Can AI Match the Human Brain? | Surya Ganguli | TED

翻译人员: Yip Yan Yeung 校对人员: Lening Xu

So what the heck happened in the field of AI in the last decade?

过去的十年里， AI 领域到底发生了什么？

It's like a strange new type of intelligence

就像我们的星球上 出现了一种奇怪的新型智能。

appeared on our planet.

But it's not like human intelligence.

但它和人类智能不一样。

It has remarkable capabilities,

它有非凡的能力，

but it also makes egregious errors that we never make.

但它也会犯下 我们不会犯的严重错误。

And it doesn't yet do the deep logical reasoning that we can do.

而且它还做不到 我们的深度逻辑推理。

It has a very mysterious surface of both capabilities and fragilities.

表面看起来它很神秘， 既有能力又脆弱。

And we understand almost nothing about how it works.

而我们对它的运作原理 几乎一无所知。

I would like a deeper scientific understanding of intelligence.

我想对智能有更深入的科学理解。

But to understand AI,

但是要理解 AI，

it's useful to place it in the historical context

将其置于生物智能的 历史背景下很有用。

of biological intelligence.

The story of human intelligence

人类智能的故事 不妨从这个小动物开始说起。

might as well have started with this little critter.

It's the last common ancestor of all vertebrates.

它是所有脊椎动物 最后一个共同祖先。

We are all descended from it.

我们都是它的后裔。

It lived about 500 million years ago.

它生活在大约 5 亿年前。

Then evolution went on to build the brain, which in turn,

然后，进化持续培育大脑，

in the space of 500 years from Newton to Einstein,

而大脑反过来在 从牛顿到爱因斯坦的 500 年间

developed the deep math and physics

发展了理解宇宙所需的 深度数学和物理学，从夸克到宇宙学。

required to understand the universe, from quarks to cosmology.

And it did this all without consulting ChatGPT.

这一切都是在没有咨询 ChatGPT 的前提下完成的。

And then, of course, there's the advances of the last decade.

当然，还有过去十年里的进步。

To really understand what just happened in AI,

要真正了解 AI 领域刚刚发生了什么，

we need to combine physics, math,

我们需要将物理学、数学、

neuroscience, psychology, computer science and more,

神经科学、心理学、 计算机科学等相结合，

to develop a new science of intelligence.

发展出一门新的智能科学。

The science of intelligence

智能的科学

can simultaneously help us understand biological intelligence

可以同时帮助我们 理解生物智能，

and create better artificial intelligence.

又创造更好的人工智能。

And we need this science now,

我们现在需要这门科学，

because the engineering of intelligence

因为智能的创造力 远超我们对它的理解能力。

has vastly outstripped our ability to understand it.

I want to take you on a tour of our work in the science of intelligence

我想带大家浏览一下 我们在智能科学上的研究，

that addresses five critical areas in which AI can improve --

包括 AI 可以改进的五个关键领域——

data efficiency, energy efficiency, going beyond evolution,

数据效率、能源效率、超越进化、

explainability and melding minds and machines.

可解释性以及 人类思维和机器的融合。

Let's address these critical gaps one by one.

让我们一一介绍这些关键缺口。

First, data efficiency.

首先，数据效率。

AI is vastly more data-hungry than humans.

AI 比人类更渴望数据。

For example, we train our language models on the order of one trillion words now.

举个例子，我们现在用大约一万亿个单词 来训练我们的语言模型。

Well, how many words do we get?

那我们会用多少个词呢？

Just 100 million.

只要 1 亿个。

It's that tiny little red dot at the center.

就是中间那个小红点。

You might not be able to see it.

你可能都看不见。

It would take us 24,000 years to read the rest of the one trillion words.

我们需要 24,000 年 才能读完一万亿字中剩下的部分。

OK, now, you might say that's unfair.

好吧，你可能会说这不公平。

Sure, AI read for 24,000 human-equivalent years,

AI 的阅读时间确实 相当于人类 24,000 年，

but humans got 500 million years of vertebrate brain evolution.

但是人类经历了 5 亿年的 脊椎动物大脑进化。

But there's a catch.

但是有一个问题。

Your entire legacy of evolution is given to you through your DNA,

你的整个进化遗传是通过DNA进行的，

and your DNA is only about 700 megabytes,

而你的 DNA 只有大约 700 兆字节，

or equivalently, 600 million [words].

相当于 6 亿个词。

So the combined information we get from learning and evolution

因此，与 AI 相比， 我们从学习和进化中

is minuscule compared to what AI gets.

获得的综合信息微不足道。

You are all incredibly efficient learning machines.

你们都是非常高效的学习机器。

So how do we bridge the gap between AI and humans?

那么，我们如何弥合 AI 与人类之间的鸿沟呢？

We started to tackle this problem by revisiting the famous scaling laws.

我们从重新审视著名的规模法则 开始解决这个问题。

Here's an example of a scaling law,

举一个规模法则的例子，

where error falls off as a power law with the amount of training data.

误差随着训练数据量的增加 而按照幂律分布降低。

These scaling laws have captured the imagination of industry

这些规模法则激发了行业的想象力，

and motivated significant societal investments

并促进了对能源、计算和 数据采集的重大社会投资。

in energy, compute and data collection.

But there's a problem.

但是有一个问题。

The exponents of these scaling laws are small.

这些规模法则的指数很小。

So to reduce the error by a little bit,

如果要将误差减少一点，

you might need to ten-x your amount of training data.

可能需要将训练数据量增加十倍。

This is unsustainable in the long run.

从长远来看，这是不可持续的。

And even if it leads to improvements in the short run,

即使它能在短期内带来改进，

there must be a better way.

也必须找到更好的方法。

We developed a theory that explains why these scaling laws are so bad.

我们研究了一种理论来解释 为什么这些规模法则如此糟糕。

The basic idea is that large random datasets are incredibly redundant.

基本逻辑是大型随机数据集非常冗余。

If you already have billions of data points,

如果你已经有了上亿个数据点，

the next data point doesn't tell you much that's new.

那么下一个数据点 并不会给你什么新信息。

But what if you could create a nonredundant dataset,

但是，如果你能创建一个非冗余数据集，

where each data point is chosen carefully

每个数据点都是精心挑选的，

to tell you something new, compared to all the other data points?

给你一些其他数据点 不能提供的新信息，会怎么样呢？

We developed theory and algorithms to do just this.

我们开发了理论 和算法来做到这一点。

We theoretically predicted and experimentally verified

通过理论预测并经实验验证，

that we could bend these bad power laws down to much better exponentials,

我们可以把这些不佳的幂律 改良成更好的指数，

where adding a few more data points could reduce your error,

通过再增加几个数据点减少误差，

rather than ten-xing the amount of data.

而不是增加十倍的数据量。

So what theory did we use to get this result?

那么我们是用什么理论 得出这个结果的呢？

We used ideas from statistical physics, and these are the equations.

我们使用了统计物理学的思想， 用了这些方程。

Now, for the rest of this entire talk,

在这场演讲剩下的时间里，

I'm going to go through these equations one by one.

我将逐一介绍这些方程。

(Laughter)

（笑声）

You think I'm joking?

你们以为我在开玩笑吗？

And explain them to you.

给大家解释这些方程。

OK, you're right, I'm joking. I'm not that mean.

没错，我在开玩笑。 我没那么坏。

But you should have seen the faces of the TED organizers

但你该看看我刚说要这么做的时候 TED 组织者是什么样的表情。

when I said I was going to do that.

Alright, let's move on.

好吧，我们继续。

Let's zoom out a little bit,

我们退后一步看，

and think more generally

更宽泛地思考一下 如何让 AI 减少对数据的需求。

about what it takes to make AI less data-hungry.

Imagine if we trained our kids

想象一下，如果我们像预训练 大语言模型一样训练孩子，

the same way we pretrain our large language models,

by next-word prediction.

也就是预测下一个单词。

So I'd give my kid a random chunk of the internet and say,

我会随便找一个 网上的片段给我的孩子，说：

"By the way, this is the next word."

“你看，这是下一个词。”

I'd give them another random chunk of the internet and say,

我再随便找一个网上的片段，说：

"This is the next word."

“这是下一个词。”

If that's all we did,

如果我们只做这些，

it would take our kids 24,000 years to learn anything useful.

那我们的孩子需要 24,000 年 才能学到有用的东西。

But we do so much more than that.

但我们做的远不止于此。

For example, when I teach my son math,

比如，在我教儿子数学时，

I teach him the algorithm required to solve the problem,

我会教他解出答案所需的算法，

then he can immediately solve new problems

然后他立即就可以解答新问题

and generalize using far less training data than any AI system would do.

并归纳总结，需要的训练数据 比任何 AI 系统都少。

I don't just throw millions of math problems at him.

我没有只丢给他几百万道数学题。

So to really make AI more data-efficient,

因此，要真正提高 AI 的数据效率，

we have to go far beyond our current training algorithms

我们必须大大突破当前的训练算法，

and turn machine learning into a new science of machine teaching.

将机器学习转变为 一门新的机器教学科学。

And neuroscience, psychology and math can really help here.

神经科学、心理学和数学 在这方面确实有用。

Let's go on to the next big gap, energy efficiency.

我们来看下一个大缺口—— 能源效率。

Our brains are incredibly efficient.

我们的大脑非常高效。

We only consume 20 watts of power.

我们只消耗 20 瓦的功率。

For reference, our old light bulbs were 100 watts.

作为参考， 我们以前的灯泡是 100 瓦。

So we are all literally dimmer than light bulbs.

我们其实比灯泡都暗。

(Laughter)

（笑声）

But what about AI?

但是 AI 呢？

Training a large model can consume as much as 10 million watts,

训练一个大模型 可能要消耗高达 1000万瓦特，

and there’s talk of going nuclear to power one-billion-watt data centers.

而且有传言称要使用核能 为 10 亿瓦特的数据中心供电。

So why is AI so much more energy-hungry than brains?

为什么 AI 比大脑更耗能呢？

Well, the fault lies in the choice of digital computation itself,

问题在于选择了数字计算，

where we rely on fast and reliable bit flips

计算过程中的每个步骤 都依赖快速可靠的位翻转。

at every intermediate step of the computation.

Now, the laws of thermodynamics

热力学定律

demand that every fast and reliable bit flip must consume a lot of energy.

决定了每次快速可靠的位翻转 都必须消耗大量能量。

Biology took a very different route.

生物学走了一条截然不同的道路。

Biology computes the right answer just in time,

生物学及时计算出了正确的答案，

using intermediate steps that are as slow and as unreliable as possible.

经历的中间步骤都相当缓慢又不可靠。

In essence, biology does not rev its engine any more than it needs to.

也就是说生物学的引擎转速 不会超过其需要的速度。

In addition, biology matches computation to physics much better.

此外，生物学可以更好地 适配计算和物理学。

Consider, for example, addition.

比如加法。

Our computers add using really complex energy-consuming transistor circuits,

我们的计算机利用非常复杂 又耗能的晶体管电路进行加法运算，

but neurons just directly add their voltage inputs,

但是神经元只需要直接将电压输入相加，

because Maxwell's laws of electromagnetism already know how to add voltage.

因为麦克斯韦的电磁定律 已经知道了如何让电压相加。

In essence, biology matches its computation

本质上，生物学将它的计算 与宇宙的自然物理学相匹配。

to the native physics of the universe.

So to really build more energy-efficient AI,

要真正构建更节能的 AI，

we need to rethink our entire technology stack,

我们需要重新考虑整个技术栈，

from electrons to algorithms,

从电子到算法，

and better match computational dynamics to physical dynamics.

更好地将计算的模式 与物理的模式相匹配。

For example, what are the fundamental limits

例如，在一定的能耗预算下，

on the speed and accuracy of any given computation,

任何计算的速度和准确率的 基本极限是什么？

given an energy budget?

And what kinds of electrochemical computers can achieve

什么样的电化学计算机 可以达到这些基本极限呢？

these fundamental limits?

We recently solved this problem for the computation of sensing,

我们最近为感知计算 解决了这个问题，

which is something that every neuron has to do.

每个神经元都在进行感知计算。

We were able to find fundamental lower bounds or lower limits on the error

我们找到了随能耗预算变化的 误差的下界或下限。

as a function of the energy budget.

That's that red curve.

就是这条红色曲线。

And we were able to find the chemical computers that achieve these limits.

我们还找到了可以达到 这个极限的化学计算机。

And remarkably, they looked a lot like G-protein coupled receptors,

值得一提的是它们 长得很像 G 蛋白偶联受体，

which every neuron uses to sense external signals.

每个神经元都用它来感知外界信号。

So this suggests that biology can achieve amounts of efficiency

这表明生物学可以实现

接近于由物理定律设定的 基本极限的效率。

that are close to fundamental limits set by the laws of physics itself.

Popping up a level,

再往上一层，

neuroscience now gives us the ability to measure not only neural activity,

神经科学现在让我们 不仅可以测量神经活动，

but also energy consumption across, for example, the entire brain of the fly.

还可以测量苍蝇整个大脑的能耗。

The energy consumption is measured through ATP usage,

能量消耗是通过 ATP 的用量来测量的，

which is the chemical fuel that powers all neurons.

ATP 是为所有神经元 提供动力的化学燃料。

So now let me ask you a question.

现在我要问大家一个问题。

Let's say in a certain brain region, neural activity goes up.

假设在某个大脑区域， 神经活动增加了。

Does the ATP go up or down?

ATP 会上升还是下降？

A natural guess would be that the ATP goes down,

凭直觉的猜测会是 ATP 下降，

because neural activity costs energy, so it's got to consume the fuel.

因为神经活动消耗能量， 所以它肯定会消耗燃料。

We found the exact opposite.

我们发现事实恰恰相反。

When neural activity goes up,

当神经活动增加时，

ATP goes up and it stays elevated

ATP 会升高并保持较高的水平，

just long enough to power expected future neural activity.

刚好足以为未来可预期的 神经活动提供动力。

This suggests that the brain follows a predictive energy allocation principle,

这表明大脑遵循预测能量分配原理，

where it can predict how much energy is needed, where and when,

它预测什么时候、 在哪里需要多少能量，

and it delivers just the right amount of energy at just the right location,

并在合适的一段时间内、 在合适的位置提供正好的能量。

for just the right amount of time.

So clearly, we have a lot to learn from physics, neuroscience and evolution

很明显，在构建 更节能的 AI 这一方面，

我们可以从物理学、神经科学 和进化中学到很多东西。

about building more energy-efficient AI.

But we don't need to be limited by evolution.

但我们不必被进化限制住。

We can go beyond evolution,

我们可以超越进化，

to co-opt the neural algorithms discovered by evolution,

采纳进化带来的神经算法，

but implement them in quantum hardware that evolution could never figure out.

但把它们用在 进化永远无法搞明白的量子硬件上。

For example, we can replace neurons with atoms.

比如，我们可以用原子代替神经元。

The different firing states of neurons

神经元的各种放电状态

correspond to the different electronic states of atoms.

对应原子的各种电子态。

And we can replace synapses with photons.

我们可以用光子代替突触。

Just as synapses allow two neurons to communicate,

就像突触能让两个神经元通信一样，

photons allow two atoms to communicate through photon emission and absorption.

光子能让两个原子 通过光子发射和吸收进行通信。

So what can we build with this?

那我们可以用它来做些什么呢？

We can build a quantum associative memory out of atoms and photons.

我们可以用原子和光子 建立‌量子联想记忆。

This is the same memory system

这就是约翰·霍普菲尔德（John Hopfield） 最近获得诺贝尔物理学奖的存储系统，

that won John Hopfield his recent Nobel Prize in physics,

but this time, it's a quantum-mechanical system built of atoms and photons,

但这里它是一个由原子和光子 构成的量子力学系统，

and we can analyze its performance

我们可以分析其性能

and show that the quantum dynamics yields enhanced memory capacity,

并证明量子动力学可以增强存储容量、

robustness and recall.

稳健性和召回率。

We can also build new types of quantum optimizers built directly out of photons,

我们还可以建造直接由光子构建的 新型量子优化器，

and we can analyze their energy landscape

我们可以分析它们的能量分布，

and explain how they solve optimization problems in fundamentally new ways.

解释它们如何以全新的方式 解决优化问题。

This marriage between neural algorithms and quantum hardware

神经算法和量子硬件之间的这种结合

opens up an entirely new field,

开辟了一个全新的领域，

which I like to call quantum neuromorphic computing.

我称之为“量子神经形态计算”。

OK, but let's return to the brain,

我们再说回大脑，

where explainable AI can help us understand how it works.

可解释的 AI 可以帮助我们 理解它的运作原理。

So now, AI allows us to build

AI 让我们能够构建 非常准确但复杂的大脑模型。

incredibly accurate but complicated models of the brain.

So where is this all going?

会发展成什么样呢？

Are we simply replacing something we don't understand, the brain,

我们是不是只是 把我们不懂的东西，也就是大脑，

with something else we don't understand, our complex model of it?

换成了另一个我们不懂的东西， 也就是它的复杂模型？

As scientists, we'd like to have a conceptual understanding

作为科学家， 我们希望从概念上了解

of how the brain works,

大脑的运作原理，

not just have a model handed to us.

而不是丢给我们一个模型。

So basically, I'd like to give you

所以我想给大家介绍一个

an example of our work on explainable AI, applied to the retina.

我们在可解释 AI 上的工作， 用于视网膜。

So the retina is a multilayered circuit of photoreceptors

视网膜是一个由感光器组成的多层回路，

going to hidden neurons, going to output neurons.

联结着隐藏神经元， 联结着输出神经元。

So how does it work?

它是如何运作的呢？

Well, we recently built the world's most accurate model of the retina.

我们最近建立了 世界上最精准的视网膜模型。

It could reproduce two decades of experiments on the retina.

它可以复现二十年以来的视网膜实验。

So this is fantastic.

太棒了。

We have a digital twin of the retina.

我们有视网膜的数字孪生。

But how does the twin work?

但是数字孪生如何运作？

Why is it designed the way it is?

为什么这么设计它？

To make these questions concrete,

为了让这些问题更具体，

I'd like to discuss just one

我想谈一谈我刚说到的 二十年以来的其中一个实验。

of the two decades of experiments that I mentioned.

And we're going to do this experiment on you right now.

现在我想跟大家进行这个实验。

I'd like you to focus on my hand, and I'd like you to track it.

请看着我的手，跟着它。

OK, great. Let's do that just one more time.

很好。再来一次。

OK.

好的。

You might have been slightly surprised when my hand reversed direction.

我的手转向的时候， 你可能会有一些意外。

And you should be surprised,

你确实应该有些意外，

because my hand just violated Newton's first law of motion,

因为我的手刚刚违反了 牛顿第一运动定律，

which states that objects that are in motion tend to remain in motion.

即运动中的物体往往会保持运动。

So where in your brain is a violation of Newton's first law first detected?

你大脑中的哪一块 最先发现违反了牛顿第一定律呢？

The answer is remarkable. It's in your retina.

答案很不寻常。 是你的视网膜。

There are neurons in your retina that will fire

视网膜中有神经元会在 且仅在违反牛顿第一定律时放电。

if and only if Newton's first law is violated.

So does our model do that?

我们的模型能做到这一点吗？

Yes, it does. It reproduces it.

可以。它复现了这个现象。

But now, there's a puzzle.

但这样就出现了一个谜题。

How does the model do it?

模型是怎么做到的？

Well, we developed methods, explainable AI methods,

我们开发了一些方法， 可解释的 AI 方法，

to take any given stimulus that causes a neuron to fire,

接受任何导致神经元放电的刺激，

and we carve out the essential subcircuit responsible for that firing,

然后我们创建了 导致这种放电的核心子电路，

and we explain how it works.

并解释其运作原理。

We were able to do this not only for Newton's first law violations,

我们的模型可以在 违反牛顿第一定律时做到这一点，

but for the two decades of experiments that our model reproduced.

也可以在我们的模型所复现的 二十年来的实验中做到。

And so this one model reproduces two decades' worth of neuroscience

这一个模型复现了 神经科学二十年以来的成果，

and also makes some new predictions.

还做出了新的预测。

This opens up a new pathway to accelerating neuroscience discovery

这为使用 AI 加速神经科学发现 开辟了一条新途径。

using AI.

Basically, build digital twins of the brain,

构建大脑的数字孪生，

and then use explainable AI to understand how they work.

再使用可解释的 AI 了解它们的运作原理。

We're actually engaged in a big effort at Stanford

我们在斯坦福大学投入了大量精力，

to build a digital twin of the entire primate visual system

构建整个灵长类动物视觉系统的数字孪生

and explain how it works.

并解释其运作原理。

But we can go beyond that

但是我们可以超越这一点，

and use our digital twins to meld minds and machines,

使用我们的数字孪生 融合人类思维和机器，

by allowing bidirectional communication between them.

通过构建它们之间的双向交流。

So imagine a scenario where you have a brain,

想象一下这样的场景： 你有一个大脑，

you record from it, you build a digital twin.

把它录下来， 构建一个数字孪生。

Then you use control theory to learn neural activity patterns

然后，你使用控制理论 学习神经活动模式，

that you can write directly into the digital twin to control it.

再将这些模式直接写入 数字孪生中来控制它。

Then, you take those same neural activity patterns

然后，你把这些神经活动模式 写入大脑来控制大脑。

and you write them into the brain to control the brain.

In essence, we can learn the language of the brain,

我们其实可以学习大脑的语言，

and then speak directly back to it.

再直接回复它。

So we recently carried out this program in mice,

我们最近在小鼠身上 进行了这个实验，

where we could use AI to read the mind of a mouse.

使用 AI 读懂小鼠的思想。

So on the top row, you're seeing images that we actually showed to the mouse,

上面一行是我们给小鼠看的图片，

and in the bottom row,

下面一行是我们从小鼠脑中 解码出来的图片。

you're seeing images that we decoded from the brain of the mouse.

Our decoded images are lower-resolution than the actual images,

解码图片的分辨率 低于实际图片，

but not because our decoders are bad.

但不是因为我们的解码器不好。

It's because mouse visual resolution is bad.

这是因为小鼠的视觉分辨率很差。

So actually, the decoded images

解码后的图像其实

show you what the world would actually look like

向你展示了如果你是一只小鼠， 你会看到怎样的世界。

if you were a mouse.

Now, we can go beyond that.

我们还可以更进一步。

We can now write neural activity patterns into the mouse's brain,

我们现在可以将神经活动模式 写入小鼠的大脑，

so we can make it hallucinate

这样我们就可以让它产生幻觉，

any particular percept we would like it to hallucinate.

产生我们希望它有的某种幻觉。

And we got so good at this

我们在这方面做得非常出色，

that we could make it reliably hallucinate a percept

我们可以让它可靠地产生幻觉，

by controlling only 20 neurons in the mouse's brain,

只需要控制小鼠大脑中的 20 个神经元，

by figuring out the right 20 neurons to control.

找出应该控制的那 20 个神经元。

So essentially, we can control what the mouse sees

也就是说我们可以控制小鼠看到什么，

directly, by writing to its brain.

只需直接写入它的大脑。

The possibilities of bidirectional communication

大脑和机器之间 进行双向通信的可能性是无限的。

between brains and machines are limitless.

To understand, to cure and to augment the brain.

理解、治愈、增强大脑。

So I hope you'll see that the pursuit of a unified science of intelligence

我希望你们能看到，追求跨越

大脑和机器的统一智能科学

that spans brains and machines

can both help us better understand biological intelligence

既可以帮助我们更好地理解生物智能，

and help us create more efficient, explainable

也可以帮助我们创造 更高效、更可解释

and powerful artificial intelligence.

和更强大的人工智能。

But it's important that this pursuit be done out in the open

但是，这一追求必须公开进行，

so the science can be shared with the world,

这样科学才能与世界共享，

and it must be done with a very long time horizon.

而且必须在很长的时间内完成。

This makes academia the perfect place to pursue a science of intelligence.

这使学术界成为 追求智能科学的理想场所。

In academia, we're free from the tyranny of quarterly earnings reports.

在学术界，我们不必承受 季度收益报告的压榨。

We're free from the censorship of corporate legal departments.

我们不必接受公司法务部门的审查。

We can be far more interdisciplinary than any one company.

我们比任何一家公司更具跨学科性。

And our very mission is to share what we learn with the world.

而我们的使命就是与世界分享 我们学到的东西。

For all these reasons, we're actually building a new center

出于这些原因，我们正在斯坦福 建立一个新的智能科学中心。

for the science of intelligence at Stanford.

While there have been incredible advances in industry

虽然工业界在创造智能上 取得了惊人的进步，

on the engineering of intelligence,

now increasingly happening behind closed doors,

但封闭也越来越多，

I'm very excited about what the science of intelligence can achieve

我非常看好智能科学 在开放环境下所能取得的成就。

out in the open.

You know, in the last century,

上个世纪，

one of the greatest intellectual adventures

人类智慧最伟大的冒险之一

lay in humanity peering outwards into the universe

是人类向外凝视宇宙、

to understand it, from quarks to cosmology.

理解宇宙，从夸克到宇宙学。

I think one of the greatest intellectual adventures of this century

我认为，本世纪人类智慧 最伟大的冒险之一

will lie in humanity peering inwards,

将是人类向内凝视，

both into ourselves and into the AIs that we create,

凝视我们自己， 也凝视我们创造的 AI，

in order to develop a deeper, new scientific understanding of intelligence.

对智能形成更深入的科学新理解。

Thank you.

谢谢。

(Applause)

（掌声）