视频库 / REC_011
--:--
点击播放，这里会跟随视频显示当前句的中英字幕。
第 11 期 · 回应 Ⅰ·11「人类还是特殊的吗？」
AI and human evolution | Yuval Noah Harari
Yuval Noah Harari章节 · 点击跳转视频
0:00
从中世纪史到 AI：新物种的竞争
1:33
AI 是智能体，不是工具
2:43
对齐难题：AI 模仿行为而非指令
5:44
信息不等于真相，权力不等于智慧
8:07
最智能也最妄想的物种
9:07
1835 年的铁路：技术影响的时滞
11:31
金融与宗教：AI 最先改写的领域
14:34
无用阶级与反技术决定论
16:42
军备竞赛与信任危机的优先级
20:03
没有「那个 AI」：百万智能体的社会实验
23:01
数字移民：以光速抵达的浪潮
本期讲者
尤瓦尔·赫拉利以色列历史学家，希伯来大学教授，牛津大学中世纪军事史博士。著有《人类简史》《未来简史》《今日简史》及 2024 年新作《Nexus》，聚焦信息网络与 AI 对人类社会的影响。
主持人商业会议现场访谈主持人，面向企业领袖听众提问，并引用此前采访 Google CEO Sundar Pichai 的经历。
01从中世纪史到 AI：新物种的竞争
0:00
So you are a milit you studied the military history of the Middle East. Did you ever expect to now be the foremost expert on all things AI and whether we are doomed as humanity? I'm not the foremost expert, but I no I didn't expect to be talking about AI with such such an audience. Uh as you said I was originally a specialist in medieval military history and it's but the Middle Ages are coming back in many ways. Okay. Okay. We're going to get into that and then let me feel free to get into that as I ask you this first question. You call uh artificial intelligence or alien intelligence as you refer to it through throughout your writing the rise of a new species that could replace replace homo sapiens. Yeah. Sapiens your prior book. What does it mean to be human right now?
0:50
um to be aware that for the first time we have real competition on the planet that we have been the most intelligent specy by far for tens of thousands of years and this is how we got from being an insignificant ape in a corner of Africa to being the absolute rulers of the planet and of the ecosystem. And now we are creating something that could compete with us in the very near future. Mhm. The most important thing to know about AI is that it is not a tool like all previous human inventions. It is an agent.
02AI 是智能体，不是工具
1:33
An agent in the sense that it can make decisions independently of us. It can invent new ideas. It can learn and change by itself. All previous human inventions, you know, whether a printing press or the atom bomb, they are tools that empower us. They needed us. They need us because a printing press cannot write books by itself. And it cannot decide which books to print. An atom bomb cannot invent the next more powerful bomb. And an atom bomb cannot decide what to attack. An AI weapon can decide by itself which target to attack and design the next generation of weapons by itself. So this is why you argue that AI is potentially or it sounds like you're saying is by the way now but you've written in the past potentially more momentous in the invention of the telegraph, the printing press, even writing. But the way you talk about it in Nexus is that it is a baby. Yeah, because it learns from us.
03对齐难题：AI 模仿行为而非指令
2:43
And therefore your argument is that we especially the powerful leaders in this room have a lot of responsibility because how we act is how AI will be. You cannot expect to lie and cheat and have benevolent AI. Yeah, explain that. Yeah, there is a big discussion around the world about AI alignment. Okay, we are creating these increasingly super intelligent, very powerful new agents. How do we make sure that these agents remain aligned with human goals and uh uh with the benefit of humanity that they do what is good for us? And so there is a lot of research and a lot of efforts focused on the idea that if we can design these AIs in a certain way, if we can teach them certain principles, if we can code into them certain goals, then we will be safe. But there are two main problems with this approach. First of all, again the very definition of AI is that it can learn and change by itself. If you have a machine that can act automatically but only following pre-programmed orders, then you know
3:58
it's a coffee machine. It can do something automatically, produce coffee, but it cannot decide or invent anything by itself. It's not an AI. So when you design an AI, by definition, this thing is going to do all kinds of things which you cannot anticipate. If you can anticipate everything it will do, it is by definition not an AI. So that's one problem. The other even bigger problem that we can think about AI like you said like a baby or a child and you can educate a child to the best of your ability and he or she will still surprise you for better or worse. No matter how much you invest in their education, they are independent agents.
4:47
they might eventually do something which will surprise you and even horrify you. Uh the other thing is everybody who has any knowledge of education knows that in the education of children it matters far less what you tell them and what you do. It matters far more what you do. Yeah. If you tell your kids don't lie and they your kids watch you lying to other people they will copy your behavior not your instructions. Now if we have now this big projects to educate the AIs, not to lie, but the AIs are given access to the world and they watch how humans behave and they see some of the most powerful humans on the planet including their parents lying.
04信息不等于真相，权力不等于智慧
5:44
The AI will copy the behavior. people who think that I can run say this huge AI corporation and while I'm lying I will teach my AIS not to lie it will not work it will copy your behavior your one of your central arguments is that we as society at large have focused way too much on power and you also make the argument that some disagree with or call counterintuitive that more information is great for democracies Because you say all information is not true information etc. Most information is not the truth.
6:21
Right? There's a huge confusion between information and truth. Yes, sign information is true and you get information to get to know the truth. But generally the truth is a very very small subset of most of the of all the of the information in the universe. So if we are focusing too much on power and that's a very important distinction. You say this is why we have failed as people largely to answer actually the biggest questions of life. If we can be more productive, we can be richer, we can have stronger militaries, but many of us can't answer the question as you write, who are we?
6:59
What should we aspire to? And what is a good life? Essentially, we are accumulating power, not wisdom. Yeah. How can we change it? Um, and that's the big problem of human history. You know, for thousands of years, we are extremely good in acquiring more power. Again, this is how we transform ourselves from an insignificant ape in East Africa into the ruler of the world. We can fly to the moon. We can split the atom. But we don't seem to be significantly happier than we were in the stone age. Uh, we don't know how to translate power into happiness. Again, you look at the most powerful people on the planet, they don't seem to be the most the happiest people on the planet. So that there is a very Do you want to ask them? There are many of them. I'm not necessarily referring to the people in this room. I don't I I I want to clarify. I don't think there is a contradiction, okay, between power and happiness. I don't think that as you acquire more power, you necessarily become miserable. No, but there is no there it can go
05最智能也最妄想的物种
8:07
together. But it doesn't necessarily go together. And as a specy, we have not been particularly good in translating power into happiness or even into knowledge and wisdom. Again, we tend to confuse intelligence with with knowledge and with truth. But um we are the most intelligent species on the planet. We are also the most delusional species. destructive you argue and and self-destructive. Yeah. The kind of things that people believe no other animal on the planet will believe such nonsense except if I look at my own country like you would not find any animal that believes that if you go and kill other members of your species you will be rewarded after death by entering paradise. No chimpanzeee will believe that. No horse would believe that. No wolf will believe that.
061835 年的铁路：技术影响的时滞
9:07
Millions of people believe that. And they believe it so strongly that they actually go and kill people in the expectation that as a result they will be rewarded in paradise with whatever. You um we we took a a really interesting poll this morning asking the leaders in this room how consequential they think AI uh has been so far in their business. And actually only a the businesses they lead and only a small portion said significantly. Most it was moderately or not at all. Yeah. Can you speak to them uh as if we were sitting here 36 months from now? Is there any world in which AI doesn't have a significant impact on their business? Um I it depends on their business. But in most fields again the question is one of of time scale. You know, I've been talking to a lot of the people who lead the AI revolution and many of them say, you know, we are already in the middle of the AI revolution. We still haven't seen anything really major. And that's just the difference between how historians view time and how CEOs and entrepreneurs
10:14
view time. For an entrepreneur, two years is a long time. For historians, it's nothing. It's like imagine that we are now sitting in London and the year is 1835. The first railway has been opened between Manchester and Liverpool 5 years ago and we have now this conference in London in 1835 and people saying you know all this talk about railways changing the world the industrial revolution this is nonsense. We have had railways for ages, five years. And look, okay, so there is a some changes that the that they now have people traveling the with the trains or they they uh uh move around cold more easily. But nothing major happened because there is a time lag between the invention of a technology and the moment when you see the actual social and political consequences. Yeah. So we now know that the industrial revolution and trains they completely transformed everything geopolitics the way people fight wars the economy family structure but it just took more than five years the same is likely to happen with AI
07金融与宗教：AI 最先改写的领域
11:31
um in all fields from the obvious to the less obvious like I think that one of the first fields we'll see major changes is finance. Okay. Uh that AI is going very quickly to take over the financial system. We have some bankers in the room. So tell us more. Uh because finance is the ideal playing ground for AI. It's purely anformational real. If you want to have an AI self-driving vehicle on the road, which have been promised again and again and we are still not there. The problem Whimo. Yeah. But you go around London, you don't see these tens of thousands of of self-driving vehicles yet. I just passed my first driving lesson. Uh, congratulations. Okay. Um, and you still need to learn how to drive. Okay. So, um, the problem is that for driving, you need to deal with the messy physical world of pedestrians and and holes in the roads and whatever. But in finance, it's only information in, information out. It's much easier for an AI to master that. And what happens to finance once AIs uh for instance start inventing
12:43
new financial devices that the human brain is simply incapable of dealing with because it's mathematically too complex. Um we are going to see AI changing even things like religion. How at least religions which are based on texts like Judaism, Islam, Christianity, they give ultimate authority to the text. Yeah. Not to any human being. Now until today, humans were nevertheless the main authority in these religions because the texts could not speak. The Bible could not interpret itself. The Bible could not answer your questions. So you needed a human being as an intermediary. What happens when you have an AI text that can speak for itself? No Jewish rabbi can know all the texts of Judaism because there are too many of them. For the first time in history, there is something on the planet that is able to remember every single word in every writing of every rabbi in the last 2,000 years and talk back to you and explain and defend its views.
13:55
So I have friends who are now working on building religious AIS that are meant to either uh augment or replace human religious leaders especially in textbased religions. If it's if the religion is not based on texts it doesn't give authority to a text. It's a different story. Okay. But I go and we're going to questions next. I'm going to come first to Matia Moore if you want to raise your hand and we'll get you a microphone. But I go and talk to my pastor at our church when I am going through a difficult time. I am never going to talk to chat GPT like that.
08无用阶级与反技术决定论
14:34
Mhm. It's it's an individual choice. The question you think some will I know that already millions of people do it. I mean I know people who go for now AIs to get psychological counseling. Yes. That AI is their best friend. like teenagers something happened in school they consult with they tell the AI what happened and ask for advice about relationships so let me get back to and then the question is next let me get back to what you've said though about replacing jobs this is really important and you write and talk a lot about a what you're worried about a uh what it becomes a useless class that's what you've talked about and I I it was five or six years ago I interviewed Google CEO Sundar Pachai in Oklahoma and one of their data centers. We need many many more of them now to to power AI. And I remember asking him about AI. It was people were talking about a lot then and he essentially told me and I'm paraphrasing here that if AI proves to be quote in his words very disruptive to too many American jobs, he essentially
15:37
said they would be open to slowing it down. Okay. I'm talking about this not just Google, just writ large these companies right now. Mhm. If we are headed for what you're talking about a potential useless class, many interestingly white collar jobs. So it's getting a little bit more attention perhaps than when it replaced blue collar jobs which is a whole issue in and of itself. The what what what do we do to make sure we as a survi society not only survive but thrive? Um I I want to emphasize that AI has enormous positive potential as well as dangerous potential. And I don't believe in historical or in technological determinism. You can use the same technology to create completely different kinds of societies. We saw it in the 20th century that you use exactly the same technology to build communist totalitarian regimes and liberal democracies. That's right. It's the same with AI. We have a lot of choices about what to do with it. If again provided we remember that for the first time we are
09军备竞赛与信任危机的优先级
16:42
dealing with agents and not tools. So it makes it much more complicated but we do have still most of the agency is in our hands and the question of how we develop the technology and even more importantly how we how we deploy it. Uh we can make a lot of choices there. We have agency in how we move forward. We don't have a choice that it has come. Yeah, we have power in how we use it and go through it. Absolutely. The main problem is that now the companies and countries that lead the AI revolution has been locked into an an arms race uh situation. So even if they know that it would be better to slow down, to invest more in safety, to be careful about this or that potential development, they are constantly afraid that if we slow down and they don't slow down, they will take over the world. So let let's get to some questions here. Matiam Moore with Emotion Network. Hi. Yes, right here.
17:48
Hi. No. Uh so congratulations because I mean your books are really eye opening and so you gave really lot of answers through your books and also today uh we as a company we have a conference called tech emotion because we strongly believe in the power of mixing technology and innovation with emotion creativity and culture but and in in what you're saying this is a lot of a mix of this so what I think it's really interesting what you are saying about also the effect on religion the fact of on the soul of the people and this is also related to what she was saying about the purpose of the life of the people that is going to be destroyed or changed a lot by AI by all these innovation. So how do you think it's possible to get a future where people is going to be more satisfied and more happy and also find more purpose what they do with the difficulties that this world is changing. So I know this is a big question but I think that is the most important thing in in our life much more than business more than anything else.
18:49
Yeah. So it's a very big subject. The most important thing I only have time to talk about one thing. So the most important thing is that we need to solve our own human problems instead of relying on the AI to do it for us. And the most and the key problem is the is the problem of trust and cooperation. At the present moment, trust is collapsing all over the world, both between countries and within societies. And the hope that okay, humans can no longer trust each other. So the international system and the trade system and everything is collapsing, but the AI will save us will. So no, it will not. In a world in which humans compete with each other ferociously and cannot trust each other, the AI produced by such a world will be a ferocious, competitive, untrustworthy AI. It's not possible for humans as they engaged in in in this ferocious competition to create benevolent, trustworthy AI. It will just not happen. So if you think about it's just a question of priority.
10没有「那个 AI」：百万智能体的社会实验
20:03
We have now this big human trust problem and we have the issue how do how do we develop AI. Too many people think okay let's first solve the how do we develop AI problem and then this will solve the human trust problem. It will not work. We need to get our priorities the other way. first solve the human trust problem then together we can create benevolent AI. Of course this is not what is happening right now in the world. Do we have one more? Yes. Right here. Thanks very much. You quick question from me. So you know in human history there's been organizing principles and and you write about that so much in your books and there's been in some senses at least geographically monolithic organizing principles like religion and the church was one of those but when we talk about AI we're not talking about something that is monolithic right the there is no the AI this is really effectively going to be multiple plethoras of AIs manifesting themselves absolutely and in that context you know when you when you describe um AI at
21:12
replacing religion in some sense. I think the real question for me is when you have no single organizing principle, there is no the AI that gets developed with any kind of intent whether that intent is benevolent or otherwise and there's all of these competing AIs that that that are effectively evolving fast. What what does that world look like? That's a very very important point. I mean the AI will not be one big AI. We are talking about potentially millions or billions of new AI agents with different characteristics and produced by different companies, different countries uh everywhere in the military, in the financial system, in the religious system. So you'll have a lot of religious AIs competing with each other which AI will be the authoritative uh AI rabbi for which currents of Judaism. and the same in Islam and the same in in Hinduism and in Buddhism and so forth. So you will have competition there and in the financial system. Um and we just have no idea what the outcome will be. We have thousands of
22:17
years of experience with uh human societies. What happens when millions of humans compete for economic power, for religious authority? It's very complex, but we at least have some experience in how these things develop. We have zero experience what happens in AI societies when millions of AIs compete with each other. We just don't know. Now you this is not something you can simulate in the AI labs. If OpenAI, for instance, wants to check the safety or the potential outcome of its latest AI model, it cannot simulate history in a laboratory.
11数字移民：以光速抵达的浪潮
23:01
It can check for all kinds of of failures in the system. But it cannot tell in advance what happens when you have millions of copies of these AIs in the big world outside developing in unanticipated ways interacting with each other and with billions of human beings. So it's the in a way it's the biggest social experiment in human history. We are all part of it and nobody has any idea how it will develop. Uh you know one analogy to keep in mind we now have this uh uh immigration crisis in the US in Europe elsewhere lots of people worried about immigrants.
23:50
Why are people worried about immigrants? There are three main things that come to people's mind. They will take off jobs. They come with different cultural ideas. They will change our culture. They may have political agendas. They might try to take over the country politically. These are the three main things that people keep coming back to. Now, you can think about the AI revolution as simply a wave of immigration of millions and billions of AI immigrants that will take people's jobs that have very different cultural ideas and that might try to gain some kind of political power.
24:39
And these AI immigrants, these digital immigrants, they don't need visas. They don't cross the sea in some rickety boat in the middle of the night. They come at the speed of light. And um I I look for instance at farright parties in in in in Europe and they talk so much about the human immigrants, sometimes with justification, sometimes without justification. They talk they don't talk almost at all about the wave of of digital immigrants that is coming into Europe. And I think they should be much if they care about the sovereignity of their country, if they care about the economic and cultural future of their country, they should be far more worried about the digital immigrants than about the human immigrants. You've all this has been remarkable. Thank you very very much. Thank you.
25:36
Great. Thank you very very much. I'll see you after.
视频总结 · 一句话概括与核心要点
一句话概括
Harari 认为 AI 是人类历史上第一个"能自主决策、自我学习"的智能体（agent）而非工具，其行为会模仿人类而非听从指令，因此人类若不先解决自身的信任与合作危机，就不可能造出可信的 AI，而 AI 对金融、宗教、就业乃至国家主权的冲击将像一场"以光速抵达的数字移民潮"。
核心要点
- AI 是智能体而非工具，这是它与一切前代发明的本质区别。 印刷机不能自己写书或决定印什么，原子弹不能自己设计下一代炸弹、也不能决定攻击谁；而 AI 武器可以自主选择目标并自行设计下一代武器。人类在数万年里凭借智力优势从非洲一角的猿类成为地球统治者，如今第一次面临真正的竞争对手。
- "对齐"路线存在两个结构性缺陷。 第一，AI 的定义就是能自我学习和改变——如果你能预知它的一切行为，那它只是台咖啡机，不是 AI；第二，AI 像孩子一样是独立主体，再好的教育也会带来意想不到、甚至令人恐惧的结果。
- AI 学的是人类的行为而不是人类的说教。 教育学常识：孩子看到父母对别人撒谎，会复制行为而非指令。AI 拥有观察世界的权限，若看到最有权势的人（包括"父母"——AI 公司领导者）在撒谎，就会照做。"我一边撒谎、一边教我的 AI 不要撒谎"注定失败。
- 信息≠真相，人类积累了权力却没有积累智慧。 真相只是宇宙中全部信息的一个极小子集。人类能登月、能分裂原子，却不比石器时代更幸福，也答不出"我们是谁、该追求什么、什么是好的生活"。人类是最聪明的物种，也是最善于自我欺骗的物种——没有任何黑猩猩、马或狼会相信"杀死同类后能进天堂"，但数百万人相信并付诸行动。
- 现场民调显示多数企业领袖认为 AI 影响"一般或没有"，Harari 用 1835 年的伦敦类比反驳。 曼彻斯特—利物浦铁路已开通 5 年，当时的人也会说"铁路改变世界是胡扯"，但铁路最终重塑了地缘政治、战争方式、经济与家庭结构。技术发明与社会政治后果之间存在时滞；企业家眼中的两年很长，历史学家眼中却微不足道。
- 金融将是 AI 最先攻陷的领域，因为它是纯信息博弈。 自动驾驶要应对行人、路坑等混乱的物理世界，承诺多年仍未大规模落地；而金融"信息进、信息出"，AI 极易掌控，并可能发明出人脑在数学上根本无法理解的金融工具。
- AI 将动摇以文本为权威的宗教（犹太教、伊斯兰教、基督教）。 这些宗教把最高权威赋予经文而非任何人，但过去经文不会说话、不能自我解释，人类只好充当中介。如今 AI 能记住过去 2000 年每位拉比的每一个字，并与你对话、辩护自己的观点——Harari 的朋友正在开发旨在增强或取代人类宗教领袖的宗教 AI。他还指出，已有数百万人把 AI 当心理咨询师和最好的朋友，青少年向 AI 倾诉校园与感情问题。
- 技术不决定命运，但军备竞赛正在剥夺选择空间。 20 世纪同一套技术既造出极权政权也造出自由民主。AI 同样如此，人类仍握有大部分能动性——尤其是在"如何部署"上。但领先的公司与国家已被锁进军备竞赛：即便知道该放慢、该加大安全投入，也担心"我慢下来而对手不慢，对手就会统治世界"。
- 优先级必须颠倒：先解决人类互信，再造 AI。 当前全球范围内国家之间、社会内部的信任正在崩塌。指望"人类互不信任，但 AI 会拯救我们"是错的——在残酷竞争、互不信任的世界里造出来的 AI，必然是残酷、好斗、不可信的 AI。
- 未来不是"一个 AI"，而是数百万、数十亿个 AI 智能体相互竞争，人类对此零经验。 不同公司、不同国家将在军事、金融、宗教等领域推出各自的 AI，甚至会有多个"AI 拉比"争夺不同教派的权威。人类对"数百万人如何竞争权力与权威"有数千年经验，对"数百万 AI 相互竞争"则一无所知；OpenAI 等实验室可以测试系统故障，却无法在实验室里模拟历史——这是人类历史上最大的社会实验。
结论与值得注意的细节
- Harari 提出的最尖锐比喻：AI 革命本质上是一场"数字移民潮"。人们担忧移民的三件事——抢走工作、带来不同文化、谋求政治权力——AI 全部符合，而且它们不需要签证、不用深夜坐破船渡海，而是以光速抵达。他直指欧洲极右翼政党整天谈论人类移民，却几乎不谈数字移民；若真在乎国家主权和经济文化前途，应更担心后者。
- 他强调自己并不认为权力与幸福相矛盾，二者可以并存，只是人类作为物种在"把权力转化为幸福与智慧"这件事上表现糟糕。
- 关于"无用阶级"：主持人提到五六年前 Google CEO Pichai 曾表示若 AI 对美国就业过于破坏性，愿意放慢步伐；Harari 的回应是 AI 有巨大正面潜力，关键在于选择，而当前最大障碍正是军备竞赛心态。
- 主持人本人表示遇到困难时会找教堂牧师而绝不会找 ChatGPT，Harari 的回应是"这是个人选择，但已有数百万人在这么做"。
- 开场的小细节：Harari 本是中世纪军事史专家，他调侃道"中世纪正在以多种方式回归"。
核心句型 · 9
1. It is not X. It is Y.
“It is not a tool like all previous human inventions. It is an agent.”
用两个短句先否定再断言，制造强烈对比。适合提出核心定义或纠正误解。仿写时保持两句都简短、平行。
2. If you can …, it is by definition not …
“If you can anticipate everything it will do, it is by definition not an AI.”
以定义为依据的归谬式论证。by definition 强调结论来自概念本身而非经验。适合哲学性或概念性辩论。
3. It matters far less what … and far more what …
“It matters far less what you tell them and what you do. It matters far more what you do.”
用 far less / far more 强化对比程度，比 less/more 更有力。可用于强调优先级的场合。
4. This is how we got from being … to being …
“This is how we got from being an insignificant ape in a corner of Africa to being the absolute rulers of the planet”
叙述从起点到终点的巨大跨越，get from A to B 后接 being + 名词短语，描述身份转变。适合概括历史或成长历程。
5. There is a huge confusion between X and Y.
“There's a huge confusion between information and truth.”
指出两个常被混淆的概念，后文随即区分。是学术和评论文写作中引出概念辨析的标准开头。
6. For X, … is a long time. For Y, it's nothing.
“For an entrepreneur, two years is a long time. For historians, it's nothing.”
两个平行短句对比不同视角，第二句用 it's nothing 收得干脆。适合展示立场或尺度差异。
7. Imagine that we are now sitting in … and the year is …
“Imagine that we are now sitting in London and the year is 1835.”
引入历史思想实验的标准句式，把听众带入具体时空再类比现实。演讲中用来打破当下视角的局限。
8. You can think about X as simply …
“You can think about the AI revolution as simply a wave of immigration of millions and billions of AI immigrants”
提出一个重新框定（reframing）的类比；simply 暗示这一视角化繁为简。适合把陌生概念翻译成熟悉图景。
9. We have zero experience (with) what happens when …
“We have zero experience what happens in AI societies when millions of AIs compete with each other.”
zero 比 no 更强调彻底空白。用于强调未知领域的不确定性，后接 what happens when 从句描述情境。
生词精讲 · 71 · 按出现顺序
foremost
/ˈfɔːrmoʊst/
adj.
0:00
最重要的，首屈一指的
doomed
/duːmd/
adj.
0:00
注定失败/毁灭的
medieval
/ˌmediˈiːvəl/
adj.
0:00
中世纪的
insignificant
/ˌɪnsɪɡˈnɪfɪkənt/
adj.
0:50
无足轻重的，微不足道的
ecosystem
/ˈiːkoʊˌsɪstəm/
n.
0:50
生态系统
agent
/ˈeɪdʒənt/
n.
0:50
（能自主行动的）主体，智能体；AI 领域术语
printing press
n. phr.
1:33
印刷机
empower
/ɪmˈpaʊər/
v.
1:33
赋能，使更强大
momentous
/moʊˈmentəs/
adj.
1:33
意义重大的，划时代的
benevolent
/bəˈnevələnt/
adj.
2:43
仁慈的，善意的
alignment
/əˈlaɪnmənt/
n.
2:43
对齐；AI 领域指使 AI 目标与人类价值一致
pre-programmed
/ˌpriːˈproʊɡræmd/
adj.
2:43
预先编程的
anticipate
/ænˈtɪsɪpeɪt/
v.
3:58
预见，预料
by definition
phr.
3:58
从定义上讲，本质上
for better or worse
phr.
3:58
不论好坏，无论结果如何
horrify
/ˈhɔːrɪfaɪ/
v.
4:47
使震惊，使恐惧
counterintuitive
/ˌkaʊntərɪnˈtuːɪtɪv/
adj.
5:44
反直觉的
at large
phr.
5:44
整体上，总体而言（society at large 整个社会）
subset
/ˈsʌbset/
n.
6:21
子集
distinction
/dɪˈstɪŋkʃən/
n.
6:21
区别，区分
aspire to
phr. v.
6:59
渴望，追求（某目标）
accumulating
/əˈkjuːmjəleɪtɪŋ/
v.
6:59
积累，积聚
split the atom
phr.
6:59
裂变原子，实现核裂变
miserable
/ˈmɪzərəbəl/
adj.
6:59
痛苦的，悲惨的
delusional
/dɪˈluːʒənəl/
adj.
8:07
妄想的，有错觉的
self-destructive
/ˌselfdɪˈstrʌktɪv/
adj.
8:07
自我毁灭的
nonsense
/ˈnɑːnsens/
n.
8:07
胡说，荒谬的想法
consequential
/ˌkɑːnsɪˈkwenʃəl/
adj.
9:07
影响重大的，有重要后果的
poll
/poʊl/
n.
9:07
民意调查，投票
time scale
n. phr.
9:07
时间尺度
for ages
phr.
10:14
很久了（此处反讽：五年而已）
time lag
n. phr.
10:14
时滞，时间差
geopolitics
/ˌdʒiːoʊˈpɑːlətɪks/
n.
10:14
地缘政治
playing ground
n. phr.
11:31
活动场地，施展领域
pedestrians
/pəˈdestriənz/
n.
11:31
行人
messy
/ˈmesi/
adj.
11:31
混乱的，杂乱无章的
ultimate authority
n. phr.
12:43
最高权威
intermediary
/ˌɪntərˈmiːdieri/
n.
12:43
中介，中间人
rabbi
/ˈræbaɪ/
n.
12:43
拉比（犹太教教士、宗教导师）
augment
/ɔːɡˈment/
v.
13:55
增强，扩充
pastor
/ˈpæstər/
n.
13:55
牧师
counseling
/ˈkaʊnsəlɪŋ/
n.
14:34
心理咨询，辅导
paraphrasing
/ˈperəfreɪzɪŋ/
v.
14:34
转述，意译
disruptive
/dɪsˈrʌptɪv/
adj.
14:34
颠覆性的，破坏性的
writ large
phr.
15:37
放大来看，整体而言（正式用语）
white collar
adj. phr.
15:37
白领的（脑力劳动）
in and of itself
phr.
15:37
就其本身而言
determinism
/dɪˈtɜːrmɪnɪzəm/
n.
15:37
决定论
totalitarian
/toʊˌtælɪˈteriən/
adj.
15:37
极权主义的
agency
/ˈeɪdʒənsi/
n.
16:42
能动性，自主行动的能力
deploy
/dɪˈplɔɪ/
v.
16:42
部署，投入使用
locked into
phr.
16:42
被困于，陷入（无法脱身）
arms race
n. phr.
16:42
军备竞赛
eye opening
adj. phr.
17:48
令人大开眼界的
ferociously
/fəˈroʊʃəsli/
adv.
18:49
凶猛地，残酷地
untrustworthy
/ʌnˈtrʌstwɜːrði/
adj.
18:49
不可信赖的
organizing principles
n. phr.
20:03
组织原则
monolithic
/ˌmɑːnəˈlɪθɪk/
adj.
20:03
单一整体的，铁板一块的
plethoras
/ˈpleθərəz/
n.
20:03
过多，大量（plethora of）
manifesting
/ˈmænɪfestɪŋ/
v.
20:03
显现，表现出
authoritative
/əˈθɔːrəteɪtɪv/
adj.
21:12
权威的
currents
/ˈkɜːrənts/
n.
21:12
（思想、宗教的）流派，潮流
simulate
/ˈsɪmjəleɪt/
v.
22:17
模拟
unanticipated
/ˌʌnænˈtɪsɪpeɪtɪd/
adj.
23:01
未预料到的
analogy
/əˈnælədʒi/
n.
23:01
类比
political agendas
n. phr.
23:50
政治议程，政治意图
take over
phr. v.
23:50
接管，夺取控制权
rickety
/ˈrɪkəti/
adj.
24:39
摇摇欲坠的，破旧不牢的
far-right
/ˌfɑːrˈraɪt/
adj.
24:39
极右翼的
sovereignty
/ˈsɑːvrənti/
n.
24:39
主权（原文拼作 sovereignity）
justification
/ˌdʒʌstɪfɪˈkeɪʃən/
n.
24:39
正当理由
理解自测 · 11 题 · 是真懂了，还是以为自己懂
1. 赫拉利认为 AI 与印刷机、原子弹等以往发明的根本区别是什么？
根本区别在于 AI 是智能体（agent）而非工具（tool）。在开场部分他指出，以往的一切发明都需要人类：印刷机不能自己写书、也不能决定印什么书；原子弹不能发明下一代更强的炸弹，也不能决定攻击谁。而 AI 武器可以自主选择目标并设计下一代武器。也就是说，AI 能独立决策、发明新想法、自我学习和改变，这是「工具赋能人类」与「智能体可能取代人类」之间的分水岭，也是全场后续论证的基础。
2. 他用哪个历史场景来回应「AI 目前对企业影响不大」的现场调查结果？
他虚构了 1835 年在伦敦召开的一场会议。当时利物浦—曼彻斯特铁路已开通五年，与会者可能会说「铁路改变世界是胡扯，我们有铁路都五年了，不过是人坐火车、运煤更方便」。但事后看，铁路和工业革命彻底改变了地缘政治、战争方式、经济和家庭结构，只是花了远超五年的时间。他借此说明技术发明与其社会政治后果之间存在时滞，而企业家的两年与历史学家的两年是完全不同的尺度。
3. 赫拉利预测哪两个领域会最先被 AI 显著改变？各自的原因是什么？
一是金融，二是宗教。金融被点名是因为它「纯信息进、信息出」，不像自动驾驶要面对行人、路坑等混乱的物理世界，所以 AI 更容易掌握；他还预测 AI 会发明人脑在数学上无法处理的金融工具。宗教方面，他专指犹太教、伊斯兰教、基督教等「文本宗教」：这些宗教把最高权威赋予经典，但经典不能自我解释，人类因而充当中介；而 AI 能记住两千年来所有拉比的著作并回答、辩护自己的解释，可能取代这一中介角色。
4. 在结尾的类比中，人们担心移民的三件事是什么？AI 如何对应？
三件事是：抢走工作、带来不同文化观念从而改变本土文化、可能有政治议程并试图夺取政治权力。赫拉利认为 AI 革命可以被理解为数百万乃至数十亿「数字移民」的浪潮，在这三点上完全对应：AI 会取代工作、带来非常不同的「文化观念」、并可能谋求某种政治权力。而且这些数字移民不需要签证，不用半夜乘破船渡海，而是以光速抵达。这一框架出现在问答环节末尾，目的是把抽象的技术风险翻译成公众和政治家熟悉的语言。
5. 赫拉利反对「靠设计目标来保证 AI 安全」的两个理由分别是什么？它们之间有什么关系？
第一个理由是定义层面的：AI 的本质就是能自我学习和改变，如果一个系统的所有行为都可以被预知，那它只是咖啡机，不是 AI；因此「完全可控的 AI」在概念上自相矛盾。第二个更大的理由是学习方式：AI 像孩子一样，学的是「你做什么」而不是「你说什么」，若它观察到最有权力的人（包括其「父母」，即 AI 公司）在撒谎，它会复制这种行为。两者的关系是递进的：第一点说明无法靠编码穷尽 AI 的行为，第二点说明即使去教育它，教育的实质内容也取决于人类自身的品行，把对齐问题从技术问题转化为伦理和政治问题。
6. 为什么赫拉利说人类是「最智能也最妄想的物种」？这一判断如何支撑他关于「权力与智慧」的论点？
他的例子是：没有任何黑猩猩、马或狼会相信「杀死同类后死后会进入天堂得到奖赏」，但数百万人相信，而且相信到真的去杀人。这说明智能不等于清醒或真理，人类的认知能力同时也是生产虚构和自我毁灭信念的能力。这一判断支撑了前文的核心论点：数千年来人类极擅长积累权力（上月球、裂变原子），却并未把权力转化为幸福、知识或智慧；我们常把智能与知识、真相混为一谈，而实际上真相只是信息的极小子集。因此问题不在于能力不足，而在于我们回避「我们是谁、该追求什么」这些根本问题。
7. 赫拉利反对技术决定论，但同时又说各方被「锁进军备竞赛」。这两个说法矛盾吗？请梳理他的逻辑。
不矛盾，两者处于不同层面。反技术决定论是说技术本身不决定社会形态：20 世纪同样的技术既造出极权政权也造出自由民主，AI 同样可以被用来建设完全不同的社会，人类在开发尤其是部署方式上仍握有大部分能动性。军备竞赛描述的则是当下的现实困境：领先的公司和国家即使知道该放慢、多投入安全，也担心「我慢了别人不慢就会被统治」。前者是可能性空间，后者是现实中的博弈结构，恰恰说明能动性并未被技术剥夺，而是被人类之间的不信任所束缚。这也自然引向他对信任问题的强调。
8. 赫拉利为什么坚持「先解决人类信任问题，再造 AI」而不是相反的顺序？他的推理链是什么？
他的推理链是：（1）AI 从人类行为中学习，而非从指令中学习；（2）当前世界各国、各社会之间信任正在崩塌，人类在进行凶猛的竞争；（3）在这种环境下造出的 AI 必然是凶猛、竞争性、不可信的 AI，因为它复制的是造它的人的行为；（4）因此指望「先造出强 AI，再用它来修复人类的信任问题」在逻辑上不成立，等于指望一个从不信任中学出来的系统去创造信任。结论是优先级必须倒过来：先解决人类的信任与合作问题，然后才可能共同创造出善意、可信的 AI。他也坦承这并非当前世界正在发生的事。
9. 为什么赫拉利认为 OpenAI 等实验室的安全测试无法预知 AI 的真实后果？这与观众关于「没有单一 AI」的提问有何关系？
观众指出，与宗教等历史上单一的组织原则不同，AI 不是「那一个 AI」，而是许多相互竞争、快速演化的 AI。赫拉利认同并推进：未来会有数百万乃至数十亿具有不同特性、来自不同公司和国家的 AI 智能体，分布在军事、金融、宗教各领域，甚至各教派会有相互竞争的 AI 拉比。人类对「人类之间竞争」有数千年经验，但对「AI 之间的社会」经验为零。实验室可以检测单个模型的各类故障，却无法模拟历史，即无法复现数百万副本在真实世界中以不可预料的方式互相互动、并与数十亿人互动的过程。所以这是史上最大的社会实验，且没有人能提前知道结果。
10. 如果有人反驳说：「AI 是从海量文本训练的，并不是只看几个 CEO 的行为，因此 CEO 撒谎不会决定 AI 的品行」，赫拉利可能如何回应？
基于他的论证，他可能有两层回应。第一，他的论点并不局限于少数 CEO，而是指 AI「被给予接触世界的权限」，观察的是整个人类社会的行为模式，包括当下各国和社会之间信任崩塌、凶猛竞争的整体状态；训练数据越广，AI 学到的越是人类真实的行为，而非教科书上的规范。第二，AI 公司作为「父母」拥有特殊影响力：它们决定训练目标、奖励什么、部署到什么场景，如果这些机构本身在军备竞赛中撒谎、走捷径，这种行为结构会通过激励设计传递给 AI。他很可能承认技术细节可讨论，但坚持核心结论：无法在一个不可信的社会里造出可信的 AI。
11. 赫拉利的「数字移民」类比，用于分析欧洲以外的地区（如中国或美国）是否同样成立？可能有什么局限？
类比的三要素——抢工作、改变文化、谋求政治权力——在美国同样适用：美国也有激烈的移民争论，而 AI 对白领就业的冲击正是主持人提到的焦点。但类比有局限。首先，「移民」暗示 AI 来自外部，而对美国这样的 AI 生产国来说，AI 更像本国培育的产物，责任主体是国内企业而非「他者」，此时类比反而可能掩盖问责对象。其次，在没有类似移民政治争论的社会里，这一修辞的动员力会减弱。第三，赫拉利本人也强调各国被锁进军备竞赛，一国若因此「限制数字移民」可能担心落后，这恰是移民类比无法涵盖的博弈维度。因此类比更适合作为唤起警觉的修辞，而非完整的政策分析框架。
下载便签
手机：长按图片保存
← 上一期 · REC_010ICM 2026 Public Lectures - Terence Tao
下一期 · REC_012 →Do LLMs Understand? AI Pioneer Yann LeCun Spars with DeepMind’s Adam Brown.
苏菲拉底
THE SOPHIE LAB · ASK THE BEST MINDS THE BIG QUESTIONS
内容仅供学习 · thesophielab.com