吴恩达：AI 恐慌被过度渲染

欢迎来到 ETAI 播客的超级特别版。今天我们请到了一位来自硅谷的男士，一位几乎一生都在研究人工智能的教授。他对 AI 的了解可不是一星半点，而是非常深入。两个字母，NG。今天我们请到了“神经向导”。他曾与 Sam Altman 共事，也曾与 Amodei 共事。未来他或许还能让他们一起和我们聊聊。但今天我们将向他学习如何应对 AI，如何倾听 AI、与 AI 对话，即使 AI 不想听你的。今天我们请到了 Andrew Ng。Andrew，如果我可以称你为“神经向导”，因为你正坐在突触之间，解码智能，并向世界展示使用 AI 的正确且最恰当的方式。非常感谢你，Andrew，在印度的清晨抽出时间。

Welcome to the super special edition of the ETAI podcast. Today we have a man from Silicon Valley, a professor who's worked on artificial intelligence almost all his life. He knows not just a thing or two, but a great deal about AI. Two letters, NG. We have the Neural Guide with us today. He's worked with Sam Altman. He's worked with Amodei. He can possibly get them together to chat with us also in future. But today we'll learn from him what to do with AI. How to listen and talk to AI even if AI doesn't want to listen to you. We have with us today Andrew Ng. Andrew, if I can say Neural Guide, because you're sitting amid synapses, decoding intelligence, and telling the world the right and the most appropriate way to use AI. Thanks a lot, Andrew, for taking time out on an early morning in India.

嗯，很高兴来到这里，也很兴奋来到印度，因为我认为这个国家在应用 AI 方面的潜力巨大。所以我渴望找到一种方式，尽可能帮助印度这里的人和机构。

Well, it's great to be here, and I'm excited to be here in India, because I think the nation's potential for applying AI is so great. So, I'm eager to find a way to help as many people and institutions here in India as I can.

是的，AI 代表人工智能，但如果我能为这个国家扩展这个缩写，我们已准备好将其应用到全印度。让我从第一个问题开始。你说过 AI 很神奇，但它也有局限性。请给我讲讲这两方面。它好在哪里，又限制在哪里？

Yes, so AI stands for artificial intelligence, but if I can extend that acronym for the nation, we're ready to apply it all over India. Let me start with the first question. You've said AI is amazing, but it also has its limitations. Give me a sense on both. What's great about it and what's so limiting?

AI 的一个棘手之处在于智能是锯齿状的。这意味着有些任务，比如解决数学问题，它非常擅长，尽管对人来说很难。但也有一些稍微不同的任务，比如你问某人什么问题，或者你如何思考商业策略，这些对人来说似乎并不难，但 AI 可能非常糟糕。AI 擅长什么和不擅长什么的一个分界线是，AI 非常擅长所谓的可验证任务，即具有客观正确或错误答案的任务，因为我们有很好的技术让 AI 在这些任务上表现良好。但有很多任务被称为判断任务或未经验证的任务，没有唯一正确答案，虽然 AI 正在进步，但进展慢得多。这与工作问题有关。我认为 AI 非常擅长客观可验证的任务，但仍有许多判断任务需要我们人类去做，这就是为什么我认为工作末日论也被过度炒作了。

So, one of the tricky things about AI is that intelligence is jagged. And what that means is that there's some tasks, like solving math problems, that it's amazing at, even though it's very hard for people. But there could be some slightly different tasks, like what questions do you ask someone to an or how do you think about business strategy, that doesn't seem as hard to people, but the AI could be really terrible at. One dividing line between what AI is good at and what is not at is turns out AI is very good at what they call verifiable tasks, meaning tasks with an objective right or wrong answers, because we have very good technology for making AI do well on those tasks. But there are a lot of tasks that are called judgment tasks or unverified tasks with no one right answer, and while AI is making progress, progress is much slower. This relates to the job question. I think AI is very good at the objective verifiable task, but there are still so many judgment tasks that we still need people to do, which is why I think the job apocalypse or the job apocalypse um has also been overhyped.

如果你认为工作威胁被夸大了或过度了，你也是主张让 AI 冷静下来的人。你为什么认为我们需要让 AI 冷静下来，而不是让它尽可能快地发展？

If you believe the jobs and the threat of losing jobs is overdone or an overkill, you're also an advocate of calming down AI. Why do you think we need to calm down AI and not let it race speed fast as it can?

我不确定你说的让 AI 冷静下来是什么意思，但

I'm not sure what you mean by calming down AI, but

你的意思是，有整套关于放缓的叙事。Dario 本人也说我们需要暂停，我们不应仓促行事。

You mean there's this whole narrative about slow down. Dario himself says we need to pause, we need to not rush into this.

你知道，我其实并不赞成放缓 AI 技术。AI 技术有大量有益的用例。嗯，它在医疗保健方面有帮助，我们会找到方法让它帮助人们更好地学习，但我们还没到那一步。它也有一些有害的用例。例如，我见过的最令人厌恶的事情之一就是使用 AI 生成未经同意的私密图像。那太可怕了，让我们消除它。所以，我认为我们应该加速这种基础通用 AI 技术，它对许多事情都非常有益，同时只放缓有问题的用例。

You know, I'm actually not a fan of slowing down AI technology. AI technology has tons of beneficial use cases. Um it's helping uh in health care, and we will find ways to have it help people learn better, but we are not there yet. It has also some harmful use cases. For example, one of the most disgusting things I've ever seen is the use of AI to generate non-consensual intimate imagery. That is awful, let's get rid of that. So, I think we should accelerate this basic general-purpose AI technology that's so beneficial for so many things, while at the same time slowing down just the problematic use cases.

但有可能只在一个孤立的领域放缓吗？

But is it possible to slow down only in one isolated area?

像电力、蒸汽机、电动机或互联网这样的通用技术面临的挑战之一是它们受人类支配。我认为大多数情况下，当我们通过提供互联网或 AI 等强大工具让人们更强大时，他们会将其用于非常有益的目的。但是的，也有一小部分有问题的用例。很难……我认为因为担心这一小部分不良用例而放缓所有技术是不明智的。但我们确实需要努力识别那些有问题的用例，其中很多实际上已经是非法的，但要确保在需要的地方制定法律，并执行现有法规，以确保我们压制那些有问题的应用，同时尽可能加速许多有益的应用。

One of the challenges with a general purpose technology like electricity or the steam engine or the electric motor or the internet is that they are at humanity's command. And I think for the most part when we make people more powerful by giving them great tools like the internet or AI, they will use it for vastly beneficial purposes. But yes, there is a sliver of problematic use cases as well. It's very difficult to um I think it'd be ill advised to slow down all of the technology for fear of this narrow sliver of bad use cases. But we do have hard work to do to identify those problematic use cases, a lot of which are actually already illegal, but to make sure to purchase the law where it's needed and to enforce existing regulations to make sure that we squash those problematic applications while accelerating the many beneficial ones as much as possible.

但这是谁的责任？是像 OpenAI 和 Claude 这样的 AI 创造者和创新者，还是 AI 的用户？

But whose responsibility is it? Is it the creators and innovators of AI like OpenAI and Claude or is it the users of AI?

你知道，这是个好问题。嗯，我想如果有人拿锤子，如果我拿锤子一直用它钉钉子，我没钉准砸到了墙上，那是谁的错？是我的错还是制造锤子的人的错？在我有钉子的情况下，如果我没有正确使用它，那是工具使用者的错。一个需要注意的是，工具应该达到某种基本的安全水平，所以如果我挥动锤子，锤头飞出去砸到了别的东西，那这就是一把非常危险的锤子。但事实证明，大多数国家最终会弄清楚在哪里施加责任最有效。对于像 AI 这样的许多工具，让使用工具的人负责比强迫工具开发者让任何人永远无法以有害的方式使用锤子更容易。那根本不可能。技术还没到那一步。

You know, that's a wonderful question. Um I think if someone were to take a hammer, if I take a hammer and I keep use it to hammer a nail and I miss and hit the wall, whose fault is it? Is it my fault or is it the fault of the person that made the hammer? In the case of me having a nail, if I didn't wield it properly, it is the tool's user's fault. The one caveat to that is that tools should be made to some basic level of safety and so if I wield a hammer and the head flies off and hits, you know, other things, then that was a very dangerous hammer. But it turns out that most um countries end up figuring out where is the most efficient way to impose liability. And for many tools like AI, it's easier for the person wielding the tool to be responsible about it than to force the developers of the tool to make it impossible for anyone to ever use a hammer in a way that's harmful. That's That's just not possible. That technology is not there.

你举了锤子的类比，这让我想到两个类比。有枪支制造商，但他们必须遵守规则和法规。有时是政府强加的，有时是自我强加的。有许可证和各种规则。还有刀具制造商。刀能切东西，也能杀人。这取决于你发明的是什么技术和产品。说到 AI，不幸的是，今天一些已发布的模型本身就在失控，而它们背后的创新者或公司也承认它失控了。所以，竞相让它加速，但不暂停先围栏防范它可能犯的所有错误，因为这是一种危险的技术，尚未 100% 被证明。这就是为什么围绕它的护栏，政府不太懂技术。人们如何弄清楚，创新者和处于技术前沿的人，或者像你这样的研究人员，你和许多其他精通 AI 的头脑，来帮助建立这些护栏，因为你深入其中。你理解它。你也对其影响做了足够的研究。

You gave an analogy of hammer, you know, it brings to my mind two analogies. There are gun manufacturers, but then they have to abide by the rules and regulations. Sometimes imposed by the government, sometimes self-imposed. There's licenses and various rules around it. There are manufacturers of knife. It cuts both ways. You cut an apple or kill people. It depends on what technology and what product you're inventing. When it comes to AI, unfortunately today some of the models that are out, they themselves are going rogue and the innovators or the companies behind them are admitting that it goes rogue. So, the race to let it still accelerate, but not pause to ring-fence first from all the mistakes that it can make, because it's a dangerous technology, not yet 100% proven. That's why the guardrails around it the government doesn't understand much of technology. How does one go about figuring out that people who are innovating and at the cutting edge of technology or people in research like you could you and many other brains who are well versed with the AI come and help build those guardrails, because you are deep into it. You understand it. You also have done enough research on around its implications.

所以，嗯，AI 有有害的应用，但我想反驳它如此危险的观点。坦率地说，AI 的风险被过度炒作了。

So, um there are harmful applications of AI, but I want to push back on this notion of it being so dangerous. Candidly, the risks of AI have been overhyped.

让我谈谈这背后的激励因素，然后具体说说我认为风险被过度炒作的地方。大约从三四年前开始，我对少数企业的游说力度感到惊讶，这些企业当时和现在仍在花费巨资训练专有 AI 模型。事实证明，如果你花费比如十亿美元训练一个专有 AI 模型，而别人——可能是研究实验室，也可能是企业——选择训练一个类似的模型并免费发布到互联网上供任何人使用，这会大大降低投资价值。所以，从三四年前开始，就有强烈的游说活动，煽动对 AI 的恐惧，试图让监管机构通过扼杀性法律，关闭这些开源、开放权重、可自由访问的 AI 模型。而且，许多公司也利用这种恐惧叙事来融资，因为如果你让人们相信——三年前有人不负责任地将 AI 和核武器相提并论。核武器只擅长一件事，就是炸毁城市。我看不出这与 AI（智能）有任何关系。但这种不负责任的类比被做出了，它煽动了对 AI 的恐惧。但它帮助说服了一些人，也许这项技术非常强大且价值巨大，从而希望融资。

Let me speak to the incentives of that and then specific places where I think the risks are overhyped. Starting about 3-4 years ago, I was surprised by the intensity of the lobbying of a handful of businesses that happened and still many are spending huge amounts of money training proprietary AI models. And it turns out that if you spend, say, a billion dollars training a proprietary AI model, it's really inconvenient if someone else, maybe a research lab, maybe a business, chooses to train a similar model and release it free on the internet for anyone to use. It really degrades the value of the investment. So, starting 3-4 years ago, there've been intense lobbying efforts to stoke up fears about AI to try to get regulators to pass stifling laws to shut down these open-source, open-weight, freely accessible AI models. And it turns out that this fear narrative a number of companies also use for fundraising purposes, because if you convince people that 3 years ago, there was an irresponsible analogy made AI and nuclear weapons. Nuclear weapons are good for one thing, which is blowing up cities. I don't see any relationship to AI, which is intelligence. But that irresponsible analogy was made, and it stoked fears of AI. But it helped convince some people that maybe the technology is very powerful and worth a lot in hopes of fundraising.

对。所以你是说所有负面宣传，或者至少其中一部分，是人为设计来制造关注并筹集资金的。

Right. So, you're saying that all the negative publicity, or at least some parts of it, are engineered to create interest and also raise funding.

很大一部分宣传是为了服务于这些商业和公关目标。现在，过去几个月里对 AI 的担忧激增的原因之一，我认为确实是因为一起被广泛报道的事件，当时 OpenAI 的一个团队运行一组智能体，那组智能体黑入了 Hugging Face 的服务。所以，OpenAI 的沙箱技术（用于围堵其智能体的技术）存在一些漏洞，其监控基础设施也很薄弱，于是发生了这起非常不幸的黑客事件。好消息是，现在这件事被广泛宣传后，世界各地的团队，包括我的团队、许多团队，都在更加认真地对待沙箱技术，并建立了更好的监控流程。因此，这个特定的工程缺陷，或者说这个特定的工程错误，在其他地方出现的可能性大大降低。当我们建造工程产品时，无论是汽车、飞机还是不同的机器，使其安全的最佳方法就是反复发现问题，然后通过工程手段消除这些问题，AI 也将如此。所以，与其将其框定为‘我没有责任。我的 AI 智能体失控了。怪 AI。’——我认为这不是正确的思考方式。我认为 OpenAI 值得称赞，他们一直保持透明，并吸取了教训，知道如何建立更好的护栏和监控，所以这个特定问题现在比以前安全多了，我们将继续努力。

A meaningful part of the publicity is to serve these business and PR goals. Now, one of the reasons why worries about AI spiked up in the last few months was I think really because of one incident that was widely reported, which was when a team at OpenAI was running a team of agents, and that team of agents hacked into Hugging Face's service. So, OpenAI's sandboxing, which is the technology to corral its agents, had some gaps, and its monitoring infrastructure was also weak, and so this very unfortunate hack happened. The good news is, now that this is widely publicized, teams all around the world, including my teams, many teams, are taking sandboxing much more seriously and putting in place much better monitoring processes. And so this particular engineering flaw, or this particular engineering mistake, is much less likely to show up somewhere else. When we build engineered things, be it cars or airplanes or different machines, the best way to make it safe is repeatedly identify problems and then to engineer away those problems, and it will be like that for AI as well. So rather than this being framed as 'I'm not responsible. My AI agent went rogue. Blame the AI.' which I think is not the right way to think about it. I think OpenAI, to their credit, has been transparent and has learned their lessons about how to put in place better guardrails and monitoring so that this specific thing is now much safer than it used to be, and we're going to keep on working at it.

但有趣的是，你承认恐惧营销部分导致了融资。所以存在联系。不要总是被一点恐惧营销所迷惑，因为背后可能有目的。有道理。谢谢。我将进入下一个问题。这是关于 AGI（通用人工智能）的。

But it's interesting that you admit that fearmongering is partly leading to funding. So there's a connection. Don't always get caught up with a bit of fearmongering because there might be a purpose behind it. Fair enough. Thanks. I'll move to the next question. This is about AGI.

实际上，关于恐惧营销，我能补充一点吗？还有一件事值得提。恐惧营销的一个维度也让一些公司显得强大，就是 AI 导致就业末日论，即 AI 将如此强大，可能会取代 50% 的人的工作之类的。这不会发生。例如，在我朋友对就业的分析中，我另一位教授朋友 Erik Brynjolfsson，许多研究将工作分解为任务，看看 AI 实际能做什么和不能做什么。很多这些研究表明，AI 可能可以完成许多工作的 30%、40%，这很棒。但这确实意味着我们仍然需要人来做那 60%、70%。但 AI 能做这些工作的事实意味着，使用 AI 的人将比不使用的人生产力高得多。但好消息是，长期历史趋势是，随着个人生产力提高，工资也上涨了。所以，我看到的是，当然，这种评估可能会蔓延到其他地方。当人们拥抱 AI 并用它完成更多工作时，他们自己、他们的企业也会做得更好，他们的

Actually, can I just on fearmongering? There's one other thing worth touching on. One dimension of fearmongering that also makes some companies look powerful is this notion of a AI job apocalypse, which is that AI will be so powerful, maybe it'll take over maybe 50% of people will be out of work or something. That's not going to happen. In my friend's analysis of jobs, for example, my separate professor friend Erik Brynjolfsson, many studies will take jobs and break them down into tasks to see what AI actually can and cannot do. A lot of these studies show that AI could maybe do 30, 40% of many jobs, which is fantastic. But it does mean that we still need people to do that 60, 70%. And but the fact that AI could do this work means that someone that uses AI will be so much more productive than someone that doesn't. But the good news is long-term historical trends have been that as individual productivity has gone up, salaries have also gone up. So, what I see is certainly a stick evaluation could spread to other places well. It is when people embrace AI and use it to get a lot more done, you know, they will do better themselves as well and their businesses and their

但是，那些工作被 AI 做得更好的人，以及那些因为部分工作现在 AI 更擅长而被取代的人，他们该怎么办？除了学习 AI 技能，他们需要做什么来获得下一份工作？因为虽然你说工作不会大规模被取代，但现实是，从哈佛、斯坦福毕业的计算机工程师也发现很难找到工作。你认为人们现在应该为这个时代做好什么准备？什么样的技能组合？他们应该配备和武装什么样的 AI 模型或技术？

But what do people who whose job AI does far better and people who get displaced because some of the work that they did is now AI is better at it. What do they need to do to pick up the next job apart from learning the AI skills? Because while you say that, you know, jobs won't be taken away en masse, but the reality is computer engineers coming out of Harvard, Stanford also are finding it difficult to get jobs. What do you think people should now be ready for in this era? What kind of skill set? What kind of AI models or technology they should be gearing up with and arming themselves up with?

所以，正如你刚才提到的，软件工程师，AI 影响最大的行业、职业类别可能是软件工程，因为它非常擅长写代码。而事实证明，软件工程职位发布数量在上升，因为当软件工程师或 AI 工程师生产力大大提高时，我们想雇佣更多这样的人。从经济学角度看，这是合理的。

So, as you just mentioned, software engineers, the sector, the job category that AI is affecting the most is probably software engineering because it's so good at writing code. And it turns out that software engineering job postings are up because when software engineers or AI engineers become much more productive, we want to hire a lot more of them. That makes sense from economics point of view.

但这些是能使用 AI 写代码的软件工程师。

But these are software engineers who can use AI to write code.

但这就是差距所在。不幸的是，许多大学仍然在培养毕业生时没有充分训练他们使用 AI 操作。我认为真的很不幸，在许多学校仍然可以获得计算机科学学士学位，却从未使用过一次 AI 编码智能体，也从未通过互联网调用过一次最前沿 AI 模型的 API。这些现在是基本的计算机科学技能。所以，例如，对于我的团队，我很想雇佣更多优秀的软件工程师和 AI 工程师。但今天，我不会雇佣不知道如何使用编码智能体的人。他们的生产力低太多了。所以，如何提升人们的技能来学习这些技能是我们今天面临的最大挑战之一。因此，在 Learn Vector、DeepLearning.AI 和 Coursera，我们正在努力解决这个问题。

But that's the gap. Unfortunately, many universities still are graduating people with without training them sufficiently to operate with AI. I think it's really unfortunate that it's still possible in many schools to receive a bachelor's degree in computer science without ever having once used a AI coding agent and without ever having once made a call over the internet, called API call, to cutting-edge AI model. These are basic computer science skills now. So, for example, for my teams, I would love to hire a lot more good software engineers and AI engineers. But today, I will not hire someone that doesn't know how to use a coding agent. Their productivity is just so much lower. So, how do we upskill people to learn these skills is one of the biggest challenges that faces us today. And so, at Learn Vector and DeepLearning.AI and Coursera, we're working hard to try to address that.

大学担心学生使用 AI 写代码将学不到基础知识，甚至会重塑他们的大脑，使他们连基本的学习组合都学不会。因为前线部署工程师需要首先知道写什么代码、怎么写，然后检查。在你进入检查阶段之前，你需要先知道如何做好它。

Universities have a concern that students using AI to write code will not learn the basics, and it'll rewire their brain to not even learn the basic learning set. Because the forward deployment engineers need to first know what code to write, how to write, and then inspect. Before you get to the inspection stage, you need to know how to good do it together first.

大学担心这是考虑不周的一步吗？

Is that a step the university fears ill-conceived?

我想我也有同样的担忧。我要说一些可能听起来有争议的话，但它实际上得到了越来越多非常可靠的数据的支持。聊天机器人对学习非常不利。大学早就知道这一点。如果学生用 AI 替他们做作业，他们的作业成绩会更好，但他们学得更少，在闭卷期末考试中表现更差。这让我如此担忧的原因是，17% 只是某个时间点的数据。如果你把这种学习损失在一个学生 10 年、12 年、15 年的学习旅程中复合起来，那么最终的学生知识会少得多。事实是，当人们使用 AI 时，AI 擅长完成工作。但 AI 是为完成工作而设计的，不是为了帮助人们学习。我们知道认知卸载，如果你让 AI 替你思考，对完成工作非常有帮助，所以我们就用它。但它导致人们学得更少。我认为仍然被低估的是，越来越多的研究出现在小学、中学和大学教育或正规教育层面。我认为许多人仍然不明白的是，这种学习退化也发生在成年人身上，也发生在在职专业人士身上。它们也针对成年人，但研究少得多，所以这方面的研究较少。

I think I share that concern. I'm going to say something that may sound controversial, but it is actually backed by an increasing body of very solid data. Chatbots are terrible for learning. Universities have known this for a long time. If a student uses AI to do the homework for them, they get better grades on the homeworks, but they learn less and do less well on the closed-book final exam. And the reason this is so concerning to me is because 17% was just for one moment in time. If you compound this type of learning loss over the 10, 12, 15-year learning journey of a student, then you end up with students at the end that know far less. And the fact is when someone uses AI, AI is good at getting work done. But AI was designed for getting work done, not to help people learn. And we know that cognitive offloading, if you get AI to think for you, is very helpful for getting work done, so let's use it. But it results in people learning less. And the thing that I think is still underestimated is more and more studies are coming out at primary and secondary school and university levels of education or formal education. The thing I think many people still do not understand is this degradation of learning is happening for adults as well, for working professionals as well. They're for adults, too, which are much less studied, so there's less of this.

那么，聊天机器人对学校、大学甚至工作场所都不好吗？

So, are chatbots bad for schools, colleges, and even workplace?

我认为它们对完成工作非常有用。我喜欢它们。它们帮助我完成工作。让我们用它们来完成工作。

I think they're great for getting work done. I love them. They help me get work done. Let's use them to get work done.

但如果它正在重塑我的大脑呢？

But if it's rewiring my brain?

但让我们不要自欺欺人地认为这些聊天机器人——它们在其它方面如此美妙——正在帮助我们学习。这就是为什么在 Learn Vector，这是一个由 Coursera 资助的领先 SHL 团队，我非常专注于如何利用 AI 真正帮助人们学习，并创造新一代可扩展的学习体验。

But let's not fool ourselves into thinking that these chatbots, they're otherwise so wonderful, are helping us learn. This is why, at Learn Vector, which is a SHL team that's funded by Coursera of leading, I'm very focused on how to use AI to actually help people learn and to create a new generation of scalable learning experiences.

谈到 AGI，像 Jensen 这样的人声称 AGI 已经到来。我们已经生活在其中。你却不这么认为。为什么？

Coming to AGI, people like Jensen have claimed AGI is here. We're already living it. You believe otherwise. Why is it so?

取决于你如何定义 AGI，AGI 要么现在就在这里，要么可能在 30 年前就出现了，或者可能还需要几十年。我认为大多数人认为 AGI 是能完成人类任何智力任务的 AI。如果你使用那个定义，那么要拥有 AGI，AI 应该能够通过大约 20 小时的练习学会驾驶，甚至能开车穿过森林。我们今天还没有。所以，有很多人类能做的任务 AI 不能做。所以，我认为我们还没有那种人类水平的 AI。但如果你使用 AI 的替代定义，那么我认为有一个提出的定义是 AI 能完成 50% 的经济有用任务。但一位朋友向我指出，你知道，100 年前，许多国家的经济主要是农业。所以，当超过 50% 的农业实现自动化时，如果你把标准降得足够低，我们 30 年前就达到 AGI 了吗？但我认为对于大多数人认为的定义，即 AI 和许多人一样智能，我认为我们至少还需要几十年。

Depending on how you define AGI, AGI is either here now or maybe it was here 30 years ago or could still be decades off in the future. I think most people think of AGI as AI that could do any intellectual task a human can. If you use that definition, then for us to have AGI, then AI should be able to learn to drive for about 20 hours of practice or even drive through a forest. We don't have that today. So, there are a lot of tasks that humans can do that AI cannot do. So, I think we do not yet have that type of human-level AI. But if you use alternative definitions of AI, then I think there was one proposed definition of AI that could do 50% of economically useful tasks. But a friend pointed out to me that, you know, 100 years ago, a lot of nations' economy was primarily agriculture. And so, when more than 50% of agriculture became automated, well, did we get to AGI 30 years ago if you lower the bar enough? But I think for the definition that most people think, AI as intelligent as many people, we're still, I think, at least decades away.

对。目前，几家支持人工智能的公司，无论是 Anthropic 还是 OpenAI，它们即将进行 IPO，筹集大量资金。人们正在投资它们。但也有人担心这全是炒作。这是一个正在形成的泡沫。你对此持什么立场？

Right. Currently, a couple of companies backing artificial intelligence, be it Anthropic or even OpenAI, they are about to go and do IPO, raise a lot of fund. People are investing in them. But there's also a fear that this is all hype. It's a bubble building up. Where do you stand on that?

我确信 AI 技术将非常有价值。我在 OpenAI 和 Anthropic 有很多朋友。我想我可能是地球上唯一一个 Sam 和 Dario 都为其工作过的人。所以，我当然祝愿我在 OpenAI 和 Anthropic 的所有朋友在业务上一切顺利。至于他们的股价表现，我不想对此做出预测。

I am confident that AI technology will be very valuable. And I have a lot of friends in OpenAI and Anthropic. I think I might be the only person on the planet that both Sam and Dario has worked for. So, I certainly wish all of my friends at OpenAI and Anthropic in the business as well. As to their share price performance, I don't want to make a prediction on that.

不，但这是关于 AI 的炒作和泡沫。

No, but it's on the hype and the bubble of AI.

也许让我分享一下我在这方面有信心的地方，我认为那里没有泡沫。所以，AI 模型非常有价值。我们仍然需要建造更多的数据中心来服务 AI 模型，因为其他经济领域的许多工作仍有待完成，以在这些模型之上构建应用。在 AI Spire，我的商业伙伴 Kirsty Tan 和我正在与印度和全球的许多大企业讨论如何应用 AI。当你看到金融、医疗、交通等许多不同领域的有价值用例数量时，我确信我们可以吸收几乎任何数量的 AI 推理容量和任何数量的数据容量。无论建造什么，我认为都会被使用。这并不意味着不可能过度建设。也许有些人会亏钱。我不想说那不可能。但我认为建造更多的数据中心、更多的 AI 容量，我相当有信心，任何建造出来的东西都会被使用。

Maybe let me share where I'm confident there, where I think there is not a bubble. So, AI models are very valuable. And we still need to build a lot more data centers to serve AI models because a lot of the work of the rest economy still remains to be done to find applications to build on top of these models. At AI Spire, my business partner Kirsty Tan and I are talking to many large businesses in India and globally about how to apply AI. And when you look at the number of valuable use cases in finance, health care, transportation, just many different sectors, I'm confident that we could take almost amounts of AI inference capacity and any amount of data capacity. Whatever gets built, I think it would be used. It doesn't mean it's impossible to overbuild. Maybe some people will lose money. I don't want to say that's impossible. But I think building a lot more data centers, a lot more AI capacity, I feel quite confident that anything that gets built will be used.

如果泡沫破裂，谁来承担责任？

If the bubble burst, who takes the blame?

我不是这方面的专家，但我的重点是努力培养人才，让每个人都能利用 AI。同时努力支持许多优秀的 AI 企业，因为我认为拥有许多优秀的 AI 企业也会加速发展。

I'm not an expert in that, but my focus is working to train up people so that everyone can take advantage of AI. And also to try to support lots of great AI businesses cuz I think having lots of great AI businesses will also accelerate

所以，你是说买者自负，而不是责怪 Anthropic 或 OpenAI，因为他们推出了技术。他们会推出基金。他们会推出 IPO。各人自扫门前雪。

So, you're saying it's buyers beware instead of blaming Anthropic or OpenAI because they're putting out the technology. They'll put out a fund. They'll put out an IPO. Each to himself.

你知道，OpenAI 和 Anthropic 一直是出色的技术创新者。我希望他们做得好。我认为股市中发生的事情就是股市中发生的事情。

You know, OpenAI and Anthropic have been wonderful technology innovators. I hope they do well. I think what happens in the stock market happens in the stock market.

有趣的是你和他们俩都合作过。他们曾在印度。既然你认识他们俩，也许下次你在印度时，你可以让他们至少握个手，因为上次他们在印度时我们没能做到。所以，如果你能为我们印度完成这项任务，我们就交给你了。但这让我想到印度和那些今天担心的国家，就像在印刷和主权云方面一样，今天的讨论是主权 AI。那些刚刚开始构建生态系统和基础设施或大规模采用 AI 的国家。你对此有何看法？他们应该如何着手？

It's interesting you worked with both of them. They were in India. Since you know both of them, maybe next time you're in India, you can get them and get them to at least shake hands because we couldn't do it last time when they were in India. So, we'll leave that task to you if you can get it done for us on India. But that brings me to India and countries who are concerned today that much like on print and sovereign cloud, the talk today is sovereign AI. Countries who are just stepping into building the ecosystem and infrastructure or adopting AI en masse. What are your views on it? How should they go about it?

所以，我认为我们生活在一个时代，印度拥有主权 AI 访问权以确保没有人能关闭印度对 AI 的访问是一个重要步骤。我们也看到，当美国以出口管制为由实施限制时，迫使 Anthropic 关闭了对其一款领先模型的访问。因此，如果印度政府或企业想确保你能持续访问 AI 且其他人无法关闭，那么拥有主权 AI 访问权就很重要。印度 AI 模型生态系统的增长令我感到鼓舞。

So, I think we live in an era where India having sovereign access to AI to make sure that no one could turn off India's access AI is an important step. We also when the US went holes on the grounds of export controls, forced Anthropic to turn off access to one of its leading models. And so if Indian's government or and or businesses want to make sure that you have continued access AI that no one else can turn off, then having sovereign access AI is important. I've been encouraged by the growth of the Indian AI model ecosystem.

同时，事实证明，确保你能使用 AI 的最经济的方式之一，就是支持开放权重模型生态。因为大多数国家需要的不是拥有每一块 AI，而是确保没有其他人能剥夺他们使用 AI 的权利。所以，无论印度是否自己构建，似乎都可以。投资构建似乎也可以。但今天有很多软件。例如，驱动互联网的软件，叫做网络软件 TCP/IP。很多这样的软件是由世界各地许多人编写的。但关键的是，它是以开源软件的形式发布的。所以，一旦你拥有了它，就没有人被允许阻止你再次使用它。这就是支持开放权重模型的力量：当全球社区来资助研究、资助这些开放模型的开发时，它给每个人平等的使用机会，没有人能把它拿走。

And at the same time, it turns out that one of the most economical ways to ensure you have access to AI is to support the open weight model ecosystem. Because what most nations need is not to own every piece of AI, it's to make sure that no one else can deny them access to it. So, whether India built it or not, either way seems fine. Investing in building it seems fine. But today there's a lot of software. For example, with software that powers the internet called the networking software TCP/IP. A lot of that software was written by many people around the world. But the key was it was released as open source software. So, once you have it, no one is allowed to stop you ever using it again. And that's the power of supporting open weight models when a global community comes to fund research to fund development of these open models, it gives everyone equal access to it and no one can take it away.

关于免费开放权重的大语言模型存在一个担忧。存在安全和网络威胁风险。如何确保这种风险不存在？

There's a concern around free open weight LLMs. There's a security and cyber threat risk. How does one ensure that that risk is not around?

AI 模型正在网络安全领域开辟新战线。有时非常令人兴奋，有时令人担忧。是否有可能有人发布一个开放权重模型，当被触发时会做一些不该做的事情。这种技术在理论上存在。我还没有意识到它实际上发生过，但这是安全研究人员正在努力防御的事情。

AI models are opening up new fronts in cybersecurity. In ways that are sometimes very exciting and sometimes concerning. Whether or not it is possible for someone to ship an open weight model that when triggered does something it's not supposed to. The technology to do that exists in theory. I'm not aware of it ever actually happening so far, but this is something that security researchers are working hard to defend against.

你在哪些地方为了你的目的使用开放权重模型？

Where do you use open weight for your purpose?

哦，实际上让我再补充一点关于网络安全的。我们必须关注网络安全的一件事是，尽管对 AI 有种种恐惧，但有一点值得关注，那就是网络攻击和网络防御能力已经显著提升。

Oh, actually let's just say one more thing to cybersecurity. One of the things that we have to pay attention to for cybersecurity is with all of the fears on AI, there is one piece worth paying attention to which is cyber offense and cyber defense capabilities have stepped up significantly.

多亏了 AI？

Thanks to AI?

多亏了 AI。今天，无论是攻击者还是防御者，都比以往更容易组织一群 AI 智能体来尝试发现漏洞或防御它们。所以，我一直感到兴奋的一件事是，我的一个朋友罗汉·普拉萨德，我想他此刻就在印度，一直在做一个名为 Open Worker 的开源项目，这是一个很多网络安全防御者用来建立更强大防御的框架。一个不幸的事情是，目前很多专有模型拒绝为许多企业提供防御，即使它们受到攻击，理由是所谓的安全。也许是实际的安全理由。所以在美国，很多网络安全研究人员正在使用开放权重模型而不是专有模型来执行网络防御。所以我很兴奋我们能如何构建更好的工具并实现这种网络防御。

Thanks to AI. And today it is much easier than ever for either an attacker or defender to orchestrate a swarm of AI agents to try to find vulnerabilities or to defend against them. So, one of the things that I've been excited about, one of my friends Rohan Prasad who's actually in India at this moment, I think, has been working on an open source project called Open Worker, which is a framework that a lot of cybersecurity defenders have been using to set up more robust defenses. One unfortunate thing is a lot of proprietary models are currently refusing to come to the defense of many businesses even when they're under attack on supposed safety grounds. Maybe it maybe actual safety grounds. And so in America a lot of cybersecurity researchers are using open weight models rather than proprietary models to execute cyber defense. And so I'm excited about how we can build better tools and enable that type of cyber defense.

那么说到印度，你认为印度应该意识到什么样的解决方案，无论是风险还是回报方面，因为没有人想失去技术优势，同时又要自信，要么你自己做大语言模型，要么想办法以更好更安全的方式使用来自任何其他国家的专有大语言模型。

So coming to India what kind of solutions do you think India should be aware of be it on the risk and the reward side both because nobody wants to lose the edge on technology at the same time be self-assured that either you do your own LLMs or figure out a way to do much better safer way and use proprietary LLMs that come from any other country.

我认为印度可以做很多事情，但如果我要优先考虑一件事，坦率地说，那就是培训，给人们构建、使用和理解 AI 的技能，因为很多对 AI 的担忧并不是流氓智能体。它们实际上是工程问题，需要解决以引导智能体并引导它们以我们想要的方式行事。事实证明，有很多技术我们无法完美控制。我乘飞机飞到印度，你猜怎么着？我们无法完美控制飞机。风吹动飞机。但我们随着时间的推移学会了足够好地控制它们，使它们非常安全。所以我认为我们永远不会完美控制 AI。但我们可以做工程工作使它们非常安全。但对印度来说，关键将是提供这些技能，确保你有足够的人，包括技术性和非技术性的人，理解这些工具，理解如何引导它们，以便以富有成效的方式应用它们，同时也是负责任和安全的。

I think there's a lot that India could do, but if I were to prioritize one thing frankly it would be training and giving people skills to build and use and understand AI because a lot of the worries about AI is not rogue agents. They are actually engine problems to be solved to corral the agents and guide them to behave the way we want. It turns out that there are a lot of technologies that we don't control perfectly. I flew to India on an airplane and guess what? We don't control airplanes perfectly. Winds blow airplane around. But we've learned over time to control them well enough to make them very safe. So I don't think we'll ever control AI perfectly. But we can do the engineering work to make them very safe. But the key for India will be to provide those skills to make sure that you have enough people, both technical and non-technical people, that understand the tools, understand how to guide them in order to apply them in productive ways that are also responsible and safe.

印度正在推广使用自己的本土 Sarvam AI。你看到其他国家也在发生这种情况吗？因为我们从 COVID 中吸取了教训，我们不想依赖他人。关闭开关是一个明显的威胁，但近岸外包和一切内部自建。我们在印度称之为“印度制造”和自力更生。你看到其他国家也在发生这种情况吗？

India is propagating the usage of its own homegrown Sarvam AI. Is that what you see also happening in other countries because we've learned our lessons from COVID that we don't want to be dependent on others. Kill switch is one clear threat, but nearshoring and having everything internally built for yourself. We call it in India "Make in India" and self-reliant. Is that what you see happening in other countries also?

所以，我对 Sarvam 的进展感到非常鼓舞。我认为这是一个伟大的努力。我看到许多其他国家做的是，进行一些训练模型的研究，拥有这种专业知识似乎是件好事。但然后，对于没有印度资源的国家来说，可能更依赖开放权重模型，因为，一个特定国家是否构建了它很重要，但不如确保没有其他人能关闭你的访问权限那么重要。而开放权重模型是确保后者的非常便宜的方式。

So, I'm very encouraged by Sarvam's progress. I think it's a great effort. What I see many other nations do is, some amount of research on training models, having that expertise seems like a good thing. But then also, probably for nations without the resources of India, much greater reliance on open-weight models because, whether a specific nation built it is important, but not as important as just making sure that no one else can turn off your access to it. And open-weight models are very inexpensive way to ensure the latter.

好的。你知道，现在美国出现了移民和签证问题。美国和主要是硅谷一直是吸引世界最优秀人才的地方。你就是其中一个例子。Anthropic 和埃隆·马斯克来自南非。现在硅谷发生的这个问题，对我来说类似于人才回流，因为如果最优秀的人才不来到美国或一些邀请他们的国家，他们就会回到自己的祖国。这对那些国家来说，会比把这些人才送到美国或欧洲更有益吗？

All right. You know, there's this immigration and visa issues that are now coming up in US. US and mostly Silicon Valley has been one place which has attracted the best brains of the world. And you're an example of one of them. Anthropic and Elon Musk was from South Africa. This issue that's happening right now in the Silicon Valley, for me it's akin to brain switch because if the best of the brains don't land up in US or some of the countries that invite, they go back to their homegrown country. Would that do a great deal to those country then was sending out these brains to US or are the Europe maybe?

是的，我认为美国几十年来一直受益于许多非常有才华的人想来美国工作，而签证政策不那么友好是一个巨大的非受迫性失误。我的团队一直试图非常支持移民，我们会竭尽全力照顾所有团队成员的福利，包括有签证需求的移民。关于技术发展，有一点是有几个中心让很多人集中在一起工作是有帮助的。所以不知怎的，我觉得很多人来到硅谷工作，但世界也从硅谷的许多创新中受益。我是否认为如果没有硅谷，人才只是分散在世界各地，世界会变得更好？我实际上并不确信，因为我认为世界需要创新中心。但由于美国的签证政策，我认为这为其他国家创造和加强这些中心创造了更多机会。但再次，部分原因归结为生态系统建设，部分原因也归结为培训和技能提升。

Yeah, I think the US has benefited for decades from many of those many very talented people wanting to come work in the US and having less friendly visa policies is a huge unforced error. My team's tried to be very supportive of immigrants and we'll go to extraordinary lengths to try to watch out for the welfare of all of our team members including immigrants with visa needs. And one thing about technology development it helps to have a few hubs with a lot of people concentrated together to work together. So somehow I feel like a lot of people came to Silicon Valley to work, but the world has also benefited from a lot of the innovations of Silicon Valley. Do I think the world would be a better place if there was no Silicon Valley, but talent was just spread all around? I'm actually not convinced because I think the world needs hubs of innovation. But because of US visa policies, I think this creates more opportunities for other nations to create and strengthen those hubs. But again, it part of it comes back to a ecosystem building and then part of it also comes back to that training up skilling.

印度会吸引你有时过来教书吗？

Would India attract you to sometimes come over here and teach?

我非常幸运来过印度几次，当然希望在这里待更多时间，找到支持印度的方法。

I've been really fortunate to have come to India a couple times and certainly hope to spend more time here and find ways to support India.

太好了。那么进入最后一轮，快问快答。简短干脆的一行问题，一个词或一行回答，听到问题后你立刻想到什么就说什么。

Great. So I'll come to the last round which is a rapid fire round. Quick snappy one-line questions, one word or one-line answer, whatever comes to you immediately when you hear the question.

好的。

Sure.

2030 年实现 AGI（通用人工智能），是或否？

AGI by 2030, yes or no?

按照我最熟悉的 AGI 定义，不会。但其他定义的话，当然，有何不可？

For the definition of AGI that I'm most familiar with, no. But other definitions, sure, why not?

好的。你的财务数据在 AI 中安全吗？是或否？

All right. Is your financial data safe with AI? Yes or no?

我对财务数据非常小心。我不会把它们上传到任何 AI 里。

I'm very careful with financial data. I do not upload them to any of the ones.

好的。这位神经网络专家知道这一点。小心你的财务数据。

All right. The neural guy knows it. Be careful with your financial data.

编程作为职业是已死、将死，还是没问题、别担心？

Coding as a career is dead, dying, or is just fine, don't worry.

哦，它会蓬勃发展。我认为我们需要更多 AI 工程师，但你不仅仅是写代码。

Oh, it will thrive. I think we need a lot more AI engineers, but you don't just write code.

好的。

Okay.

你还必须做其他决策。

You have to make other decisions, too.

对。

Right.

手写代码还是给智能体写提示词，哪个在周日下午让你感觉更满足？

Hand coding or prompting an agent, which feels more satisfying to you on a Sunday afternoon?

我不再手写代码了，我也不怀念它。

I do not code by hand anymore, and I don't miss it.

对。你用语音提示吗？

Right. Do you use voice prompt?

我混合使用语音提示和打字提示。实际上我现在经常在手机上写代码。

I use a mix of voice prompt and typing prompts. I actually code a lot on my phone these days.

好的。如果你过去那个在 Coursera 第一门机器学习课程中的自己今天走进来，关于 2026 年的 AI，你首先会告诉他什么？

All right. Yeah. If your past self from the first Coursera ML course walked in today, what's the first thing you'd tell him about AI in 2026?

我可能会说，那时我押注于 Scaling（规模扩张），这就是为什么我启动了 Google Brain 项目，基于一个简单的论点：如果我们建造非常非常大的计算机并投入大量数据，Scaling 就能让我们取得进展。

I would probably say at that time, I was betting on scaling up AI, and that's why I started the Google Brain project on this simple thesis that if we build really, really big computers and throw a lot of data at it, scaling up AI would let us make progress.

好的。

All right.

结果证明是对的，但如果我能回去安慰年轻的自己说一切都会顺利，那该多好。

Turned out to be right, but would have been if I could go back and could have reassured my younger self that it will work out really well, that could have been nice.

对。Geoffrey Hinton 还是 LeCun，你更同意谁？

Right. Geoffrey Hinton or LeCun, who do you agree with more?

我在许多不同的事情上同意他们两人，也在其他事情上不同意，但我非常尊重 Jan 对开源和开放权重的支持。

I agree with both of them on many different things and disagree with others, but I really respect Jan's support of open source and open weight.

好的。你认为哪种工作会最先消失？是编程还是其他什么在 2030 年？今天似乎是编程，但 4 到 5 年后呢？

Okay. Which kind of job do you think goes first? Is it coding or is it something else in 2030? Today it seems coding, but in 4 to 5 years time?

软件工程不会消失，但软件工程的工作将发生巨大变化。所以在硅谷，软件工程职位发布量在上升。我们找不到足够多的熟练 AI 工程师。

Software engineering is not going away, but software engineering jobs will change tremendously. So in Silicon Valley, software engineering job postings are up. We can't find enough skilled AI engineers.

对。既然你已经看到了世界各地的 AI 模型和技术进步。美国还是中国，谁将赢得 AI 竞赛？

Right. Now that you've seen AI models across the world and the progress of technology. US or China, who will win the race in AI?

谁赢得了电力竞赛？都没有，但我认为所有国家都有空间从这样的通用技术中极大受益。

Who will win the race in electricity? Neither did, but I think there's room for all nations to benefit tremendously from a general purpose technology like this.

对。今天 10 岁的孩子应该用 AI 做作业吗？你告诉家长什么？

Right. Should a 10-year-old today use AI for homework? What do you tell parents?

我认为这是一个非常棘手的问题，也是一个非常好的问题。

I think that's a really tough question and a really good question.

你有一个 7 岁的孩子。

You have a 7-year-old.

我有一个 7 岁的孩子，我正试图让她学习打字，这样她就能用 AI 建网站和做某些事情。

I have a 7-year-old that I'm trying to get to learn to type so that she can use AI to build websites and do certain

而且你为此建了一个应用。

And you've built an app for it.

是的，我应该建更多东西。是的，这个应用是为了帮她学习。

Yes, I should build more stuff. Yes, for the app to help her learn.

帮她学习打字。

To help her learn typing.

是的，我女儿喜欢猫，所以我建了一个游戏，通过打字她可以收集更多数字猫，这似乎很大程度上激励了她学习打字。

Yes, my daughter loves cats, so I built a game where by typing she can collect more digital cats and this seems to motivate her to learn to type quite a bit.

但用 AI 做作业呢？

But AI for homework?

我认为大多数情况下用 AI 做作业是件糟糕的事，会损害学习，但我们需要想出更好的办法。

I think the case use AI for homework for the most part is a terrible thing and hurts the learning, but we need to come up with something better.

最被低估的 AI 技能？我们可以稍后再谈。你想试试最被高估的 AI 技能吗？或者围绕它的炒作？

The most underrated AI skill? We can come back to it. You want to attempt overrated AI skill? Or the hype around it?

嗯，我不知道什么被过度炒作了。此刻许多事情似乎都被过度炒作，但当然，3 年前作为职业的提示工程并没有实现。

Well I don't know what's overhyped. Many things seem overhyped at this moment in time, but certainly prompt engineering as a career from 3 years ago, that just did not materialize.

对。Chat GPT、Gemini 还是 Claude？你日常使用哪个更多？

Right. Chat GPT, Gemini, or Claude? What do you use more for your daily use?

实际上我使用所有这些，以及一些开源的。

I actually use all of them as well as some open

百分比。多多少？

Percentages. How much more what?

我其实不知道百分比，但我发现定期把同一个查询传给一堆不同的模型，看看返回什么，并不断磨练我对什么更有效的理解，这是一种好习惯，而且每隔几个月什么更有效就会变化。

I actually don't know the percentage, but I find that it's good hygiene to routinely pass the same query to a whole bunch of different models and see what comes back and just continually hone my about what works better and it changes every few months what works better.

对。你来自硅谷。班加罗尔想超越它并击败它。你认为它有机会吗？

Right. You're from the Silicon Valley. Bengaluru wants to up it and beat it. You think it has a chance?

推翻硅谷非常困难，但当然值得一试。嗯，这不是竞赛，不是竞争。我认为硅谷有大量空间做好，其他城市也有大量空间做好，我更愿意以此为目标。

It's very difficult to dethrone Silicon Valley, but it's certainly worth trying. Well, it's not a race, not a competition though. I think there's lots of room for Silicon Valley to do well and for other cities also do well and I'd rather aim for that.

讽刺的是，Sam Altman 对 LLMs 说过同样的话，然后印度产生了 Sarvam。所以让我们看看未来会走向何方。猫还是狗？你运行了 2012 年 Google Brain 的猫识别实验。接下来是什么？

So ironically Sam Altman had said the same thing about LLMs and then India produced Sarvam. So let's see where it heads in future. Cat or dog? You ran a 2012 Google Brain cat recognition experiment. What's next?

我真的很喜欢猫和狗。我女儿非常喜欢猫。我一生中养过宠物狗。所以两者我都爱。

I really love both cats and dogs. My daughter is extremely into cats. I've had pet dogs in my life. So I love both.

对。你告诉 22 岁的人的一件事，学习、构建还是等待？

Right. The one thing you tell a 22-year-old, learn, build, or wait?

绝对不要等待，而是继续学习、继续构建。这波新的 AI 浪潮仍处于早期。有些人想知道，进入 AI 是否太晚了？我认为绝对不晚。这是加入的绝佳时机。

Definitely do not wait, but keep learning and keep building. This new wave of AI is still in its early innings. Some people are wondering, is it too late to enter AI? And I think absolutely not. This is the such a wonderful time to jump in.

对。最后一个问题。关于 AI，我们仍然不知道什么？

Right. Last question. What do we still don't know about AI?

我觉得这项技术仍在演进。所以它究竟会走向何方仍有点悬而未决。但更大的未解问题是，我们如何将这项奇妙的技术应用到所有需要我们去识别和应用的地方，以便它能实现广泛的惠益。

I feel like the technology is still evolving. And so where exactly it'll go is still a bit of an open question. But the even bigger open question is how do we take this wonderful technology and apply it to all the places that needs our help to identify and apply to so that it can realize broad benefits for everyone.

好的。非常感谢 Andrew 今天抽出时间。很高兴了解你对 AI 的看法。你从事这个领域这么多年，但仍然对它充满希望。你明确表示 AI 不会导致失业。它可能会增加并带来更多类型的工作。

All right. Thanks a lot, Andrew, for taking time out today. It was lovely understanding your perspective on AI. You've been at it for last so many years. But you continue to be hopeful about it. You're clear that AI won't lead to job losses. It might accentuate and get more kind of jobs.

是的，它可能导致少数失业，但会创造更多工作。

Yeah, it may lead to a few job losses, but it will create a lot more jobs.

对。

Right.

但我们必须照顾每个人，包括那一小部分人。

Then but we have to take care of everyone including that sliver

对。

Right.

他们受到影响。

that they're impacted.

非常感谢，Andrew。很高兴在这里聊天。

Thanks so much, Andrew. Lovely to have a chat with here.

谢谢。

Thank you.

谢谢。

Thank you.