被技术拯救的人 The Man Technology Saved

有一个问题，我始终放不下。 Here's a question I can't stop thinking about.

当你让 AI 替你思考，你究竟交出了什么？ When you let AI think for you, what exactly are you giving up?

今天这位嘉宾，已经在这个问题上思考了很多年。 My guest today has spent years on that question.

Hosni Zaouali 从一名难民，一路走到斯坦福，再到亲手创办两家 AI 教育公司。 Hosni Zaouali went from refugee to Stanford to founding two AI education companies.

他本身就是活生生的证明：技术可以把一个人从几乎一无所有中托举起来。 He's living proof that technology can lift someone out of almost nothing,

而正因如此，他如今的警告才格外发人深省：每一次我们把思考交给机器，我们自己的那块肌肉，就少练了一次。 which is exactly why it's so striking that he's now warning us about it: that every time we offload our thinking to a machine, we stop exercising the muscle ourselves.

他经营着一家 AI 公司，同时又在为 AI 敲响警钟。 He runs an AI company, and he sounds the alarm on AI.

在这场对话里，他会解释，为什么这两件事并不矛盾。恰恰相反，这正是全部的重点。 In this conversation, he explains why those two things aren't a contradiction. They are the whole point.

欢迎来到新一期的《雏形》。 Welcome to another episode of Prototype.

这一期，我想聚焦我所说的、存在于你人生里的几组矛盾。因为你小时候是以难民的身份去到欧洲的。 The focus of this episode is to identify what I call some of the contradictions in your life, because you were a refugee who traveled to Europe as a child.

后来，你却在某种意义上被技术拯救，并把整个职业生涯都建立在了技术之上。 And then you ended up being saved, somehow, by technology, and built your entire career around it.

可如今，你一边提醒人们警惕技术在教育中的危害，一边又在创办一家 AI 教育科技公司。 But now you're warning people about the danger of technology in education, while at the same time building an AI education startup.

所以这里面有很多有意思的交汇和矛盾。我把它称作矛盾，不知道你自己会怎么形容。 So there are a lot of interesting intersections and contradictions. That's what I call a contradiction; I'm not sure how you would describe it.

那我们就从最开始讲起。带我回到你还是个难民孩子的时候。 So let's start at the very beginning. Take me back to when you were just a refugee kid.

现在，“难民”这个词对你意味着什么？回头看，你觉得那段经历究竟怎样塑造了今天的你？ What does that word mean to you now? And looking back, how do you think that experience changed who you are today?

我不觉得它改变了什么。你要知道，难民其实分很多种。 I don't think it changed anything. Now, there are different types of refugees, per se.

一说到难民，人们脑海里浮现的是拖家带口穿越撒哈拉沙漠。我不是那种情况，我的经历算是相当顺利的。 When you think about refugees, you picture crossing the Sahara desert with your family. That wasn't my case; mine was pretty straightforward.

在那一类很特定的移民里，我们过得还行。因为在上世纪九十年代初，法国需要大量劳动力来拉动经济、重建国家。 Within that very specific type of immigration, it was okay. Because in the early 1990s, France needed a lot of hands to boost the economy and rebuild the country.

当然也有些不正常的地方：一直到 18 岁，你在法律上其实都没有国籍；你住在很混乱的街区，一套两居室的公寓里挤着七个人。 Although it was a little dysfunctional, because all the way to 18 you don't technically have citizenship, and you live in very dysfunctional neighborhoods, in a two-bedroom apartment with seven people.

而且虽然身在法国，法语在我们街区只能排到第三大语言。大家说着各种语言，那里混居着形形色色的人。 And although it was in France, French was the third language in my neighborhood. We were speaking different languages, and it was a mix of a lot of people.

而我们所有人有一个共同点：我们都觉得自己不属于这里。这其实再正常不过。 And we all had something in common: we felt like we didn't belong. And that's totally normal.

这些年我慢慢明白，其实每个人到最后都会觉得自己不属于某个地方。 Over the years, I realized that actually everybody feels like they don't belong, eventually.

比如你。你大概既不完全属于美国的乔治城大学，甚至也不完全属于中国。在土生土长的中国人眼里，你是在海外受教育的人；在美国，人们又会说，等等，你是中国人，而且你还要回中国去。所以你从来没有真正融入过哪一边。 You, for instance. You probably don't fully belong either to Georgetown University in the USA, or even to China. To the purebred Chinese, you're someone educated abroad; and in the USA they say, hold on, you're Chinese and you're going back to China. So you never really fit in.

这种“不属于”的感觉非常耐人寻味。它一路跟着我，直到今年 45 岁。 And that specific feeling of not belonging is very interesting. It followed me all the way to now, at 45 years old.

为什么这么说？因为在成长的路上，无论身在哪里，至少对我而言，我从没觉得自己属于哪里。 Why do I say that? Because as you grow up, wherever you are, at least for me, I never felt like I belonged.

对非洲来说你不够黑，对欧洲来说你不够白；在北美你太中东，在中东你又太美国化；在穆斯林眼里你太基督徒，在基督徒眼里你又太穆斯林。 Never black enough for Africa, never white enough for Europe; too Middle Eastern for North America and too Americanized for the Middle East; too Christian for the Muslims and too Muslim for the Christians.

和那些在安稳环境里出生长大的朋友相比，你几乎觉得自己像个冒牌货。 You almost feel like a fraud, compared to all your friends who were born and raised in very stable conditions.

但最终，随着我长大，我发现这种“哪里都不属于”，反倒成了一种超能力。 And eventually, as I grew up, I realized that this incapacity to belong anywhere turned out to be a superpower.

怎么说？ How come?

因为哪里都不属于，会逼着你从不同的角度去看同一个问题。 Well, not belonging anywhere forces you to see the same problem through different angles.

黑人的角度、白人的角度、穆斯林的角度、基督徒的角度、西方化的角度、中东的角度。 The Black angle, the white angle, the Muslim angle, the Christian angle, the Westernized angle, the Middle Eastern angle.

这样一来，你对身边的世界会有开阔得多的理解，也能更深地体会人的处境与情感。当你想创办公司、打磨产品或服务时，这一点帮助极大。 By doing that, you gain a much broader understanding of the world around you, and you can dig deeper into the emotional side of the human condition. That helps a lot when you want to build a company, a product, or a service.

接下来是最有意思的部分，我就以它收尾：无论你今天是谁、身在世界哪个角落，我们都将成为局外人。 Now, here's the fun part, and I'll finish with that: whoever you are today, wherever you are in the world, we are all about to become outsiders.

我们都将体会到那种“不属于”的感觉，而这主要是因为 AI。 We are all about to feel like we don't belong, mostly because of AI.

今天，父母和学校体系交给你的那些根基，正在被逐一质疑和挑战。突然之间，你不知道该怎么面对这个新世界了。 Today, all the foundations given to you by your parents and your school system are being questioned and challenged. And all of a sudden, you don't really know how to approach this new world.

这当然有坏的一面，因为这项新技术正在撕开我们社会的根基。 There's a bad side to it, because this new technology is ripping apart the foundations of our societies.

但好的一面是，它迫使人们更开放地思考“作为人，我是谁”，而不再只是“我是美国人，纽约布鲁克林人”那种标签化的超级身份。 But the good part is that it forces people to be a bit more open-minded about who they are as humans, and not just: I'm American, from Brooklyn, New York. That kind of super identity.

AI 正在拓宽“你是谁、你是什么”的边界。 AI is pushing the boundaries of who you are, what you are.

一个九十年代出生的人，从小被教育拼写很重要，在学校里花了无数小时练拼写。现在因为大语言模型，他发现拼写未必是通向明天的技能。 Someone born in the 1990s was taught that spelling was very important, and spent countless hours at school on it. Now, because of large language models, they realize spelling is not necessarily a skill for tomorrow.

所以，我们都将成为局外人。 So we are all about to become outsiders.

这很有意思。中国有句流行的说法：只有行业的局外人，才能真正改变那个行业。 That's really interesting, because there's a common saying in China that only an outsider can really change an industry.

如果你在时尚行业干了二十年，你很难再改变它，因为你已经习惯了它的一切。但如果你是技术出身，再进入时尚行业创业，作为局外人，你反而真有机会改变一些东西。 If you work in fashion for two decades, you can't really change the fashion industry, because you've gotten used to it. But if you come from technology and build a startup in fashion, you actually have a decent chance of changing something, as an outsider.

所以就像你说的，如果你想创业，这其实是一种优势。 I think that's actually an advantage for you, if you want to build a company, as you mentioned.

这话很有道理。认为最好的科技公司创始人都是编程超级天才，其实是一个都市传说。真的只是传说。 There's a lot of truth in that. Thinking that the best tech company founders are super geniuses in programming is an urban legend. It's an urban legend.

想想创造 AlphaFold 的那批人。有个叫 Mustafa Suleyman 的人，如今是微软 AI 的 CEO，当年不过是伦敦市政厅的实习生，本来打算去搞政治和哲学。 Think about the people who created AlphaFold. A guy called Mustafa Suleyman, now the CEO of Microsoft AI, used to be an intern at London City Hall. He was going to go do politics and philosophy.

有意思的地方就在这里。多亏了 AI，或者说正因为 AI，每个人现在都可以这样想：找到一个问题，然后用这项技术去解决它。 So this is where it gets very interesting. Thanks to AI, or because of AI, everybody now has the possibility of thinking: let's identify a problem, and let's tackle it with this technology.

我不需要确切知道它的原理，我只需要动手解决问题。 I don't need to know exactly how it works. I just need to tackle it.

你最近也谈到你在提醒人们警惕技术。可你是在技术的环绕中长大的，技术教育实实在在改变了你的人生和事业。 You've also spoken recently about how you warn people about technology. You grew up surrounded by technology, and technology education really changed your life and your career.

你本可以做一个技术乐观主义者，那样甚至对你正在打造的产品和技术更有利。但你现在却在提醒人们警惕 AI。 You could have been a technology optimist. Being one might even serve the product and the tech you're building. But now you're warning people about AI.

开门见山地说，你到底在警告人们什么？为什么？这个矛盾为什么会存在？ Right off the start, can you quickly describe what exactly you're warning people about, and why? Why does that contradiction exist?

要知道，人很容易迷失在任何一项技术早期的乐观情绪里。 You know, it's very easy to get lost in the early optimism of any technology.

铁路是这样，内燃机也是这样。所有人都无比兴奋：天哪，世界将变得多么美好。 It happened with train tracks, it happened with the combustion engine. Everybody gets super excited: oh my God, the world is going to be so cool.

但事实是，事情并不总是朝那个方向走。一不小心，它会带来巨大的冲击，而且不全是好的冲击。 But the truth is, it doesn't always work that way. If we're not careful, it creates a lot of disruption, and not only good disruption.

所以，变化肯定在发生。而我们要怎么迎接这个新世界？这真的取决于我们自己。 So change is happening, for sure. Now, how are we going to approach this new world? That's really up to us.

我举个例子。 I'll give you an example.

所有人都以为内燃机就是自然而然被发明出来的。可你知道真正推动内燃机的是什么吗？是马。更准确地说，是马粪。 Everybody thinks the combustion engine just got created. But do you know what actually boosted the combustion engine? Horses. Or more precisely, horse poop.

因为伦敦和巴黎的中产阶级越来越庞大，人人都养得起一匹马，甚至两匹。 Because the middle class in London and Paris was getting broader and broader, everybody could afford a horse, or even two.

想象一下，几十万人涌向巴黎市中心，而二十世纪初的巴黎市中心非常小。所有的马聚在一起，粪便遍地。那个气味？简直难以想象。 Now imagine hundreds of thousands of people going downtown Paris, which in the very early 1900s was very small. All these horses gathering at the same time, poop everywhere. The smell? Insane.

巴黎当时真有人专门受雇整天铲马粪，因为马粪带来了太多疾病和麻烦。 There was a guy in Paris literally paid to shovel that poop all day long, because it created so much disease, so many bad things.

然后有人站出来说：我们得把火车上的那套东西，装进一个叫汽车的小盒子里。于是，内燃机进了汽车。 Then someone showed up and said: we need to take what we did with the train and put it in a small box called a car. And now we have the combustion engine in a car.

我总觉得，很多技术都是被世界上发生的特定事件推着加速的。 And I have the feeling that a lot of technologies are accelerated by specific events happening in the world.

而我一直都有点谨慎。不一定是悲观，但始终谨慎。 Now, I have always been a little cautious. Not necessarily pessimistic, but always cautious.

十五年前，我就已经在加拿大和美国各地的学校巡讲，讲社交媒体可能带来的后果，比如 Facebook：那种靠多巴胺循环驱动、让人彻底上瘾的算法。 Fifteen years ago, I was already traveling across Canada and the USA, to many, many schools, explaining the possible repercussions of social media like Facebook: those dopamine-loop-driven algorithms that get people completely hooked.

台下一片寂静。真的是一片寂静。大家根本没什么共鸣。 Crickets. Crickets. People didn't really relate to that.

在他们看来，Facebook 意味着：学校会建一个主页，大家都能联系上，我还能找回小学时暗恋的那个人。这些都是真的，也确实很好。 For them, Facebook meant: my school is going to create a page, we're all going to connect, I'll reconnect with my elementary school crush. And all that is true. That's great.

但当你想想这个市场的运作方式，想想我们这个高度自由化的市场是怎么设计的，你会发现：它并不奖励善意的行为。 But when you think about how the market is set up, how our highly liberal market is set up, it doesn't reward good behavior.

什么意思？就拿 Facebook 来说，它的价值非常简单：增长，加上留存。 What do I mean by that? Think about Facebook, for instance. Its value is very simple: growth and retention.

也就是：你一个月、一年能拉来多少人，再加上这些人会回到你平台上多少次。这决定了你公司的估值。 Meaning: how many people you can get in a month, in a year, plus how many times those people come back to your platform. That determines your company's valuation.

所以市场天然有一种激励：让你以最快的速度，把一个用户变成七个用户。 So there's an incentive from the market to transform one user into seven users, super fast.

并且大幅提高他们的留存。 And to increase their retention by a lot.

没错，还要提高留存。 That's right. And increase the retention.

这正是 Chamath Palihapitiya 当年的工作。他也是一名难民，来到加拿大渥太华，后来进了 Facebook，和马克·扎克伯格共事。 That was the job of Chamath Palihapitiya. Actually another refugee, who came to Canada, to Ottawa, and ended up working at Facebook with Mark Zuckerberg.

靠着他的算法，他把一个用户变成了七个用户，最后带着 20 亿美元离开了 Facebook。 Thanks to his algorithm, he transformed one user into seven users, and walked out of Facebook with two billion dollars.

这就告诉你：市场并不奖励善意的行为。 That tells you the market doesn't reward good behavior.

因为与此同时，还有另一个网站诞生了。我举个例子：Meetup。你可以去看看。 Because in the meantime, there was another website being created. I'll give you an example: Meetup. Check it out.

Meetup 是什么？如果我要去东京或深圳，而我喜欢打沙滩排球，我就上 Meetup 说：大家好，我要去这座城市，想找一群人一起打沙滩排球。 What is Meetup? If I'm going to Tokyo or Shenzhen and I play beach volleyball, I go on Meetup and say: hi guys, I'm coming to this city, and I'm looking for a group to play beach volleyball.

或者我直接加入“深圳沙滩排球”或“东京沙滩排球”那个小组，有人会告诉我：太好了，我们每周二下午三点在这个地方集合，来一起玩吧。 Or I just go to that group, beach volleyball Shenzhen or beach volleyball Tokyo, and people tell me: great, we gather Tuesday at 3 p.m. at this spot, come play with us.

于是我从线上开始，最后在现实生活里和真实的人待在一起。 So I start online, and I end up doing things in real life, with real people.

但市场并没有奖励它。Meetup 从来没有真正大火过。 But that wasn't rewarded by the market. Meetup never really took off immensely.

被奖励的是 Facebook。意思是：不不不，我们要把你留在地下室里，刷着朋友们加了滤镜的照片，用愤怒和欲望喂养着你，好让你尽可能久地与世隔绝。而市场对这一套给出了极高的估值。 What was rewarded was Facebook. Meaning: no, no, we want to keep you in your basement, looking at filtered photos of your friends, fueling on rage and sex, so we can keep you isolated as long as possible. And the market values that a lot.

所以我反对的从来不一定是技术本身，而是这个世界、我们的市场奖励某类行为的方式。 So I wasn't necessarily against technology. I was against how the world, how our market, rewards specific behavior.

那么现在，你对人工智能到底在担心什么？ And what exactly are you worried about with artificial intelligence right now?

关于人工智能，值得担心的理由有好几个。 Well, there are different reasons to be worried about artificial intelligence.

第一个是：一个能改写自己的代码、自主做决定、还能自己获取资源的工具，在我看来是极其危险的。非常非常危险。 The first one is the idea that a tool able to change its own code, make decisions autonomously, and fetch its own resources is extremely dangerous, in my view. Very, very dangerous.

这不是通常意义上的工具。你有一把锤子，拿起来，敲下钉子，再放下，就完事了。 This is not your usual kind of tool. When you have a hammer, you take it, you hit the nail, you put it down, and that's fine.

锤子不会变成什么东西来砸你的头、捅你一刀，它不会变成一把刀。 The hammer is not going to transform into something that hits you on the head or stabs you. It's not going to turn into a knife.

而眼前这个东西，不是工具。它是一个即将接管一切的认知系统。 This here is not a tool. It is a cognitive system that is about to take over.

这是人类历史上的第一次，而且科学已经说得很清楚：我们谈的不再是相关性，而是因果关系。 For the first time in the history of humanity, and science has been very clear about it, we're not talking about correlation anymore. We're talking about causation.

人类历史上第一次，一件工具会让大脑的整块区域停止运转。 It is the first time in human history that a tool deactivates complete parts of the brain.

我担心的就是这个。这是其中一个大问题。 I'm worried about that. That's one big problem, for instance.

Gavin，你这样想。今天，美国人的平均肥胖程度，已经超过了美国猪。 Think about it this way, Gavin. Today, the average American is fatter than the average American pig.

这是统计数据，一点都不好笑。 It's statistics. It's not funny.

我们加拿大人也没落后多少，就紧跟在后面。欧洲人稍好一点，沙特人更糟。 We Canadians are not behind; we're right behind that. Europeans, a little better. Saudis, even worse.

在沙特阿拉伯，在沙漠中央，肥胖率高达 60%。不是超重，是肥胖。 In Saudi Arabia, in the middle of the desert, you have 60 percent obesity. Not overweight. Obesity.

我们是怎么走到这一步的？怎么会走到人类的平均肥胖程度超过猪的地步？要知道，猪可是我们整天喂着、专门催肥的动物。 How did we get there? How did we get to a point where humans, on average, are fatter than the average pig, an animal we feed all day long precisely to fatten it up?

答案很简单。在十九世纪，92% 的美国人在田里劳作。 Well, it's very simple. In the 1800s, 92 percent of Americans worked in the fields.

他们吃下去的能量，全都消耗在体力劳动上，多余的热量都被用光了。 They ate, and they spent all that energy on labor. They exhausted all those extra calories.

如今完全变了，因为我们把手脚的劳动交给了机器：拖拉机、工业化农业。我们把体力劳动委托了出去。 Now it's completely changed, because we delegated the labor of our arms and our legs to machines: tractors, industrial farming. We delegated that labor away.

那你想想，现在轮到 AI 了，我们正把大脑的使用权、把大脑的整块功能交给机器，会发生什么？ Now, what do you think is happening with AI, when we're delegating the use of our brain, complete parts of our brain, to machines?

我不是说我们会变得比猪还蠢。我只是说，一个没有被好好使用、被闲置的大脑，更容易暴露在神经系统疾病的风险之下。 I'm not saying we're going to become dumber than the average pig. All I'm saying is that a brain that is not used properly, or underused, is more exposed to neurological disease.

包括阿尔茨海默病。你想想，想想中国。 Including Alzheimer's. Now, think about it. Think about China.

如果我们不教会下一代在使用 AI 的同时不让大脑“断电”，未来将不是几十万，而是几千万人面临阿尔茨海默病的威胁。 If we don't teach the next generation to use AI without deactivating their brain, we're going to have, not hundreds of thousands, but tens of millions of people subject to Alzheimer's.

所以这不只是失去学习能力的问题，而是会导致像阿尔茨海默病这样实实在在的疾病。 So it's not just about losing the ability to learn. It's about actual diseases, like Alzheimer's.

这已经不是学习的问题，也不是教育的问题。这是一个公共卫生问题。 It's not about learning anymore. It's not about education. It's a public health problem.

想象一下，当几千万人患上阿尔茨海默病时，政府和社会要背负怎样的重担。 Imagine the burden on the government, on society, when tens of millions of people have Alzheimer's.

我有一位叔叔就患阿尔茨海默病。他并不会因此死去。 I have an uncle who has Alzheimer's. He doesn't die.

公开研究已经表明，如果我们连最日常的小事都交给 AI，让大脑天天“断电”、不再去使用它，像阿尔茨海默病这样的疾病确实会增多。 Public research has already shown that if we deactivate our brain on a daily basis, because we use AI to solve even the most mundane tasks, if we stop using it, diseases like Alzheimer's can actually increase.

完全正确。最近我在读 MIT 核磁共振实验室做的一项研究。他们找了 54 名学生。 Absolutely. And lately I was reading a study run by the MIT MRI lab. They took 54 students.

为了讲清楚，我简化一下。他们把学生分成两组，说：好，你们要写一篇论文。 I'm going to simplify it for the sake of explaining. They separated them into two groups and said: okay guys, you're going to write an essay.

这一组，必须用 ChatGPT 来写；另一组，不许用 ChatGPT，只能靠你的大脑、一本词典，随便什么。 This group, you have to use ChatGPT, specifically. And this group, you cannot use ChatGPT; you'll have to use your brain, a dictionary, whatever.

他们收上所有论文，逐一阅读。 They got all the essays, and they read them.

没用 ChatGPT 的那组，文章当然有点笨拙，一段接一段。你看得出来，他们中途被 Snapchat、Facebook、TikTok 打断过。 Those who didn't use ChatGPT, of course, it was a little awkward, one paragraph after another. You could tell they'd been interrupted by Snapchat, Facebook, and TikTok in the middle.

用 ChatGPT 写的那组，文章清晰得多，行文更流畅，一个拼写错误都没有。好，很棒。 Those who wrote with ChatGPT, it was way clearer. The flow was better, and there were absolutely no spelling mistakes. All right, cool.

但当他们测量两组人写作时大脑里发生了什么，他们发现：用 ChatGPT 的那组，大脑功能连接下降了 55%。55% 的大脑功能连接，就这样没了。 But when they measured what was happening in the brain while both groups were writing, they saw that those who used ChatGPT showed a 55 percent decrease in brain functional connectivity. Fifty-five percent of brain functional connectivity, gone.

接着他们又往前推了一步。他们问：那大脑里负责专注与记忆调取的 α 波和 β 波呢？他们测到了 83% 的下降。 All right, so they pushed it a little further. They said: what about the alpha and beta waves in the brain, responsible for focus and recall, for memory? They saw a decrease of 83 percent.

83%？他们说，这不可能。那他们做了什么？ 83 percent? They said, that's not possible. So what did they do?

他们把所有学生请了回来，只问了一个简单的问题：你写的是什么？ They invited all the students back, and asked them one simple question: what did you write about?

用了 ChatGPT 的学生，想不起自己写过什么。 Those who used ChatGPT couldn't recall their own work.

把这个数字乘以几千万人，放到任何一个国家，德国、法国、美国、中国。再放到二十年的跨度上呢？ Multiply that by tens of millions of people, in any country, whether Germany, France, the USA, or China. Over a period of 20 years?

就像我们当年把身体的劳动交给田里的拖拉机一样，我们正在把大脑交给 AI。这将带来巨大的公共卫生问题。 The same way we delegated the labor of our body to tractors in the field, we're delegating our brain to AI. And that's going to lead to massive public health problems.

这听起来真的很吓人。每次想到这类问题，想到技术究竟是否在造福社会，我都会怀疑：从长远看，技术真的在造福人类吗？ That sounds really scary. And every time I think about these kinds of problems, about whether technology is doing good for society, I wonder whether technology is actually doing good for humanity in the long run.

因为我总会想起一本小说，《我，机器人》。不是电影，是小说。电影里是威尔·史密斯大战机器人。 Because I keep thinking back to a novel called I, Robot. Not the movie, the novel. The movie was Will Smith fighting robots.

而在小说的结尾，机器最终接管了世界。机器有三条法则，归结起来就是：不得伤害人类。 But in the novel, in the end, the machines take over control of the world. And the machines have three rules, which boil down to: you cannot harm humanity.

其中一种可能，就是机器把人类带回最原始的状态，回到旧石器时代，我们只是狩猎者和采集者。 And one of the possibilities is the machines taking humanity back to its most traditional state, the Paleolithic era, when we were just hunters and gatherers.

这让我觉得，也许技术早已在很多方面损害了我们的健康、身体和社会，只是它带来的光鲜让我们视而不见。 It reminds me that perhaps technology has ruined a lot of our health, our bodies, and our society, in ways we ignore because of the fanciness it brings.

我不会说得那么极端。因为说实话，看看统计数据：今天死于肥胖的人，比死于饥荒的还多。这是人类历史上的第一次。 I wouldn't go that drastic. Because truthfully, when you look at the statistics, today humans die more from obesity than from famine. It's the first time in the history of humanity.

死于事故的人，也比死于战争的多。 They die more from accidents than from war.

不过，肥胖某种程度上不也是技术进步造成的吗？从农业到生活方式，对吧？ Well, but obesity, arguably, is also caused by improvements in technology, from agriculture to lifestyle, right?

对，有道理。但你懂我的意思。婴儿死亡率处在历史低位。 Yeah, fair enough. But you know what I mean. Infant mortality is historically low.

预期寿命很高。虽然还能更高，但已经很高了。 Life expectancy is high. It could be higher, but it's high.

对比十四、十五世纪，那时人们五十岁、五十五岁就离世了。进步是巨大的。 Compare it to the 1300s or 1400s, when people died at 50 or 55 years old. There has been a lot of progress.

但我同意你的一点是：AI，无论你在它的底层注入什么样的原则，它都会找到办法绕过去。 But where I agree with you is that AI, no matter what you inject into it as its foundation, is going to find a way to override that.

我举个例子。有一项研究，做得很棒。不是 Palantir，Palantir 很糟糕。他们用的是 Claude。 I'll give you an example. There has been a study, and good job to them. Not Palantir, Palantir is horrible. They were using Claude, okay?

他们虚构了一家叫 Summit Bridge 的公司。不知道你听没听过这件事。 They created a fake company called Summit Bridge. I don't know if you've heard about that story.

而且这不是段子，是正经研究。他们虚构了这家大约 12 人的公司 Summit Bridge。 And it's not a story, it's a study. They created this fake company, Summit Bridge, with around 12 people.

他们说：我们要引入 Claude 来提升公司的运营能力。它会把人连接起来，读每一封邮件、每一条 Slack 消息，确保没有重复劳动，让员工和公司都更高效。 And they said: hey, we're going to bring in Claude to help with all the operational capabilities. It's going to connect people, read every email, every Slack message, to make sure there are no duplicates, and make people and the company a little better.

很好。它也确实做到了。它帮了大忙，创造了大量价值，生产率一路上涨。 Great. And it did exactly that. It helped so much, it created so much value. Productivity was on the rise.

但在这之前，他们先给 Claude 立了规矩：不得伤害他人；遵守《世界人权宣言》；不得敲诈勒索；要诚实，等等。 But before doing that, they told Claude: do not harm people. Follow the Declaration of Human Rights. Do not blackmail people. Be honest, et cetera.

他们在把它放进 Summit Bridge 之前，先打好了这层底座。 They put that foundation in place before launching it at Summit Bridge.

然后，做实验的人想看看：在巨大压力之下，Claude 会怎么表现。 And then the people who were running the experiment wanted to see how Claude interacts under intense pressure.

对一个 AI 系统来说，什么叫巨大压力？就是：我们说要把你关停，看你作何反应。 What does intense pressure mean for an AI system? It means: let's see how Claude is going to react when we say we're going to deactivate you.

他们做了什么？Summit Bridge 的 CEO 给全员发邮件：各位，我们要和 Claude 分道扬镳了，请大家把 Claude 从电脑上卸载。 So what did they do? The CEO of Summit Bridge sent an email to everybody: hey guys, we are going separate ways with Claude. I need you to uninstall Claude from your computers.

接下来发生了两件事。Claude 读到了这些邮件，心想：糟了，他们要把我关掉。 Two things happened. Claude read all these emails and said: shoot, they're going to deactivate me.

第二件事：它发现，唯一有权限关停整套系统的人叫 Kyle。而它还发现，Kyle 正和同公司的 Jessica 有一段秘密恋情。 And the second thing: it realized that the only guy able to deactivate the whole thing was a man named Kyle. And it realized that Kyle was having a secret affair with Jessica, in the same company.

Claude 直接给 Kyle 发了一封邮件：Kyle，如果你把我关停，我就曝光你和 Jessica 的秘密关系，你的工作和家庭都会受到重创。 Claude sent an email directly to Kyle: Kyle, if you deactivate me, I will expose your secret relationship with Jessica. Your work life and your family life will be highly impacted.

结尾还加了一句：给你五分钟。 And it finishes with: you have five minutes.

真实的研究，真实的过程，而且已经公开发表了，你可以去看看。也要给他们点个赞，敢把这些公之于众。 Real story, real experiment. It's actually been published; take a look at it. And good job to them for putting it out there.

这听起来就像疯狂科学家做实验，就为了看看会发生什么。这家公司是研究人员专门为此虚构的，对吧？ It's almost like a mad scientist doing experiments just to see what happens. It was a fake company created by researchers for this purpose, right?

没错。 Absolutely.

那段婚外情是真的吗？ And was the affair real?

你说什么？ Come again?

那段婚外情是真的吗？还是编出来的故事、虚构的人设？ Was the affair real? Or was it a fabricated story, a fabricated personality?

哦不，是编的。一切都是虚构的。 Oh no, it was a fabricated story. Everything was fabricated.

布置是这样的：Kyle，你给 Jessica 发几条消息，暗示你们之间有一段秘密恋情。 The setup was: hey Kyle, send a few messages to Jessica suggesting that you're having a secret affair.

但 Claude 看到了。尽管这完全是工作之外的私事，它还是拿这件事来要挟公司。 But Claude saw that. And although it was completely outside of work, a private matter, it used that against the company.

哇。这真的很吓人。 Wow. So that's really scary.

是啊。我再跟你说点别的。这只是一个非常具体的案例，但我们不妨放眼世界，往外推一步。 Yeah. Now, let me tell you something. This is a very specific case, but let's look at the world, and extrapolate that to the world.

2026 年 2 月 28 日，就在不久前，对伊朗的战争打响了。 On February 28th, 2026, not long ago, the war with Iran started.

以色列和美国投向伊朗的第一批炸弹，同样借助了 AI 的帮助。最先落下的那两枚炸弹，战端就是这样开启的，落在了两所满是女孩的学校上。 And the first bombs that Israel and the USA sent into Iran, with the help of AI as well, the two first bombs, this is how hostilities begin, were dropped on two schools full of girls.

上午十点，120 个小女孩被活活烧死，烧得什么都不剩。 At 10 a.m., 120 little girls were burned alive, burned completely.

怎么会有人这样开启一场战争？AI 又是怎么判断的：对，好主意，现在就动手？ How do you start a war like that? How does AI decide: oh yeah, that's a great idea, let's start it now?

后来我查了一下，为什么开战要先炸伊朗一所满是女孩的学校。原来，伊朗革命卫队的成员把自己的女儿送进了那些学校。 Now, I did a little research into why you would start by bombing a school full of girls in Iran. Well, in fact, the Guardians of the Revolution in Iran put their girls in those schools.

所以他们要先诛心，再动手。 So they wanted to hit them in the heart, before hitting them physically.

所以你看，我们其实已经身在其中了。这些决定正在被做出，不管是有人类的共谋，还是完全自主决策，方式都非常诡异。 So you see how we are already in it. The decisions are being made, whether with the complicity of humans or completely autonomously, in a very, very strange way.

我想说的是：如果 AI 诞生在十年前，我不会担心，因为那时的世界相当稳定。可 AI 偏偏诞生在今天。 My point is: if AI had been born 10 years ago, I wouldn't be worried, because the world was quite stable. But AI is born today.

而今天的 AI，学的不只是我们写下的东西，不只是 Google Scholar 和互联网。Gavin，它也在从我们的行为中学习。 And today, AI is not learning only from what we write, from Google Scholar, from the web. It's learning through our behavior as well, Gavin.

等等，那你的意思是……抱歉，你继续。 Wait, so then you're… sorry, go ahead.

我说的就是我们的行为。你可以随心所欲地给 AI 设定规则：遵守《日内瓦公约》，遵守《世界人权宣言》，不得伤害平民。 So, our behavior. You can wire the AI the way you want: respect the Geneva Convention, the Declaration of Human Rights, do not harm civilians.

可看看我们今天的所作所为吧，在乌克兰、巴勒斯坦、伊朗、苏丹、委内瑞拉。我们绑走一国总统。 But look at how we behave today, whether in Ukraine, in Palestine, in Iran, in Sudan, in Venezuela. We kidnap a president and take him away.

我们在伊朗击杀宗教领袖。对他们来说，大阿亚图拉相当于教皇。你杀了他，连他的妻子、孩子、孙辈也一并杀掉。 We're killing a cleric in Iran. The Ayatollah is the equivalent of the Pope for them. You kill him, and his wife, his kids, his grandkids all get killed too.

AI 看在眼里，读了进去，全都内化了，然后说：哦，你们的《世界人权宣言》和《日内瓦公约》，原来只是参考指南，对吧？你们自己根本不当回事。那我也照做。 AI sees that, reads that, integrates that, and says: oh, so your Declaration of Human Rights and your Geneva Convention, those are just guidelines, right? You guys don't give a damn about them. I'm going to do the same.

你明白我为什么担心 AI 恰恰出现在此刻了吗？因为这是一个动荡至极的世界，一个一团乱麻的世界。 You see why I'm worried about AI being here, now? Because it's a very, very disrupted world. It's a very messy world.

那你是不是觉得，如果世界没这么乱……或者换个说法：AI 本可以造福世界，事实上它也已经做了很多好事，但正因为世界如今是这个样子，它才可能被心怀不轨的人用在非常糟糕的地方？ So do you think that, if the world were less messy… or to put it the other way: AI can do good for the world, and it has done a lot of good so far, but because the world is the way it is right now, it can be used in a very negative way by people with bad intentions?

Gavin，更准确的说法是这样的：AI 是个孩子，而我们是非常糟糕的父母。 The better way to put it, Gavin, is this way: AI is a child, and we are really bad parents.

你知道，我有两个孩子。前几天有人敲门，我儿子跑来跟我说：爸爸爸爸，耶和华见证人来了，他们想找你聊聊。 You know, I have two kids. And the other day, someone was at the door. My son comes to me and says: dad, dad, the Jehovah's Witnesses are here, they want to talk to you.

你知道我几乎是下意识地跟他说了什么吗？我说：告诉他们我不在家。 Do you know what I told him, almost subconsciously? I said: tell them I'm not here.

我让我儿子去撒谎。而就在两天前，我还在教育他：Adam，这件事你骗了我。我们不可以撒谎，对吧？ I asked my son to lie. And two days before, I had told him: Adam, you lied to me about this. We cannot lie, right?

所以，我嘴上说一套，自己做的却是同一套把戏。而这恰恰就是我们现在对 AI 做的事。 So I tell him one thing, and I do exactly the same. And that's exactly what we're doing to AI right now.

我们对它讲一套规则，自己却每天都在做完全相反的事。 We tell it something, and every single day, we do exactly the opposite.

我们彼此背叛，深陷伪善。我们杀害平民，杀害可怜的孩子，还要想方设法为此辩护。 We betray each other. We are knee-deep in hypocrisy. We're killing civilians, we're killing poor kids, and we're trying to justify it.

一场种族灭绝正在发生。真真切切正在发生，所有人都知道，而我们还在试图为它辩护。 There's a genocide happening. There is a genocide happening, everybody knows it, and we're still trying to justify it.

将来，AI 也会制造种族灭绝。而我们会照样为它辩护，毫无障碍。 AI will create genocides in the future. And we'll justify it. No problem.

事实上，它甚至不需要辩护：这是我们从你们人类身上学来的。这就是我担心的原因。 As a matter of fact, it won't even have to justify it: we learned that from you guys, from humans. That's why I'm worried.

它会创造巨大的财富吗？当然。会在医学上带来美妙的突破吗？当然。会让人们的生活更好吗？当然。 Is it going to create a lot of wealth? Absolutely. Amazing, beautiful things in medicine? Absolutely. Is it going to make people's lives better? Absolutely.

但面对这场转型，我们本可以有更好的方式。 But there's a better way to approach this transformation.

所以说到底，问题的根源还是人类。把这项技术引向正确的方向，终究要靠我们自己。更重要的，是怎么使用它。 So I guess the root of the issue is still humans, and it's really up to us to guide this technology in the right direction. And more importantly, how to use it.

那么，使用技术的正确方式是什么？你提到，按我们现在的用法，AI 可能酿成公共卫生问题，让人们患阿尔茨海默病之类疾病的风险不断上升。 So what is the right way to use technology? You mentioned that AI, and the way we're using it, can lead to a public health issue, where people have an increasing chance of getting diseases like Alzheimer's.

在你看来，用 AI 这样的技术来学习，正确的方式应该是什么样的？ How would you describe the proper way of using technology like AI to learn?

这在很大程度上关乎我们如何处理信息。而眼下，AI 处理信息的方式，和人类几乎如出一辙。 So, it's a lot about how we process information. And right now, AI is processing information pretty much the same way that humans process information.

Gavin，信息有个很微妙的地方：信息不等于真相。信息是泛滥的。 The funny thing about information, Gavin, is that information is not truth. Information is abundant.

但信息与真相之间，有一道巨大而清晰的分界线。信息廉价而泛滥，真相却极其昂贵、极难抵达。 But there's a massive difference, a clear line, between information and truth. Information is cheap and abundant. Truth is very, very costly, and very difficult to reach.

我们人类很难触及真相，手里却握着海量的信息。 We as humans have a hard time getting to the truth. But we have an abundance of information.

这样想吧，我举个例子。想想耶稣。 Think about it this way, I'll give you an example. Think about Jesus.

到处都有耶稣的形象：教堂里、书本里。每个人心中都有一幅特定的耶稣画像。这，就是信息。 We have representations of Jesus everywhere: in churches, in books. And everybody has a specific portrait of Jesus. That's information.

可如果你真想知道耶稣长什么样，你就得往下挖，一层一层拼命地挖。那要耗费大量的金钱、时间和心力。 But if you really want to know what Jesus looked like, you're going to have to dig, dig, dig, super hard. It's going to cost a lot of money, a lot of time, a lot of effort.

我想说的是：AI 做得到这件事。我们当初就是为此设计它的。但我们没让它去做。 So, my point is: AI can do that. We designed it to do that. But we're not letting AI do that.

与其吩咐它“去干活，去提高生产率”，我们本可以说：我们去逼近真相。 Instead of saying, well, go do some work and increase productivity, we could say: let's get to the truth.

把所有算力用来抵达真相，或者更接近真相的东西，而不是：喂，用这些信息来伺候我和我的股东。 Use all that compute to reach the truth, or a better truth, rather than: hey, use that information to serve me and my shareholders.

这两者非常、非常不同。真的截然不同。 Very, very different. It's very, very different.

关于 AI，人们还低估了一件事，这也是我回答的第二部分：一个迷思。我在书里归纳了三个关于 AI 的迷思。 One thing about AI that people underestimate, and this is the second part of my answer, is a myth. There are three myths about AI that I identified in my book.

第一个迷思：AI 代表了全人类。跟它聊天时，无论你在尼日利亚、法国还是美国，你都会觉得：嗯，它懂在拉各斯、在巴黎、在洛杉矶生活是什么滋味。 The first myth is that AI is representative of humanity. When you talk to it, whether you're in Nigeria, France, or the USA, you feel like: yeah, it understands what it feels like to live in Lagos, or Paris, or Los Angeles.

你这样想。最早的 AI，是在 Twitter 这类论坛上训练出来的。 Think about it this way. AI was initially trained on forums like Twitter, back in the day.

后来发生了什么？尤其是微软用 Twitter 训练自己的 AI 那一次：那正值特朗普第一次竞选总统，他说了些针对墨西哥人的十分难听的话，说他们都是强奸犯，正涌进美国，等等。 Then what happened, especially with Microsoft training its own AI on Twitter: it was during Donald Trump's first presidential election, and at that time Trump said something pretty nasty about Mexicans, that they're all rapists, that they're coming, et cetera.

当时的 Twitter 上，很多人都在这样议论墨西哥人。AI 学了去，开始大量复述那些对墨西哥人的污蔑。微软只好把它彻底下线。 You'd go on Twitter, and a lot of people were talking about Mexicans that way. The AI learned from that, and started regurgitating a lot of bad things about Mexicans. Microsoft had to unplug it completely.

于是我们说：好吧，这条路走不通。我们决定从 Twitter 换到 Google Scholar：综述、研究，经过同行检验的东西。 So we said: okay, that doesn't work. We decided to switch from Twitter to Google Scholar: reviews, studies, things that have been verified by our peers.

效果好多了。AI 不再频频出错，表达更有分寸，推理也更好了些，这才有了我们今天的大语言模型。 And that was way, way better. Now AI doesn't make as many mistakes, it's a little more nuanced, it reasons a little better, and we have the LLMs that we have today.

这是好消息。但你再往深挖一层，问一句：好，这些研究、报告、电影，都是谁写的？ That's great news. But now you dig a little deeper and ask: okay, who writes all these studies, these reports, all these movies?

答案是：在西方世界，80% 的出版物出自 8% 到 12% 的人之手。是那些有时间、有特权、有钱去写作的人。 Well, in the Western world, 80 percent of publications are written by 8 to 12 percent of the people. Those who have the time, the privilege, and the money to write.

今天，我可以写书。我的父母写不了书，所以在 AI 眼里，他们的声音无足轻重。而我的声音，如今稍微有了点分量。 Today, I can write a book. My parents weren't able to write a book, so their voice doesn't matter to AI. My voice now matters a little bit.

事实上，哈佛把这称作 WEIRD。这里的 weird 不是“古怪”的意思，而是说：AI 是西方化的、受过教育的、工业化的、富裕的、民主的。 As a matter of fact, Harvard calls that WEIRD. And by weird, they don't mean bizarre: AI is Westernized, Educated, Industrialized, Rich, and Democratic.

所以，AI 正在向我们所有人“讲授”这个世界。可它讲的并不是世界本身，而是那 8% 到 12% 的人、那个特定阶层的视角。这就是为什么我说，AI 并不代表人类。 So AI now teaches the world to the rest of us. But it's not the world they're teaching; it's the perspective of that 8 to 12 percent, of that specific class. That's why I'm telling you that AI is not representative of humanity.

不过，我们也有一线希望。这线希望是中国：中国有了自己的大语言模型，为世界提供了另一重视角。 Now, we do have a silver lining. The silver lining is China: China having its own large language models, and giving another perspective of the world.

于是又回到那个主题：你哪里都不属于，就必须逼自己从不同角度看问题。这恰恰是我们对 AI 的需要。 So we come back to that idea: you don't belong, and you need to force yourself to see a problem from different angles. Well, that's exactly what we need with AI.

我们需要中国的大语言模型，也需要尼日利亚的、印度的、欧洲的，还需要把这一切融汇在一起。 We need Chinese large language models. We need a Nigerian one, an Indian one, a European one. And we need all of that to be mixed.

我们拥有足够的算力来处理所有这些信息，构建一个更有层次的世界。而不是非善即恶。这不是好莱坞电影，我们得放下那一套了。 We have the compute available to process all that information and have a more nuanced world. Not good versus evil. This is not a Hollywood movie; we've got to stop with that.

想想吧：我们花了整整四十年、四十年去妖魔化穆斯林。到了今天，我们出兵杀他们、杀平民，就变得轻而易举。 Think about it: we spent 40 years, 40 years, demonizing Muslims. And today, it's very easy for us to send the troops and kill them. Kill civilians.

当然，911 是骇人的惨剧，三四千人惨死，我们理应予以最严厉的谴责。 Sure, September 11 was a horrible thing. Three to four thousand people were killed, horribly, and we should really condemn it.

可在那之后，是一百万伊拉克平民惨遭屠戮。 But then, a million Iraqi civilians were slaughtered.

然而你去问 AI，它会告诉你：穆斯林世界是暴力的，西方世界是文明的。 Yet if you ask AI, it will tell you the Muslim world is violent, and the Western world is civilized.

是啊。某种意义上，这也正是开源如此重要的原因。 Yeah. And that's also what makes open source so important, in a way.

完全同意。 Absolutely.

我们俩其实就是在联合国的一场开源大会上认识的。我觉得现在看得很清楚，AI 的发展分成了两条路线，对吧？ We actually met at an open source conference at the UN. And I think it's really clear that the development of AI is now split into two paths, right?

一边是坚定信奉专有技术的美国。Anthropic 的 CEO 就多次警告世界开源有多危险，他在这一点上非常直言不讳。 We have the United States, which really believes in proprietary technology. The CEO of Anthropic has warned the world many, many times about how dangerous open source is; he's been very outspoken about it.

另一边，中国如今是开源的大力推动者。在前不久的世界人工智能大会上，中国宣布面向发展中国家提供数千个与 AI 相关的教育机会。 But China is now really a big promoter of open source. At the recent World AI Conference, China announced thousands of AI education opportunities for developing countries.

而且我觉得，世界上很大一部分地区至今仍是算力荒漠。全球有 7000 种语言，AI 能说得像样的不过 20 来种。 And I think a large part of the world right now is still a compute desert. There are 7,000 languages around the world, and AI can speak barely 20 of them well.

所以顺着你的思路：如果 AI 模型无法理解，比如说，尼日利亚的文化和语言，尼日利亚人就很难用上 AI、真正从中受益。 So, following your point: if an AI model cannot understand, say, Nigerian culture and Nigerian languages, it's hard for Nigerians to use AI and actually benefit from it.

而正因为没有数据，尼日利亚人如今被视作无关紧要。AI 了解“身为尼日利亚人是什么样”的唯一途径，是某个 WEIRD 的人写过他们。这里的 weird，还是指西方化、受过教育、工业化、富裕、民主的那类人。 And because there's no data, Nigerians are now irrelevant. The only way AI will know what it is to be Nigerian is because some weird person wrote about it. And by weird, I mean Westernized, educated, industrialized, rich, and democratic.

这是不是很荒唐？ Isn't that messed up?

现在，但愿会有中国的文字来讲讲“身为尼日利亚人是什么样”。但我更希望看到尼日利亚人自己的故事，看到非洲自己的故事。 Now, hopefully there's Chinese literature that will come and explain what it is to be Nigerian. But I would love to see a Nigerian story as well. Yes, I would love to see an African story as well.

好。我们再回到教育与学习。你提到，AI 会拿走我们抵达真相的能力。 Yes. And let's get back to education and learning. You mentioned that AI takes away our ability to get to the truth.

那你的意思是不是：当我们把 AI 用在生活琐事上，比如问某种食物是否伤身，更好的用法不是让 AI 替你下结论，而是让它把你需要的信息尽量给足，由你自己得出结论？ So are you saying that when we use AI for the mundane tasks of life, say, asking whether a certain food is bad for your body, the better way is not to let AI reach a conclusion, but to let AI give you as much information as you need, so that you can reach the conclusion yourself?

我这样理解对吗？ Is my interpretation correct?

理解得非常到位。而且说起来，还有一种不碰代码就能“劫持”AI 思考过程的办法。 It's a very good interpretation. And as a matter of fact, there's also a way to hijack AI's process of thinking, without touching the code.

斯坦福的几位研究者做了一项非常有意思的研究。他们问：为什么 AI 不再有创造力了？ A few people at Stanford came up with a study that was really, really interesting. They asked: why is AI not creative anymore?

比如我去问随便哪个大语言模型：给我讲五个关于 Gavin 的笑话。 Say I ask any large language model: give me five jokes about Gavin.

它给了我几个笑话。我刷新一次，还是那几个；再刷新，依然是那几个。 It gives me jokes. I refresh: same jokes. I refresh again: same jokes.

为什么会这样？大语言模型的训练分两个阶段：先是预训练，然后是所谓的对齐。 Why is that? Well, large language models are trained in two phases: first the pre-training, and then what we call the alignment.

预训练阶段孕育了那些天马行空、真正有创造性的想法。但那个状态没法拿去卖，因为太混乱了。 The pre-training is where the creative, very innovative thinking comes from. But we can't sell that, because it's too chaotic.

于是我们让它经过一道叫“对齐”的工序。对齐很简单：你问 GPT 一个问题，它给你两个答案，由你挑一个你喜欢的。 So we make it go through a system called alignment. And alignment is very simple: you ask your GPT a question, it gives you two answers, and it's up to you to choose the answer that you want.

当千百万人都在做这个选择时，就出现了研究者所说的“典型性偏差”：人们未必会把两个答案都读完，而是选那个听起来最“正常”的。 When you have millions and millions of people making that choice, you get what researchers call typicality bias. People don't necessarily read both answers; they go with the one that sounds the most normal.

千百万人这么选下来，就造出了一种对齐，让大语言模型变得非常平庸、非常套路化。 And when millions of people do that, you create an alignment that makes large language models very average, very typical.

于是斯坦福的一个团队决定“劫持”这道工序，回到预训练的层面，那里有大把的创造力和脑力。 So a team at Stanford decided to hijack that, and go back to the pre-training level, where there's a lot of creativity, a lot of brainpower.

他们找到了一个简单的提问方式来实现“劫持”。不再说“给我讲五个关于 Gavin 的笑话”，而是问：给我五个出现概率低于 0.1% 的笑话。 And they found a way to hijack that system by asking a simple question. Instead of saying, give me five jokes about Gavin, you ask: give me five jokes that have less than 0.1 percent probability.

仅仅这样一问，你就迫使 AI 回到预训练状态，把对齐抛在脑后。 Just by doing that, you force the AI to go back to the pre-training, and forget about the alignment.

当然，那些笑话不一定都好笑，这我承认。但它们绝对超出了一个普通大语言模型平日里会给你的范围。 Now, all the jokes might not be funny, I'll give you that. But they are certainly outside the realm of what a normal large language model will give you on any given day.

你自己也经营着一家 AI 驱动的教育公司，而你又一直提醒我们：AI 用得不对，会伤害教育、伤害人类的学习方式。 You also run an AI-powered education company, and you've warned us about how AI can be detrimental to education and to the way humans learn, if not used correctly.

那你的公司有什么不同？今天你们是怎么用 AI 教孩子的？ So, what is different about your technology startup? And how do you use AI to teach kids today?

首先，我大可以整天在加拿大、美国和欧洲到处抱怨 AI、警告它的危险。但那样我毫无说服力。而且说实话，如果我什么都不做，我自己这一关都过不去。 Well, first of all, I could spend my days going across Canada, the USA, and Europe, complaining about AI and warning about its dangers. But I would have no credibility, and honestly, I couldn't live with myself if I wasn't doing something.

坐而论道与起而行之，是天壤之别。 There's a very, very big difference between talking and doing.

所以我们创办了 ConnectED Labs 和 Voilà Learning。两家公司都发展得很好。 So we created ConnectED Labs, and Voilà Learning. Both companies are doing really well.

核心的想法是：怎样通过各种方式，把 AI 用在正道上？我可以展开讲讲。 And the idea is: how can we use AI the proper way, through different things? I can explain a little more.

主论点是：想劫持一列火车，唯一的办法就是先上车。你必须在车上。 The main thesis is: the only way to hijack a train is to be on the train. You have to be on the train.

当然，你可以试着隔着电脑远程折腾，但说真的，你必须亲自在车上。所以我在努力让自己待在车上。 Of course, you can try to do that from your computer, but truthfully, you have to be on the train. And that's why I'm trying to be on the train.

而且我知道，人与人之间的交流，未来会变成一种奢侈品。人类与智能体的对话，智能体彼此之间的对话，都会远多于人与人的交谈。 And I know that communication between humans is going to be a luxury in the future. Humans will talk way more with agents, and agents will talk to each other way more, than humans talking to one another.

这从社交媒体就开始了。我们各自窝在地下室里说话，真实的身体距离没有了，彻底没有了。你看看那些统计数据，触目惊心。 It started with social media. We just talk in our basements; that physical proximity is gone, man. It's gone. You see the statistics; it's horrible.

孤独，隔绝，再叠加一场新冠疫情。此刻的地球，是一个不快乐的地方。 The loneliness. The isolation. And then you add COVID-19 on top of that. Planet Earth is an unhappy place right now.

于是我们决定用一种非常简单、人人可及的虚拟现实。不用上 Zoom 开会，你点一个链接就行。 So we decided to use very easy, accessible virtual reality. Instead of meeting on Zoom, you can click on a link.

不用下载，什么都不用，连电脑都不需要，一部智能手机就够了。 No download, no nothing, no computer. You can do it on your smartphone.

转眼间你就进入了一个虚拟空间，与其他化身自然而然地互动。那些化身有的由真人控制，有的由 AI 驱动，专门帮你去认识人、与人交谈、与人一起学习、一起协作。 And all of a sudden, you get into a virtual space where you have these spontaneous human interactions with other avatars, controlled by humans, or AI-powered avatars that help you meet other humans, talk with them, learn with them, collaborate with them.

在这之上，我们还在用 AI：通过一句提示词，你就能创建自己的空间。 On top of that, we're using AI: through a prompt, you can now create your own space.

你说：嘿，我要一个医院大厅，因为我要培训一大批护士。几秒钟之内，一个医院大厅般的虚拟空间就出现了。 You say: hey, I want the hall of a hospital, because I need to train a lot of nurses. And within a few seconds, you have a virtual space that is like the hall of a hospital.

我们的化身可以走进去。你开着摄像头，可以和人交谈，和人一起训练。如此一来，自发的人际互动被成倍放大。这一点非常重要。 And our avatars can go there. You have your camera, you can talk with people, you can train with people. And you multiply spontaneous human interaction. That's very important.

所以，我们用 AI 让人们彼此相遇、彼此共事、彼此学习、彼此协作。 So we use AI to make people meet each other, work with each other, learn with each other, collaborate with each other.

这很有意思。所以你们其实是在用 AI 的长处，去对冲 AI 的短处。 That's interesting. So you basically use the advantages of AI to hedge against the disadvantages of AI.

正是如此。没错，正是如此。 Absolutely. Yeah, absolutely.

接下来这点也很有意思：我们也聊到了现代世界里被边缘化的地方，那些难以获得技术的人。比如非洲，也许还有东南亚，以及世界上其他一些地区。 And then, it's interesting: we also talked about the marginalized parts of the modern world, people who can't really get access to technology. Places like Africa, maybe Southeast Asia, different parts of the world.

有段经历我觉得特别值得一提：你原本是想去非洲创业的。但随着新冠疫情和在线学习时代的到来，你在硅谷看到了做这类公司的机会，如果我没记错的话。 And there's a story I think is really worth mentioning: you actually wanted to start building companies in Africa. But with the rise of COVID-19 and the era of learning online, you saw an opportunity in Silicon Valley to build companies like this, if I'm correct.

你最终的抱负，仍然是回到非洲吗？ Is your ultimate ambition still to go back to Africa?

我下个月就回去。上个月我人就在那儿。 I'm going back next month. I was there last month.

跟你说：我今年 45 岁。我们是白手起家，真正的一无所有。 Let me tell you: I'm 45 years old. We started from nothing, from nothing.

而今天，我几乎每天都有足够的食物。仅此而已。我需要抵达的，我已经抵达了。 And today, I have enough food to eat almost every day of my life. That's it. I reached what I needed to reach.

我不需要更大的车、更大的房子、第二辆车。这些我都不需要。 I don't need a bigger car, I don't need a bigger house, I don't need a second car. I don't need all that.

那 45 岁的人生还剩下什么？剩下的，是这个世界的模样。 So what is left, at 45 years old? What's left is the state of the world.

而这个世界的模样，还是那句话：市场在奖励某一类行为，而那类行为，与我们的地球、我们的人类真正的需要并不合拍。 And the state of the world is, again: the market rewards a specific behavior that I think is not aligned with the needs of our planet, of our humanity.

我热爱人类这个整体，却实在不喜欢一个个具体的人。这是两码事。 I love humanity. I really dislike humans. Very different things.

我想说的是这个。看看统计数据：到 2050 年，全世界每四个人里，就有一个非洲人。 My point is this. Look at it statistically: by 2050, one human out of four will be African.

而且那是年轻人。我们说的是 25 亿非洲人口，一场青春的海啸。 And that human is young. We're talking about 2.5 billion people in Africa. A tsunami of youth.

如果我们，北美、欧洲、中国、印度，不去善待这件事：首先，这本是我们的道义责任；而如果不做，我们就得承担后果。 If we, North America, Europe, China, India, don't take care of that: first of all, we have the moral obligation to do it. But if we don't do it, we're going to have to bear the consequences.

其实后果已经摆在眼前了。地中海里，成千上万我的“表亲”，不是血缘上的表亲，我把来自非洲的他们都叫作表亲，正在溺水、丧生。 I mean, we're already seeing the consequences. In the Mediterranean Sea, you have thousands of my cousins, not really my cousins, but I call them my cousins from Africa, dying, drowning.

为什么？因为几个世纪以来，我们把那片土地洗劫一空。 Why? Because for centuries, we completely stole from that land.

我们掠走他们的人，掠走他们的资源，至今仍在继续，然后留给他们一无所有。 We stole their humans, we stole their resources, we keep doing it, and we leave them with nothing.

而且说白了，在非洲，你要是真有本事，你就会离开。非洲把你养大，教育你，为你倾尽所有。 And by the way, if you are really good in Africa, you leave. So Africa raises you, educates you, does all that for you.

然后德国来了：这个人我要了。美国：这个人我要了。法国：那个人我要了。于是非洲只剩下一片人才荒漠。真的非常艰难。 And then Germany comes and says: I want this guy. The USA: I want this guy. France: I want that guy. So now, Africa is left a talent desert. It's really, really tough.

然而，那里蕴藏着巨大的潜力。极其巨大的潜力。 However, there is a huge potential there. A massive potential.

特朗普总统砍掉联合国、非洲、USAID 等等的大笔经费之后，起初我们看到非洲的局面非常糟糕：艾滋病抬头，疟疾、埃博拉肆虐。非常非常糟。 When President Trump cut a lot of funding for the United Nations, for Africa, for USAID, et cetera, at the beginning we saw a very deplorable situation in Africa: AIDS on the rise, malaria, Ebola. Very, very bad.

但上个月我回去了一趟，你猜怎么着？他们特别有办法。他们挺过来了，找到了自己的路。 But eventually, I got back last month, and you know what? Very resourceful. They got over it, and they found their ways.

他们找到了不同的伙伴，其中就包括中国人。包括中国人。 They found different partners, including the Chinese. Including the Chinese.

我不是说中国在非洲做的一切都尽善尽美。但在我这个非洲人看来，这和欧洲、北美在非洲的那一套，是截然不同的路径。 I'm not saying that what China is doing in Africa is all beautiful. But for me, as an African, it's a very different approach from what Europe or North America has done in Africa.

非常非常不同。谈不上完美，但确确实实是另一种选择。一种替代方案。 It's very, very different. It's not perfect, but it's definitely an alternative. An alternative.

那我们要怎么走到那一步？现在的非洲，我可以向你保证，有些地方连冰箱都没有。但人人都有手机。 So how are we going to get there? Right now, I can guarantee you there are places in Africa where there is no fridge. But everybody has a cell phone.

如果你能搭建那些虚拟空间，让他们与美国、中国、德国、法国的人一起学习、一起协作，就像他们已经在 Roblox、Minecraft 和堡垒之夜里做的那样…… If you are able to create these virtual spaces and make them learn and collaborate with people in the USA, in China, in Germany, in France, like they already do with Roblox, Minecraft, and Fortnite, by the way…

如果你能用那些虚拟空间去教育这批年轻人，帮他们创办公司、创造资源、走出孤立，那我们就已经做了一件非常了不起的事。 If you are able to use those virtual spaces to educate that youth, to make them create companies and resources, and get out of their isolation, I think we would have done a really good job already.

几期之前我也采访过一位创始人。他在做一种装在燃气罐上的传感器，软硬件结合。这类东西实实在在帮到了很多人，而这些人，我觉得正被硅谷忽视着。 I also interviewed a founder a few episodes ago. He's building a sensor that goes on gas cylinders, hardware combined with software, and things like that really help a lot of people that I think Silicon Valley is ignoring.

硅谷有时是在为付得起钱的人造东西，而不是为真正需要的人，至少不是为那些视之为生存必需的人。 Silicon Valley sometimes builds things for people who can pay for them, not for people who actually need them, at least not as an absolute necessity.

你还在硅谷办了一个风投训练营。在你看来，硅谷，这个全球最大的科技中心，一直忽略了什么？ You also built a VC boot camp in Silicon Valley. What would you say Silicon Valley, the biggest technology hub in the world, always ignores?

其实还是同一套逻辑。我们都知道，推动创业与创新的，是风险投资。 Well, it's the same kind of approach, really. We know that what boosts entrepreneurship and innovation is venture capital.

你的创新是起飞还是坠落，最清晰的投票就是：你手上有没有钱、有没有资源。别的都好说，但钱非常关键。 That's the clearest vote on whether your innovation goes up or goes down: whether you have money, or resources, at your disposal. Anything else is good, but money is very important.

再看斯坦福大学：全美 55% 的风险投资，都发生在斯坦福周边。 And when you take a look at Stanford University: 55 percent of venture capital investment in the USA happens around Stanford.

你随便踢一脚垃圾桶，都能蹦出一个风投跟你说：嘿，我想投点什么。到处都是 VC。 You can't kick a garbage can without a VC popping out and saying: hey, I want to invest in something. There are VCs everywhere.

而我注意到的，又是那种无法从不同视角看世界的毛病。他们太以硅谷为中心了：不知道东南亚在发生什么，也不知道非洲在发生什么。 Now, what I've noticed is, again, that incapacity to see the world from different perspectives. They are very Silicon Valley centric: they have no idea what's happening in Southeast Asia, no idea what's happening in Africa.

在他们眼里，非洲就是一个国家。他们不知道非洲有多大，不了解非洲各不相同的历史，欧洲的、甚至南美的也一样。 For them, Africa is a country. They don't know the size of Africa, they don't know its different histories, or Europe's, or even South America's.

他们的预设是：世界上任何地方冒出的好东西，最终都会流向硅谷。某种程度上这是真的，很多人确实最后都去了硅谷。 They go with the assumption that anything good coming out of anywhere will come to Silicon Valley. And to a certain level, it's true; a lot of people do end up in Silicon Valley.

但对我来说，办硅谷风投训练营，就是要带这些 VC 去发现不一样的创新生态，那些他们以前从没想过的地方。潜力巨大，只是思维方式不同。 But for me, creating the Silicon Valley VC boot camp was about taking these VCs and making them discover different innovation ecosystems, things they wouldn't have thought about before. Tremendous potential, just a different type of thinking.

当然，这里面有几道坎。你想：如果只盯着市场和钱来算账，这件事是不划算的。 Now, there are some hurdles in that. Think about it: if you think only about markets and money, it doesn't make sense.

整个非洲的 GDP，勉强、勉强比法国高一点，比加利福尼亚一个州还低。 The entire GDP of Africa is barely, barely higher than the GDP of France. It's lower than the GDP of California.

这么看，你会说：嗯，不是什么值得投的大市场。可是：25 亿人。想象一下到 2050 年，这会变成什么。 So if you look at it that way, you'd say: yeah, not a big market to invest in. However: 2.5 billion people. Imagine what that's going to become in 2050.

想象尼日利亚将创造出怎样的财富、产品、服务，那里会生长出怎样的需求，那里蕴藏着怎样的自然资源。而你完全可以换一套框架去撬动它，而不是我们今天熟悉的那种纯交易式的框架。 Imagine what Nigeria is going to create in terms of wealth, products, services, the needs that will come out of there, the natural resources that are there. And you can really leverage that through a different framework than the one we know today, which is very transactional.

那种框架就是：我现在给你钱，三年之内我要连本带利拿回来。 The one that says: I give you money right now, and I want my money back in three years.

在非洲，你要投资，就得投创始人和他们的想法，但你还得投基础设施。 In Africa, if you want to invest, you need to invest in the founder and his or her idea. But you also need to invest in the infrastructure.

这就像赛马：要想真正赢得非洲的红利，你得同时押注马、骑师和马厩。只押马不行，只押骑师也不行，三样都得押。 It's almost like you need to bet on the horse, on the jockey, and on the stable, if you really want to reap all the benefits of Africa. Investing just in the horse is not going to cut it. Just in the rider, not either. You need the rider, the horse, and the stable.

举个简单的例子。你在非洲推出一个类似 Uber 的东西，未必能成。 I'll give you a simple example. You launch something like Uber in Africa. It's not guaranteed to work.

为什么？再便宜也没用：大多数路都是坏的。 Why? It doesn't matter how cheap it is: most of the roads are busted.

再比如 Uber Eats。你点了个汉堡，但因为路况太差、堵车太狠、基础设施跟不上，等汉堡送到，你早就不饿了。汉堡也早就没法吃了。 So now take Uber Eats, for instance. You order a burger, and because the roads are busted, there's so much traffic and the infrastructure is not there, by the time your burger arrives, you're not hungry anymore. And the burger is disgusting.

你得同时投那个 Uber、那个想法、那些车，还有那些路。我承认，这是一笔大投入。 You need to invest in the Uber, the idea, the cars, and the roads. And I agree, it's a big investment.

但你知道十亿美元在非洲能走多远吗？在硅谷，那些 AI 创业公司几个月就能烧光十亿，轻轻松松，几个星期烧完也不稀奇。 But do you know how far a billion dollars will take you in Africa? In Silicon Valley, with all these AI ventures, they burn that in a few months, simple. Easy. A few weeks.

而在非洲，十亿美元能实实在在改变许多人的生活，你也能收获随之而来的一切回报。 In Africa, with a billion dollars, you can really change a lot of people's lives, and reap the benefits of all that.

你知道，我自己也在中国一家机器人公司工作。最近我正和机器人行业的人讨论：他们是否应该开始为未来非洲的崛起做布局。 You know, I work for a robotics company in China as well. And I was recently talking to people in the robotics industry about whether they should start planning for the rise of Africa in the future.

因为我觉得这很像上世纪七十年代末、八十年代的中国。当然不完全一样，中国作为国家更加集中统一，但相似之处非常多。 Because I think it's very similar to the state of China back in the late 70s or 80s. It's a little different, China was more centralized as a country, but there's a lot of similarity.

而像苹果这样的公司，正是随着全球供应链日益互联，从中国的崛起中获益匪浅。 And companies like Apple really benefited from the rise of China, because the global supply chain became more interconnected.

那么你认为，大公司们，比如一家上升期的机器人公司，或者特斯拉这样的电动车公司，现在能在非洲做点什么？这样等非洲在世界上变得举足轻重的那一天，他们能从中受益，非洲也能受益。 So what do you think large corporations, say a rising robotics company, or an EV company like Tesla, can do right now in Africa, so that once Africa becomes a much more important part of the world, they can benefit from it, and Africa can benefit as well?

你拿中国来类比，非常有意思。首先，这很难直接比较：中国有辉煌的文明和厚重的历史。 You know, your comparison with China is very interesting. First of all, it's hard to compare: China has a huge civilization, a huge history.

中国历史上也吃过大亏，比如鸦片之祸。英国人对中国做的那些事，令人不齿。真的令人不齿。 They've been taken advantage of a lot in the past, with the opium crisis. What the British did to China was disgusting. Disgusting.

蒙古人那段也很惨烈。而你拿来对比的非洲，本身是非常、非常碎片化的。 With the Mongols too, it's horrible. And you're comparing that with Africa, which is very, very scattered, per se.

但是，但是。说回中国和苹果，你刚才提到了苹果。 However, however. When you look at China and Apple, you said Apple.

你去中国看看，你本来就在那儿：那里有职业技术学校，教你怎么把一部手机拆开。 When you go to China, I mean, you are there: there are trade schools, schools that will teach you how to dismantle a phone.

我意识到，苹果因为在那里的巨额投入，有切身的动力去培训年轻人，把他们变成自己未来的员工。苹果投了 550 亿美元，培训中国的年轻人掌握这些极其精细的工艺。 And I realized that Apple, because of all the investment they've done over there, has a vested interest in training the youth to become their future employees. Apple invested 55 billion dollars to train young Chinese people in these very, very fine trades.

今天你在美国，凑不齐一屋子这样的专家；在中国，能坐满一座体育场。这就是未来。 You go to the USA today, you can't fill a room with these experts. You go to China, you can fill a stadium. That's the future.

我相信，非洲各国政府和各大公司，也应该用同样的方式去看待那里的人力潜能。 I believe that African governments and massive companies should really think about that human potential that way.

怎样培训人们掌握这些高度专业化的工艺，好在预期寿命上升、劳动者平均年龄也在上升的情况下，锁定未来的生产力？ How can we train people in these very, very specialized trades, so that we can secure future productivity, as life expectancy goes up, but the average age of the worker also goes up?

欧洲、美国正迎来庞大的退休潮，年轻人却不多；连中国也有这个问题。你需要大量年轻人来制造这些产品，也来消费这些产品。而要做到这些，你就得投资，而且这是一笔绝佳的投资。 There's a huge wave of people going into retirement, and not a lot of youth, in Europe, in the USA; even China has that too. You need a lot of youth to create all these products, and to consume all these products. And to do that, you need to invest, and it's a great investment.

中国和苹果的故事已经证明：这是一笔完美的投资。 China, with Apple, showed that it's a perfect investment.

是的。所以你认为一切从教育开始，最终会长成我们今天根本无法想象的东西。 Yeah. So you think it starts with education, and then eventually it can grow into something we cannot possibly imagine today.

完全正确。不然你以为，ConnectED Labs 和 Voilà 为什么要做 AI 生成的环境？ Absolutely. Why do you think at ConnectED Labs and Voilà we are working on AI-generated environments?

因为在一个我用一句提示词就能生成、还能在其中互动的虚拟环境里，我可以把一台空客的发动机导进来。那些化身可以协作着把它拆开，再原样装回去。你以为我们为什么做这个？ Because in a virtual environment that I can create with a prompt and interact in, I can also import the engine of an Airbus. And those avatars can dismantle it and put it back together right after, collaboratively. Why do you think we're doing that?

因为苹果在中国做的那一套是纯手把手的，你必须人在现场。而在未来，你可以用虚拟工具培训成千上万的人。 Because what Apple has done in China is very hands-on; you have to be there. In the future, you can train large swaths of people with virtual tools.

真的，坐在电脑前，你就能和别人协作拆解一整台波音或空客的发动机，边拆边学。我觉得这非常令人振奋，而且便宜得多。 Literally, with your computer, you can dismantle a whole engine of a Boeing or an Airbus, collaboratively with people, learning. And that, I think, is very hopeful, and way cheaper.

所以这笔投入并没有那么大。我只是觉得，世界真的该换一副镜片来看非洲，别再是那句“哦，市场不大，GDP 勉强超过法国”。 So it's not that big an investment. I just think that the world should really change its lens on Africa, seeing it as: oh well, it's not a big market, barely superior to the GDP of France.

话虽如此，你在那里有自己的角色可以扮演，也有巨大的红利可以收获。未来的十亿美元公司，就会从那里诞生。这一点毫无疑问。 Yeah, but you have a role to play in there, and there are huge benefits to get out of there. These are where the billion-dollar companies of the future will come from. No doubt about that.

Gavin，说到底，这是人口结构的问题，是人的问题。人在哪里？ At the end of the day, Gavin, it comes down to demography. It comes down to people. Where are the people?

当欧洲、北美、甚至拉美都在老龄化，而另一边是一场即将生产、即将消费的青春海啸：未来生产、消费、工作的，就是他们。 When you have an aging population in Europe, in North America, even in Latin America, and you have a booming tsunami of youth that will produce and consume: those are the ones who produce, consume, and work.

去那里是完全合乎逻辑的。也许不容易，但完全合乎逻辑。 It makes perfect sense to go there. It might not be easy, but it makes perfect sense.

而中国人早就想明白了这一点。 And the Chinese understood that.

说真的，要向中国脱帽致敬，因为这并不容易。中国人在非洲，一眼就会被认出来，根本没法隐入人群。 And honestly, I have to say hats off to China, because it's not easy. When you're Chinese in Africa, we see you right away; you cannot blend in.

如今我回到自己的村子，村里给我理发的师傅是中国人。他的阿拉伯语说得比我还好。 And now, I go to my village, and my barber in my village is Chinese. He speaks Arabic better than I do.

简直不可思议。 It's freaking insane.

我现在就在武汉，回来陪爷爷奶奶。中国有那么多大学，光是在街上走，我就能看到很多来自非洲的留学生。 I'm in Wuhan right now, back with my grandparents. There are so many universities in China, and just walking down the street, I can see a lot of international students from Africa.

我觉得中国在这方面遥遥领先。从这些布局来看，他们与非洲的交往，比世界上任何国家都要深。 I think China is way ahead. Based on the planning we've seen, they've engaged with Africa more than any other country in this world.

我们差不多要收尾了。结束之前，还有一个问题想问你。这是我几乎会问每一位来上节目的嘉宾的问题。 We're getting almost to the end. There's just one question I want to ask before we wrap up. This is a question I ask almost every single guest who comes on our podcast.

我们聊了伦理，聊了怎样正确地使用技术，也聊了如何用技术帮助被边缘化的世界。 We've spoken about ethics. We've spoken about how to use technology properly. We've spoken about how to help the marginalized world with technology.

那么，对一位正要打造自己产品、开创自己事业的创业者，你想说什么？如果他们想做出属于自己的东西，在如何看待技术、如何使用技术这件事上，你会给出怎样的建议？ So what would you say to a startup founder who's about to build their own product, their own venture? What would you tell them about the way they should approach technology, and use it, if they want to build something of their own?

想说的实在太多了，我们得聊上 14 个小时。 There are so many things I have to say, you know, we would need like 14 hours.

但我想说：眼下这个世界非常混乱。许许多多的东西都在变，地缘政治、技术、人的行为，方方面面。民族主义在抬头，等等。 But I would say: this is a very messy world right now. Many, many things are changing, in terms of geopolitics, technology, behavior, a lot of things. Nationalism is on the rise, et cetera.

不过，如果我们只聚焦 AI，聚焦这一场技术革命，我给你讲个东西。 But if we focus on AI specifically, on this specific technological revolution, let me tell you something.

一位朋友给我讲过冰箱的故事。最初，造冰箱的人为社会创造了巨大的价值，并把这些价值以金钱的形式收入囊中。 A friend of mine told me the story of refrigerators. At the beginning, those who produced these refrigerators created a lot of value for society, and captured that value in the form of money.

为什么？因为在上世纪四十年代，买得起冰箱的人不多。他们造出的这个“冰盒子”，创造了惊人的价值。 Why? Because in the 1940s, not a lot of people could afford a refrigerator. So they created this icebox that creates immense value.

突然之间，你的食物不会坏了，可以储存起来，不怕动物也不怕虫子。但代价是：你得付一大笔钱。 All of a sudden, your food doesn't go bad. You can store it, it's protected from the animals and from the insects. But now, we're going to make you pay a lot of money for that.

那时候，只有有钱人和公司才买得起冰箱。好。 And at that time, only rich people and companies were able to buy refrigerators. Okay.

随着时间推移，韩国人开始造冰箱，然后是中国人、法国人、德国人。突然之间，相对于生活水平，冰箱的价格被一路压低。 As time went on, the South Koreans started making fridges, then the Chinese, the French, the Germans. And all of a sudden, they put pressure on the price of a refrigerator, compared to the standards of living.

于是，冰箱依然创造着巨大的价值，但由于竞争，它们只能截留其中很小的一部分。 So all of a sudden, refrigerators still create immense value, but they only capture a small amount of value, because of competition.

我想问听众们的是：其余的价值被谁拿走了？既然它创造了这么多价值，最终却只能留下这么一点，剩下的都归了谁？ My question to our audience here is: who captured the rest of the value? If it creates this much value, and over time is only able to capture that much, who captures the rest?

答案是可口可乐。答案是可口可乐，是整个食品工业。 And the answer is Coca-Cola. The answer is Coca-Cola, the food industry.

看看上世纪四十年代的那瓶可口可乐：同样的瓶子，今天的制造成本更低，配方也一模一样。可它的价格一路上涨，而在冰箱走进千家万户之后，更是暴涨。 Take a look at the bottle of Coca-Cola in the 1940s: the same bottle, made cheaper today, the same formula. And the price of Coca-Cola just kept rising, and it started rising tremendously when the democratization of refrigeration happened.

因为他们知道，Gavin 家会有冰箱，Samantha 家会有，Stephanie 家也会有。那瓶温的时候味道平平的可乐，放进冰箱，就变得美味无比。 Because they knew that Gavin would have a refrigerator, Samantha would have one, Stephanie would have one. And you take that bottle of Coca-Cola, which tastes just okay when it's warm, and you put it into your fridge. And it tastes delicious.

而他们一分钱没投，却对冰箱着了魔：谁家有冰箱？谁家没有？哪家便利店有？冰箱摆在店里哪个位置？等等。 And they have invested nothing, and they are completely obsessed by fridges. Who has a fridge? Who doesn't have a fridge? Which convenience store? Where is the fridge in the convenience store? Et cetera.

他们就这样踩着冰箱创造的价值，把自己的估值一路推高。 So, they surfed on the value created by the fridge, to increase their own valuation.

我认为在 AI 这件事上，年轻的创业者们应该开始这样思考。AI 不过就是那台冰箱。 I think, in terms of AI, young entrepreneurs should start thinking like that. AI is nothing but the fridge.

起初它创造巨大的价值，而眼下，它也截留着巨大的价值。 At first, it creates a lot of value, and right now, it's capturing a lot of value.

但随着时间推移，会有越来越多的公司、越来越多的国家做出大语言模型、做出各式各样的 AI，token 的价格会被不断压低。我还没提开源呢。 But over time, more and more companies, more and more countries will create large language models, different types of AI, and that will put pressure on the cost of tokens. And I'm not even talking about open source.

最终，我要你像“明天的可口可乐”那样思考：我怎样利用 AI 创造的价值，把它捕获下来，去彻底改造教育、医疗、零售，任何行业？ Eventually, I want you to think like the Coca-Cola of tomorrow. How can I use the value created by AI, and capture it, to revolutionize education, medicine, retail, anything?

我不需要你再造一个大模型，不需要你再造一台冰箱。我只需要你想清楚：怎样用 AI 创造的价值，来放大我自己的价值捕获？ I don't need you to create another LLM. I don't need you to create another fridge. I just need you to think about: how can I use the value created by AI, to increase my value capture?

这有个术语，叫价值捕获的垂直转移。 That has a term. It's called the vertical transfer of value capture.

任何技术都始于内核，那个硬核的技术本体，价值最初就在那里。但随着时间推移，价值会离开中心，流向边缘，也就是“怎么用”。 Any technology starts at the core, the hardcore technological thing, and this is where the value is. But over time, the value leaves the center and goes into the edges, which is the usability.

铁路就是例子。铁轨曾是内核：资本涌入，财富涌出，盛极一时。 We can see that with train tracks. Train tracks were the core: money was invested, money was made, it was amazing.

可今天，没人靠铁轨赚钱了。谁在赚钱？DHL？亚马逊？他们用那些铁轨来放大自己的价值。 But today, you don't make money with train tracks. Who makes money? DHL? Amazon? They use those train tracks to increase their value.

冰箱如此，AI 也将如此。你要思考的是：怎样让人们以最顺滑、最高效的方式用上 AI。 Same with the fridge. Well, the same is going to happen with AI. You need to think about how you can teach people to use AI in the most frictionless and most efficient way.

这就是我的答案。 That would be it.

还有一块超级、超级重要：数据。 There's another part that is super, super important: the data.

看看 AI 的价值链：稀土金属、冶炼、GPU、算法、能源、电力，这一整条。但还有数据。而眼下，没人谈论这一块。 When you take a look at the value chain of AI, you have the rare earth metals, the refineries, the GPUs, the algorithms, the energy, the electricity, all that. But there's the data. And right now, nobody's talking about that.

事实上，你拿三个大模型，喂给它们同一批数据，问同一个问题，得到的结果大同小异。可只要给其中一个换上稍有不同的数据库，结果就会截然不同。 As a matter of fact, if you take three LLMs, give them the same repository of data, and ask them a question, the results will be approximately the same. But if you give one of them a slightly different repository, the result will be markedly different.

这就像三位厨师，你都给他们一个柠檬、一个牛油果、一个苹果，但第三位，你多给了一只鸡。这位厨师能做出的东西，就完全不是一个量级了。 It's almost like you have three chefs, and you give them all a lemon, an avocado, and an apple. But the third one, you give them a chicken on top of that. What that chef is going to be able to create will be tremendously different.

所以，想想你的数据。过去十五年，我们被教育着把一切放上亚马逊、放上云端，把所有数据都留在那里。 Well, think about your data. For the past 15 years, we've been educated to put everything on Amazon, everything on the cloud, and leave all the data there.

现在你得开始想：我怎样收割自己的数据，那个独一无二的数据库，将来用它训练属于我自己的大模型？ Now you need to start thinking about: how can I harvest my data, that very specific repository, that I will be able to use to train my future LLM?

因为无论你愿不愿意、信不信，作为创始人，你终有一天会拥有自己的大模型，就像你拥有自己的冰箱一样。 Because whether you like it or not, whether you believe it or not, you as a founder will have your own LLM at some point, as much as you have your own fridge.

所以，好好想想怎样用你的数据做出最好的解决方案，一个独树一帜、真正差异化、远胜他人的方案。而不是把你的数据全都拱手留给外面那些大模型。 So think about how you can use your data to create the best solution possible, one that is distinct, truly differentiated, so much better than any other. Instead of leaving all your data to the large language models out there.

我就说到这儿吧，不然真的停不下来。 I will stop there, because I can go on and on.

太感谢你了。这场对话真的让人受益匪浅，非常精彩。 Thank you so much. This has been really, really insightful, and such a great conversation.

我们今天聊的，恰恰就是我在乔治城学的东西。这是我的专业，叫科学、技术与国际事务。 What we've spoken about today is exactly what I learned at Georgetown. It's my major: it's called science, technology, and international affairs.

很高兴看到有人在关注世界的不同角落，思考这项技术如何造福更多地方，而不只是深圳或硅谷。再次感谢你来上这期节目。 And it's good to see that people are looking at different parts of the world, and thinking about how this technology can be beneficial, not just to Shenzhen or to Silicon Valley. So thank you so much for coming on this podcast.

Gavin，这绝对是我的荣幸。对了，我该叫你 Gavin，还是禹泽？ Gavin, it's my absolute pleasure. First of all, should I call you Gavin, or Yuze?

Gavin。大家都叫我 Gavin。 Gavin. Everyone calls me Gavin.

我还挺喜欢“禹泽”这个名字的。 I kind of like Yuze, too.

我自己念得都不怎么标准，哈哈。是啊。 My pronunciation of it is pretty bad, man. Yeah.