourse about AI, and many of them are ill-informed, and that's dangerous, right? That's unfair, that's dangerous, it tends to harm people who are not in the position of power. And I have a responsibility to communicate. And third, I also feel Stanford, especially as one of America's higher institutions, has a responsibility to help make the world better, to help our policy makers, to help civil society, to help companies, to help entrepreneurs, to educate, to inform, and to give insights. And all this is the messiness of meeting the real world, and I feel I shouldn't shy away from that, I should take on that responsibility.

是的，当然。你是最懂 AI 的人之一。我们需要你告诉我们需要注意哪些障碍，以及如何确保我们善用 AI 而非滥用，并采取相应步骤。

Yeah, for sure. You're one of the most knowledgeable people about AI. We need you to tell us what are the roadblocks that we need to look out for and how can we make sure that we use AI for good and not for bad, and take the steps to do that.

那我们接下来聊聊计算机视觉。你是一位计算机视觉 AI 科学家。最初是什么让你对此感兴趣？什么是计算机视觉 AI？

So let's talk about computer vision next. So you are a computer vision AI scientist. So what first got you interested in this and what is computer vision AI?

是的，一句话来说，计算机视觉 AI 是 AI 的一部分，是让计算机看见并理解所见内容的特定部分。这非常深刻。当人类睁开眼睛，我们看到的不仅是颜色和阴影，我们看到了意义，对吧？比如我现在看着自己凌乱的桌子，有手机、杯子、显示器、过敏药，还有很多意义。不仅如此，我们还能构建，即使我们不是最好的艺术家，人类自文明诞生以来就在描绘世界、雕刻世界、建造桥梁和纪念碑，创造了视觉世界。所以看见、视觉创造和理解的能力在人类身上是与生俱来的，如果计算机也有这种能力，那该多好？这就是计算机视觉。

Yeah, well, in one sentence, computer vision AI is part of AI, it's the specific part of AI that makes computers see and understand what it sees. And this is very profound. When humans open our eyes, we see the world not only in colors and shades, we see it in meaning, right? Like I'm looking at my messy desk right now, it has cell phones, it has a cup, it has a monitor, it has my allergy medicine, and it has a lot of meaning. And more than that, we can also construct, especially even if we're not the best artists, humans since the dawn of civilization have been drawing about the world, have been sculpting about the world, have been building bridges and monuments, and have created the visual world. So the ability to see and visually create and understand is so innate in humans, and wouldn't it be great if computers have that ability? And that is what computer vision is.

而且，当我想到意识时，一切有意识的东西都有眼睛。这总是让我毛骨悚然：虫子有眼睛，鱼有眼睛，而且它们的眼睛看起来像我们的眼睛，比如鱼的眼睛看起来像我们的眼睛，这太诡异了。所有这些生物都有眼睛，如果 AI 开始有眼睛，那不就意味着它们是有生命、有感知的吗？

And you know, when I think about consciousness, everything that has consciousness has eyes. And I always this always freaked me out: bugs have eyes, fish have eyes, and the eyes look like our eyes, like fish eyes look like our eyes, and that's so scary weird. The fact that all these living things have eyes, if AI starts to have eyes, wouldn't it just be that they're living and sentient at that point?

首先，Hala，你触及了非常非常深刻的东西，因为从进化角度来说，视觉感知是最古老的之一。5.4 亿年前，动物开始发展眼睛；最初只是一个收集光的小孔，但后来进化成鱼、章鱼、大象和我们人类的眼睛。所以你确实触及了非常深刻的东西：这极其本能，嵌入在我们智能的发展中。当然，你也问了一个哲学上非常深刻的问题：有眼睛的东西都有意识吗？实际上……

So first of all, Hala, you touched on something really, really profound, because visual sensing is one of the oldest evolutionarily speaking. So 540 million years ago, animals started developing eyes; it was a pinhole that collects light, but it evolved into the kind of eyes the fish, the octopus, the elephant, the eyes we have. So you actually touch on something really profound: this is extremely innate, embedded into our development of our intelligence. And of course, you also ask a philosophically really profound question: everything that has eyes has consciousness? Actually...

神经科学家或神经哲学家可能会说——你应该请一位来跟你辩论——比如，一只小虾用眼睛做事，它有意识吗？或者它有感知吗？老实说，我没有答案。你怎么衡量意识？就因为虾能看到石头并爬来爬去，就意味着它只是感官反射，还是有更深层的意识？我不知道。所以，仅仅因为机器有了眼睛，它就会发展出意识吗？这是一个我们可以讨论的话题，但我只是想确保我们至少达成共识：看见本身并不意味着有意识。但我们拥有的那种视觉智能——就像我刚才描述的，去理解、创造、构建、呈现一个具有如此视觉复杂性的世界——至少在人类身上，它确实需要意识。

A neuroscientist or neurophilosopher would probably say—you should invite one to debate with you—for example, does a tiny shrimp using its eyes to do things have consciousness? Or does it have perception? I don't have an answer honestly. How do you measure consciousness? Just because the shrimp can see the rock and climb around, does it mean it's just a sensory reflex, or does it have deeper consciousness? I don't know. So just because machines have eyes, does it develop consciousness? It's a topic we can talk about, but I just want to make sure we are at least on the same page that seeing itself doesn't mean it has consciousness. But the kind of visual intelligence we have—like I just described, to understand, to create, to build, to represent a world with such visual complexity—at least in humans, it does take consciousness.

你说的每一点都很有趣。就连那个虾的例子——确实如此。尽管它在导航、在岩石周围游来游去，但这并不意味着它真的有意识。按你的说法，这可能只是反射。这让机器最终拥有眼睛这件事变得不那么可怕了。那么，你现在是如何在计算机中复制像视觉这样的生物过程的呢？

Everything you're saying is just so interesting. Even that shrimp example—it's true. Even though it's navigating, swimming around rocks and whatever, doesn't mean it's actually conscious. It could be, to your point, just all reflexes. And that makes it a little less scary if machines end up having eyes. So how are you replicating biological processes like vision in computers now?

是的，再次强调，我认为很多计算机视觉都受生物学启发，至少在两个方面。一个是算法本身。整个神经网络算法——事实上，在 20 世纪 50 年代和 60 年代，计算机科学家受到视觉神经科学家的启发，当时他们正在研究猫的视觉系统。他们发现了层级神经元，正是这一点启发了计算机科学家构建神经网络算法。所以动物大脑中的视觉结构是当今 AI 技术的基础灵感。这是第一个方面。第二个灵感来自功能性——看的能力。我们看到了什么？例如，人类并不擅长看颜色。我们看到的颜色足够丰富，但事实是有无限波长定义了无限颜色，而我们可能只有几十种颜色。所以显然，我们不像用机器记录波长那样看颜色。但另一方面，我们看到意义，看到情感，看到所有这些。能够将这种功能构建到机器中是非常鼓舞人心的。这是生物启发的另一部分——功能性启发。有了这个，我认为有很多可以想象的。例如，首先，视力受损的患者——如果我们用人工视觉系统帮助他们理解我们看到的丰富世界，那将非常有帮助。机器——对吧？我不知道，你家里有 Roomba 吗？是的，对吧。所以它几乎是在看——不是像我们那样看，但它是在看和映射。但有一天我希望我不仅有 Roomba，还有清洁机器人，对吧？那么它需要以更复杂的方式看我的房子。最重要的是，例如救援机器人。有很多情况会让人类处于危险中，或者人类已经处于危险中，你想救援人类但不想让更多人类陷入危险。想想福岛核泄漏事件——人们不得不牺牲自己进去阻止泄漏。如果机器人能做到这一点，那将非常棒，这需要视觉，需要更深层次的视觉智能。

Yes, so again, I think a lot of computer vision is biologically inspired, and it's inspired in at least two areas. One is the algorithm itself. The whole neural network algorithm—in fact, back in the 1950s and 60s, computer scientists were inspired by vision neuroscientists when they were studying the cat's visual system. They discovered hierarchical neurons, and it's because of that it inspired computer scientists to build neural network algorithms. So the animal visual structure in the brain is very much the foundational inspiration for today's AI technology. So that's one area. The second inspiration comes from functionality—the ability to see. What do we see? Humans are not that good at seeing color, for example. We see color richly enough, but the truth is there are infinite wavelengths that define infinite colors, but we have only probably dozens of colors. So clearly we're not seeing just colors in the same way as if I use a machine to register wavelength. But on the other hand, we see meaning, we see emotion, we see all these things. And it's just incredibly inspiring that we can build this functionality into machines. And that is another part of biological inspiration—it's the functional inspiration. With that, I think there is a lot to imagine. For example, first of all, visually impaired patients—if we help them with an artificial visual system to understand the rich world we see, it will be tremendously helpful. Machines—right? I don't know, do you have a Roomba in your house? Yeah, right. So it almost is kind of seeing—it's not seeing the same way we are, but it's kind of seeing and mapping. But one day I hope I not only have a Roomba but also a cleaning robot, right? Then it needs to see my house in a much more complex way. And most importantly, for example, rescue robots. There are so many situations that put humans in danger, or humans are already in danger and you want to rescue humans but you don't want to put more humans in danger. Think about that Fukushima nuclear leak incident—people had to really sacrifice to go in there to stop the leak. It would be amazing if robots could do that, and that needs vision, it needs visual intelligence in much deeper ways.

这太有趣了，你这么说很有帮助，因为我的第一反应是，我们为什么要给机器人这么多权力？我们正在失去作为人类的力量。但按你的说法，它可以帮助人类。我知道这就是你一直在谈的——以人为本的 AI。那么你能用自己的话定义什么是以人为本的 AI 吗？

That's so interesting, and it's helpful for you to say that because my first reaction is like, why are we giving robots this much power? We're losing our power as humans. But to your point, it can help humans. And I know that's a whole thing you talk about—human-centered AI. So can you define what human-centered AI is in your own words?

是的，以人为本的 AI 是一个开发和利用 AI 的框架，这个框架将人类、人类价值、人类尊严置于中心，这样我们就不会开发对人类有害的技术。所以这实际上是一种以仁慈的方式看待或使用技术的方式。我并非天真——我知道技术是一把双刃剑。我知道这把双刃剑可能被坏人有意或无意地以坏的方式使用。所以以人为本的 AI 真正试图强调的是，我们有集体责任专注于 AI 的良好发展和良好使用。这实际上受到我在工业界休假期间的启发，当时我作为教授，看到了……

Yeah, human-centered AI is a framework of developing and using AI, and that framework puts humans, human values, human dignity in the center, so that we're not developing technology that's harmful to humans. So it's really a way to see technology or use technology in a benevolent way. Now I'm not naive—I know technology is a double-edged sword. I know that double-edged sword can be used intentionally or unintentionally by bad actors in bad ways. So human-centered AI is really trying to underscore that we have a collective responsibility to focus on the good development and good use of AI. And it was really inspired by my time in industry when I was on sabbatical as a professor, seeing the...

早在 2018 年，AI 就已经打开了商业机会的闸门。当企业开始使用 AI 时，它会影响到每个人的生活，对吧？所以我回到斯坦福，和同事们一起意识到，作为一个思想领导机构，作为美国高等教育、培养下一代学生的地方，我们真的应该有自己的观点，去发展并保持在技术发展的前沿。这就是我们如何制定以人为本的 AI 框架的。

Incredible business opportunities that are already opening the floodgates of AI back in 2018. And knowing that when businesses start to use AI, it impacts the lives of every individual, right? So I went back to Stanford, and together with my colleagues, we realized as a thought leadership institution, as American higher education, a place to educate the next generation of students, we should really have a point of view to develop and stay at the forefront of the development of this technology. This is how we formulated the human-centered AI framework.

人们对 AI 最大的恐惧之一是 AI 会取代我们所有的工作。现在，AI 可能会创造很多工作，我在播客中和其他嘉宾也讨论了很多。但您建议我们如何创造工作，并确保 AI 不会夺走所有工作？

One of the biggest fears that people have with AI is that AI is going to replace all of our jobs. Now, AI is probably going to create a lot of jobs, and I've talked a lot about that with other guests on the podcast. But how do you suggest that we make jobs and take into consideration making sure that AI doesn't take all the jobs?

嗯，有几点，Hala。首先，我们为什么要有工作？思考这一点非常重要。我认为工作是人类繁荣的一部分，因为我们需要它转化为经济回报，从而为家庭带来繁荣。同时，它也是人类尊严的一部分。它超越了金钱，对很多人来说是意义，是生活和自尊的意义。从这个角度看，我们必须认识到，在人类历史上，工作一直在变化。技术以及其他因素会创造、摧毁、改变、转化工作。但不变的是对人类繁荣和尊严的需求。所以，当我们思考 AI 及其对工作的影响时，重要的是要触及工作的核心本质和意义，以及技术能做什么。以人类尊严为例，我做了很多 AI 在医疗保健方面的研究，我很清楚，临床医生和医护人员所做的许多工作是人类关爱人类的一部分，那种情感纽带、尊严和尊重是永远无法被替代的。同样清楚的是，美国的医护人员，尤其是护士，过度疲劳、工作过度。如果技术能成为一种积极的力量，帮助他们更好地照顾病人，减轻他们的工作量，尤其是一些重复性的、吃力不讨好的工作，比如不断记录病历，或者每天走几英里取药等等——如果这些工作任务可以通过机器增强，那确实是为了保护人类的繁荣和尊严，同时增强人类的能力。从这个角度看，我认为 AI 有很多机会发挥积极作用。但同样，这取决于我们如何真正——首先，取决于我们如何设计 AI。在我的实验室，我们做了一项非常有趣的研究。我们试图创建一个大型机器人项目，来完成大约一千项人类日常任务。但在项目开始时，对我们来说非常重要的一点是，我们创造的机器人要执行那些人类希望得到帮助的任务。例如，买结婚戒指——我认为即使我们有世界上最好的机器人，谁想让机器人挑选结婚戒指呢？或者打开圣诞礼物——打开盒子并不难，但人类的情感、喜悦、家庭纽带、那个时刻，并不是关于打开一个傻盒子。所以我们实际上让人们为我们对成千上万的任务进行排序，告诉我们他们希望机器人做哪些任务。例如，清洁马桶——每个人都希望机器人帮忙。所以我们专注于那些人类更希望机器人帮助的任务，而不是那些人类在乎并想自己做的任务。这就是以人为本的 AI 的思考方式：我们如何创造对人类有益、受人类欢迎的技术，而不是我直接告诉你我要用机器人取代你在乎的一切。另外一层，为了结束这个话题，是政策层面。经济社会的福祉非常重要，技术专家并非无所不知，我们也不应该觉得自己无所不知。我们应该与公民社会、法律界、政策界、经济学家合作，去理解工作和任务以及 AI 影响的细微之处和深刻性。这也是为什么我们斯坦福以人为本 AI 研究所设立了一个数字经济实验室。我们与政策制定者合作，思考这些问题。我们努力为他们提供信息，并帮助以积极的方式推动这些议题。

Yeah, so several things, Hala. First of all, why do we have jobs? It's really important to think about it. I think jobs are part of human prosperity because we need that to translate into financial rewards so that we have the prosperity for our family. And we need it also as part of human dignity. It's beyond money; it's the meaning for many people, it's the meaning of life and self-respect. So from that point of view, I think we have to recognize job shift throughout human history. Technology makes, and also other factors, creates, destroys, morphs, transforms jobs. But what doesn't change is the need for human prosperity and human dignity. So I think when we think about AI and its impact on jobs, it's important to go to the very core of what jobs are and mean, and what technology can do. When it comes to human dignity, for example, I do a lot of healthcare research with AI, and it's so clear to me that many of the jobs that our clinicians and healthcare workers do are part of humans caring for humans, and that emotional bond, that dignity, that respect can never be replaced. What is also clear to me is that American healthcare workers, especially nurses, are over-fatigued, overworked. And if technology can be a positive force to help them take care of patients better, to reduce their workload, especially some of the repetitive, thankless work like constant charting or walking miles and miles a day to fetch pharmacy medicines and all that—if those parts of the job, the tasks, can be augmented by machines, it is truly intended to protect human prosperity and dignity but augment human capabilities. So from that point of view, I think there is a lot of opportunity for AI to play a positive role. But again, it depends on how we truly—first of all, it depends on how we design AI. In my lab, we did a very interesting research. We were trying to create a big robotics project to do around a thousand human everyday tasks. But at the beginning of this project, it was very important to us that we are creating robots to do these tasks that humans want help for. For example, buying a wedding ring—I don't think even if we have the best robot in the world, who wants a robot to choose a wedding ring? Or opening Christmas gifts—it's not that hard to open a box, but the human emotion, the joy, the family bond, the moment is not about opening a silly box. So we actually ask people to rank for us thousands and thousands of tasks and tell us which tasks they want robots for. For example, cleaning toilet—everybody wants robot help. So we focus on those tasks that humans prefer robotic help, rather than those tasks that humans care about and want to do themselves. And that is a way of thinking about human-centered AI: how do we create technology that is beneficial, welcomed by humans, rather than I just go in and tell you I'm using robots to replace everything you care about. Another layer, just to finish this topic, is the policy layer. Economic social well-being is so important, and technologists don't know it all, and we shouldn't feel we know it all. We should be collaborating with civil society, legal world, policy world, economists to try to understand the nuance and the profoundness of jobs and tasks and AI's impact. And this is also why our Human-Centered AI Institute at Stanford has a digital economy lab. We work with policymakers and think about these issues. We try to inform them and provide information and help move these topics forward in a positive way.

我觉得您谈到了很多——您的以人为本 AI 框架有三个层面，对吧？AI 是跨学科的，AI 需要确保我们拥有人类尊严并用于人类福祉，还有一个是关于智能的。您能详细说明一下您以人为本 AI 框架的三个支柱吗？

I feel like you're touching on a lot—you have three aspects to your human-centered AI framework, right? So AI is interdisciplinary, AI needs to be trying to make sure that we have human dignity and using it for human good, and then there's also one about intelligence. Can you break down your three pillars of your human-centered AI framework?

是的，以人为本 AI 框架的三个支柱实际上是关于 AI 的思想领导力，并专注于像斯坦福这样的高等教育机构能做什么。我们谈到的第一个是跨学科——认识到 AI 的跨学科性质，欢迎多利益相关方的研究、教育、政策推广，以确保 AI 以有益的方式嵌入我们今天和明天的社会结构。第二个是你提到的：专注于增强人类，创造增强人类能力、人类福祉和人类尊严的技术，而不是剥夺。第三个是关于继续从人类智能中汲取灵感，开发与人类兼容的 AI 技术。因为人类智能非常复杂、非常丰富。我们谈论很多情感、意图、同情心，而今天的 AI 缺乏这些，离这些还很远。受此启发可以帮助我们创造。顺便说一句，今天的 AI 还有一个比人类差得多的方面：它消耗大量能量。人类大脑大约以 20 瓦特工作，比你家里最暗的灯泡还要暗，但我们能做很多事情——我们可以建造金字塔，可以提出 E=mc²，可以创作优美的音乐等等。今天的 AI 非常非常耗能，体积庞大。所以人类智能中有很多可以启发下一代 AI 做得更好的地方。

Yeah, the three pillars of the human-centered AI framework is really about thought leadership in AI and focusing on what a higher education institute like Stanford can do. One we talked about is that interdisciplinary—recognizing the interdisciplinary nature of AI, welcoming the multi-stakeholder studies, research, education, policy outreach to make sure that AI is embedded in the fabric of our society today and tomorrow in a benevolent way. The second one is what you said: focusing on augmenting humans, creating technology that enhances human capability and human well-being and human dignity rather than taking away. The third one is about continuing to be inspired by human intelligence and develop AI technology that is compatible with humans. Because human intelligence is very complex, it's very rich. We talk a lot about emotion, intention, compassion, and today's AI lacks most of that; it's pretty far from that. Being inspired by this can help us to create. And by the way, there's another thing about today's AI that is far worse than humans: it draws a lot of energy. Humans' brain works around 20 watts, that is dimmer than the dimmest light bulb in your house, yet we can do so many things—we can create the pyramids, we can come up with E=mc², we can write beautiful music and all that. AI today is very, very energy-consuming; it's bulky, it's huge. So there's a lot in human intelligence that can inspire the next generation AI to do better.

每次我做 AI 相关的节目，我都觉得自己学到了很多以前没有意识到的东西。而且，我们在节目中和其他人讨论过，很多人害怕 AI 获得顶级智能，它会比人类聪明得多，会接管世界，会控制人类。您对此有恐惧吗？

Every time I have an AI episode, I feel like I learn so much that I didn't really realize before. And you know, we've had conversations with other people on the show about how a lot of people are scared of AI getting like apex intelligence, that it's going to be so much smarter than humans, it's going to take over the world, it's going to control humans. Do you have any fears around that?

我确实有恐惧。我想，生活在 2024 年，谁没有恐惧呢？作为世界公民，我认为我们的文明、我们的物种，总是由光明与黑暗的斗争所定义。

I do have fears. I think, you know, who lives in 2024 and doesn't have fears? And as a citizen of the world, I think our civilization, our species, is always defined by the struggle of dark and light.

我认为我们的 DNA 中既有令人难以置信的善意，也有令人难以置信的恶意。AI 作为一种技术，可以被恶意所利用。从这个角度看，我确实感到恐惧。我应对恐惧的方式是努力建设性地提供帮助，倡导这项技术的善意使用，并用它来对抗恶意。归根结底，我对 AI 的任何希望都不在于 AI 本身，而在于人类。借用马丁·路德·金博士的话，历史的长弧虽长，但终将弯向正义与善意。但抛开抽象思考，我认为我们还有工作要做。老实说，如果 AI 落入坏人之手，如果 AI 只集中在少数有权势的人手中，它可能会走向非常糟糕的方向。我们甚至不需要等到超级智能出现。就拿今天的汽车来说：想象一下，一个坏人负责建造美国 50%的汽车，他只想让所有汽车的刹车失灵，或者加装一个传感器，说‘如果你看到行人，就撞上去’。实际上，今天的技术就能做到这一点，不需要超级智能。但我们没有看到这种反乌托邦的场景，首先是因为人性总体上是善良的。我们的汽车工厂工人、汽车制造企业的领导者，没有人会想到这样做。我们还有法律。如果有人试图作恶，我们有社会约束。我们还努力教育民众向善。所以这一切都是艰苦的工作，我们需要在 AI 领域付出这种努力，以确保它不会作恶。

I think we have incredible benevolence in our DNA, but we also have incredible badness in our DNA. AI as a technology can be used by the badness. So from that point of view, I do have fear. The way I cope with fear is to try to be constructively helpful, to advocate for the benevolent use of this technology, and to use this technology to combat the badness. At the end of the day, any hope I have for AI is not about AI, it's about humans. To paraphrase Dr. King, the arc of history is long but it does bend towards justice and benevolence in general. But to come down from that abstract thinking, I think we have work to do. Honestly, because if AI is in the hands of bad actors, if AI is concentrated only in a few powerful people's hands, it can go very wrong. We don't need to wait for superintelligence. Even today's cars: imagine there's a bad person who is in charge of building 50% of America's cars, and that person just wants to make all the car brakes malfunction or add a sensor and say, 'If you see a pedestrian, run it over.' Actually, today's technology can do that. You don't need superintelligence. But the fact that we don't have that dystopian scenario is first of all, human nature is by and large good. Our car factory workers, our business leaders in building cars, nobody thinks about doing that. We also have laws. If someone is trying to do harm, we have societal constraints. We also try to educate the population towards good things. So all this is hard work, and we need that hard work in AI to ensure it doesn't do bad.

我想举个例子。我和斯蒂芬·沃尔夫勒姆聊过，因为那次采访我还记忆犹新。他说了一些让我对 AI 以及它可能变得非常聪明这件事感到些许安心的话。他说：‘我们生活在 AI 中。我们生活在自然中。自然如此复杂，我们无法控制它。它有非常简单却又极其复杂的过程。我们可以随心所欲地预测，但我们永远无法真正知道自然要做什么。我们已经生活在一个每天与自然互动的世界里，我们必须接受这样一个事实：我们无法控制它，它在某种程度上比我们更聪明。’他说：‘这也许就是未来 AI 的样子。它会在那里，它会成为自己的系统。’您对此有何看法？

I just want to give an example. When I was talking to Stephen Wolfram, because the interview is fresh in my head, he said something that made me feel a little bit at ease with AI and the fact that it could get really smart. He said, 'We're living in AI. We live in nature. Nature is so complex, we can't control it. It has simple processes that are really, really complex. We can predict it all we want, but we'll never really know what nature is going to do. And already we live in a world where we're interacting with nature every day, and we have to just deal with the fact that we don't control it and it's smarter than us to a degree.' He's like, 'That's what maybe AI will be like in the future. It will be there, it will be its own system.' What are your thoughts on that?

这是一个非常有趣的表述。好吧，我第一次听到这种说法。我喜欢他说的，人类在面对复杂和强大的事物时，仍然有办法与之共存。我同意。自然在某种意义上就是 AI，因为自然不可编程，而且我认为自然没有集体意图。地球并不想变得更大或更蓝。所以从这个角度看，两者非常非常不同。但我欣赏他的说法。我还认为，用他的类比来说，我们也与其他人类生活在一起，有些人比我们更强壮、更聪明、更优秀。但总体而言，我们的世界并不是每个人都在互相残杀。总体而言，这正是我们看到黑暗的地方，而这与 AI 无关。人性中有黑暗，我们会互相伤害。希望在于——不仅仅是希望，而是工作——当我们创造出与人类智能相似的机器时，我们应该防止它们对我们彼此造成类似的伤害，并努力激发出我们更好的一面。

That's a very interesting way to put it. Okay, first time I heard that. I like his way of saying that humans, in the face of complexity and powerful things, we still have a way to cohabit with it. I agree. Nature is AI in the sense that nature is not programmable, and I don't think nature has a collective intention. It's not like the Earth wants to be a bigger Earth or bluer Earth. So from that point of view, it's very, very different. But I appreciate the way he says that. And I also think, using his analogy, we also live with other humans, and there are humans who are stronger than us, smarter than us, do better than us. But yet, by and large, our world is not everyone killing each other. By and large, now this is where we do see the darkness, and this has nothing to do with AI. Human nature has darkness, and we harm each other. The hope is, it's not just the hope, the work is that when we create machines that resemble our intelligence, we should prevent them from doing similar harms to us and to each other, and try to bring out the better part of ourselves.

在我们结束这次访谈之前，因为需要准时结束，我想问您几个问题。首先，对于所有年轻的创业者——您现在和很多年轻创业者以及想成为创业者的人交流——您对他们如何拥抱这个 AI 世界有什么建议？

As we wrap up this interview, because I need to get you out on time, I wanted to ask you a couple of questions. First off, to all the young entrepreneurs you're talking to a lot of young entrepreneurs right now and people who want to be entrepreneurs, what's your advice to them about how to embrace this AI world?

首先，我希望你读过我的书《我看见的世界》，因为这本书是写给年轻人的，关于年轻人的。它是一位科学家的成长故事，但书的真正主题是找到你的北极星，找到你的热情，并排除万难相信它，追随它。这正是创业精神的核心：你相信能带给世界一些东西，并且不顾一切想要实现它，这应该是你的北极星。就 AI 而言，它是一个极其强大的工具。所以根据你正在做的业务和产品，它要么能赋能你，要么是你核心产品的重要组成部分，要么让你保持竞争力。它的横向性如此之强，以至于对大多数创业者来说，如果你对 AI 一无所知，那么自我教育就很重要，因为 AI 可能对你有利，也可能对你的竞争对手有利。所以了解这一点很重要。

First of all, I hope you read my book 'The Worlds I See' because the book is written to young people for young people. It's a coming-of-age of a scientist, but the true theme of the book is finding your North Star, finding your passion, and believing in that against all odds, and chasing after the North Star. That is the core of what entrepreneurship is about: you believing in bringing something to the world and against all odds you want to make it happen, and that should be your North Star. In terms of AI, it's an incredibly powerful tool. So depending on what business and products you're making, it either can empower you, or it's an essential part of your core product, or it keeps you competitive. So it's so horizontal that for most entrepreneurs out there, if you don't know anything about AI, it is important to educate yourself because it's possible that AI will play either in your favor or in your competitor's favor. So knowing that is important.

好的，时间差不多了，我再问您最后一个问题。这实际上是关于愿景的。让我们想象一个 10 年后，也就是 2034 年的世界，那里是以人为中心的 AI。同时，也让我们想象一个 10 年后的世界，也许不是以人为中心的 AI，也许它落入了坏人之手。我们来谈谈这两个世界，然后我们就结束。

Okay, and since we're just about time here, I'm just going to ask you one last question. This is really about visioning. Let's vision a world 10 years from now, 2034, where there is human-centered AI. And let's also try to visualize a world 10 years from now where maybe it's not human-centered AI, maybe it got in the bad hands of some folks. Let's talk about those two worlds, and then we'll close it out.

以人为中心的 AI 世界，我认为它至少离我们生活的北美世界不远，尽管我知道我们并不完美。我们仍然有强大的民主，我们仍然相信个人尊严，以及大体上的自由市场资本主义，我们作为个体被允许追求自己的幸福和繁荣，并相互尊重。而 AI 帮助我们进行更好的科学发现，拥有自动驾驶汽车，帮助那些不能开车的人或减少交通拥堵，让生活更轻松，让教育更个性化，赋能我们的教师和医疗工作者，发现疾病的治疗方法，缓解人口老龄化问题，提高农业效率，寻找气候解决方案。AI 在这个世界上可以做很多事情，而我们仍然拥有良好的基础。至于反乌托邦的世界，AI 可以被用作推翻民主的邪恶工具。虚假信息是一种极其有害的方式，会损害我们现有的民主和公民生活。如果 AI 完全集中在权力手中，无论是国家权力还是个人权力，它会使社会其他部分更加受制于该权力的意志甚至愤怒。无论是不是 AI，我们在人类历史上已经看到，集中的权力总是坏的，而集中的权力加上强大的技术并不是好事。

The world that is human-centered AI, I think it's not too far from at least the North American world we live in, even though I know we're not perfect. We still have a strong democracy, we still believe in individual dignity, and by and large free market capitalism, that we are allowed as individuals to pursue our happiness and prosperity and respect each other. And AI helps us to do better scientific discovery, to have self-driving cars, to help people who can't drive or reduce traffic, to make life easier, to make education more personalized, to empower our teachers and healthcare workers, to discover cures for diseases, to alleviate our aging population problems, to make agriculture more effective, to find climate solutions. There is so much AI can do in the world that we still have the good foundation. Now the dystopian world is AI can be used as a bad tool to topple democracy. Disinformation is an incredibly harmful way of harming democracy and the civil life we have right now. If it's completely concentrated in power, whether it's state power or individual power, it makes the rest of the society much more subject to the will and possibly wrath of that power. Whether it's AI or not, we have seen in human history that concentrated power is always bad, and concentrated power using powerful technology is not a recipe for good.

可以了解更多关于你和你的工作。谢谢你，Hala。感谢你推广我的书，请持续关注斯坦福以人为本人工智能研究所的新闻通讯和网站。太棒了。我们会把所有链接放在节目说明中。李博士，感谢您做客《年轻与成长》播客。谢谢你，Hala。

Can learn more about you and everything that you do. Thank you, Hala. Thank you for promoting my book and please constantly check with STF for Human Center AI Institute newsletter and website. Amazing. We'll stick all those links in the show notes. Dr. Lee, thank you for joining us on Young and Profiting podcast. Thank you, Hala.