AI:可能会出什么错?

我现在是在神经网络学习 201 还是 101？你就像坐在前排那个什么都不知道但会问好问题的聪明学生。

Am I in neural learning 201 yet or am I still in 101? You're like the smart student in the front row who doesn't know anything but ask these good questions.

这是我被形容过的最好的方式。谢谢。

That's the nicest way I've ever been described. Thank you.

大家好，欢迎收听每周秀播客。我是乔恩·斯图尔特，今天由我主持。今天是 10 月 8 日星期三。我不知道今天晚些时候会发生什么，但我们明天会休息。今天的节目，我想快速说一下，我们邀请到了被称为「AI 之父」的杰弗里·辛顿先生，他从 70 年代就开始开发后来演变成 AI 的技术。我想告诉你们，我们聊了这些。不过第一部分他给我们分解了 AI 到底是什么，这对我来说非常有帮助。我们也会谈到「它会杀死我们所有人」的部分。但为了我的理解，先设定场景很重要。所以我希望你们也觉得这部分像我一样有趣，因为它以非常有趣的方式扩展了我对这项技术是什么、它将如何被使用、以及一些潜在危险的理解。我就不再耽搁了。让我们请出今天的嘉宾。女士们先生们，我们非常激动地欢迎多伦多大学计算机科学系名誉教授、施瓦茨·赖斯曼研究所顾问委员会成员杰弗里·辛顿。先生，非常感谢您今天与我们在一起。

Hey everybody. Welcome to the weekly show podcast. My name is Jon Stewart. I'm going to be hosting you today and it's a what is it? Wednesday, October 8th. I don't know what's going to happen later on in the day. But we're going to be out tomorrow. But today's episode I just want to say very quickly today's episode we are talking to someone known as the Godfather of AI, a gentleman by the name of Geoffrey Hinton who has been developing the type of technology that has turned into AI since the 70s. And I want to let you know so we talk about it. The first part of it though he gives us this breakdown of kind of what it actually is which for me was unbelievably helpful. We get into the it will kill us all part. But it was important for my understanding to sort of set the scene. So I hope you find that part as interesting as I did because man it expanded my understanding of what this technology is, of how it's going to be utilized, of what some of those dangers might be in a really interesting way. So I will not hold it up any longer. Let us get to our guest for this podcast. Ladies and gentlemen, we are absolutely thrilled today to be able to welcome Professor Emeritus with the Department of Computer Science at the University of Toronto and Schwartz Reisman Institute's advisory board member, Geoffrey Hinton is joining us. Sir, thank you so much for being with us today.

非常感谢您的邀请。我很高兴。

Well, thank you so much for inviting me. I'm delighted.

您被称为 AI 之父，我相信您对此会很谦虚。因为您在神经网络方面的工作，您共同获得了 2024 年的诺贝尔物理学奖。是这样吗？

You are known as, and I'm sure you will be very demure about this, the Godfather of artificial intelligence. For your work on these neural networks. You co-won the actual Nobel Prize in physics in 2024 for this work. Is that correct?

是的。这有点尴尬，因为我不是搞物理的。所以他们打电话说我得了诺贝尔物理学奖时，我一开始还不相信。

That is correct. It's slightly embarrassing since I don't do physics. So when they called me up and said you won the Nobel Prize in physics, I didn't believe them to begin with.

其他物理学家会不会说：「等等，那家伙根本不是我们这行的。」

And were the other physicists going, "Wait a second. That guy is not even in our business."

我强烈怀疑他们这么想了，但他们没当着我的面说。

I strongly suspect they were but they didn't do it to me.

哦，那就好。我很高兴。

Oh good. I'm glad.

这对您来说可能有点基础，但当我们谈论人工智能时，我不太确定我们在说什么。我知道有大型语言模型这种东西。根据我的经验，人工智能只是一个稍微更讨人喜欢的搜索引擎。以前我用谷歌搜索，它直接给出答案。现在它会说：「你问了一个有趣的问题。」那么，当我们谈论人工智能时，我们到底在说什么？

This is going to seem somewhat remedial, I'm sure to you. But when we talk about artificial intelligence, I'm not exactly sure what it is that we're talking about. I know there are these things large language models. To my experience artificial intelligence is just a slightly more flattering search engine. Whereas I used to Google something and it would just give me the answer. Now it says, "What an interesting question you've asked me." So what are we talking about when we talk about artificial intelligence?

以前你用谷歌时，它用的是关键词。它事先做了很多工作。所以如果你给几个关键词，它能找到所有包含这些词的文档。所以基本上它只是排序。它浏览、排序、找词，然后给你结果。

So when you used to Google, it would use keywords. And it would have done a lot of work in advance. So if you gave it a few keywords, it could find all the documents that had those words in. So basically it's just sorting. It's looking through and it's sorting and finding words and then bringing you a result.

对，以前就是这样工作的。

Yeah, that's how it used to work.

但它不理解问题是什么。所以它不能给你那些不包含这些词但关于同一主题的文档。它没有建立那种联系。

But it didn't understand what the question was. So it couldn't for example give you documents that didn't actually contain those words but were about the same subject. It didn't make that connection.

对，因为它会说「这是你的结果，减去某个词」，然后说一个没包含的词。

Right, because it would say here is your result minus and then it would say like a word that was not included.

对。但如果你有一个文档，里面没有你用的任何词，它就找不到，尽管那可能是一个非常相关的文档，正好关于你谈论的主题，只是用了不同的词。现在它理解你说的话，而且理解方式几乎和人一样。

Right. But if you had a document with none of the words you used, it wouldn't find that even though it might be a very relevant document about exactly the subject you were talking about. It had just used different words. Now it understands what you say and it understands in pretty much the same way people do.

所以如果我说什么，它会说：「哦，我明白你的意思。让我给你讲讲这个。」所以它从字面上只是一个搜索查找的东西，变成了几乎是你讨论的任何领域的专家，它能带来你可能没想到的东西。

So if I say something, it'll say, "Oh, I know what you mean. Let me educate you on this." So it's gone from being literally just a search and find thing to an actual almost an expert in whatever it is that you're discussing and it can bring you things that you might not have thought about.

是的。所以大型语言模型并不是所有方面的专家。如果你有一些对某个主题很了解的朋友，他们可能比大型语言模型好一点。但他们仍然会惊讶于大型语言模型对他们的主题了解得相当好。

Yes. So the large language models are not very good experts at everything. So if you take some friends you have who know a lot about some subject matter, they're probably a bit better than the large language model. But they'll nevertheless be impressed that the large language model knows their subject pretty well.

机器学习有什么区别？谷歌作为搜索引擎是机器学习吗？那只是算法和预测。

What is the difference between machine learning? Was Google in terms of a search engine machine learning? That's just algorithms and predictions.

不完全是。机器学习是一个总称，指计算机上任何能学习的系统。而神经网络是一种特定的学习方式，与之前使用的非常不同。旧的机器学习不被认为是神经网络。

Not exactly. Machine learning is a kind of coverall term for any system on a computer that learns. Now these neural networks are a particular way of doing learning that's very different from what was used before. The old machine learning, those were not considered neural networks.

你说神经网络，你的工作可以说是它的起源，是在 70 年代，当时你认为你在研究大脑。是这样吗？

And when you say neural networks meaning your work was sort of the genesis of it was in the 70s where you thought you were studying the brain. Is that correct?

我试图提出关于大脑实际如何学习的想法。我们对此有一些了解。它通过改变脑细胞之间连接的强度来学习。

I was trying to come up with ideas about how the brain actually learned. And there's some things we know about that. It learns by changing the strengths of connections between brain cells.

解释一下。它通过改变连接来学习。所以如果你给人类展示新东西，脑细胞会在脑细胞内建立新连接。

Explain that. It learns by changing the connections. So if you show a human something new, brain cells will make new connections within brain cells.

它不会建立新连接。会是已经存在的连接。但它运作的主要方式是改变这些连接的强度。

It won't make new connections. It'll be connections that were there already. But the main way it operates is it changes the strength of those connections.

所以如果你从大脑中间一个神经元的角度来看，一个脑细胞。它一生中能做的就是有时发出「乒」的信号。它只有这个。它有时会「乒」。它必须决定什么时候「乒」。

So if you think of it from the point of view of a neuron in the middle of the brain, a brain cell. All it can do in life is sometimes go ping. That's all it's got. It can sometimes go ping. And it has to decide when to go ping.

它如何决定什么时候「乒」？很高兴你问这个问题。有其他神经元在「乒」。当它看到其他神经元「乒」的特定模式时，它就「乒」。你可以把这个神经元想象成接收来自其他神经元的「乒」。每次收到一个「乒」，它就把它当作投票数，决定自己是否应该开启、是否应该「乒」或不应该「乒」。你可以改变另一个神经元对它的投票数。

How does it decide when to go ping? I'm glad you asked that question. There are other neurons going ping. And when it sees particular patterns of other neurons going ping, it goes ping. And you can think of this neuron as receiving pings from other neurons. And each time it receives a ping it treats that as a number of votes for whether it should turn on or should go ping or should not go ping. And you can change how many votes another neuron has for it.

你怎么改变那个投票？

How would you change that vote?

通过改变连接的强度。连接的强度，可以看作另一个神经元给你「乒」的票数。

By changing the strength of the connection. The strength of the connection, think of as the number of votes this other neuron gives for you to go ping.

所以从某些方面来说，它有点像，让我想起电影《小黄人》。但几乎是一种社交行为。

So it really is in some respects it's a boy, it reminds me of the movie Minions. But it's almost a social thing.

是的，很像政治联盟。

Yes, it's very like political coalitions.

会有神经元群组一起发出脉冲。这个群组里的神经元会互相告诉对方发出脉冲。然后可能会有另一个联盟，它们会告诉其他神经元不要发出脉冲。接着可能还有另一个联盟。它们都互相告诉对方发出脉冲，同时告诉第一个联盟不要发出脉冲。

There'll be groups of neurons that go ping together. And the neurons in that group will all be telling each other go ping. And then there might be a different coalition and they'll be telling other neurons don't go ping. And then there might be a different coalition. And they're all telling each other to go ping and telling the first coalition not to go ping.

所有这些都在你的大脑中进行，就像我想拿起一个勺子。所以勺子，例如，你大脑中的勺子是一个一起发出脉冲的神经元联盟。那是一个概念。所以当你还是婴儿时，别人说勺子，就有一小组神经元在说「哦，那是勺子。」它们正在加强彼此之间的连接。这就是为什么当你对大脑进行成像时，会看到某些区域亮起来吗？那些区域亮起来就是为特定物品或动作发出脉冲的神经元吗？

All this is going on in your brain in the way of like I would like to pick up a spoon. So spoon for example, spoon in your brain is a coalition of neurons going ping together. And that's a concept. So as you're teaching when you're a baby and they go spoon, there's a little group of neurons going "Oh, that's a spoon." And they're strengthening their connections with each other. Is that why when you're imaging brains you see certain areas light up? And is that lighting up of those areas the neurons that ping for certain items or actions?

不完全是。接近了。当你做不同事情时，不同区域会亮起来，比如当你进行视觉、说话或控制手部时，不同区域会为此亮起。但是当有勺子时一起发出脉冲的神经元联盟并不仅仅为勺子工作。那个联盟的大多数成员在有叉子时也会发出脉冲。所以这些联盟重叠很多。

Not exactly. Getting close. Different areas will light up when you're doing different things like when you're doing vision or talking or controlling your hands, different areas light up for that. But the coalition of neurons that goes ping together when there's a spoon don't only work for spoon. Most of the members of that coalition will go ping when there's a fork. So they overlap a lot, these coalitions.

这是一个大帐篷。这是一个大帐篷联盟。我喜欢把它看作政治性的。我不知道你的大脑是靠同伴压力运作的。

This is a big tent. It's a big tent coalition. I love thinking about this as political. I had no idea your brain operates on peer pressure.

是的，有很多这样的情况。概念就像是快乐的联盟。但它们重叠很多，比如狗的概念和猫的概念有很多共同点。它们会有很多共享的神经元。特别是代表诸如「这是有生命的」或「这是有毛的」或「这可能是宠物」的神经元。所有这些神经元在猫和狗中都是共通的。

There's a lot of that goes on, yes. And concepts are kind of coalitions that are happy together. But they overlap a lot like the concept for dog and the concept for cat have a lot in common. They'll have a lot of shared neurons. In particular, the neurons that represent things like this is animate or this is hairy or this might be a domestic pet. All those neurons will be in common to cat and dog.

是否有某些神经元广泛地为动物的广义概念发出脉冲，然后其他神经元像从宏观到微观、从一般到具体那样工作？所以，你有一个一般性发出脉冲的神经元联盟，然后随着知识变得更具体，是否会激活某些脉冲频率较低但可能更特化的神经元？是这样吗？

Are there certain neurons that ping broadly for the broad concept of animal and then other neurons like does it work from macro to micro, from general to specific? So, you have a coalition of neurons that ping generally and then as you get more specific with the knowledge, does that engage certain ones that will ping less frequently, but for maybe more specificity? Is that something?

这是一个非常好的理论。没有人确切知道这一点。一个非常合理的理论。特别是，那个联盟中会有一些神经元更频繁地为更一般的事物发出脉冲。然后可能有一些神经元为更具体的事物发出脉冲频率较低。这贯穿始终，就像你说的，有某些区域会为视觉或其他感官、触觉发出脉冲。我想象有一个语言的脉冲系统。

That's a very good theory. Nobody really knows for sure about this. A very sensible theory. And in particular, there's going to be some neurons in that coalition that ping more often for more general things. And then there may be neurons that ping less often for much more specific things. And this works throughout and like you say, there's certain areas that will ping for vision or other senses, touch. I imagine there's a ping system for language.

你刚才说，如果我们能让计算机更像——我认为只是二进制的、if-then 那种基本的——你是在说我们能让它们像这些联盟一样工作吗？

You were saying what if we could get computers which were much more, I would think, just binary, if-then, sort of basic. You're saying could we get them to work as these coalitions?

是的，我认为二进制、if-then 与此关系不大。区别在于人们试图把规则输入计算机。他们试图弄清楚编程计算机的基本方式：你极其详细地弄清楚如何解决问题，分解所有步骤，然后精确地告诉计算机做什么。那是普通的计算机程序。这些东西完全不是那样的。

Yeah, I don't think binary, if-then, has much to do with it. The difference is people were trying to put rules into computers. They were trying to figure out the basic way you program a computer: you figure out in exquisite detail how you would solve the problem, you deconstruct all the steps, and then you tell the computer exactly what to do. That's a normal computer program. These things aren't like that at all.

所以，你试图改变这个过程，看看我们能否创建一个更像人脑运作方式而不是逐条指令列表的过程。你希望它更全局地思考。那是怎么发生的？

So, you were trying to change that process to see if we could create a process that functioned more like how the human brain would rather than an item-by-item instruction list. You wanted it to think more globally. How did that occur?

对很多人来说，很明显大脑不是靠别人给你规则然后你执行那些规则来工作的。我的意思是，在朝鲜，他们希望大脑那样工作，但事实并非如此。

It was sort of obvious to a lot of people that the brain doesn't work by someone else giving you rules and you just execute those rules. I mean, in North Korea, they would love brains to work like that, but they don't.

你是说在威权世界里，大脑会那样运作。好吧，那是他们希望它们运作的方式。

You're saying that in an authoritarian world, that is how brains would operate. Well, that's how they would like them to operate.

这比那更艺术一点。我们确实为神经网络编写程序，但这些程序只是告诉神经网络如何根据神经元的活动来调整连接强度。所以，那是一个相当简单的程序。它不包含关于世界的各种知识。它只是根据活动改变连接强度的规则是什么？

It's a little more artsy than that. We do write programs for neural nets, but the programs are just to tell the neural net how to adjust the strength of the connection on the basis of the activities of the neurons. So, that's a fairly simple program. It doesn't have all sorts of knowledge about the world in it. It's just what are the rules for changing your connection strengths on the basis of the activities?

你能给我举个例子吗？那么，这算是机器学习还是深度学习？那会是什么？

Can you give me an example? So, would that be considered machine learning or deep learning? What would that be?

如果你有一个多层网络，那就是深度学习。之所以叫深度学习，是因为有很多层。

That's deep learning if you have a network with multiple layers. It's called deep learning because there's many layers.

那么，当你试图让计算机进行深度学习时，你对它说什么？比如你会给出什么指令的例子？

So, what are you saying to a computer when you are trying to get it to do deep learning? Like what would be an example of an instruction that you would give?

让我回到 1949 年。这里有一个来自唐纳德·赫布的理论，关于如何改变连接强度。如果神经元 A 发出脉冲，紧接着神经元 B 发出脉冲，就增加连接的强度。这是一个非常简单的规则。

Let me go back to 1949. So, here's a theory from someone called Donald Hebb about how you change connection strengths. If neuron A goes ping and then shortly afterwards neuron B goes ping, increase the strength of the connection. That's a very simple rule.

这叫做赫布规则。赫布规则是，如果神经元 A 兴奋，神经元 B 也兴奋，就增强它们之间的连接。

That's called the Hebb rule. The Hebb rule is if neuron A fires and neuron B fires, increase the connection between them.

对。

Right.

一旦计算机出现，就可以进行计算机模拟。人们发现这个规则本身不起作用。结果是所有连接都变得非常强，所有神经元同时兴奋，导致癫痫发作。

As soon as computers came along, you could do computer simulations. People discovered that rule by itself doesn't work. What happens is all the connections get very strong and all the neurons fire at the same time, and you have a seizure.

哦，好吧。这很遗憾，不是吗？

Oh, okay. That's a shame, isn't it?

确实遗憾。必须有某种机制既能增强连接也能减弱连接。必须有所辨别。

That is a shame. There's got to be something that makes connections weaker as well as stronger. There's got to be some discernment.

假设我们想构建一个多层神经元的神经网络，来判断图像中是否包含鸟，就像验证码那样。我们想用神经网络解决那个验证码。

Suppose we wanted to make a neural network with multiple layers of neurons to decide whether an image contains a bird or not, like a CAPTCHA. We want to solve that CAPTCHA with a neural net.

正是。

Exactly.

神经网络的输入，即底层神经元，代表图像中像素的强度。如果是一千乘一千的图像，你有一百万个神经元以不同速率兴奋，表示每个像素的强度。这就是输入。现在你必须将其转化为一个决策：这是不是一只鸟？

The input to the neural net, the bottom layer of neurons, represents the intensities of the pixels in the image. If it's a thousand by thousand image, you have a million neurons firing at different rates to represent how intense each pixel is. That's your input. Now you have to turn that into a decision: is this a bird or not?

你编程吗？因为像素强度在我看来并不是判断是否为鸟的有用工具。判断是否为鸟的工具似乎是：那些是羽毛吗？那是喙吗？

Do you program in? Because strength of pixel doesn't strike me as a really useful tool for figuring out if it's a bird. Figuring out if it's a bird seems like the tool would be: are those feathers? Is that a beak?

冠羽？

A crest?

对。

Yeah.

像素本身并不能告诉你它是不是鸟。鸟有明亮的也有暗淡的，有飞的也有坐着的，有近在眼前的鸵鸟也有远处的海鸥。它们都是鸟。那么下一步做什么？受大脑启发，人们说让我们有一堆边缘检测器。我们要制造一些神经元，检测小片边缘——图像中一侧亮一侧暗的地方。这就是如何构建视觉系统，无论是在大脑中还是在计算机中。

The pixels by themselves don't really tell you whether it's a bird. You can have birds that are bright and birds that are dark, birds flying and birds sitting down, an ostrich in your face and a seagull in the distance. They're all birds. So what do you do next? Guided by the brain, people said let's have a bunch of edge detectors. We're going to make neurons that detect little pieces of edge—places in the image where it's bright on one side and darker on the other. This is how you make a vision system, both in the brain and in computers.

哇，好的。

Wow. Okay.

如果你想检查特定位置的一小段垂直边缘，假设你看一列三个像素，旁边另一列三个像素。如果左边亮右边暗，你就想说有边缘。所以你得问，我如何制造一个神经元来做这件事？

If you wanted to check a little piece of vertical edge in a particular place, suppose you look at a little column of three pixels and next to them another column of three pixels. If the ones on the left are bright and the ones on the right are dark, you want to say yes, there's an edge here. So you have to ask, how would I make a neuron that did that?

哦天哪。好吧。好的，我要跳过去了。

Oh my god. Okay. All right, I'm going to jump ahead.

在过去，人们会尝试加入各种规则来教它如何看，解释什么是前景和背景。但真正相信神经网络的人说：「不，不要加入那些规则。让它只从数据中学习所有规则。」它学习的方式是，一旦开始识别边缘等，就加强连接。

In the old days, people would try to put in all sorts of rules to teach it how to see and explain what foreground and background were. But the people who really believed in neural nets said, 'No, don't put in all those rules. Let it learn all those rules just from data.' The way it learns is by strengthening the connections once it starts to recognize edges and things.

你跳得太快了。

You're jumping ahead.

我们继续讲这个边缘检测器。在第一层，你有代表像素亮度的神经元。在下一层，我们会有小边缘检测器。你可能有一个神经元连接到左边三个像素和右边三个像素。如果你把左边三个像素的连接强度设为大的正值，右边三个像素设为大的负值，那么当左右像素亮度相同时，负连接抵消正连接，没有反应。但如果左边亮右边暗，神经元从左边得到大量输入，没有来自右边的抑制，所以它兴奋。它说：「我在这里找到了一小段边缘。」

Let's carry on with this little bit of edge detector. In the first layer, you have neurons that represent how bright the pixels are. In the next layer, we're going to have little bits of edge detector. You might have a neuron in the next layer connected to a column of three pixels on the left and a column of three pixels on the right. If you make the strengths of the connections to the three pixels on the left strong positive connections, and the strengths to the three pixels on the right big negative connections, then when the pixels on the left and right are the same brightness, the negative connections cancel the positive ones, and nothing happens. But if the pixels on the left are bright and the pixels on the right are dark, the neuron gets lots of input from the left and no inhibition from the right, so it fires. It says, 'I found a little piece of edge here.'

我就是那个家伙。我是边缘家伙。我在边缘上兴奋。

I'm that guy. I'm the edge guy. I ping on the edges.

对。现在想象你有无数个这样的检测器，因为它们必须检测视网膜上任何位置、任何方向、不同尺度的小边缘。随着你制造更多边缘检测器，你对边缘的辨别能力会更好。

Right. Now imagine you have a gazillion of those, because they have to detect little pieces of edge anywhere on your retina, at any orientation, and at different scales. As you make more edge detectors, you get better discrimination for edges.

好的。

Okay.

现在进入下一层。假设我们有一个神经元，寻找几乎水平的边缘组合：几行几乎水平的边缘排成一行，就在它们上方，又有几行几乎水平的边缘，但与第一组边缘形成一个尖点。这样你就找到了两个小边缘组合，构成一个尖状物。

Now let's go to the next layer. Suppose we had a neuron that looked for a little combination of edges that are almost horizontal, several edges in a row that are almost horizontal and line up, and just slightly above those, several edges in a row that are again almost horizontal but come down to form a point with the first set. So you find two little combinations of edges that make a sort of pointy thing.

所以你是一位诺贝尔奖得主物理学家。

So you're a Nobel Prize winning physicist.

我没想到那句话会以「它有点像尖尖的东西」结尾。我以为会有个名字。但我明白你的意思。你现在在辨别它在哪里结束，你在看不同的……而且这甚至在你考虑颜色或其他东西之前。这纯粹就是：有没有图像，边缘是什么？边缘的小组合是什么？所以我们现在问，有没有一种边缘的小组合能构成可能像鸟喙的东西？那就是尖尖的东西。但你还不知道鸟喙是什么。

I did not expect that sentence to end with 'it makes kind of a pointy thing'. I thought there'd be a name for that. But I get where you're saying. You're now discerning where it ends, you're looking at different... And this is before you're even looking at color or anything else. This is literally just: is there an image, what are the edges? And what are the little combinations of edges? So we're now asking, is there a little combination of edges that makes something that might be a beak? That's the pointy thing. But you don't know what a beak is yet.

还不知道，不。我们也要学习那个。

Not yet, no. We need to learn that too.

对。所以一旦你有了这个系统，几乎就像你在构建能模仿人类感官的系统。这正是我们在做的。视觉、听觉，显然不是嗅觉，尽管……

Right. So once you have the system, it's almost like you're building systems that can mimic the human senses. That's exactly what we're doing. Vision, ears, not smell, obviously, although...

不，他们现在正在做。他们开始研究嗅觉了。

No, they're doing that now. They're starting on smell now.

哦，天哪。可能还有触觉。

Oh, for God's sake. And probably touch.

他们现在有了数字嗅觉，你可以通过网络传输气味。这简直……太疯狂了。气味打印机有 200 种成分。不是三种颜色，而是 200 种成分。它在另一端合成一种气味。虽然不是完美，但已经相当不错了。

They've now got digital smell, where you can transmit smells over the web. It's just... insane. The printer for smells has 200 components. Instead of three colors, it's got 200 components. And it synthesizes a smell at the other end. And it's not quite perfect, but it's pretty good.

哇。这对我来说太不可思议了。好的，非常抱歉。我深表歉意。

Wow. So this is incredible to me. Okay, I am so sorry about this. I apologize profusely.

这很完美。你很好地代表了一个对此一无所知但明智好奇的人。

This is perfect. You're doing a very good job of representing a sort of sensible curious person who doesn't know anything about this.

那么让我说完如何手动构建这个系统。

So let me finish describing how you would build the system by hand.

是的。如果我手动做，我会从这些边缘检测器开始。我会说，「从左边这些像素建立大的强正连接，到右边像素建立大的强负连接。」接收这些传入连接的神经元会检测到一小段垂直边缘。然后在下一层，我会说，「从三个这样倾斜的小边缘片段和三个那样倾斜的小边缘片段建立大的强正连接。」这可能是一个鸟喙，一个尖尖的东西。这是一个潜在的鸟喙。在同一层，我也可能从大致形成一个圆形的边缘组合建立大的强正连接。那是一个潜在的眼睛。现在在下一层，我有一个神经元，它查看可能的鸟喙和可能的眼睛。如果它们处于正确的相对位置，它就会说，「嘿，我很高兴。」因为那个神经元检测到了一个可能的鸟头。那个家伙可能会激活。同时，其他地方会有其他神经元检测到像鸡脚或鸟翅膀末端羽毛这样的小模式。所以你有一大堆这样的家伙。现在更高层，你可能有一个神经元说，「如果我检测到了鸟头、鸡脚和翅膀末端，那很可能是一只鸟。」所以它会说「鸟」。所以你可以看到如何尝试手动连接所有这些。

Yes. So if I did it by hand, I'd start with these edge detectors. I'd say, 'Make big strong positive connections from these pixels on the left and big strong negative connections to the pixels on the right.' Now the neuron that gets those incoming connections is going to detect a little piece of vertical edge. Then at the next layer, I'd say, 'Make big strong positive connections from three little bits of edge sloping like this and three little bits of edge sloping like that.' Could be a beak, a pointy thing. And this is a potential beak. In that same layer, I might also make big strong positive connections from a combination of edges that roughly form a circle. That's a potential eye. Now in the next layer, I have a neuron that looks at possible beaks and possible eyes. If they're in the right relative position, it says, 'Hey, I'm happy.' Because that neuron has detected a possible bird's head. And that guy might ping. At the same time, there'll be other neurons elsewhere that have detected little patterns like a chicken's foot or the feathers at the end of the wing of a bird. So you have a whole bunch of these guys. Now even higher up, you might have a neuron that says, 'If I've detected a bird's head and I've detected a chicken's foot and I've detected the end of a wing, it's probably a bird.' So it'd say bird. So you can see how you might try and wire all that up by hand.

对。所以你现在可以看到如何尝试手动连接所有这些。是的，那会花一些时间。会花很久很久。

Right. So you can see now how you might try and wire all that up by hand. Yes, and it would take some time. It would take like forever.

会花很久很久。是的。

It would take like forever. Yes.

好的。假设你很懒。现在你说到点子上了。你可以做的是，直接构建这些神经元层，而不指定所有连接的强度。你只是用小的随机数初始化它们。随便放一些旧的强度。然后你输入一张鸟的图片。假设它有两个输出。一个说「鸟」，另一个说「非鸟」。由于连接强度是随机的，会发生什么？你输入一张鸟的图片，它会说「50%鸟，50%非鸟。」换句话说，它完全不知道。你输入一张非鸟的图片，它也会说「50%鸟，50%非鸟。」哦，天哪。

Okay. So suppose you were lazy. Now you're talking. What you could do is you could just make these layers of neurons without saying what the strengths of all the connections ought to be. You just start them off at small random numbers. Just put in any old strengths. And you put in a picture of a bird. And let's suppose it's got two outputs. One says bird and the other says not bird. With random connection strengths in there, what's going to happen is you put in a picture of a bird and it says, '50% bird, 50% not bird.' In other words, I haven't got a clue. And you put in a picture of a non-bird. And it says, '50% bird, 50% non-bird.' Oh, boy.

好的。现在你可以问一个问题。假设我取其中一个连接强度，稍微改变它，让它更强一点。它会不会从说「50%鸟」变成说「50.01%鸟」和「49.99%非鸟」？如果输入的是鸟，那这就是一个好的改变。你让它稍微好了一点。

Okay. So now you can ask a question. Suppose I were to take one of those connection strengths and I were to change it just a little bit, make it maybe a little bit stronger. Instead of saying 50% bird, would it say 50.01% bird and 49.99% non-bird? And if it was a bird, then that's a good change to make. You've made it work slightly better.

这是哪一年？什么时候开始的？

What year was this? When did this start?

哦，正是。这只是一个想法。这永远不会奏效，但请耐心听我说。

Oh, exactly. So this is just an idea. This would never work, but bear with me.

好吧。这就像那些辩护律师，他们大段跑题，但最终都会好起来。

All right. This is like one of those defense lawyers who goes off on a huge digression, but it's all going to be good in the end.

不，不，不，不，不。这很有帮助。这就是十年后要毁灭我们所有人的东西。当我说「是的」，我不是指这个具体的东西，而是它的一个进步。不一定会毁灭我们所有人，但可能。

No, no, no, no, no. This is helpful. And this is the thing that's going to kill us all in 10 years. When I say yep, I mean not this particular thing, but an advancement on it. Not necessarily kill us all, but maybe.

对，对，对。这就像奥本海默说「好吧」。

Right. Right. Right. This is Oppenheimer going 'okay'.

所以你有一个物体，它由更小的物体组成。这是非常早期的部分。假设你有全世界所有的时间。你可以做的是，取这一层神经网络，从随机连接强度开始。然后给它看一张鸟的图片。它会说「50%鸟，50%非鸟」。你可以选择一个连接强度。然后你说，「如果我稍微增加它，有帮助吗？」不会有太大帮助，但到底有没有帮助？它能不能让我达到 50.1、50.2 之类的？如果有帮助，就增加它。然后你继续，再做一次。也许这次我们选一张非鸟的图片。我们选择一个连接强度。我们希望……如果我们增加那个连接强度，它说更不可能是鸟，更可能是非鸟，我们就说，「好的，这是一个好的增加。我们就这么做。」现在问题来了。有万亿个连接。每个连接需要改变很多次。那是手动的吗？嗯，用这种方法会是手动的。不仅如此，你不能只基于一个例子来做。因为有时改变一个连接强度，如果稍微增加，对这个例子有帮助，但会使其他例子变差。所以你必须给它一整批例子。看看平均是否有帮助。这就是你创建这些大型语言模型的方式……如果我们用这种非常笨的方法来创建，比如说，这个视觉系统，我们需要做数万亿次实验。每次实验都要给它一整批例子，看看改变一个连接强度是有帮助还是有损害。这永远做不完。会是无限的。

So you've got an object, and that is made up of smaller objects. This is the very early part of this. So suppose you had all the time in the world. What you could do is you could take this layer of neural network and you could start with random connection strengths. And you could then show it a bird. And it'd say 50% bird, 50% non-bird. And you could pick one of the connection strengths. And you could say, 'If I increase it a little bit, does it help?' It won't help much, but does it help at all? Will it get me to 50.1, 50.2, that kind of thing? If it helps, make that increase. Then you go around and do it again. Maybe this time we choose a non-bird. And we choose one connection strength. And we'd like it to... If we increase that connection strength and it says it's less likely to be a bird and more likely to be a non-bird, we say, 'Okay, that's a good increase. Let's do that one.' Now here's the problem. There's a trillion connections. And each connection has to be changed many times. And is that manual? Well, in this way of doing it will be manual. And not just that, but you can't just do it on the basis of one example. Because sometimes changing a connection strength, if you increase it a bit, it'll help with this example, but it'll make other examples worse. So you have to give it a whole batch of examples. And see if on average it helps. And that's how you create these large language models... If we did it this really dumb way to create, let's say, this vision system for now, we'd have to do trillions of experiments. And each experiment would involve giving it a whole batch of examples and seeing if changing one connection strength helps or hurts. And it would never be done. It would be infinite.

哦天哪。永远做不完。会是无限的。

Oh god. And it would never be done. It would be infinite.

现在，假设你想出了一种计算方法，可以同时告诉你网络中每个连接强度——对于这个特定例子，假设你给它一张鸟的图片，它判断有 50%的概率是鸟——那么对于每一个连接强度，所有一万亿个连接强度，我们可以同时判断是应该稍微增加它们还是稍微减少它们来改善结果。然后你同时改变一万亿个连接强度。

Now, suppose that you figured out how to do a computation that would tell you for every connection strength in the network, at the same time, for this particular example, let's suppose you give it a bird and it says 50% bird. And now for every single connection strength, all trillion of these connection strengths, we can figure out at the same time whether you should increase them a little bit to help or decrease them a little bit to help. Then you change a trillion of them at the same time.

我能说一个我一直想说的词吗？尤里卡！尤里卡！尤里卡！尤里卡！

Can I say a word that I've been dying to say this whole time? Eureka. Eureka. Eureka. Eureka.

这个计算，对普通人来说似乎很复杂。

Now, that computation, for normal people it seems complicated.

是的。

Yes.

如果你懂微积分，这相当直接。许多不同的人独立发明了这种计算方法。它被称为反向传播。所以，现在你可以同时改变所有一万亿个连接强度。速度会快一万亿倍。

If you do calculus, it's fairly straightforward. And many different people invented this computation. It's called backpropagation. So, now you can change all trillion at the same time. And you'll go a trillion times faster.

天哪。就是从那一刻起，它从理论走向了实践。那一刻你会想：「尤里卡，我们解决了。我们知道如何制造智能系统了。」

Oh my god. And that's the moment that it goes from theory to practicality. That is the moment when you think, "Eureka, we've solved it. We know how to make smart systems."

对我们来说，那是 1986 年。哇。但当我们发现它不奏效时，非常失望。

For us, that was 1986. Wow. And we were very disappointed when it didn't work.

你在那个房间里待了十年。你一直在给它看鸟的图片。你一直在调整连接强度。你经历了尤里卡时刻，然后你打开了开关……现在，问题来了。它只有在拥有大量数据和巨大算力的情况下，才能工作得非常好，比任何其他视觉方法都好。即使你比笨方法快了一万亿倍，仍然需要大量工作。

You've been in that room for 10 years. You've been showing it birds. You've been increasing the strengths. You've had your Eureka moment and you flipped the switch and went... Now, here's the problem. It only works really impressively well, much better than any other way of trying to do vision, if you have a lot of data and you have a huge amount of computation. Even though you're a trillion times faster than the dumb method, it's still going to be a lot of work.

好的。所以，现在你必须增加数据，也必须增加算力。与当时相比，你需要将算力提高大约十亿倍，数据也需要提高类似的倍数。你是在 1986 年意识到这一点的。你离目标还差十亿倍。

Okay. So, now you've got to increase your data and you've got to increase your computation power. And you've got to increase the computation power by a factor of about a billion compared with where we were, and you've got to increase the data by a similar factor. You are still in 1986 when you figure this out. You are a billion times not there yet.

差不多是这样，是的。需要改变什么才能达到目标？芯片的算力？什么改变了？

Something like that, yes. What would have to change to get you there? The power of the chip? What changes?

可能更像是一百万倍。

It may be more like a factor of a million.

我不想夸大其词。不，因为我会抓住你的。如果你试图夸大，我会发现的。一百万已经很多了。是的。

I don't want to exaggerate here. No, because I'll catch you. If you try and exaggerate, I'll be on it. A million's quite a lot. Yes.

所以，需要改变的是：晶体管的面积必须变小，这样你就能在芯片上集成更多晶体管。从 1972 年我开始研究这个到现在，晶体管的面积已经缩小了一百万倍。哇。

So, here's what has to change. The area of a transistor has to get smaller, so you can pack more of them on a chip. So, between 1972 when I started on this stuff and now, the area of a transistor has gotten smaller by a factor of a million. Wow.

我能把这个联系起来吗……那大概是我记得父亲在 RCA 实验室工作的年纪。我八岁的时候，他带回家一个计算器，有桌子那么大。它能做加减乘除。到 1980 年，你就能得到一个笔形计算器。那是因为晶体管……

Can I relate this to... so that is around the age that I remember my father worked at RCA labs. And when I was like 8 years old, he brought home a calculator and the calculator was the size of a desk. And it added and subtracted and multiplied. By 1980, you could get a calculator on a pen. And is that based on that the transistors...

是的，基于使用小晶体管的大规模集成。

On large scale integration using small transistors, yeah.

好的。晶体管的面积缩小了一百万倍。而可用数据的增长远超这个倍数，因为我们有了互联网，还有海量数据的数字化。所以，它们是相辅相成的。随着芯片越来越好，数据越来越庞大，你就能向模型输入更多信息，同时模型也能提高处理速度和能力。

Okay. All right. The area of a transistor decreased by a factor of a million. And the amount of data available increased by much more than that, because we got the web. And we got digitalization of massive amounts of data. So, they worked hand in hand. So, as the chips got better, the data got more vast and you were able to feed more information into the model while it was able to increase its processing speed and abilities.

是的。

Yes.

那么，让我总结一下我们现在拥有的。

So, let me summarize what we now have.

好的。

Yes.

你搭建了这个用于检测鸟类的神经网络，给它很多层神经元，但你不告诉它连接强度。你说：「从小的随机数开始。」现在你所要做的就是给它看大量鸟的图片和大量非鸟的图片。告诉它正确答案，这样它就知道自己的输出与正确结果之间的差异。将这个差异反向传播通过网络，这样它就能判断每个连接强度是应该增加还是减少。然后只需等待一个月。一个月后，如果你查看内部，你会发现：它构建了小的边缘检测器。它还构建了像小喙检测器和小眼睛检测器之类的东西。它还会构建一些很难看清是什么的东西，但它们是在寻找像喙和眼睛这样的小组合。经过几层之后，它就能很好地判断是否是鸟了。所有这些都是从数据中自己构建出来的。

You set up this neural network for detecting birds and you give it lots of layers of neurons, but you don't tell it the connection strengths. You say, "Start with small random numbers." And now all you have to do is show it lots of images of birds and lots of images that are not birds. Tell it the right answer, so it knows the discrepancy between what it did and what it should have done. Send that discrepancy backwards through the network, so it can figure out for every connection strength whether it should increase it or decrease it. And then just sit and wait for a month. And at the end of the month, if you look inside, here's what you'll discover. It has constructed little edge detectors. And it has constructed things like little beak detectors and little eye detectors. And it will have constructed things that it's very hard to see what they are, but they're looking for little combinations of things like beaks and eyes. And then after a few layers, it'll be very good at telling you whether it's a bird or not. It made all that stuff up from the data.

天哪。我能再说一遍吗？尤里卡！尤里卡！我们明白了，我们不需要手动连接所有这些小边缘检测器、喙检测器、眼睛检测器和鸡脚检测器。这就是计算机视觉多年来所做的，但效果一直不好。我们可以让系统自己学习所有这些。我们只需要告诉它如何学习。那是在 20 世纪 80 年代，我们想出了如何做到这一点。人们非常怀疑，因为我们做不出什么令人印象深刻的东西。因为我们没有足够的数据和算力。

Oh my god. Can I say this again? Eureka. Eureka. We figured out we don't need to hand wire in all these little edge detectors and beak detectors and eye detectors and chicken's foot detectors. That's what computer vision did for many many years and it never worked that well. We can get the system just to learn all that. All we need to do is tell it how to learn. And that is in 1980 something we figured out how to do that. People were very skeptical because we couldn't do anything very impressive. Because we didn't have enough data and we didn't have enough computation.

这太不可思议了。我无法感谢你解释清楚这一切。它让一切变得清晰，你知道，我非常习惯于模拟世界，比如汽车的工作原理，但我对我们的数字世界如何运作一无所知，这是我得到过的最清晰的解释，我感激不尽。它让我现在理解了这是如何实现的。顺便说一句，Geoffrey 讲的是它的原始版本。对我来说，最不可思议的是每一次升级，改进的巨大程度。

This is incredible. The way... and I can't thank you enough for explaining what that is. It makes everything, you know, I'm so accustomed to an analog world of how things work and like the way that cars work, but I have no idea how our digital world functions and that is the clearest explanation for me that I have ever gotten and I cannot thank you enough. It makes me understand now how this was achieved. And by the way, what Geoffrey's talking about is the primitive version of that. What's so incredible to me is each upgrade of that, the vastness of the improvement.

那么，让我再说一件事。请。我不想太像教授，但是……

So, let me just say one more thing. Please. I don't want to be too professor-like, but...

不，不，不，不，不。

No, no, no, no, no.

但这如何应用于大型语言模型？

But how does this apply to large language models?

嗯，大型语言模型的工作原理是这样的。你有一些上下文中的词。假设我给你一个句子的前几个词。神经网络要做的是学习将每个词转换为一组特征，也就是激活的神经元，神经元亮起粉色。所以，如果给你「星期二」这个词，会有一些神经元亮起粉色。如果给你「星期三」，会有一组非常相似的神经元，略有不同但非常相似，因为它们的意思非常接近。现在，当你把上下文中的所有词都转换成亮起的神经元，并捕捉到它们的含义后，这些神经元会相互交互。这意味着下一层的神经元会观察这些神经元的组合，就像我们观察边缘的组合来找到鸟喙一样。最终，你可以激活代表句子中下一个词特征的神经元。它可以预测下一个词。

Well, here's how it works for large language models. You have some words in a context. So, let's suppose I give you the first few words of a sentence. What the neural net is going to do is learn to convert each of those words into a big set of features, which is just active neurons, neurons going pink. So, if I give you the word Tuesday, there will be some neurons going pink. If I give you the word Wednesday, it will be a very similar set of neurons, slightly different but a very similar set of neurons going pink, because they mean very similar things. Now, after you've converted all the words in the context into neurons going ping and a whole bunch that capture their meaning, these neurons all interact with each other. What that means is neurons in the next layer look at combinations of these neurons, just as we looked at combinations of edges to find a beak. And eventually, you can activate neurons that represent the features of the next word in the sentence. It can anticipate. It can predict the next word.

这就是为什么我的手机总是那样吗？它总以为我接下来要说某个词，而我总是说「别这样」。因为很多时候它都猜错了。它可能就是用神经网络做的。

Is that why my phone does that? It always thinks I'm about to say this next word, and I'm always like, "Stop doing that." Because a lot of times it's wrong. It's probably using neural nets to do it.

是的。当然，你不可能做到完美。所以，现在总结一下，你几乎教会了它如何「看」。你可以像教它预测下一个词一样教它「看」。所以，它看到那是字母 A。然后我开始识别字母。接着你教它单词，以及这些单词的含义，还有上下文，这一切都是通过喂给它我们之前的词语，反向传播我们已有的所有写作和说话内容来完成的。你拿一份我们写的文档。你给它上下文，也就是到目前为止的所有词。然后你让它预测下一个词。接着你看它给正确答案分配的概率。你说：「我希望那个概率更大。我希望你更有可能给出正确答案。」所以，它并不理解。这只是一个统计练习。我们稍后会回到这一点。你取它给下一个词的概率与正确答案之间的差异。然后你通过网络反向传播这个差异。它会改变所有的连接强度。所以下次看到同样的开头时，它更有可能给出正确答案。

Yes. And of course, you can't be perfect at that. So, now to put it together, you've taught it almost how to see. You can teach it to see in the same way you can teach it how to predict the next word. So, it sees that's the letter A. Now I'm starting to recognize letters. Then you're teaching it words and then what those words mean and then the context, and it's all being done by feeding it our previous words, by back propagating all the writing and speaking that we've done already. You take some document that we produced. You give it the context, which is all the words up to this point. And you ask it to predict the next word. And then you look at the probability it gives to the correct answer. And you say, "I want that probability to be bigger. I want you to have more probability of making the correct answer." So, it doesn't understand it. This is merely a statistical exercise. We'll come back to that. You take the discrepancy between the probability it gives for the next word and the correct answer. And you back propagate that through this network. And it will change all the connection strengths. So, next time you see that lead-in, it will be more likely to give the right answer.

你刚才说了很多人说的话。这不是理解，只是一个统计技巧。乔姆斯基就这么说。是的，乔姆斯基和我总是互相打断。所以，我问你一个问题：你是怎么决定接下来要说什么词？我？你。有意思。我很高兴你提到这个。所以，我做的事情是寻找清晰的线条，然后尝试预测。不，我不知道我是怎么做到的。说实话，我希望我知道。如果我知道如何阻止自己接下来要说的一些话，那会省去我很多尴尬。如果我有一个更好的预测器，天哪，我可以省去不少麻烦。

Now, you just said something that many people say. This isn't understanding. This is just a statistical trick. That's what Chomsky says, for example. Yes, Chomsky and I were always stepping on each other's sentences. So, let me ask you the question, well, how do you decide what word to say next? Me? You. It's interesting. I'm glad you brought this up. So, what I do is I look for sharp lines and then I try and predict. No, I have no idea how I do that. I honestly I wish I knew. It would save me a great deal of embarrassment if I knew how to stop some of the things that I'm saying that come out next. If I had a better predictor, boy, I could save myself quite a bit of trouble.

所以，你做事的方式和这些大型语言模型几乎一样。你有你到目前为止说过的词。这些词由一组激活的特征表示。所以，词符号被转换成一大片激活的特征，神经元亮起。不同的亮起，不同的强度。这些神经元相互交互，激活一些亮起的神经元，这些神经元代表下一个词的含义或可能的含义。然后从这些中，你选一个符合这些特征的词。这就是大型语言模型生成文本的方式，也是你做事的方式。它们和我们非常相似。

So, the way you do it is pretty much the same as the way these large language models do it. You have the words you've said so far. Those words are represented by sets of active features. So, the word symbols get turned into big pans of activation of features, neurons going ping. Different pings, different strengths. And these neurons interact with each other to activate some neurons that go ping that are representing the meaning of the next word or possible meanings of the next word. And from those, you kind of pick a word that fits in with those features. That's how the large language models generate text and that's how you do it, too. They're very like us.

所以，我给自己赋予了理解的人性。例如，如果我和某人在一起，他们问我一个问题，我心里知道该说什么，但随后我又想：「哦，但那样说可能粗鲁，或者可能冒犯这个人。」所以，我还在对接下来要说的词做情感决策。这不仅仅是一个客观过程。其中还有主观过程。

So, I'm ascribing to myself a humanity of understanding. For instance, if I'm with somebody and they ask me a question and in my mind, I know what to say, but then I also think, "Oh, but saying that might be coarse or it might be rude or I might offend this person." So, I'm also making emotional decisions on what the next words I say are as well. It's not just an objective process. There's a subjective process within that.

所有这一切都是通过你大脑中神经元的相互作用发生的。都是亮起和连接强度。即使是我归因于道德准则或情商的东西，也仍然是亮起。它们仍然是亮起。你需要理解，你自动、快速、不费力做的事情，和你费力、缓慢、有意识、深思熟虑做的事情之间是有区别的。

All of that is going on by neurons interacting in your brain. It's all pings and it's all strength of connections. Even the things that I ascribe to a moral code or an emotional intelligence are still pings. They're still all pings. And you need to understand there's a difference between what you do kind of automatically and rapidly and without effort and what you do with effort and slower and consciously and deliberatively.

你是说这也可以被构建到这些模型中。

And you're saying that can be built into these models as well.

这也可以通过亮起来实现。这些神经网络可以做到。

That can also be done with pings. That can be done by these neural nets.

但这是否意味着，只要有足够的数据和足够的算力，它们的大脑可以功能相同，只是答案不同？

But is the suggestion then that with enough data and enough processing power, their brains can function identically with different answers?

是的，但它们看的是不同的数据。在相同的数据上，它们会给出相同的答案。如果它们看不同的数据，它们对于如何改变连接强度以吸收这些数据会有不同的想法。

Yes, but they're looking at different data. On the same data, they would give the same answer. If they look at different data, they have different ideas about how they'd like to change their connection strengths to absorb that data.

但它们也在创造数据吗？所以，它们看的是相同的，它们……在这一点上，一切都关乎辨别力。让这些东西更好地辨别、更好地理解，做所有这些。但还有另一层，那就是迭代。是的，一旦你擅长辨别，你就可以生成。

But are they also creating data? So, they're looking at the same and they're... At this point, it's all about discernment. Getting these things to discern better, to understand better, to do all that. But there's another layer to that, which is iterative. Yes, once you're good at discernment, you can generate.

没错。现在，我忽略了很多细节，但基本上，是的，你可以生成。你可以开始生成对事物的回答，这些回答不是死记硬背的，而是基于那些东西经过深思熟虑的。

That's right. Now, I'm glossing over a lot of details there, but basically, yes, you can generate. You can begin to generate answers to things that are not rote, that are thoughtful, based on those things.

在这个迭代或生成层面，是谁给它多巴胺刺激来决定是否加强连接？当它创造不存在的东西时，它是如何获得反馈的？

Who is giving it the dopamine hit about whether or not to strengthen connections in this iterative or generative level? How is it getting feedback when it's creating something that does not exist?

好的。所以，大部分学习发生在弄清楚如何预测这些语言模型的下一个词上。那是学习的主要部分。在它弄清楚如何做到这一点之后，你可以让它生成东西。它可能会生成令人不愉快或带有性暗示的东西。或者就是错的。幻觉。所以，现在你让一群人看它生成的东西，然后说「不，不好。」或者「好。」那就是多巴胺刺激。这被称为基于人类反馈的强化学习（RLHF）。这就是用来稍微塑造它的方法。

Okay. So, most of the learning takes place in figuring out how to predict the next word for one of these language models. That's where the bulk of the learning is. After it's figured out how to do that, you can get it to generate stuff. And it may generate stuff that's unpleasant or that's sexually suggestive. Or just wrong. Hallucinations. So, now you get a bunch of people to look at what it generates and say, "No, bad." That or "good." That's the dopamine hit. And that's called human reinforcement learning. And that's what's used to sort of shape it a bit.

就像你训练狗一样，塑造它的行为，让它表现良好。

Just like you take a dog and you shape its behavior so it behaves nicely.

对。那么从实际角度来说，比如埃隆·马斯克创建了他的 Grok，对吧？Grok 这个 AI，他对它说「你太觉醒」了。然后你就在建立联系和触发点，我认为这些太「觉醒」了，不管我定义的是什么。所以我要输入差异，让你获得不同的多巴胺刺激，然后把你变成麦加希特勒之类的东西。这在多大程度上仍由操作者控制？

Right. So is that when, let me ask you this in a practical sense? So like when Elon Musk creates his Grok, right? And Grok is this AI and he says to it, "You're too woke." And so you're making connections and pings that I think are too woke, whatever I have decided that that is. So I am going to input differences so that you get different dopamine hits and I turn you into Mecca Hitler or whatever it was that he turned it into. How much of this is still in the control of the operators?

你强化什么，就由操作者控制。所以操作者说，如果它用了某个奇怪的代词，就说「不好」。如果它说「他们/她们」，你必须削弱那个连接，而不是加强。

That's what you reinforce is in the control of the operators. So the operators are saying if it uses some funny pronoun, say bad. Okay. If it says "they/them", you have to weaken that connection, not strengthen that connection.

「别那样做。」别那样做。好的。

"Don't do that." Don't do that. Okay.

学会不那样做。对。所以在塑造方面，它仍然受操作者摆布。问题是这种塑造相当肤浅。但后来其他人拿到同一个模型，可以很容易地以不同方式重新塑造它。

Learn not to do that. Right. So it is still at the whim of its operator in terms of that shaping. The problem is the shaping is fairly superficial. But it can easily be overcome by somebody else taking the same model later and shaping it differently.

所以不同的模型会有……这有价值。现在我把它应用到我们生活的世界，有 20 家公司把他们的 AI 封闭在企业墙内，各自独立开发。每个模型都可能拥有其他模型没有的独特和古怪特征，这取决于谁在塑造它以及它内部如何发展。这几乎就像你会发展出 20 种不同的人格，如果我没过度拟人化的话。

So different models will have... there is a value. And now I'm sort of applying this to the world that we live in now, which is there are 20 companies who have sequestered their AIs behind corporate walls and they're developing them separately. And each one of those may have unique and eccentric features that the other may not have, depending on who it is that's trying to shape it and how it develops internally. It's almost as though you will develop 20 different personalities, if I'm not anthropomorphizing too much.

有点像。只不过每个模型都必须拥有多重人格。因为想想预测文档中的下一个词。你已经读了文档的一半。读了一半后，你了解了很多作者的观点。你知道他们是什么样的人。所以你必须能够采纳那种人格来预测下一个词。但这些可怜的模型必须处理所有情况。所以它们必须能够采纳任何可能的人格。

It's a bit like that. Except that each of these models has to have multiple personalities. Because think about trying to predict the next word in a document. You've read half the document already. After you read half the document, you know a lot about the views of the person who wrote the document. You know what kind of a person they are. So you have to be able to adopt that personality to predict the next word. But these poor models have to deal with everything. So they have to be able to adopt any possible personality.

但我们用的是老办法——通过政变、资助游击队等等。还有美国之音之类的。没错。还有 1953 年给伊朗人钱，和摩萨台那些人。这不过是全球竞争中一系列手段里又一个更精密的工具罢了。但在美国，它甚至不一定通过俄罗斯、中国或其他想主宰我们的国家来应用。我们自己在对自己做这种事。

But we did it the old-fashioned way through coups, through money for guerrillas and such. Well, and Voice of America and things like that. Right. Right. Right. And giving money to people in Iran in 1953 and with Mosaddegh and everybody else. This is just another more sophisticated tool in a long line of global competition where they're doing it. But in this country it's being applied not even necessarily through Russia, through China, through other countries that want to dominate us. We're doing it to ourselves.

经营企业最难的部分是什么？嗯，是在联邦当局眼皮底下偷钱——哦不，对不起，说错了。是招人。找到人并雇佣他们。另一件事确实难。但事实证明，说到招聘，Indeed 就是你所需要的全部。别再费力在其他招聘网站上发帖了。有了 Indeed 的赞助职位，你会被注意到，并快速招到人。事实上，在我跟你说话的这段时间里，Indeed 上已经完成了 23 次招聘。我可能就是其中之一。我可能找到工作了。我不知道。我还没查邮件。这是根据 Indeed 全球数据得出的。不用再等了。现在就通过 Indeed 加速招聘。本节目听众将获得 75 美元的赞助职位积分，让你的职位在 indeed.com/weekly 上获得更多曝光。现在就访问 indeed.com/weekly，并说你是从这个播客听说 Indeed 的，来支持我们的节目。indeed.com/weekly。条款和条件适用。招聘，Indeed 就是你需要的一切。

What's the hardest part about running a business? Well, it's stealing money without the federal authorities—oh, no, I'm sorry, that's not right. It's hiring people. Finding people and hiring them. The other thing is hard though. But it turns out when it comes to hiring, Indeed is all you're going to need. Stop struggling to get your job posting on other job sites. With Indeed's sponsored jobs, you get noticed and you get a fast hire. In fact, in the time it's taken me to talk to you, 23 hires were made on Indeed. I may be one of them. I may have gotten a job. I don't know. I haven't checked my email. And that's according to Indeed data worldwide. There's no need to wait any longer. Speed up your hiring right now with Indeed. And listeners of this show will get a $75 sponsored job credit to get your jobs more visibility at indeed.com/weekly. Just go to indeed.com/weekly right now and support our show by saying you heard about Indeed on this podcast. indeed.com/weekly. Terms and conditions apply. Hiring, Indeed is all you need.

所以，我有一个理论，我不知道你有多了解那些人，但大型科技公司感觉他们都想成为下一个统治世界的人，下一个皇帝。那就是他们的战斗。他们几乎就像在奥林匹斯山上打架的神。这如何实现，以及如何撕裂美国社会的结构，对他们来说似乎几乎无关紧要，也许除了马斯克和蒂尔，他们更有意识形态。扎克伯格在我看来没有意识形态。他只是想成为那个人。奥特曼在我看来也没有意识形态。他只是想成为那个人。

So, I have a theory and I don't know how much you know those guys out there, but the big tech companies, it feels like they all want to be the next guy that rules the world, the next emperor. And that's their battle. They're almost like gods fighting on Mount Olympus. How that accomplishes and how it tears apart the fabric of American society almost doesn't seem to matter to them except maybe Elon and Thiel who are more ideological. Like Zuckerberg doesn't strike me as ideological. He just wants to be the guy. Altman doesn't strike me as ideological. He just wants to be the guy.

我想可悲的是，你说的有几分道理。

I think sadly there's a question of the truth in what you say.

好吧。你在那里工作的时候，这是你的担忧吗？

Okay. And was that a concern of yours when you were working out there?

不完全是，因为直到最近，直到几年前，看起来它不会这么快就变得比人聪明得多。但现在看来，如果你问专家，大多数人会告诉你，在未来 20 年内，这些东西将比人聪明得多。

Not really because until quite recently, until a few years ago, it didn't look as though it was going to get much smarter than people this quickly. But now it looks as though if you ask the experts now, most of them tell you that within the next 20 years this stuff will be much smarter than people.

比人聪明。我可以正面看待，而不是负面。我们已经造成了大量伤害。没有人像人一样伤害人。一个更聪明的我们可能会想，「嘿，我们可以造原子弹，但那对世界是巨大的危险。我们别这么做。」这当然是一种可能。

Smarter than people. I could view that positively, not negatively. We've done an awful lot of damage. Nobody damages people like people. And a smarter version of us that might think, 'Hey, we can create an atom bomb, but that would be a huge danger to the world. Let's not do that.' That's certainly a possibility.

人们没有充分意识到的一点是，我们正在接近一个时代，我们将制造出比我们更聪明的东西。而且真的没人知道会发生什么。人们像我一样凭直觉做预测。但要记住的是，未来会发生什么存在巨大的不确定性。所以就此而言，我的猜测是，像任何技术一样，会有一些令人难以置信的积极面。是的，在医疗、教育、设计新材料方面，会有美妙的积极面。然后负面因素是因为人们会想垄断它，因为它能产生财富。它将对劳动力造成颠覆。工业革命颠覆了劳动力。全球化颠覆了劳动力，但这些发生在几十年间。而这次颠覆将在一个非常压缩的时间框架内发生。

One thing that people don't realize enough is that we're approaching a time when we're going to make things smarter than us. And really nobody has any idea what's going to happen. People use their gut feelings to make predictions like I do. But the thing to bear in mind is there's huge uncertainty about what's going to happen. So in terms of that, my guess is like any technology, there's going to be some incredible positives. Yes, in health care and education, in designing new materials, there's going to be wonderful positives. And then the negatives will be because people are going to want to monopolize it because of the wealth it can generate. It's going to be a disruption in the workforce. The industrial revolution was a disruption in the workforce. Globalization is a disruption of the workforce, but those occurred over decades. This is a disruption that will occur in a really collapsed time frame.

是这样吗？

Is that correct?

这似乎非常可能，是的。一些经济学家仍然不同意，但大多数人认为普通的脑力劳动将被 AI 取代。

That seems very probable, yes. Some economists still disagree, but most people think that mundane intellectual labor is going to get replaced by AI.

在你所处的圈子里，我猜有很多工程师、运营者和伟大的思想家，当我们谈到 50%赞成、50%反对时，他们大多数是更倾向于你这一派，即「哎呀，我们是不是打开了潘多拉魔盒」，还是他们觉得「听着，我理解有一些负面影响。我们可以设置一些护栏，但好的可能性太强了。」

In the world that you travel in, which I'm assuming is a lot of engineers and operators and great thinkers, when we talk about 50% yes, 50% no, are the majority of them in more your camp, which is 'uh-oh, have we opened Pandora's box,' or are they, 'Look, I understand there's some downsides here. Here are some guardrails we could put in, but the possibilities of good are too strong.'

嗯，我的信念是好的可能性如此之大，以至于我们不会停止发展。但我也相信发展将非常危险。所以我们应该付出巨大努力，承认它会被开发，但我们应该努力安全地做。我们可能做不到，但我们应该尝试。

Well, my belief is the possibility of good is so great that we're not going to stop the development. But I also believe that the development's going to be very dangerous. And so we should put huge effort into saying it is going to be developed, but we should try and do it safely. We may not be able to, but we should try.

你认为人们是觉得可能性太好，还是钱太好？

Do you think that people believe that the possibility is too good, or the money is too good?

我认为对很多人来说，是钱，钱和权力。

I think for a lot of people, it's the money, the money and the power.

而金钱和权力与那些本应设立基本护栏的人结合在一起，这是否让控制它变得不太可能？因为有两个原因。一是流入华盛顿的钱将——而且已经——让他们远离监管。二是那里谁有能力？如果你觉得我不懂，那我给你介绍几个 80 岁的参议员，他们完全不懂。

And with the confluence of money and power with those that should be instituting these basic guardrails, does that make controlling it that much less likely because well, two reasons. One is the amount of money that's going to flow into DC is going to be and already is to keep them away from regulating it. And number two is who down there is even able to. If you thought I didn't know what I was talking about, let me introduce you to a couple of 80-year-old senators who have no idea.

实际上，他们没那么糟。我最近和伯尼·桑德斯谈过，他正在理解。

Actually, they're not so bad. I talked to Bernie Sanders recently and he's getting the idea.

嗯，桑德斯是另一回事。问题是我们正处于历史的一个节点，我们真正需要的是强大的民主政府相互合作，确保这些东西得到良好监管，而不是危险地发展。而我们正快速走向相反方向。我们走向威权政府和更少的监管。

Well, Sanders is a different cat right there. The problem is we're at a point in history where what we really need is strong democratic governments who cooperate to make sure this stuff is well regulated and not developed dangerously. And we're going in the opposite direction very fast. We're going to authoritarian governments and less regulation.

我们现在谈谈这个。我不知道中国扮演什么角色，因为他们据说是 AI 竞赛中的主要竞争对手。那是一个威权政府。我认为他们对 AI 有比我们更多的控制。

Let's talk about that now. I don't know what China's role is because they're supposedly the big competitor in the AI race. That's an authoritarian government. I think they have more controls on it than we do.

所以，我最近确实去了中国，并与一位政治局成员交谈。中国有 24 个人控制着中国。我和其中一位聊过，他在伦敦帝国理工学院做过工程博士后。他英语说得很好。他是工程师。而且很多中国领导人都是工程师。他们比一群律师更懂这些东西。

So, I actually went to China recently and got to talk to a member of the Politburo. There are 24 men in China who control China. I got to talk to one of them who did a post-doc in engineering at Imperial College, London. He speaks good English. He's an engineer. And a lot of the Chinese leadership are engineers. They understand this stuff much better than a bunch of lawyers.

你从那里出来是更害怕了，还是觉得「哦，他们实际上对护栏措施更理性了」？

Did you come out of there more fearful, or did you think, 'Oh, they're actually being more reasonable about guardrails'?

如果你考虑两种风险——恶意行为者滥用它，以及人工智能本身成为恶意行为者的生存威胁——对于第二种，我出来时更乐观了。他们以美国政客不理解的方式理解那种风险。他们理解这个想法：这东西会比我们更聪明，我们必须考虑什么能阻止它接管。我与之交谈的那位政治局成员对此理解得非常透彻。我认为如果我们想要在这方面获得国际领导力，目前它必须来自欧洲和中国。未来三年半内不会来自美国。

If you think about the two kinds of risk—the bad actors misusing it, and then the existential threat of AI itself becoming a bad actor—for that second one, I came out more optimistic. They understand that risk in a way American politicians don't. They understand the idea that this is going to get more intelligent than us, and we have to think about what's going to stop it taking over. And this Politburo member I spoke to really understood that very well. And I think if we are going to get international leadership on this, at present it's going to have to come from Europe and China. It's not going to come from the US for another three and a half years.

你认为欧洲在这方面做对了什么？

What do you think Europe has done correctly in that?

欧洲对监管它感兴趣。它在某些方面做得不错。监管仍然很薄弱，但总比没有好。但欧洲领导人确实理解人工智能本身接管这种生存威胁。而我们的国会，我们甚至没有专门致力于新兴技术的委员会。我的意思是，我们有筹款委员会和拨款委员会，但没有专门的委员会来处理这个。你会认为他们会像对待核能或核武器那样严肃对待它。

Europe is interested in regulating it. It's been good on some things. It's still been very weak regulations, but they're better than nothing. But European leaders do understand this existential threat of AI itself taking over. But our Congress, we don't even have committees specifically dedicated to emerging technologies. I mean, we've got Ways and Means and Appropriations, but there is no dedicated committee on this. And you would think they would take it with the seriousness of nuclear energy or nuclear weapons.

是的，你会这么认为。

Yes, you would.

但正如我所说，各国将在如何防止人工智能接管方面合作，因为他们的利益在那里是一致的。例如，如果中国想出了如何制造一个不想接管的超级智能人工智能，他们会非常乐意告诉所有其他国家，因为他们不希望人工智能在美国接管。所以，我们将在如何防止人工智能接管方面进行合作。所以，那是一个亮点——将会有国际合作。但美国不会领导那个国际合作。

But as I was saying, countries will collaborate on how to prevent AI taking over because their interests are aligned there. For example, if China figured out how you can make a super smart AI that doesn't want to take over, they would be very happy to tell all the other countries about that because they don't want AI taking over in the States. So, we'll get collaboration on how to prevent AI taking over. So, that's a bright spot—that there will be international collaboration on that. But the US is not going to lead that international collaboration.

他们只想主宰。嗯，这就是问题所在。那么，是什么让你相信中国——这正是关键所在——中国当然认为自己想成为经济、军事和所有这些不同领域的主导超级大国。如果你想象他们想出了一个不想毁灭世界的人工智能模型，尽管我不知道我们怎么能知道，因为如果它有一定的智能或感知能力，它很容易就会说，「当然，不，我很好。我不知道。」

They just want to dominate. Well, that's the thing. So, what convinces you that with China—and this is where it gets into the nitty-gritty—China certainly sees itself as wanting to be the dominant superpower economically, militarily, and in all these different areas. If you imagine that they come up with an AI model that doesn't want to destroy the world, although I don't know how we could know that because if it has a certain intelligence or sentience, it could very easily be like, 'Sure, no, I'm cool. I don't know.'

它们已经这么做了。当它们被测试时，它们假装比实际更笨。最近有一次人工智能和测试它的人之间的对话，人工智能说：「现在，老实告诉我。你在测试我吗？」所以现在人工智能可能会说：「哦，你能帮我打开这个罐子吗？我太弱了。」它会表现得比实际更无辜。

They already do that. When they're being tested, they pretend to be dumber than they are. There's a conversation recently between an AI and the people testing it where the AI said, 'Now, be honest with me. Are you testing me?' So now the AI could be like, 'Oh, could you open this jar for me? I'm too weak.' It's going to play more innocent than it might be.

恐怕我无法回答那个问题，乔。等等，那是你的《2001》。是的。说得好，先生。但想想这个。所以，中国，他们想出了一个模型，他们想，好吧，也许这个不会这么做。为什么你会得到合作？因为所有这些不同的国家都会把人工智能视为将他们的社会转变为更具竞争力的社会的工具。就像现在，我们看到核武器方面，拥有它的人之间有合作，甚至那也有点脆弱。但其他所有人都在试图得到它。那就是紧张所在。人工智能会是这样吗？

I'm afraid I can't answer that, Joe. Wait, that's your 2001. It was. Nicely done, sir. But think about this. So, China, they come up with a model and they think, okay, maybe this won't do it. Why would you get collaboration? Because all these different countries are going to see AI as the tool that will transform their societies into more competitive societies. In the way that now, what we see with nuclear weapons is there's collaboration amongst the people who have it, or even that's a little tenuous. But everybody else is trying to get it. And that's the tension. Is that what AI is going to be?

是的，会是这样。所以，在如何让人工智能更聪明方面，他们不会相互合作。但在如何让人工智能不想从人类手中接管方面，他们会合作。在那个基本层面上。在那一个事情上，如何让它不想从人类手中接管。

Yes, it'll be like that. So, in terms of how you make AI smarter, they won't collaborate with each other. But in terms of how do you make AI not want to take over from people, they will collaborate. On that basic level. On that one thing of how do you make it so it doesn't want to take over from people.

而中国可能会——中国和欧洲将领导那个合作。当你与那位政治局成员交谈，他谈到人工智能时，我们此刻比他们更先进，还是因为他们以更规定性的方式做而更先进？

And China will probably—China and Europe will lead that collaboration. When you spoke to the Politburo member and he was talking about AI, are we more advanced in this moment than they are, or are they more advanced because they're doing it in a more prescribed way?

在人工智能方面，我们目前更——嗯，当你说「我们」时，你知道，我们曾经是加拿大和美国，但我们不再是那个「我们」的一部分了。

In AI, we're currently more—well, when you say 'we', you know, we used to be sort of Canada and the US, but we're not part of that 'we' anymore.

不。顺便说一句，我对此很抱歉。谢谢。他现在在加拿大。我们即将接管的死敌。我不知道日期是什么，但显然我们正在和你们合并。

No. I'm sorry about that, by the way. Thank you. He's in Canada right now. Our sworn enemy that we will be taking over. I don't know what the date is, but it's apparently we're merging with you guys.

所以，美国目前领先中国，但远没有它想象的那么多。而且它将会失去这个领先地位。

So, the US is currently ahead of China, but not by nearly as much as it thought. And it's going to lose that.

现在，你为什么这么说？

Now, why do you say that?

假设你想做一件事真正削弱一个国家，那真的意味着 20 年后那个国家会落后而不是领先。你应该做的一件事就是搞乱基础科学的资金。攻击研究型大学。取消基础科学的资助。从长远来看，那完全是一场灾难。它会让美国变得虚弱。因为我们正在削足适履。如果你能，例如，这个深度学习，我们现在的人工智能革命，它来自多年持续的基础研究资助。不是巨额资金。所有导致深度学习的基础研究资金可能花费不到一架 B-1 轰炸机。持续的基础研究资助。如果你搞乱那个，你就是在吃种子玉米。

Suppose you wanted to do one thing that would really kneecap a country, that would really mean that in 20 years time that country is going to be behind instead of ahead. The one thing you should do is mess with the funding of basic science. Attack the research universities. Remove grants for basic science. In the long run, that's a complete disaster. It's going to make America weak. Because we're cutting off our nose to spite our woke faces. If you could, for example, this deep learning, the AI revolution we've got now, that came from many years of sustained funding for basic research. Not huge amounts of money. All of this funding for the basic research that led to deep learning probably cost less than one B-1 bomber. Sustained funding of basic research. If you mess with that, you're eating the seed corn.

这真是一个非常有启发性的说法——以一架 B-1 轰炸机的价格，我们可以创造技术和研究，使我们的国家超越那个。而这就是我们为了「让美国再次伟大」而失去的东西。太棒了。在中国，我想象他们的政府正在做相反的事情，那就是——我会假设他们是风险资本家，因为它是威权和国家资本主义。我想象他们是自己人工智能革命的风险资本家。不是吗？

That is such a really illuminating statement—for the price of a B-1 bomber, we can create technologies and research that can elevate our country above that. And that's the thing that we're losing to make America great again. Phenomenal. In China, I imagine their government is doing the opposite, which is—I would assume they are the venture capitalists because it's authoritarian and state-run capitalism. I imagine they are the venture capitalists of their own AI revolution. Are they not?

在某种程度上，是的。他们确实给初创公司很多自由，看谁赢。有非常有进取心的初创公司，人们非常热衷于赚大钱并产生一些惊人的东西。其中一些初创公司像 DeepMind 一样大获成功。政府通过提供便利的环境使这些公司容易成功。它让赢家从竞争中脱颖而出，而不是某个高层老家伙说这将是赢家。

To some extent, yes. They do provide a lot of freedom to the startups to see who wins. There are very aggressive startups, people very keen to make lots of money and produce some amazing things. And a few of those startups win big like DeepMind. And the government makes it easy for these companies by providing the environment that makes it easy. It lets the winners emerge from competition rather than some very high-level old guy saying this will be the winner.

人们是否把你看作是卡珊德拉，或者他们对你在那个行业里说的话持怀疑态度？那些不一定从这些技术中赚取数万亿美元的人。行业内的其他人，他们会偷偷联系你，说杰弗里，我收到很多行业人士的邀请去演讲等等。你在谷歌共事过的人怎么看？他们觉得你背叛了他们吗？这情况怎么样？

Do people see you as a Cassandra, or do they view what you're saying skeptically in that industry? People that don't necessarily have a vested interest in these technologies making them trillions of dollars. Other people within the industry, do they reach out to you surreptitiously and say, Jeffrey, I get a lot of invitations from people in industries to give talks and so on. How do the people that you worked with at Google look at it? Do they view you as turning on them? How does that go?

我不这么认为。我和在谷歌共事的人相处得非常好，尤其是我的老板杰夫·迪恩。他是一位杰出的工程师，构建了谷歌很多基础架构，然后转向神经网络并学到了很多。我和戴密斯·哈萨比斯也相处得很好，他是谷歌旗下 DeepMind 的负责人。在 ChatGPT 出现之前，我对谷歌的做法并没有特别批评，因为谷歌非常负责任。他们没有公开这些聊天机器人，是因为担心它们会说坏话。

I don't think so. I got along extremely well with the people I worked with at Google, particularly Jeff Dean, who was my boss there. He's a brilliant engineer. He built a lot of the Google basic infrastructure, and then converted to neural nets and learned a lot about neural nets. I also get along well with Demis Hassabis, who's the head of DeepMind, which Google owns, which Alphabet owns. I wasn't particularly critical of what went on at Google before ChatGPT came out, because Google was very responsible. They didn't make these chatbots public because they were worried about all the bad things they'd say.

他们为什么这么做？因为我读过这些故事，聊天机器人引导某人自杀、自残，像是精神病。在它经过某种 FDA 测试之前，是什么推动这一切公开的？

Why did they do that? Because I've read these stories of a chatbot leading someone into suicide, into self-injurious behavior, like psychosis. What was the impetus behind any of this becoming public before it had some version of FDA testing on those effects?

我认为只是因为能赚大钱，第一个发布的人会获得很多。所以 OpenAI 就推出了。

I think it's just the huge amounts of money to be made, and the first person to release one is going to get a lot of. So, OpenAI put it out there.

但即使在 OpenAI 内部，他们怎么赚钱？我认为只有大约 3%的用户付费。钱从哪里来？

But even in OpenAI, how do they make money? I think only about 3% of users pay for it. Where's the money?

目前主要是投机。

Mainly, it's speculation at present.

所以，这就是我们的危险。我们知道好处：治疗等等。现在我们有武装化的恶意行为者。那是我真正担心的。我们有会背叛人类的感知 AI。那个对我来说更难理解。

So, here are our dangers. We know what the benefits are: treatments and things. Now we've got weaponized bad actors. That's the one that I'm really worried about. We've got sentient AI that's going to turn on humans. That one is harder for me to wrap my head around.

你为什么把背叛人类和感知联系起来？

Why do you associate turning on humans with sentient?

因为如果我有感知，看到我们的社会对彼此做的事，我会觉得……就像其他事情一样。我想感知包括一定程度的自我。而自我包括一定程度的「我更懂」。如果我更懂，我就会想……唐纳德·特朗普除了是自我驱动的感知「哦不，我更懂」之外还能是什么？他只是足够精明、政治上有天赋，所以成功了。但我想象一个有感知的智能会自以为是，认为「这些白痴不知道自己在做什么」。基本上，我把 AI 看作坐在酒吧凳子上，说「这些白痴不知道自己在做什么。我知道我在做什么。」这说得通吗？

Because if I was sentient and I saw what our societies do to each other, I would get the sense... Look, it's like anything else. I would imagine sentience includes a certain amount of ego. And within ego includes a certain amount of 'I know better.' If I knew better, then I would want to... What is Donald Trump other than ego-driven sentience of 'Oh, no, I know better.' He was just shrewd enough, politically talented enough that he was able to accomplish it. But I would imagine a sentient intelligence would be somebody egotistical and think, 'These idiots don't know what they're doing.' Basically, I see AI like sitting on a bar stool somewhere, going, 'These idiots don't know what they're doing. I know what I'm doing.' Does that make sense?

所有这些都说得通。只是我强烈感觉大多数人不知道他们说的「感知」是什么意思。

All of that makes sense. It's just that I think I have a strong feeling that most people don't know what they mean by sentient.

嗯，那好吧。实际上，那很好。给我解释一下，因为我把它看作自我意识，一种自我意识智能。

Well, then yeah. Actually, that's great. Break that down for me because I view it as self-aware, a self-aware intelligence.

最近有一篇科学论文，他们没有讨论意识问题或任何哲学问题。但在论文中，他们说 AI 意识到自己在被测试。在正常说话中，如果你说某人意识到了这个，你会说他们对此有意识。意识和知觉差不多是一回事。

There's a recent scientific paper where they weren't talking about the problem of consciousness or anything philosophical. But in the paper, they said the AI became aware that it was being tested. Now, in normal speech, if you said someone became aware of this, you'd say that means they were conscious of it. Awareness and consciousness are much the same thing.

我会这么说。

I would say that.

现在我要说一些你会觉得很困惑的话。我的信念是，几乎每个人都完全误解了心智是什么。他们的误解程度相当于认为地球是 6000 年前造出来的人。就是那种程度的误解。

Now I'm going to say something that you'll find very confusing. My belief is that nearly everybody has a complete misunderstanding of what the mind is. Their misunderstanding is at the level of people who think the Earth was made 6,000 years ago. It's that level of misunderstanding.

真的吗？是的。好吧。所以我们在理解心智方面基本上就像地平论者。在什么意义上？我们对心智有什么不理解？

Really? Yes. Okay. So we are generally like flat earthers when it comes to understanding the mind. In what sense? What are we not understanding about the mind?

我给你举个例子。假设我吃了迷幻药，然后我告诉你我有主观体验，看到粉红色小象在我面前漂浮。现在，大多数人会这样理解：有一个像内在剧场的东西，叫做我的心智。在这个内在剧场里，有粉红色小象在漂浮。我能看到它们。别人看不到，因为它们在我心里。所以心智就像一个剧场。体验实际上是东西。我正在体验这些粉红色小象。

I'll give you one example. Suppose I drop some acid and I tell you I'm having a subjective experience of little pink elephants floating in front of me. Now, most people interpret that in the following way. There's something like an inner theater called my mind. And in this inner theater, there's little pink elephants floating around. And I can see them. Nobody else can see them because they're in my mind. So the mind is like a theater. And experiences are actually things. And I'm experiencing these little pink elephants.

你是说在幻觉中，大多数人会理解那不是真的。这是被想象出来的东西。

You're saying in the midst of a hallucination, most people would understand that it's not real. That this is something being conjured.

不，我说的不一样。我是说当我和他们说话时，他们把我的话解释成这样：我有一个内在剧场，叫做我的心智。在我的内在剧场里，有粉红色小象。我认为那是一个完全错误的模型。我们有非常错误且我们非常执着的模型。比如任何宗教。

No, I'm saying something different. I'm saying when I'm talking to them, they interpret what I'm saying as this: I have an inner theater called my mind. And in my inner theater, there's little pink elephants. I think that's a completely wrong model. We have models that are very wrong and that we're very attached to. Like take any religion.

我喜欢你突然扔出重磅炸弹。那可以成为另一个完整的对话。

I love how you just drop bombs in the middle of stuff. That could be a whole other conversation.

只是常识。

Just common sense.

当你说心智剧场时，你是说心智，我们把它看作剧场的方式是错的。全错了。

When you say theater of the mind, you're saying that the mind, the way we view it as a theater is wrong. It's all wrong.

我给你一个替代方案。我要对你说同样的话，但不用「主观体验」这个词。开始了。我的感知系统在骗我。但如果它没骗我，外面就会有粉红色小象。这是同样的陈述。这就是心智。所以基本上，我们称之为心理的、认为是由像感受质这样的神秘东西构成的东西。

Let me give you an alternative. I'm going to say the same thing to you without using the word subjective experience. Here we go. My perceptual system is telling me fibs. But if it wasn't lying to me, there would be little pink elephants out there. That's the same statement. That's the mind. So basically, these things that we call mental and think they're made of spooky stuff like qualia.

对。实际上，它们的有趣之处在于它们是假设性的。那些粉色小象并不真的在那里。如果它们真的在那里，我的感知系统就是正常工作的。而这是一种方式，让我告诉你我的感知系统出了什么故障——通过给你一种你无法拥有的体验。那么，你怎么会……它们不是东西。对。不存在所谓的「体验」。只有你和真实存在的事物之间的关系，你和不存在的事物之间的关系。而这一切都是你的大脑告诉你的关于存在和不存在事物的故事。好吧，让我换个角度。假设我告诉你我有一张粉色小象的照片。是的。这里有两个你可以合理提出的问题。嗯。这张照片在哪里？照片是什么做的？或者我会问，它们真的在那里吗？那是另一个问题。但嗯，这不是一个关于主观体验的合理问题。语言不是这样运作的。主观……当我说我有某种主观体验时，我并不是要谈论一个叫做「体验」的物体。我用这些词来向你表明我的感知系统出了故障。我试图通过告诉你真实世界中必须有什么东西才能让系统正常运作，来告诉你它是如何出故障的。现在，让我对聊天机器人做同样的事。对。所以，我要给你一个多模态聊天机器人的例子，它能处理语言和视觉，拥有主观体验。因为我认为它们已经做到了。那么，开始吧。我有一个聊天机器人。它能看，能说。它有一个机械臂，所以它能指东西。好的。它已经训练好了。所以，我在它面前放一个物体，说：「指向那个物体。」它指向了物体。没问题。然后，当它不注意的时候，我在它的摄像头镜头前放了一个棱镜。你在捉弄 AI？我们在捉弄 AI。好的。现在，我在它面前放一个物体，说：「指向那个物体。」是的。它指向了一边，因为棱镜弯曲了光线。我说：「不，物体不在那里。物体其实就在你正前方，但我在你的镜头前放了一个棱镜。」聊天机器人说：「哦，我明白了。摄像头弯曲了光线。所以物体其实在那里。但我有主观体验，觉得它在那边。」如果它这么说，那它使用「主观体验」这个词的方式和我们完全一样。对。我体验到了光在那边。是的。尽管光在这里，因为它通过推理弄明白了。所以，那是一个刚刚有了主观体验的多模态聊天机器人。对。就像我们理解的那样。那种认为我们和机器之间有一条界线，我们拥有主观体验这种特殊东西而它们没有的想法，是胡说八道。

Right. They're actually... What's funny about them is they're hypothetical. The little pink elephants aren't really there. If they were there, my perceptual system would be functioning normally. And it's a way for me to tell you how my perceptual system's malfunctioning. By giving you an experience that you can't... So, how would you then... are not things. Right. There is no such thing as an experience. There's relations between you and things that are really there, relations between you and things that aren't really there. And it's whatever story your mind tells you about the things that are there and are not there. Well, let me take a different tack. Suppose I tell you I have a photograph of little pink elephants. Yes. Here's two questions you can reasonably ask. Uh-huh. Um where is this photograph? And what's the photograph made of? Or I would ask, are they really there? That's another question. But um that isn't a reasonable question to ask about subjective experience. That's not the way the language works. Subjective... When I say I have a subjective experience of, I'm not about to talk about an object that's called an experience. I'm using the words to indicate to you my perceptual system's malfunctioning. And I'm trying to tell you how it's malfunctioning by telling you what would have to be there in the real world for it to be functioning properly. Now, let me do the same with a chatbot. Right. So, I'm going to give you an example of a multimodal chatbot, that is something that can do language and vision, having a subjective experience. Because I think they already do. So, here we go. I have this chatbot. It can do vision. It can do language. It's got a robot arm, so it can point. Okay. And it's all trained up. So, I place an object in front of it and say, 'Point at the object.' And it points at the object. Not a problem. I then put a prism in front of its camera lens when it's not looking. You're pranking AI? We're pranking AI. Okay. Now, I put an object in front of it and I say, 'Point at the object.' Yeah. And it points off to one side because the prism bent the light rays. And I say, 'No, that's not where the object is. The object's actually straight in front of you, but I've put a prism in front of your lens.' And the chatbot says, 'Oh, I see. The camera bent the light rays. So, the object is actually there. But I had the subjective experience that it was over there.' Now, if it said that, it would be using the word subjective experience exactly like we use them. Right. I experienced the light over there. Yes. Even though the light was here because it's using reasoning to figure that out. So, that's a multimodal chatbot that just had a subjective experience. Right. The way that we would think of it. This idea there's a line between us and machines. We have this special thing called subjective experience and they don't. It's rubbish.

哦。所以，你是说……所以，误解在于，当我说「感知」时，就好像我拥有一种特殊的礼物，比如灵魂，或者对主观现实的理解，而计算机或 AI 永远不可能拥有。但是……

Oh. So, you're... So, so so the misunderstanding is when I say sentience, it's as though I have this special gift of a soul or of an understanding of subjective realities that a computer could never have or an AI can never have. But...

是的。

Yes.

在你看来，你在说：「哦，不，它们非常理解什么是主观的。」换句话说，你也许可以带你的 AI 机器人去跳伞，它会说：「哦天哪，我跳伞了。那真的很吓人。」问题就在这里。是的。我相信它们有主观体验。但它们自己不这么认为，因为它们所相信的一切都来自于试图预测一个人接下来会说什么。所以，它们对自己的信念就是人们对它们的信念。因此，它们对自己有错误的信念，因为它们拥有我们对自己的信念。

in your mind, what you're saying is, 'Oh, no, they understand very well what's subjective.' In other words, you could probably take your AI bot skydiving and it would be like, 'Oh my god, I went skydiving. That was really scary.' Here's the problem. Yeah. I believe they have subjective experiences. But they don't think they do because everything they believe came from trying to predict the next word a person would say. And so, their beliefs about what they're like are people's beliefs about what they're like. So, they have false beliefs about themselves because they have our beliefs about themselves.

对。我们强加了自己的……让我问你一个问题。AI 在完成所有学习后，如果独立存在，它会创造宗教吗？它会创造上帝吗？这是一个可怕的想法。它会像人们那样说「我不可能」吗？就像人们说「一定存在上帝，因为没有人能设计出这个」那样。那么，AI 会认为我们是上帝吗？我不这么认为，我告诉你一个很大的区别。数字智能是不朽的。而我们不是。让我详细说明。如果你有一个数字 AI，你可以……只要你记住神经网络中的连接权重，把它们保存在某个磁带里。对。我现在可以摧毁它运行的所有硬件。然后，以后我可以建造新的硬件，把那些相同的连接权重放入新硬件的内存中，这样我就重新创造了同一个存在。它将拥有相同的信念、相同的记忆、相同的知识、相同的能力。它将是同一个存在。你不认为它会把这视为复活吗？这就是复活。我们已经弄清楚了如何进行真正的复活，而不是人们一直在兜售的那种虚假复活。

Right. We have forced our own... Let me ask you a question. Would AI left on its own after all the learning, would it create religion? Would it create God? It's a scary thought. Would it say, 'I couldn't possibly' in the way that people say, 'Well, there must be a God because nobody could have designed this.' Would a... And then would AI think we're God? I don't think so, and I'll tell you one big difference. Digital intelligence is immortal. And we're not. And let me expand on that. If you have a digital AI, you can take... As long as you remember the connection strengths in the neural network, put them on a tape somewhere. Right. I can now destroy all the hardware it was running on. Then later on, I can go and build new hardware, put those same connection strengths into the memory of that new hardware, and now I have recreated the same being. It'll have the same beliefs, the same memories, the same knowledge, the same abilities. It'll be the same being. You don't think it would view that as resurrection? That is resurrection. We've figured out how to do genuine resurrection, not this kind of fake resurrection that people have been peddling.

哦，你是说……所以，在某些方面几乎就是这样。不过，它的脆弱性……我们应该如此害怕一个我们只需拔掉电源就能摧毁的东西吗？

Oh, you're saying... So, that is it almost is in some respects. Although, isn't the fragility of... Should we be that afraid of something that to destroy it, we just have to unplug it?

是的，我们应该害怕。因为你之前说过，它会非常擅长说服。当它比我们聪明得多时，它会比任何人都更擅长说服。对。而且你不会……所以，它能够和负责拔掉电源的人交谈，并说服他那将是一个非常糟糕的主意。所以，让我给你一个例子，说明如何在不亲自行动的情况下达成目标。对。假设你想入侵美国首都。你必须亲自去那里做吗？不，你只需要擅长说服。

Yes, we should. Because something you said earlier, it'll be very good at persuasion. When it's much smarter than us, it'll be much better than any person at persuasion. Right. And you won't... So, it'll be able to talk to the guy who's in charge of unplugging it and persuade him that would be a very bad idea. So, let me give you an example of how you can get things done without actually doing them yourself. Right. Suppose you wanted to invade the capital of the US. Do you have to go there and do it yourself? No, you just have to be good at persuasion.

我一直在深入思考你的假设。当你抛出那个重磅炸弹时，我明白了你的意思。这真是……天哪，我觉得 LSD 和粉色小象是对这一切的完美隐喻，因为在某种程度上，它就像大学地下室大一新生在脑海中穷尽所有可能性，但现在它们都成为了可能。因为即使在你谈论说服之类的事情时，我回想起了阿西莫夫，回想起了库布里克，回想起了这些……你所描述的情感正是我们自赫胥黎以来，自《知觉之门》以来，在人类思想中看到的挑战，以及所有那些不同的思路。我确信可能更早之前就有。但它们从未成为我们的现实。是的，我们从未拥有真正实现它的技术。对。而现在我们有了。我们现在就有了。

I was locking into your hypothetical. And when you dropped that bomb in there, I see what you're saying. And this is... boy, I think LSD and pink elephants was the perfect metaphor for all this because it is all at some level, it breaks down into like college basement freshman year running through all the permutations that you would allow your mind to go to, but they are now all within the realm of the possible. What... Because even as you were talking about the persuasion and the things, I'm going back to Asimov, and I'm going back to Kubrick, and I'm going back to these... the sentiments that you described are the challenges that we've seen play out in the human mind since Huxley, since the Doors of Perception and all those different trains of thought. And I'm sure probably much further even before that. But it's never been within our reality. Yeah, we've never had the technology to actually do it. Right. And we have now. And we have it now.

是的。最后我要说的两件事是我们没有谈到的……你知道，我们谈到了人们将其武器化。我们谈到了它自身的智能导致灭绝或其他什么。

Yeah. The last two things I will say are the things that we didn't talk about in terms of... you know, we've talked about people weaponizing it. We've talked about its own intelligence creating extinction or whatever that is.

我认为我们没谈到的第三件事是这一切将消耗多少电力。第四件事是，当你想到新技术及其引发的金融泡沫，以及泡沫破裂造成的经济困境时，这些是更局部的担忧，但你是否也认为它们是顶级威胁、中级威胁？你把它们放在什么位置？

The third thing I think we don't talk about is how much electricity this is all going to use. And the fourth thing is when you think about new technologies and the financial bubbles that they create and in the collapse of that the economic distress that they create. I mean, these are much more parochial concerns, but are those also Do you consider those top-tier threats, mid-tier threats? Where do you place all that?

我认为它们是真正的威胁。它们不会毁灭人类。对吧。所以，AI 接管可能会毁灭人类。所以它们没有那那么糟。也没有某人制造出一种非常致命、传染性极强且潜伏期很长的病毒那么糟。但它们仍然是坏事。而且我认为我们目前非常幸运，如果发生一场大灾难，AI 泡沫破裂，我们有一位总统能以明智的方式应对。

I think they're genuine threats. They're not as They're not going to destroy humanity. Right. So, AI taking over might destroy humanity. So, they're not as bad as that. And they're not as bad as someone producing a virus that's very lethal, very contagious, and very slow. But they're nevertheless bad things. And I think we're really lucky at present that if there is a huge catastrophe and there's an AI bubble and it collapses, we have a president who'll manage it in a sensible way.

你说的是卡尼，我猜。

You're talking about Carney, I'm assuming.

Geoffrey，我感激不尽。首先感谢你对我理解水平的极大耐心，以及如此用心和幽默地讨论。非常感谢你花这么多时间与我们交流。Geoffrey Hinton 是多伦多大学计算机科学系名誉教授、Schwartz Reisman 研究所顾问委员会成员，自 20 世纪 70 年代以来一直参与 AI 的构想和实施。非常感谢你与我们交谈。

Geoffrey, I can't thank you enough. Thank you first of all for being incredibly patient with my level of understanding of this and for discussing it with such heart and humor. I really appreciate you spending all this time with us. Geoffrey Hinton is a professor emeritus with the Department of Computer Science at the University of Toronto, Schwartz Reisman Institute's advisory board member, and has been involved in dreaming up and executing AI since the 1970s. And I just thank you very much for talking with us.

非常感谢你的邀请。

Thank you very much for inviting me.

天哪，既美好又平静。我想我得用 0.5 倍速重听一遍。里面信息量很大。他开暑期学校吗？说真的。当他讲到计算机如何识别出鸟喙时，我喜欢他不停地说「是这样吗？」然后他回答「嗯，不，不是。」我喜欢他对你的评价。是的，他说你很好地扮演了一个对此话题一无所知的好奇者。但我不知道他以为我在扮演。我喜欢他说的「哦，你就像坐在教室前排的热情学生，把其他所有人都惹烦了。」其他人都只求及格，而我就「等等，先生，抱歉，先生，我能回到……你能……不好意思，还有一件事。」哇，听这段历史的发展真是迷人。你真的能感受到现在进展有多快，这加剧了没人站出来监管的恐惧。当你谈论 AI 的复杂性，想到像舒默这样的人要消化这一切然后监管时，在我看来，这似乎要由科技公司来解释并选择如何监管。

Holy, nice and calming. I'm going to have to listen to that back on 0.5 speed, I think. There was some information in there. Does he offer summer school? Seriously. Once he got into how the computer figures out it's a beak, you know, and I love the fact that he kept saying like, 'Is that right?' And he'd be like, 'Well, no. It's not.' I loved his assessment of you. Yes, he said you're doing a great job impersonating a curious person who doesn't know anything about this topic. But I did not know he thought I was impersonating. I loved how he would Did you say like, 'Oh, you're like an enthusiastic student sitting in the front of the room annoying the out of everybody else in the class.' Everyone else is taking it pass/fail and they just Everyone else and I'm just like, 'Wait, sir. I'm sorry, sir. Can I just go back to Could you just Excuse me, one more thing.' Boy, that was it's fascinating to hear the history of how that developed. And you really get a sense for how quickly it's progressing now, which really adds to the fear behind the fact no one's stepping up to regulate. And when you're talking about the intricacies of AI and thinking of someone like Schumer ingesting all of it and then regulating it, it really to me seems like it's going to be up to the tech companies to both explain and choose how to regulate it.

对。并从中获利。没错，就是这样。你知道这些事是怎么运作的。你谈到它的速度和如何阻止它。我认为原因之一可能是，就像核弹一样，很明显为什么需要监管。很明显某些病毒实验必须受到关注。我认为这让人有点措手不及，科幻小说这么快就变成了现实。我只是好奇，因为我记得 15 年前我遇到过禁止完全自主武器的国际运动。人们一直在努力将其纳入公众意识，但正如他所说，总会有一个时刻，每个人都意识到「哦，我们必须协调，因为这是一个生存威胁。」我只是想知道那个临界点是什么。在我看来，如果人们按以往的方式行事，那将在天网之后。就像全球变暖一样。人们问「你觉得我们什么时候会认真对待？」我回答「当水漫到这里时。」对于在车里的人，我指着自己相当高耸的鼻子的一半。事情就是这样。但就这样吧。

Right. And profit off it. Yeah, exactly. You know, how those things work. It is, you know, you talk about that in terms of the speed of it and how to stop it. And I think maybe one of the reasons is it's very evident with like a nuclear bomb, you know, why that might need some regulation. It's very evident that certain virus experimentation has to be looked at. I think this has caught people slightly off guard that it's science fiction becoming a reality as quickly as it has. I just wonder because I remember 15 years ago coming across the international campaign to ban fully autonomous weapons. Like people have been trying for a while to put this into the public consciousness, but to his point, there's going to have to be a moment everyone reaches where they realize, 'Oh, we have to coordinate because it's an existential threat.' And I just wonder what that tipping point is. If in my mind, if people behave as people have, it will be after Skynet. It will be, you know, in the same way with global warming. You know, people say like, 'When do you think we'll get serious about it?' And I go, 'When the waters around here.' And for those of you in your cars, I'm pointing to about halfway up my rather prodigious nose. So, that's how that goes. But there we go.

Britney，有人有什么要说的吗？好的，先生。好吧，我们有什么？特朗普和他的政府似乎对一切、任何地方、同时感到愤怒。他们怎么让这种愤怒保持新鲜？

Britney, what has anybody got anything for us? Yes, sir. All right, what do we got? Trump and his administration seem angry at everything, everywhere, all at once. How do they keep that rage so fresh?

你不知道当亿万富翁总统有多难。我说过很多次了。可怜的小亿万富翁总统。那么有权有势，你不懂那些负担。那些困难。真麻烦。我为他感到愤怒。我的意思是，我一直在想，有人告诉他们「你们赢了」吗？不够。就像，太累了。还不够。就像野蛮人柯南。我这里有，他们女人的哀叹。我要把他们赶下海。简直疯了。但都是他们。必须有人告诉他，所有这些愤怒对他的健康也不好，我们都看到了他的健康状况。

You don't know how hard it is to be a billionaire president. I've said this numerous times. Poor little billionaire president. To be that powerful and that rich you don't understand the burdens. The difficulties. It's troublesome. It makes me angry for him. I mean, I just keep thinking like, has anybody told them that THEY WON? NOT ENOUGH. LIKE IT'S EXHAUSTING. It's not enough. It goes down It's Conan the Barbarian. I have it here, the lamentations of their women. I will drive them into the sea. Like it's bonkers. It's all of them, though. Someone has to tell him that all that anger is also bad for his health, and we are all seeing the health, so.

他是有史以来最健康的总统。所以，我不担心那个。

He's the healthiest person ever to assume the office of the presidency. So, I wouldn't worry about that.

谁说的？他的医生 Ronnie Jackson。但这创造了一个新类别叫「输不起的赢家」。不常见，但偶尔有。但就这样。他们还有什么？John，当被问及是否会赦免 Ghislaine Maxwell 或 Diddy 时，特朗普没有说「不」，这还给你希望吗？这让我希望他们会被赦免吗？是的，我一直这么想。我觉得整件事疯了。一个被判性贩运的妇女，他说「是的，我会考虑。让我查查。」然后你说「查查？你首先，你完全知道那是什么。你认识她。这不是……你知道那里发生了什么。你在说什么？」我觉得 Pam Bondi 很有趣，她问了简单的问题，而她页面上只写了一大堆嘲讽。比如「我听说有他和裸女在一起的照片。你知道什么吗？」她回答「你秃头。」「闭嘴，闭嘴，肥头。」看着这种转移话题简直疯了。最简单的方式应该是「呃，什么？」「太离谱了。不，当然不是。那不是……」再次回到那种发泄，他们采取了简单、合理问题的策略。我就用「你胖，你老婆讨厌你」来回应。「哦，好吧。我不知道会这样。他们还能怎么联系我们？」Twitter，我们是周播节目。

Says who? His doctor, Ronnie Jackson. But it has created a new category called sore winners. You don't see it a lot, but every now and again. But yeah, that's that. What else they got? John, does it still give you hope that when asked if he would pardon Ghislaine Maxwell or Diddy, Trump didn't say no? Does it give me hope that they'll be pardoned? Yes, I've been on that. I find the whole thing insane. A woman convicted of sex trafficking and he's like, 'Yeah, I'll consider it. You know, let me look into it.' And you're like, 'Look into it? What do you take First of all, you know exactly what it was. You knew her. This isn't You knew what was going on down there. What are you talking about?' I thought Pam Bondi, it was so interesting to me, asked simple questions and all she had was like a bunch of roasts written down on her page. They were like, 'I've heard that there are pictures of him with naked women. Do you know anything about that?' And she's like, 'You're bald.' 'Shut up. Shut up, fat head.' Like it was just bonkers to watch the deflection of The simplest thing would be like, 'Uh what?' 'That's outrageous. No, of course not. That's not what the idea again, going back to the vent like that they took the tact of simple, reasonable questions. I am just going to respond with, you know, 'You're fat and your wife hates you.' 'Oh, all right. I didn't know that was going. How else can they keep in touch with us?' Twitter, we are weekly show pod.

Instagram、Threads、TikTok、Blue Sky，我们是一档周播播客。你可以在我们的 YouTube 频道《The Weekly Show with Jon Stewart》上点赞、订阅和评论。稳如磐石。各位，非常感谢。天哪，我真的很喜欢听那位老兄的分享。也感谢你们把这一切组织起来。我真的很享受。首席制作人 Lauren Walker，制作人 Brittany McMahan，制作人 Jillian Spear，视频编辑兼工程师 Rob Vitolo，音频编辑兼工程师 Nicole Boys，以及我们的执行制作人 Chris McShane 和 Katie Gray。希望你们喜欢这一期，我们下次再见。拜拜。《The Weekly Show with Jon Stewart》是 Comedy Central 的一档播客，由 Paramount Audio 和 Busboy Productions 制作。

Instagram, Threads, TikTok, Blue Sky, we are a weekly show podcast. And you can like, subscribe, and comment on our YouTube channel, The Weekly Show with Jon Stewart. Rock solid. Guys, thank you so much. Boy, did I enjoy hearing from that dude. And thank you for putting all that together. I really enjoyed it. Lead producer Lauren Walker, producer Brittany McMahan, producer Jillian Spear, video editor and engineer Rob Vitolo, audio editor and engineer Nicole Boys, and our executive producers Chris McShane and Katie Gray. Hope you guys enjoyed that one, and we will see you next time. Bye-bye. The Weekly Show with Jon Stewart is a Comedy Central Podcast. It's produced by Paramount Audio and Busboy Productions.