重新定义天才与 AI：超越大语言模型，走向世界模型

最终定义并重新定义天才。天才将是什么？

End up defining and redefining genius. What will a genius be?

嗯，我认为几千年前甚至几个世纪前，人们认定的天才与我们现在认定的天才非常不同。而且我认为天才这个概念还会进一步演变。过去，天才可能是一种创造或发明的行为，但不一定是我们今天倾向于认为的理论层面。它更偏实用。当然在远古时期，那些懂得如何种植作物或驯养动物的人可能被视为天才。

Well, I think several thousand years ago or even a few centuries ago, what people identified as genius is very different from what we currently identify as genius. And I think there will be more evolution of that concept. In the past, genius might have been an act of creation or invention, but not necessarily at the theoretical level like we tend to think of it today. It was more practical. Certainly in the very ancient past, people who figured out how to cultivate crops or domesticate animals were probably seen as genius.

所以我们经常看到，这也是你公开分享过的观点，即 AI 很强大但不智能。当我们做出这种区分，并且在围绕 LLM 的讨论中，你认为智能和 AI 驱动的力量在哪里？

So we have often seen, and this is a thought that you have pretty openly shared, that AI is powerful but not intelligent. When we make that distinction, and there are conversations around LLM, where do you see intelligence and AI-driven power?

是的，我认为确实存在很多混淆，因为我们倾向于将能够复制某些人类功能的系统拟人化。LLM 非常有用，这一点毫无疑问。它们确实放大了人类智能，就像 20 世纪 40 年代以来的计算机技术一样。但除了少数领域，LLM 大多是信息检索系统。它们可以压缩大量人类先前产生的事实知识，并方便地访问。在某种程度上，这是印刷术、图书馆、互联网和搜索引擎的自然演变。它只是获取信息的一种更高效的方式。在少数领域，这些系统的智能能力不仅仅是检索，比如生成代码或进行一些数学运算，其中推理涉及符号操作。问题是，为什么我们有能通过律师考试和赢得数学奥林匹克竞赛的系统，却没有家用机器人？我们甚至没有自动驾驶汽车，更不用说像任何 17 岁少年那样能在 20 小时的练习中自学驾驶的汽车了。所以我们仍然缺失了一些重要的东西。

Yeah, I think there's a lot of confusion really because we tend to anthropomorphize systems that can reproduce certain human functions. LLMs are incredibly useful, no question about that. And they do amplify human intelligence, like computer technology going back to the 1940s. But LLMs, except for a few domains, are mostly information retrieval systems. They can compress a lot of factual knowledge previously produced by humans and give easy access to it. In a way, it's a natural evolution of the printing press, libraries, the internet, and search engines. It's just a more efficient way to access information. There are a few domains where the intelligent capabilities of these systems are more than just retrieval, like generating code or doing some mathematics, where reasoning involves manipulating symbols. The problem is, why do we have systems that can pass the bar exam and win mathematics olympiads, but we don't have domestic robots? We don't even have self-driving cars, and certainly not self-driving cars that can teach themselves to drive in 20 hours of practice like any 17-year-old. So we're missing something big still.

那么，我们到底在教一个 17 岁少年什么呢？

So, what are we teaching a 17 year old then?

嗯，问题是婴儿甚至动物是如何学习的？动物对物理世界的理解比我们今天任何 AI 系统都要好得多，这就是为什么我们没有智能机器人。我们主要通过观察来了解世界，当我们还是几个月大的婴儿时，然后通过互动，我们学习世界的心理模型，这使我们能够理解任何新情况，即使我们之前没有接触过，我们仍然可以处理。当今 AI 的一个热门词是世界模型。这个想法是我们发展出世界的心理模型，使我们能够提前思考、理解新情况、规划行动序列、推理并预测我们行动的后果，这绝对关键，而 LLM 并不真正做这个。

Well, the question is how does a baby learn, or even an animal? Animals have a much better understanding of the physical world than any AI systems we have today, which is why we don't have smart robots. We learn about the world mostly by observation when we are babies a few months old, and then by interaction, and we learn mental models of the world that allow us to apprehend any new situation, even if we haven't been exposed to it beforehand, we can still handle it. A big buzzword in AI today is world models. This is the idea that we develop mental models of the world that allow us to think ahead, apprehend new situations, plan sequences of actions, reason, and predict the consequences of our actions, which is absolutely critical, and LLMs don't do this really.

教授，有一种感觉是，也许 AI 将开启一个极度充裕的时代。这种充裕会让我们受益吗？

There is the sense, professor, that perhaps AI will unlock an era of radical abundance. Will this abundance benefit us?

嗯，如果你和经济学家交谈，他们会告诉我们，如果我们能衡量 AI 将带来的生产力提升，即每小时工作产出的商品量，每年可能增加约 6%。这是研究技术革命对劳动市场和经济影响的经济学家，如菲利普·阿吉翁和埃里克·布林约尔松的观点。这看起来很小，但实际上相当大。它肯定会加速科学进步和医学进步。我不相信会有一个单一的、可识别的节点，经济起飞并出现充裕。还有围绕这个的政策问题：这些好处是否会在全人类或不同国家的不同人群之间共享。这是一个政治问题，与技术无关。

Well, if you talk to economists, they tell you that if we can measure the improvement AI will bring to productivity, which is amount of goods produced per hour worked, it's going to add up to maybe 6% per year. This is from economists who have studied the effect of technological revolutions on the labor market and the economy, like Philippe Aghion and Erik Brynjolfsson. That seems small, but it's actually quite big. It's certainly going to accelerate scientific progress and progress in medicine. I do not believe there's going to be a singular identifiable point where the economy takes off and there's abundance. There's also the question of policies surrounding this: whether those benefits will be shared across humanity or different categories of people in various countries. That's a political question, nothing to do with technology.

那么如果经济学家认为这是一个繁荣，开放性会幸存吗？

So if economists see this as a boom, will openness survive?

这不会是一个事件。它将是渐进的。有一种错误的想法，认为在某个时刻我们会发现人类水平智能的秘密。我不喜欢 AGI 这个说法，因为人类智能是专门化的。所以我不喜欢人工通用智能这个短语，但这不会是一个事件。我们不会发现一个秘密。我们将持续进步，并且我们无法仅仅通过一系列测试来衡量这种进步，这些测试检验机器是否比人类更聪明，因为机器已经在大量且越来越多的狭窄任务上比人类更聪明了。这不是一个统一的标量质量度量；它是一个质量的集合。但更重要的是，智能不仅仅是技能的集合。它是一种极快学习新技能的能力，甚至是在第一次遇到新任务时无需训练就能完成。这才是智能应该被衡量的标准。所以我们无法设计一个测试来判断我们的机器是否比人类更聪明。

It's not going to be an event. It's going to be progressive. There is a false idea that at some point we're going to discover the secret of human-level intelligence. I don't like the phrase AGI because human intelligence is specialized. So I don't like the phrase artificial general intelligence, but it's not going to be an event. We're not going to discover one secret. We're going to make continuous progress, and we're not going to be able to measure that progress by just having a series of tests that test whether a machine is more intelligent than humans, because machines are already more intelligent than humans on a large and growing number of narrow tasks. It's not a uniform scalar measurement of quality; it's a collection of qualities. But what's more important is that intelligence is not just a collection of skills. It's an ability to learn new skills extremely quickly and even to accomplish new tasks without being trained to do it the first time we encounter them. That's really what intelligence should be measured at. So we're not going to be able to design a test that figures out whether our machines are more intelligent than humans.

那么如果这是关于提升技能和确保自己保持相关性，那么也许只有你才是智能的。这是否意味着，那些采用 AI 的国家，以及印度采用 AI 的速度和规模，挑战将是培养经过技能提升和再培训、具备所需技能的人才？

So if it's about upskilling and ensuring that you're relevant, then only perhaps you're intelligent. Will that then mean that the countries that adopt AI, and the pace at which India and the scale at which India has adopted AI, the challenge would be to create talent which is upskilled and reskilled and have the required skills for this?

绝对如此。所以我们与智能 AI 系统的关系将类似于商业、政治、学术或其他领域的领导者与其员工的关系。AI 将成为我们的员工。我们每个人都将成为一群智能机器的管理者。它们会执行我们的指令。它们可能比我们更聪明，但如果你是一名学者或政治家，你与比你更聪明的员工一起工作。事实上，这正是关键所在。

Absolutely. So the relationship we're going to have with intelligent AI systems is going to be similar to the relationship that a leader in business, politics, academia, or some other domain has with their staff. AI is going to be our staff. Every one of us is going to be a manager of a staff of intelligent machines. They'll do our bidding. They might be smarter than us, but certainly if you are an academic or a politician, you work with staff that are smarter than you. In fact, that's the whole point.

你需要吸引比你更聪明的人，因为这能让你更高效。在学术界，比教授更聪明的学生会反过来教教授；不是教授教研究生，而是反过来。我们有很多政客被比他们更聪明的人包围的例子。

You need to attract people who are smarter than you because that makes you more productive. For an academic, students who are smarter than their professor teach them; it's not the professor that teaches graduate students, it's the other way around. And we have a lot of examples of politicians surrounded by people who are smarter than them.

今天早些时候，莫迪总理在集会上发表讲话说，印度不惧怕人工智能；我们将其视为我们的命运和未来，即‘Bhagya’。您是否认为，在印度举办这样性质的峰会，是向全球南方发出的一个信号，而那里可能正是人工智能下一个重大创新的来源？

Earlier today, Prime Minister Modi addressed the gathering and said that India doesn't fear AI; we see this as our destiny and future, which is 'Bhagya'. Do you see that with a summit of this nature being hosted in India, it's a message to the global south, and that's where perhaps the next big innovation in AI could be coming from?

长期来看，它将来自人口结构有利的国家，即印度和非洲。青年是人类最具创造力的部分，而北方国家在这方面严重不足。未来的许多顶尖科学家，实际上现在很多也是来自印度，未来将主要来自非洲。但这首先意味着要激励年轻人学习。认为我们不再需要学习、因为人工智能会替我们做的想法是完全错误的。相反，我们将不得不学得更多。例如，在工业界，我们看到对更高学历人才的需求增加，尤其是博士水平。过去 15 年，工业界对博士级科学家的需求增长了，部分原因是人工智能，但也因为技术进步依赖于科学进步，而科学进步主要由拥有博士学位的科学家推动。因此，对于全球南方国家来说，这意味着要投资于教育和青年。

Long term, it will come from countries with favorable demographics, meaning India and Africa. Youth is the most creative part of humanity, and there's a deficit of that in the north largely. Many top scientists of the future, and in fact many of the present, are from India, and in the future will be mostly from Africa. But that means having incentives for young people to study first of all. The idea that we don't need to study anymore because AI will do it for us is completely false. On the contrary, we will have to study more. For example, in industry, we see more demand for people with more education, at the PhD level. The demand for PhD-level scientists in industry has grown in the last 15 years, in part because of AI but also because technological progress hinges on scientific progress, which is brought about by scientists who mostly have PhDs. So for countries in the global south, that means investing in education and youth.

让人工智能更易获取是印度所信奉的：民主化人工智能，‘人工智能为人人’是本次峰会的主题。您认为人工智能能变得如此易获取吗，特别是对于像我们这样拥有 14 亿人口的大国？

And making AI more accessible is something India believes in: democratizing AI, 'AI for all' is the theme of this summit. Do you think AI can become that accessible, particularly for a country as large as ours with 1.4 billion people?

是的，以各种方式。不幸的是，人工智能系统的推理成本必须降下来，才能对像印度这样国家的大多数人口变得实用。目前，推理成本太高，主要是能源成本。所以这必须降下来，但它将在教育中发挥作用。一旦我们找到最佳使用方式，人工智能将提高教育质量，而不是降低它。它还将改善农业和医疗保健。我不再在 Meta 工作，但一两年以前，我以前的同事进行了一项实验，他们给印度农民提供了智能眼镜。他们可以与人工智能助手交谈，以确定植物得了什么病、是否该收割或天气如何。它在农业中被广泛使用，帮助农民获得更好的收成并做出正确的选择。

Yes, in all kinds of ways. Unfortunately, the cost of inference for AI systems has to come down to become practical for the vast majority of the population in a country like India. Right now, inference is just too expensive, mostly due to energy costs. So this has to come down, but then it will play a role in education. AI will improve the quality of education, not degrade it, once we figure out how to use it best. It will also improve agriculture and healthcare. I don't work at Meta anymore, but there was an experiment a year or two ago run by my former colleagues where they gave smart glasses to farmers in India. They could talk to the AI assistant to figure out what disease is on their plant, whether to harvest now, or what the weather will be. It is being used a lot in agriculture, assisting farmers to ensure better produce and right choices.

当您谈到教育时，人工智能会帮助教育使国家的学生或青年更有文化，还是他们会变得更依赖人工智能？

When you say about education, will AI assist education in terms of making students or youth of the country more literate, or will they become more AI dependent?

我们依赖技术，对吧？我依赖这副眼镜，否则我看不见你。这已经伴随我们几个世纪了。我们当然会依赖人工智能，但人工智能将促进知识的获取，从而成为教育的工具。我认为对社会的影响可以从 15 世纪的观察中推断出来，当时印刷机使得印刷品生产和知识传播成为可能。它对全球社会产生了巨大影响，至少在那些允许它蓬勃发展的国家。我认为人工智能将带来类似的变革，只是让人们更容易获取知识。互联网也发挥了类似作用。如果部署得当，这将使人们信息更灵通、更聪明、能够做出更理性的决策。

We are dependent on technology, right? I'm dependent on this pair of glasses, otherwise I don't see you. That's been with us for centuries. We'll be dependent on AI, of course, but AI will facilitate access to knowledge and thereby be a tool for education. I think the effect on society might be extrapolated from what was observed in the 15th century when the printing press enabled the production of printed matters and the dissemination of knowledge. It had a huge effect on society worldwide, at least in countries that allowed it to flourish. I think it's going to be a similar transformation with AI, just more access to knowledge. The internet played a similar role. This is just going to make people more informed, smarter, able to make more rational decisions if deployed properly.

那么，如果您要定义我们正在见证的这一历史时刻，您会怎么说？它像电力的出现吗？

So if you were to define this moment which we are witnessing in history, how will you say it? Is it like the advent of electricity?

人们曾这样说过，包括经济学家。它是新的电力。我认为它更像是新的印刷机。在更广泛传播和分享知识、放大人类智能的愿景中。但对社会和国家治理方式的影响目前很难预测。我是一个乐观主义者，相信社会会找出如何最好地利用这项技术造福其民众。

People have made that claim, including economists. It's the new electricity. I think it's more like the new printing press, really. In the vision of more dissemination and sharing of knowledge and amplification of human intelligence. But the impact on society and the way countries need to be run is very difficult to predict at this point. I'm an optimist in the sense that societies will figure out how best to use this technology for the benefit of their population.

虽然我是一个乐观主义者，但我还是要问您这个问题，教授。我们是高估了变化，还是低估了已经降临的事物？

While I am an optimist, nevertheless, I'm going to ask this question to you, professor. Are we overestimating the change or underestimating what has struck us?

通常在这种类型的技术变革中，我们短期内高估变化，长期内低估变化。对于人工智能，情况略有不同，因为存在大量炒作和期望，认为向人类级人工智能和超人级人工智能的过渡将在未来几年内发生。过去 15 年人们一直在这样说，但这是错误的。事实上，过去 60 或 70 年他们一直在这样说，但这是错误的。在人工智能历史上，每当科学家发现构建智能机器的新范式时，人们就声称十年内地球上最聪明的实体将是计算机。这在过去 70 年里被证明是错误的，而且错了四五次。现在仍然是错误的。我们离那还很远。

Usually in technological shifts of this type, we are overestimating the changes in the short term and underestimating them in the long term. For AI, it's a little bit different because there's been a huge amount of hype and expectations that the transition to human-level AI and superhuman-level AI is going to be an event happening within the next few years. People have been making that claim for the last 15 years, and it's been false. In fact, they've been making it for the last 60 or 70 years, and it's been false. Every time in the history of AI that scientists have discovered a new paradigm of how to build intelligent machines, people have claimed that within 10 years the smartest entity on the planet will be a computer. That proved to be wrong four or five times in the last 70 years. It's still wrong. We're still very far from that.

我们快接近了，对吧？我们看到了隧道的尽头。但并不是说我们会在两年内拥有超级智能系统。这不会发生，因为存在这个差距。那个能像 17 岁少年一样在 20 小时练习中学会开车的机器人在哪里？尽管我们有数百万小时的人类驾驶训练数据，我们应该能训练一个 AI 系统来模仿他们。但这实际上不太奏效。它不够可靠。

We're getting close, right? We're seeing the end of the tunnel. But it's not like we're going to have superintelligent systems within two years. It's just not happening because of this gap. Where is the robot that can learn to drive in 20 hours of practice like a 17-year-old? Even though we have millions of hours of training data of people driving cars, we should be able to train an AI system to just imitate them. That doesn't actually quite work. It's not reliable enough.

可能两者共同，但主要是人类。我们设定议程。最大的困难是不要被误导，认为一个计算机系统仅仅因为能操纵语言就是智能的。我们倾向于将语言视为人类智能的缩影，但实际上，语言很容易处理，因为它是一系列离散符号，数量有限。这使得训练系统预测文本中的下一个词变得容易，这正是 LLM 的基础。现实世界要复杂得多。这在计算机科学中早已为人所知，被称为莫拉维克悖论，以机器人学家汉斯·莫拉维克命名。我正在创建的公司和过去 15 年从事的研究项目，就是为现实世界打造智能：如何处理现实世界的高维、连续、嘈杂信号，你的家猫能完美处理，但计算机还不能。这是未来几年 AI 的重大挑战。

Probably both together, but mostly humans. We set the agenda. The biggest difficulty is not to get fooled into thinking that a computer system is intelligent simply because it can manipulate language. We tend to think of language as the epitome of human intelligence, but in fact, language is easy to deal with because it's a sequence of discrete symbols of which there is only a finite number. That makes things easy when you train a system to predict the next word in a text, which is what LLMs are based on. The real world is much more complicated. It's been known in computer science for many years as Moravec's paradox, after roboticist Hans Moravec. The company I'm building and the research program I've been working on for the last 15 years is intelligence for the real world: how to deal with high-dimensional, continuous, noisy signals that the real world is, which your house cat is perfectly able to deal with, but not computers yet. That's the big challenge for the next few years in AI.

所以是 AI 必须应对现实世界，还是现实世界必须应对 AI？

So AI has to deal with the real world, or the real world has to deal with AI?

AI 必须应对现实世界。现实世界的混乱和不可预测性。

AI has to deal with the real world. The messiness of the real world, the unpredictability of the real world.

让我们从大想法开始。LeCun 教授，我们是否在创造人类已知的最聪明心智的道路上？这会在我们有生之年发生吗？

Let's begin with the big idea. Professor LeCun, are we on a path to creating the smartest mind that humanity has ever known? And will that happen in our lifetime?

也许在座有些人的有生之年，可能不是我的。这需要一段时间。但我认为我们将构建的更有趣的东西是人类智能的放大器。所以也许不是一个在所有领域超越人类智能的实体，尽管那终将发生，但至少是某种以加速进步的方式放大人类智能的东西。

Maybe in the lifetime of some people here, possibly not in mine. It will take a while. But I think the more interesting thing we're going to build is an amplifier for human intelligence. So maybe not an entity that surpasses human intelligence in all domains, although that will happen at some point, but at least something that will amplify human intelligence in ways that will accelerate progress.

那么我们会最终定义和重新定义天才吗？天才会是什么？

So will we end up defining and redefining genius? What will a genius be?

几千年前甚至几个世纪前，人们认为的天才与我们今天认为的天才非常不同。这个概念会进一步演变。过去，天才是某种创造或发明行为，但可能不是我们今天认为的理论层面。它更实际。在远古时代，那些想出如何种植作物或驯养动物的人可能被视为天才。

Several thousand years ago or even a few centuries ago, what people identified as genius is very different from what we currently identify as genius. There will be more evolution of that concept. In the past, genius was some act of creation or invention, but maybe not at the theoretical level like we think of it today. It was more practical. In the very ancient past, people who figured out how to cultivate crops or domesticate animals probably were seen as genius.

你经常分享一个观点：AI 很强大但不智能。当我们做出这种区分时，你认为智能和 AI 驱动的力量在哪里？

You have often shared the thought that AI is powerful but not intelligent. When we make that distinction, where do you see intelligence and AI-driven power?

存在很多混淆，因为我们倾向于将能复现某些人类功能的系统拟人化。LLM 非常有用，毫无疑问。它们放大了人类智能，就像 20 世纪 40 年代以来的计算机技术。但 LLM 除了少数领域外，大多是信息检索系统。它们压缩了大量人类先前产生的事实知识，并提供便捷访问。这是印刷术、图书馆、互联网和搜索引擎的自然演变。在少数领域，智能能力超越了检索，比如生成代码或某些数学。但这仍然很大程度上是涉及符号操作的推理领域。问题是：为什么我们有能通过律师考试和赢得数学奥林匹克竞赛的系统，却没有能像任何 17 岁少年一样在 20 小时练习中自学开车的家用机器人或自动驾驶汽车？我们缺少一些重要的东西。

There's a lot of confusion because we tend to anthropomorphize systems that can reproduce certain human functions. LLMs are incredibly useful, no question. They amplify human intelligence like computer technology going back to the 1940s. But LLMs, except for a few domains, are mostly information retrieval systems. They compress a lot of factual knowledge previously produced by humans and give easy access to it. It's a natural evolution of the printing press, libraries, the internet, and search engines. There are a few domains where the intelligent capabilities are more than retrieval, like generating code or some mathematics. But it's still largely domains where reasoning involves manipulating symbols. The problem is: why do we have systems that can pass the bar exam and win math olympiads, but we don't have domestic robots or self-driving cars that can teach themselves to drive in 20 hours of practice like any 17-year-old? We're missing something big.

那么我们是在教一个 17 岁少年什么呢？

So what are we teaching a 17-year-old then?

问题是婴儿甚至动物是如何学习的？动物对物理世界的理解比今天任何 AI 系统都要好得多。这就是为什么我们没有智能机器人。我们在几个月大时主要通过观察来了解世界，然后通过互动，我们学习世界的心理模型，这使我们能够理解任何新情况，即使之前没有接触过。当今 AI 的一个热门词汇是世界模型。这个想法是我们发展出世界的心理模型，使我们能够提前思考、理解新情况、规划行动序列、推理并预测我们行动的后果。LLM 并不真正做这些。

The question is how does a baby learn, or even an animal? Animals have a much better understanding of the physical world than any AI system today. That's why we don't have smart robots. We learn about the world mostly by observation when we are babies a few months old, then by interaction, and we learn mental models of the world that allow us to apprehend any new situation even if we haven't been exposed to it before. A big buzzword in AI today is world models. This is the idea that we develop mental models of the world that allow us to think ahead, apprehend new situations, plan sequences of actions, reason, and predict the consequences of our actions. LLMs don't really do this.

有一种感觉是 AI 可能会开启一个极度充裕的时代。这种充裕会让我们受益吗？

There is a sense that perhaps AI will unlock an era of radical abundance. Will this abundance benefit us?

如果你和经济学家交谈，他们会告诉你 AI 带来的生产力提升——每小时工作产生的财富量——可能每年增加约 6%。这是来自研究技术革命对劳动市场和经济影响的经济学家的观点。

If you talk to economists, they tell you that the improvement AI will bring to productivity—amount of wealth produced per hour worked—could add up to maybe 6% per year. This is from economists who have studied the effect of technological revolutions on the labor market and the economy.

像 Philip Aong 或 Eric Buffson 这样的人，这看起来很小，实际上很大，而且肯定会加速科学进步、医学进步。我不认为会有一个单一的、可识别的点，让经济起飞并实现富足。还有关于政策的问题：这些好处是否会惠及全人类，或者不同国家、不同类别的人？这是一个政治问题，与技术无关。那么，如果经济学家认为这是一场繁荣，开放性还能幸存吗？

People like Philip Aong or Eric Buffson, and that seems small, it's actually quite big, and it's certainly going to accelerate scientific progress, progress in medicine. I do not believe there's going to be a singular identifiable point where the economy is going to take off and there's going to be abundance. And there's also the question of the policies surrounding this: are those benefits going to be shared across humanity or different categories of people in various countries? That's a political question, has nothing to do with technology. So if economists see this as a boom, will openness survive?

这不会是一个事件，而是一个渐进的过程。有一种错误的想法，认为在某个时刻我们会发现人类智能的秘密。我不喜欢 AGI 这个词，因为人类智能是专业化的。所以我不喜欢通用人工智能这个说法，但这不会是一个事件。我们不会发现一个秘密。我们会持续进步，而且我们无法仅仅通过一系列测试来衡量这种进步，测试机器是否比人类更聪明，因为机器已经在大量且越来越多的狭窄任务上比人类更聪明了。这不是一个统一的标量质量测量，而是多种质量的集合。但更重要的是，智能不仅仅是技能的集合，它是一种快速学习新技能的能力，甚至是在第一次接触时就能完成新任务而无需训练。这才是智能应该被衡量的标准。所以我们无法仅仅设计一个测试来判断机器是否比人类更聪明。

It's not going to be an event. It's going to be progressive. There is a false idea that at some point we're going to discover the secret of human-level intelligence. I don't like the phrase AGI because human intelligence is specialized. So I don't like the phrase artificial general intelligence, but it's not going to be an event. We're not going to discover one secret. We're going to make continuous progress, and we're not going to be able to measure that progress by just having a series of tests that test whether a machine is more intelligent than humans, because machines are already more intelligent than humans on a large number, a growing number of narrow tasks. It's not like a uniform scalar measurement of quality; it's a collection of qualities. But what's more important is that intelligence is not just a collection of skills. It's an ability to learn new skills extremely quickly and even to accomplish new tasks without being trained to do it the first time we apprehend it. That's really what intelligence should be measured at. So we're not going to be able to just design a test that figures out whether our machine is more intelligent than humans.

那么，如果这关乎提升技能并确保自己保持相关性，那么也许只有这样才能算作聪明。这是否意味着，那些采用 AI 的国家，以及印度采用 AI 的速度和规模，挑战将在于培养经过技能提升和再培训、具备所需技能的人才？

So if it's about upskilling and ensuring that you're relevant, then only perhaps you're intelligent. Will that then mean that the countries that adopt AI, and the pace at which India and the scale at which India has adopted AI, the challenge would be to create talent which is upskilled and reskilled and have the required skills for this?

完全正确。所以我们将与智能 AI 系统建立的关系，类似于商业、政治、学术或其他领域的领导者与其员工的关系。AI 将成为我们的员工。我们每个人都将成为一群智能机器的管理者。它们会执行我们的命令。它们可能比我们更聪明，但如果你是一名学者或政治家，你与比你更聪明的员工一起工作。事实上，这正是关键所在。你需要吸引比你更聪明的人，因为这能让你更高效。对于学者来说，比教授更聪明的学生会教他们；实际上，不是教授教研究生，而是反过来。而且我们有很多例子，政治家身边围绕着比他们更聪明的人。

Absolutely. So the relationship that we're going to have with intelligent AI systems is going to be similar to the relationship that a leader in business, politics, academia, or some other domain has with their staff. AI is going to be our staff. Every one of us is going to be a manager of a staff of intelligent machines. They'll do our bidding. They might be smarter than us, but certainly if you are an academic or a politician, you work with staff that are smarter than you. In fact, that's the whole point. You need to attract people who are smarter than you because that's what makes you more productive. For an academic, students who are smarter than their professor teach them; it's not the professor that teaches graduate students, it's the other way around actually. And certainly we have a lot of examples of politicians who are surrounded by people who are smarter than them.

今天早些时候，莫迪总理在集会上发表讲话时说，印度不惧怕 AI，我们将其视为我们的命运和未来。您是否认为，在印度举办这样性质的峰会，是向全球南方发出的一个信号，也许下一个 AI 重大创新就来自那里？

Earlier today when Prime Minister Modi addressed the gathering, he said that India doesn't fear AI, we are seeing this as our destiny, future which is bhagir. Do you see that with a summit of this nature being hosted in India, it's a message to the global south and that's where perhaps the next big innovation in AI could be coming from?

从长远来看，它将来自人口结构有利的国家，这意味着印度、非洲。年轻人是 humanity 中最具创造力的部分，而北方国家在这方面严重不足。因此，未来的顶尖科学家，实际上现在许多顶尖科学家都来自印度，未来将主要来自非洲。但这意味着什么呢？首先，这意味着要激励年轻人学习。所以那种认为我们不再需要学习，因为 AI 会替我们做的想法是完全错误的。绝对完全错误。我这么说并不是因为我是教授。相反，我们将不得不学习更多。例如，我们看到一个趋势，在过去，在某些国家，印度当然如此，但在欧洲国家也是如此，在美国当然也是如此，我们看到对更高学历（如博士）人才的需求增加。例如，工业界对博士级科学家的需求在过去 15 年中增长了，部分原因是 AI，但也因为一切，因为技术进步依赖于科学进步，而科学进步是由科学家带来的，科学家大多拥有博士学位。所以对教育的需求更多，而不是更少。因此，对于全球南方国家来说，这意味着要投资于教育和年轻人。

Well, long term it's going to come from countries that have favorable demographics, and that means India, Africa. The youth is the most creative part of humanity, and there's a deficit of that in the north largely. So the top scientists of the future, and in fact many of the present, are from India, and in the future will be mostly from Africa. What does that mean though? It means having incentives for young people to study first of all. So the idea that somehow we don't need to study anymore because AI is going to do it for us is completely false. Absolutely completely false. And it's not because I'm a professor that I'm saying this. On the contrary, we're going to have to study more. We see for example a trend where in industry in the past, in certain countries it's certainly true for India, but it's also true in European countries, and certainly in the US, we see more demand for people with more education at the PhD level. For example, the demand for PhD level scientists in industry has grown in the last 15 years, in part because of AI but because of everything, because technological progress hinges on scientific progress, and scientific progress is brought about by scientists, and scientists mostly have done PhDs. So there's more demand for education, not less. And so for countries in the global south, that means investing in education and youth.

还有让 AI 更易获取，这是印度所信仰的 AI 民主化。AI 为所有人也是本次峰会的主题。您认为 AI 能变得如此易获取吗，特别是对于我们这样一个拥有 14 亿人口的大国？

And making AI more accessible, something that India believes in democratizing AI. AI for all is the theme of this summit as well. Do you think AI can become that accessible, particularly for a country as large as ours with 1.4 billion people?

是的，通过各种方式。不幸的是，AI 系统的推理成本必须降下来，才能对印度这样的国家中的绝大多数人口变得实用。目前，推理成本太高了，还有能源成本等等。实际上主要是能源成本。所以这必须降下来，但随后它将在教育中发挥作用。AI 将提高教育质量，而不是降低它，一旦我们找到最佳使用方法。它将特别改善农业和医疗保健。

Yeah, in all kinds of ways. Unfortunately, the cost of inference for AI systems has to come down to become practical for the vast majority of the population in a country like India. Right now, inference is just too expensive, and energy cost and things like that. It's mostly energy cost actually. So this has to come down, but then it will play a role in education. AI will improve the quality of education, not degrade it, once we figure out how to use it best. It will improve agriculture and healthcare in particular.

医疗保健，当然。我不再在 Meta 工作了，但几年前或一年前，我的前同事进行了一项实验，他们给印度的农民配备了智能眼镜，农民可以与 AI 助手对话，了解我的植物得了什么病，或者现在是否应该收割，或者天气会怎样。这在农业中也得到了广泛应用。

Healthcare, of course. So I don't work at Meta anymore, but there was an experiment a couple years ago or a year ago that was run by my former colleagues where they gave smart glasses to farmers in India, and they could talk to the AI assistant to figure out like what is this disease on my plant, or should I harvest now, or what's going to be the weather. And it is being used a lot in agriculture as well.

没错。它正在帮助农民确保他们的农产品变得更好，他们做出正确的选择。

That's right. It is assisting farmers to ensure that their produce gets better, they make right choices.

但说到教育，AI 是会让国家的学生或年轻人更有文化，还是会让他们更依赖 AI？

But when you say about education, will AI assist education in terms of making students or youth of the country more literate or will they become more AI dependent?

嗯，我的意思是，我们依赖技术，对吧？我依赖这副眼镜，否则我看不见你。所以这已经伴随我们几个世纪了。我们当然会依赖 AI，但 AI 会促进知识的获取，从而成为教育的工具。我认为对社会的影响可以从 15 世纪印刷术开始使印刷品生产和知识传播成为可能时观察到的情况推断出来。它对全球社会产生了巨大影响，至少在允许其蓬勃发展的国家是这样。我认为 AI 将带来类似的变革，只是更多地获取知识。互联网也扮演了类似角色。我认为如果以适当的方式部署，这只会让人们更有见识、更聪明、能够做出更理性的决策。

Well, I mean, we're dependent on technology, right? I'm dependent on this pair of glasses otherwise I don't see you. So that's been with us for centuries. We'll be dependent on AI of course, but AI will facilitate access to knowledge and thereby be a tool for education. I think the effect on society might be extrapolated from what was observed in the 15th century when the printing press started enabling the production of printed matters and the dissemination of knowledge. It had a huge effect on society worldwide, at least in countries that allowed it to flourish. I think it's going to be a similar transformation with AI, just more access to knowledge. The internet played a similar role too. I think this is just going to make people more informed, smarter, able to make more rational decisions if it's deployed in the proper way.

那么，如果你要定义我们正在见证并经历的这一历史时刻，你会怎么说？它像电力的出现吗？

So if you were to define this moment which we are witnessing in history we are living it. How will you say it? Is it like the advent of electricity?

是的，有人提出过这种说法，包括经济学家。这是新的电力。我认为它更像新的印刷术。但同样，在更广泛传播和分享知识以及放大人类智能的愿景中。但对社会和国家治理方式的影响目前很难预测。我在某种意义上是个乐观主义者，我认为社会会找出如何最好地利用技术造福其民众。

Yeah, people have made that claim, including economists. It's the new electricity. I think it's more like the new printing press really. But again, in the sort of vision of more dissemination and sharing of knowledge and amplification of human intelligence. But the impact on society and the way countries need to be run is very difficult to predict at this point. I'm sort of an optimist in the sense that I think societies will figure out how best to use the technology for the benefit of their population.

虽然我是个乐观主义者，但我还是要问你这个问题，教授。我们是高估了变化，还是低估了已经降临的东西？

While I am an optimist, nevertheless, I'm going to ask this question to you, professor. Are we overestimating the change or underestimating what has struck us?

通常在这种技术变革中，我们短期高估变化，长期低估变化。我认为 AI 有点不同，因为有很多炒作和期望，认为向人类水平或超人类水平 AI 的过渡将是一个事件，并将在未来几年内发生。过去 15 年人们一直在这么说，但这是错误的。事实上，过去 60 或 70 年他们一直在这么说，但都是错误的。在 AI 历史上，每当科学家发现构建智能机器的新范式时，人们就声称十年内地球上最聪明的实体将是计算机。这在过去 70 年里被证明是错误的四五次。现在仍然是错误的。我们离那还很远。我们正在接近，看到了隧道的尽头，但不会在两年内拥有超级智能系统。这不会发生，因为有这个差距。那个能像 17 岁少年一样在 20 小时练习中学会开车的机器人在哪里？尽管我们有数百万小时的人类驾驶训练数据，我们应该能够训练一个 AI 系统来模仿他们。但这实际上不太奏效，不够可靠。

So, usually in technological shifts of this type, we are overestimating the changes in the short term and underestimating them in the long term. I think for AI it's a little bit different because there's been a huge amount of hype and expectations that the transition to human-level AI or superhuman-level AI is going to be an event and it's going to happen within the next few years. People have been making that claim for the last 15 years and it's been false. In fact, they've been making it for the last 60 or 70 years and it's been false. Every time in the history of AI that scientists have discovered a new paradigm of how to build intelligent machines, people have claimed that within 10 years the smartest entity on the planet will be a computer. That just proved to be wrong four or five times in the last 70 years. It's still wrong. We're still very far from that. We're getting closer, we're seeing the end of the tunnel, but it's not like we're going to have super intelligent systems within two years. It's just not happening because of this gap. Where is the robot that can learn to drive in 20 hours of practice like a 17-year-old, even though we have millions of hours of training data of people driving cars? We should be able to train an AI system to just imitate them. That doesn't actually quite work. It's not reliable enough.

那么让我们试着结束这次对话：从现在起，谁来定义智能？实际上是人类、机器，还是两者一起？

So let's try and wrap up this conversation with who gets to define intelligence now onwards. Will it be actually humans, machines or both together?

可能两者一起，但主要是人类。我们设定议程。最大的困难是不要被愚弄，认为一个计算机系统仅仅因为能操纵语言就是智能的。我们倾向于认为语言是人类智能的缩影，但事实上语言很容易处理，因为语言实际上是一系列离散符号，数量有限。当你训练系统预测文本中的下一个词时，这使事情变得容易，这正是 LLM 的基础。事实证明现实世界要复杂得多。这在计算机科学中已经知道多年，称为莫拉维克悖论，以机器人学家汉斯·莫拉维克命名。我正在建立的公司和过去 15 年左右的研究项目是面向现实世界的智能：如何处理现实世界的高维连续噪声信号，家猫或松鼠完全能处理，但计算机还不能。这是未来几年 AI 的重大挑战：处理现实世界。这也是我建立公司的目的。

Probably both together, but mostly humans. We set the agenda. The biggest difficulty is not to get fooled into thinking that a computer system is intelligent simply because it can manipulate language. We tend to think of language as the epitome of human intelligence, but in fact it turns out language is easy to deal with because language is really a sequence of discrete symbols of which there is only a finite number. That turns out to make things easy when you train a system to predict what the next word is in a text, which is what LLMs are based on. It turns out the real world is much more complicated. It's been known in computer science for many years, called the Moravec's Paradox after roboticist Hans Moravec. The company I'm building and the research program I've been working on for the last 15 years or so is intelligence for the real world: how to deal with high-dimensional continuous noisy signals that the real world is, which a house cat is perfectly able to deal with, or a squirrel, but not computers yet. That's the big challenge for the next few years in AI: dealing with the real world. And that's the point of the company I'm building.

所以 AI 必须处理现实世界，还是现实世界必须处理 AI？

So AI has to deal with the real world or real world has to deal with AI.

AI 必须处理现实世界。现实世界的混乱和不可预测性。

AI has to deal with the real world. The messiness of the real world, the unpredictability of the real world.

好的。非常感谢这次对话，教授。

All right. Thank you so much for this conversation, Professor.

大家下午好。那么让我们从大想法开始，LeCun 教授。我们是否正在创造人类有史以来最聪明的头脑？这会在我们有生之年发生吗？

Good afternoon, everyone. So let's begin with the big idea here, Professor LeCun. Are we on a path to creating the smartest mind that humanity has ever known? And will that happen in our lifetime?

可能在座有些人的有生之年。可能不是我的。我们拭目以待。这需要一段时间。但我认为我们将要构建的更有趣的东西是人类智能的放大器。所以也许不是一个在所有领域都超越人类智能的实体，尽管那会在某个时候发生，但至少是某种以加速进步的方式放大人类智能的东西。

Maybe in the lifetime of some people here. Possibly not in mine. We'll see. It will take a while. But I think the more interesting thing that we're going to build is an amplifier for human intelligence. So maybe not an entity that surpasses human intelligence in all domains, although that will happen at some point, but at least something that will amplify human intelligence in ways that will accelerate progress.

那么，我们最终会定义和重新定义天才吗？天才会是什么？

So then will we end up defining and redefining genius? What will a genius be?

嗯，我认为几千年前甚至几个世纪前，人们认为的天才与我们今天认为的天才非常不同。我认为天才的概念会有更多演变。过去，天才可能是某种创造或发明行为，但可能不是我们今天倾向于认为的理论层面。它更实用。当然在远古时代，那些想出如何种植作物或驯养动物的人可能被视为天才。所以你知道，我们经常看到，而且这是你们都相当公开分享的想法：AI 很强大但不智能。

Well, I think several thousand years ago or even a few centuries ago, what people identified as genius is very different from what we currently identify as genius. And I think there will be more evolution of that concept of genius. In the past, perhaps genius was some act of creation or invention, but maybe not at the theoretical level like we tend to think of it today. It was more practical. Certainly in the very ancient past, people who figured out how to cultivate crops or domesticate animals probably were seen as genius. So you know, we have often seen and this is a thought that you have all pretty openly shared that AI is powerful but not intelligent.

当我们做出这种区分，并在围绕 LLM 的讨论中，你认为智能和 AI 驱动的力量体现在哪里？

When we make that distinction and there are conversations around LLM, where do you see intelligence and AI-driven power?

是的，我认为确实存在很多混淆，因为我们倾向于将能够再现某些人类功能的系统拟人化。LLM 非常有用，毫无疑问，它们确实放大了人类智能，就像 1940 年代以来的计算机技术一样。但 LLM 在某种程度上，除了少数领域，主要是信息检索系统。它们可以压缩大量人类先前产生的事实性知识，并提供便捷的访问。在某种程度上，这是印刷术、图书馆、互联网和搜索引擎的自然演变。这只是一种更高效的信息访问方式。有几个领域这些系统的智能能力超出了检索，比如生成代码或进行某些数学运算，我们觉得这超越了检索，但这些领域在很大程度上仍然涉及符号操作推理。问题在于，我们有能通过律师资格考试和赢得数学奥林匹克竞赛的系统，但我们没有家用机器人。我们甚至没有自动驾驶汽车，当然也没有像任何 17 岁少年那样通过 20 小时练习就能学会驾驶的自动驾驶汽车。所以，我们仍然缺少一些重要的东西。

Yeah, I think there's a lot of confusion really because we tend to anthropomorphize systems that can reproduce certain human functions. LLMs are incredibly useful, no question about that, and they do amplify human intelligence like computer technology going back to the 1940s. But LLMs, to some extent except for a few domains, are mostly information retrieval systems. They can compress a lot of factual knowledge that has been previously produced by humans and give easy access to it. In a way, it's a natural evolution of the printing press, libraries, the internet, and search engines. It's just a more efficient way to access information. There are a few domains where the intelligent capabilities of these systems are more than that, more than just retrieval. For generating code, maybe doing some type of mathematics, we get the impression that it's beyond this, but it's still to a large extent domains where reasoning has to do with manipulating symbols. The problem is that we have systems that can pass the bar exam and win mathematics olympiads, but we don't have domestic robots. We don't even have self-driving cars, and we certainly do not have self-driving cars that can teach themselves to drive in 20 hours of practice like any 17-year-old. So, we're missing something big still.

那么，我们到底在教一个 17 岁少年什么呢？

So, what are we teaching a 17 year old then?

问题是，婴儿甚至动物是如何学习的？动物对物理世界的理解比我们今天任何 AI 系统都要好得多，这就是为什么我们没有智能机器人。我们主要通过观察来了解世界如何运作，当我们还是几个月大的婴儿时，然后我们通过互动学习，并学习世界的心理模型，这使我们能够理解任何新情况，即使我们之前没有接触过。我们仍然可以处理它。当今 AI 的一个流行词是世界模型。这个想法是我们发展出世界的心理模型，使我们能够提前思考、理解新情况、规划行动序列、推理并预测我们行动的后果，这绝对至关重要，而当前系统并没有真正做到这一点。

Well, the question is how does a baby learn or even an animal? Animals have a much better understanding of the physical world than any AI systems we have today, which is why we don't have smart robots. We learn about how the world works mostly by observation when we are babies a few months old, and then we learn by interaction, and we learn mental models of the world that allow us to apprehend any new situation even if we haven't been exposed to it beforehand. We can still handle it. A big buzzword in AI today is world models. This is the idea that we develop mental models of the world that allow us to think ahead, apprehend new situations, plan sequences of actions, reason, and predict the consequences of our actions, which is absolutely critical, and current systems don't do this really.

教授，有一种感觉，也许 AI 将开启一个极度充裕的时代。这种充裕会让我们受益吗？

There is the sense, professor, that perhaps AI will unlock an era of radical abundance. Will this abundance benefit us?

如果你和经济学家交谈，他们会告诉你，如果我们能衡量 AI 将带来的生产力提升，即每小时工作产出的商品数量，可能每年增加约 6%。这是来自研究技术革命对劳动市场和经济影响的经济学家，如 Philippe Aghion 或 Erik Brynjolfsson。这看起来很小，但实际上相当大，而且肯定会加速科学进步和医学进步。我不认为会有一个单一可识别的节点，经济起飞并出现充裕。还有围绕此的政策问题：这些好处是否会惠及全人类，或不同国家的不同人群？这是一个政治问题，与技术无关。

Well, if you talk to economists, they tell you that if we can measure the improvement AI will bring to productivity, which is the amount of goods produced per hour worked, it might add up to maybe 6% per year. This is from economists who have studied the effect of technological revolutions on the labor market and the economy, like Philippe Aghion or Erik Brynjolfsson. That seems small but it's actually quite big, and it's certainly going to accelerate scientific progress and progress in medicine. I do not believe there is going to be a singular identifiable point where the economy takes off and there is abundance. There is also the question of policies surrounding this: are those benefits going to be shared across humanity or different categories of people in various countries? That's a political question, nothing to do with technology.

如果经济学家认为这是一场繁荣，开放性会幸存吗？

If economists see this as a boom, will openness survive?

这不会是一个事件。它将是渐进的。有一种错误的想法，认为在某个时刻我们会发现人类水平智能的秘密。我不喜欢 AGI 这个说法，因为人类智能是专门化的，所以我不喜欢通用人工智能这个短语。但这不会是一个事件。我们不会发现一个秘密。我们将持续进步，我们无法仅仅通过一系列测试来衡量这种进步，这些测试检验机器是否比人类更聪明，因为机器已经在大量且越来越多的狭窄任务上比人类更聪明。这不是一个统一的质量标量测量；它是多种品质的集合。但更重要的是，智能不仅仅是技能的集合。它是一种极其快速学习新技能的能力，甚至是在第一次遇到新任务时无需训练就能完成。这才是智能应该被衡量的地方。所以我们无法仅仅设计一个测试来判断机器是否比人类更聪明。

It's not going to be an event. It's going to be progressive. There is a false idea that somehow at some point we're going to discover the secret of human-level intelligence. I don't like the phrase AGI because human intelligence is specialized, so I don't like the phrase artificial general intelligence. But it's not going to be an event. We're not going to discover one secret. We're going to make continuous progress, and we're not going to be able to measure that progress by just having a series of tests that test whether a machine is more intelligent than humans, because machines are already more intelligent than humans on a large number, a growing number of narrow tasks. It's not a uniform scalar measurement of quality; it's a collection of qualities. But what's more important is that intelligence is not just a collection of skills. It's an ability to learn new skills extremely quickly and even to accomplish new tasks without being trained to do it the first time we encounter them. That's really what intelligence should be measured at. So we're not going to be able to just design a test that figures out whether machines are more intelligent than humans.

那么，如果这是关于提升技能并确保你保持相关性，那么也许只有你才是智能的。这是否意味着，采用 AI 的国家，以及印度采用 AI 的速度和规模，挑战将在于培养经过技能提升和再培训、具备所需技能的人才？

So if it's about upskilling and ensuring that you're relevant, then only perhaps you're intelligent. Will that then mean that the countries that adopt AI, and the pace at which India and the scale at which India has adopted AI, the challenge would be to create talent which is upskilled and reskilled and have the required skills for this?

绝对如此。我们将与智能 AI 系统建立的关系，类似于商业、政治、学术或其他领域的领导者与其员工的关系。AI 将成为我们的员工。我们每个人都将成为一群智能机器的管理者。它们会执行我们的指令。它们可能比我们更聪明，但当然，如果你是一名学者或政治家，你与比你更聪明的员工一起工作。事实上，这正是关键所在。你需要吸引比你更聪明的人，因为这让你更高效。对于学者来说，比教授更聪明的学生会教他们；实际上，不是教授教研究生，而是反过来。当然，我们有很多政治家被比他们更聪明的人包围的例子。

Absolutely. The relationship we're going to have with intelligent AI systems is going to be similar to the relationship that a leader in business, politics, academia, or some other domain has with their staff. AI is going to be our staff. Every one of us is going to be a manager of a staff of intelligent machines. They'll do our bidding. They might be smarter than us, but certainly if you are an academic or a politician, you work with staff that are smarter than you. In fact, that's the whole point. You need to attract people who are smarter than you because that makes you more productive. For an academic, students who are smarter than their professor teach them; it's not the professor that teaches graduate students, it's the other way around actually. And certainly we have a lot of examples of politicians who are surrounded by people who are smarter than them.

今天早些时候，莫迪总理在集会上表示，印度不惧怕人工智能；我们将其视为命运和未来，即‘Bhagya’。您是否认为，在印度举办这样性质的峰会，是向全球南方发出的信号，而那里可能正是人工智能下一个重大创新的来源？

Earlier today when Prime Minister Modi addressed the gathering, he said that India doesn't fear AI; we are seeing this as our destiny, future, which is Bhagya. Do you see that with a summit of this nature being hosted in India, it's a message to the global south and that's where perhaps the next big innovation in AI could be coming from?

从长远来看，它将来自人口结构有利的国家，这意味着印度和非洲。青年是人类最具创造力的部分，而北方国家在这方面严重不足。因此，未来的顶尖科学家，实际上现在的许多顶尖科学家，都来自印度，未来将主要来自非洲。但这意味着什么呢？这意味着要为年轻人提供学习的激励。认为我们不再需要学习，因为人工智能会替我们做，这种想法完全错误。恰恰相反，我们将不得不学习更多。我们看到一种趋势，在工业界，在印度、欧洲国家和美国等国家，对受过更多教育的人，例如博士水平的人才，需求更大。过去 15 年里，工业界对博士级科学家的需求增长了，部分原因是人工智能，但也因为其他一切。技术进步依赖于科学进步，而科学进步是由科学家带来的，他们大多拥有博士学位。因此，对教育的需求更大，而不是更小。对于全球南方国家来说，这意味着要投资于教育和青年。

Well, long term it's going to come from countries that have favorable demographics, and that means India, Africa. The youth is the most creative part of humanity, and there's a deficit of that in the north largely. So the top scientists of the future, and in fact many of the present, are from India, and in the future will be mostly from Africa. But what does that mean? It means having incentives for young people to study. The idea that we don't need to study anymore because AI is going to do it for us is completely false. On the contrary, we're going to have to study more. We see a trend in industry, in certain countries like India, European countries, and the US, there is more demand for people with more education, at the PhD level for example. The demand for PhD level scientists in industry has grown in the last 15 years, in part because of AI, but because of everything. Technological progress hinges on scientific progress, and scientific progress is brought about by scientists, mostly with PhDs. So there's more demand for education, not less. For countries in the global south, that means investing in education and youth.

让人工智能更易获取，这是印度所信仰的：民主化人工智能，人工智能为所有人，这也是本次峰会的主题。您认为人工智能能变得如此普及吗，尤其是对于像我们这样拥有 14 亿人口的大国？

And making AI more accessible, something that India believes in: democratizing AI, AI for all is the theme of this summit as well. Do you think AI can become that accessible, particularly for a country as large as ours with 1.4 billion people?

是的，以各种方式。不幸的是，人工智能系统的推理成本必须降低，才能在像印度这样的国家中对绝大多数人口变得实用。目前，推理成本太高，主要是能源成本。所以这必须降下来，但之后它将在教育中发挥作用。一旦我们找到最佳使用方式，人工智能将提高教育质量，而不是降低它。它将特别改善农业和医疗保健。我不再在 Meta 工作了，但几年前我的前同事们进行了一项实验，他们给印度农民配备了智能眼镜。农民可以与人工智能助手交谈，以了解植物得了什么病、是否该收割，或者天气会怎样。它在农业中也得到了广泛应用。

Yeah, in all kinds of ways. Unfortunately, the cost of inference for AI systems has to come down to become practical for the vast majority of the population in a country like India. Right now, inference is just too expensive, mostly due to energy cost. So this has to come down, but then it will play a role in education. AI will improve the quality of education, not degrade it, once we figure out how to use it best. It will improve agriculture and healthcare in particular. I don't work at Meta anymore, but there was an experiment a couple years ago run by my former colleagues where they gave smart glasses to farmers in India. They could talk to the AI assistant to figure out what disease is on their plant, whether to harvest now, or what the weather will be. It is being used a lot in agriculture as well.

它在农业中也得到了广泛应用。没错。它正在帮助农民确保他们的收成更好，做出正确的选择。但您谈到教育时，人工智能会帮助教育，使国家的学生或青年更有文化，还是他们会变得更依赖人工智能？

It is being used a lot in agriculture as well. That's right. It is assisting farmers to ensure their produce gets better, they make right choices. But when you say about education, will AI assist education in terms of making students or youth of the country more literate, or will they become more AI dependent?

嗯，我们依赖技术，对吧？我依赖这副眼镜，否则我看不见你。这已经伴随我们几个世纪了。我们当然会依赖人工智能，但人工智能将促进知识的获取，从而成为教育的工具。我认为对社会的影响可以从 15 世纪的观察中推断出来，当时印刷机开始使印刷品生产和知识传播成为可能。它对全球社会产生了巨大影响，至少在允许其蓬勃发展的国家是如此。我认为人工智能将带来类似的变革，只是更多地获取知识。互联网也发挥了类似的作用。我认为，如果以适当的方式部署，这将使人们更有见识、更聪明、能够做出更理性的决策。

Well, we're dependent on technology, right? I'm dependent on this pair of glasses, otherwise I don't see you. That's been with us for centuries. We'll be dependent on AI, of course, but AI will facilitate access to knowledge and thereby be a tool for education. I think the effect on society might be extrapolated from what was observed in the 15th century when the printing press started enabling the production of printed matters and the dissemination of knowledge. It had a huge effect on society worldwide, at least in countries that allowed it to flourish. I think it's going to be a similar transformation with AI, just more access to knowledge. The internet played a similar role. I think this is just going to make people more informed, smarter, able to make more rational decisions if it's deployed in the proper way.

那么，如果您要定义我们正在见证的历史时刻，我们正生活在其中。您会怎么说？它像电力的出现吗？

So if you were to define this moment which we are witnessing in history, we are living it. How will you say it? Is it like the advent of electricity?

是的，人们包括经济学家都提出过这种说法。它是新的电力。我认为它更像新的印刷机。但同样，在更多传播和分享知识以及放大人类智能的愿景中。对社会和国家治理方式的影响目前很难预测。我有点乐观，因为我认为社会会找出如何最好地利用这项技术来造福其民众。

Yeah, people have made that claim, including economists. It's the new electricity. I think it's more like the new printing press really. But again, in the vision of more dissemination and sharing of knowledge and amplification of human intelligence. The impact on society and the way countries need to be run is very difficult to predict at this point. I'm sort of an optimist in the sense that I think societies will figure out how best to use that technology for the benefits of their population.

虽然我是个乐观主义者，但我还是要问您这个问题，教授。我们是高估了变化，还是低估了已经降临的东西？

While I am an optimist, nevertheless, I'm going to ask this question to you, professor. Are we overestimating the change or underestimating what has struck us?

通常，在这种技术变革中，我们短期内高估变化，长期低估变化。我认为对于人工智能来说有点不同，因为存在大量的炒作和期望，认为向人类水平的人工智能、超人类水平的人工智能的过渡将是一个事件，并且将在未来几年内发生。人们过去 15 年一直在提出这种说法，但它是错误的。事实上，他们过去 60 或 70 年一直在提出这种说法，每次都是错误的。在人工智能的历史上，科学家们发现了构建智能机器的新范式，人们就声称 10 年内地球上最聪明的实体将是一台计算机。这在过去 70 年里被证明是错误的四五次。现在仍然是错误的。我们离那还很远。我们正在接近，看到了隧道的尽头，但并不是说我们将在两年内拥有超级智能系统。

So, usually in technological shifts of this type, we are overestimating the changes in the short term and underestimating them in the long term. I think for AI it's a little bit different because there's been a huge amount of hype and expectations that the transition to human-level AI, superhuman-level AI is going to be an event and it's going to happen within the next few years. People have been making that claim for the last 15 years, and it's been false. In fact, they've been making it for the last 60 or 70 years, and it's been false every time. In the history of AI, scientists have discovered a new paradigm of how to build intelligent machines, and people have claimed within 10 years the smartest entity on the planet will be a computer. That just proved to be wrong four or five times in the last 70 years. It's still wrong. We're still very far from that. We're getting closer, we're seeing the end of the tunnel, but it's not like we're going to have superintelligence systems within two years.

就是因为这个差距，对吧？那个能像 17 岁少年一样在 20 小时内学会开车的机器人在哪里？尽管我们有数百万小时的人类驾驶训练数据，我们应该能够训练一个 AI 系统来模仿他们。但这实际上不太奏效，不够可靠。

It's just not happening because of this gap, right? Where is the robot that can learn to drive in 20 hours of practice like a 17-year-old? Even though we have millions of hours of training data of people driving cars around, we should be able to train an AI system to just imitate them. That doesn't actually quite work. It's not reliable enough.

好的，那么让我们试着总结一下这场对话：从现在起，谁来定义智能？实际上会是人类、机器，还是两者一起？

Okay, so let's try and wrap up this conversation with who gets to define intelligence now onwards. Will it be actually humans, machines, or both together?

可能是两者一起，但主要是人类。我们设定议程。最大的困难是不要被愚弄，认为一个计算机系统仅仅因为能操控语言就是智能的。我们倾向于将语言视为人类智能的缩影，对吧？但实际上，语言很容易处理，因为语言本质上是一个离散符号序列，符号数量有限。当你训练系统预测文本中的下一个词时，这会让事情变得简单，这正是 LLM 的基础。而现实世界要复杂得多。这在计算机科学中已经知道很多年了，被称为莫拉维克悖论，以机器人学家汉斯·莫拉维克命名。所以我正在创建的公司和我过去 15 年左右的研究项目，就是为现实世界打造智能。如何应对现实世界的高维、连续、嘈杂信号，你的家猫或松鼠都能完美应对，但计算机还不行。这是未来几年 AI 的重大挑战：应对现实世界。这也是我创建这家公司的意义所在。

Probably both together, but mostly humans. We set the agenda. The biggest difficulty is not to get fooled into thinking that a computer system is intelligent simply because it can manipulate language. We tend to think of language as the epitome of human intelligence, right? But in fact it turns out language is easy to deal with, because language is really a sequence of discrete symbols of which there is only a finite number. And that turns out to make things easy when you train a system to predict what the next word is in a text, which is what LLMs are based on. It turns out the real world is much more complicated. It's been known in computer science for many years, called the Moravec Paradox after roboticist Hans Moravec. So the company I'm building and the research program I've been working on for the last 15 years or so is intelligence for the real world. How to deal with high-dimensional continuous noisy signal that the real world is, which your house cat is perfectly able to deal with, or a squirrel, but not computers yet. That's the big challenge for the next few years in AI: dealing with the real world. And that's the point of the company I'm building.

那么是 AI 必须应对现实世界，还是现实世界必须应对 AI？

So AI has to deal with the real world, or the real world has to deal with AI?

AI 必须应对现实世界，现实世界的混乱和不可预测性。

AI has to deal with the real world, the messiness of the real world, the unpredictability of the real world.

好的，非常感谢您参与这次对话，教授。

All right, thank you so much for this conversation, professor.

大家下午好。那么让我们从一个大问题开始。LeCun 教授，我们是否正走在创造人类有史以来最聪明头脑的道路上？这会在我们有生之年发生吗？

Good afternoon everyone. So let's begin with the big idea here. Professor LeCun, are we on a path to creating the smartest mind that humanity has ever known? And will that happen in our lifetime?

也许在座的一些人有生之年能看到。可能不是我的有生之年。我们拭目以待。这需要一段时间。但我认为我们将要构建的更令人兴奋的东西是人类智能的放大器。所以可能不是一个在所有领域都超越人类智能的实体，尽管那迟早会发生。但至少是一个能以加速进步的方式放大人类智能的东西。

Maybe in the lifetime of some people here. Possibly not in mine. We'll see. It will take a while. But I think the more interesting thing that we're going to build is an amplifier for human intelligence. So maybe not an entity that surpasses human intelligence in all domains, although that will happen at some point. But at least something that will amplify human intelligence in ways that will accelerate progress.

那么，我们最终会定义和重新定义天才吗？天才将是什么？

So then will we end up defining and redefining genius? What will a genius be?

嗯，我认为几千年前甚至几个世纪前，人们认为的天才与我们今天认为的天才大不相同。而且我认为天才这个概念还会进一步演变。过去，天才可能是一种创造或发明的行为，但可能不是我们今天倾向于认为的理论层面。它更实用。当然在远古时代，那些懂得如何种植作物或驯养动物的人可能被视为天才。所以，我们经常看到，而且这也是你们所有人都公开分享的想法，即 AI 强大但不智能。当我们做出这种区分时，围绕 LLM 的对话中，你认为智能和 AI 驱动的力量在哪里？

Well, I think several thousand years ago or even a few centuries ago, what people identified as genius is very different from what we currently identify as genius. And I think there will be more evolution of that concept of genius. In the past, perhaps genius was some act of creation or invention, but maybe not at the theoretical level like we tend to think of it today. It was more practical. Certainly in the very ancient past, people who figured out how to cultivate crops or domesticate animals probably were seen as genius. So you know, we have often seen, and this is a thought that you have all pretty openly shared, that AI is powerful but not intelligent. When we make that distinction, and there are conversations around LLMs, where do you see intelligence and AI-driven power?

是的，我认为确实存在很多混淆，因为我们倾向于将能够再现某些人类功能的系统拟人化。我的意思是，LLM 非常有用，这一点毫无疑问。它们确实放大了人类智能，就像 20 世纪 40 年代以来的计算机技术一样。但 LLM 在某种程度上，除了少数领域，大多是信息检索系统。它们可以压缩大量人类先前产生的事实知识，并方便地访问。在某种程度上，这是印刷术、图书馆、互联网和搜索引擎的自然演变，对吧？只是一种更高效的信息访问方式。而在少数领域，这些系统的智能能力实际上不止于此，不仅仅是检索。比如生成代码，或者做一些数学题，我们感觉它已经超越了检索，但在很大程度上，这些领域仍然是推理与符号操作相关。问题是，为什么我们有能通过律师资格考试的系統？

Yeah, I think there's a lot of confusion really, because we tend to anthropomorphize systems that can reproduce certain human functions. So what's I mean LLMs are incredibly useful. There's no question about that. And they do amplify human intelligence, like computer technology going back to the 1940s. But LLMs to some extent, except for a few domains, are mostly information retrieval systems. They can compress a lot of factual knowledge that has been previously produced by humans and can give easy access to it. In a way it's kind of a natural evolution of the printing press, the libraries, the internet and search engines, right? It's just a more efficient way to access information. And there are a few domains where the intelligent capabilities of those systems actually is more than that. It's more than just retrieval. So for generating code, maybe doing some type of mathematics that we're getting the impression that it's beyond this, but it's still to a large extent domains where reasoning has to do with manipulating symbols. The problem is that why do we have systems that can pass the bar?