Jeff Dean: AI isn't as smart as you think -- but it could be | TED

翻译人员: Jennifer Huang 校对人员: Lexi Ding

Hi, I'm Jeff.

大家好，我是杰夫。

I lead AI Research and Health at Google.

我领导着谷歌的 人工智能研究与健康部门。

I joined Google more than 20 years ago,

我在二十多年前加入谷歌，

when we were all wedged into a tiny office space,

那时候我们还挤在 一间狭小的办公室里，

above what's now a T-Mobile store in downtown Palo Alto.

就位于现在帕罗阿托市中心的 一家德国电信商店上面。

I've seen a lot of computing transformations in that time,

在那段时间，我见证了 许多计算机技术的发展，

and in the last decade, we've seen AI be able to do tremendous things.

在过去十年，我们见证了 人工智能完成许多了不起的事情。

But we're still doing it all wrong in many ways.

但我们仍在许多方面都没有做对。

That's what I want to talk to you about today.

这是我今天想讨论的主题。

But first, let's talk about what AI can do.

不过首先，我们先来聊聊 人工智能可以做什么。

So in the last decade, we've seen tremendous progress

在过去的十年间， 我们看到了人工智能

in how AI can help computers see, understand language,

在帮助计算机识别物体、 理解语言和谈话方面

understand speech better than ever before.

取得的巨大进步。

Things that we couldn't do before, now we can do.

以往的天方夜谭 现在一一成为现实。

If you think about computer vision alone,

就单单拿计算机视觉来说，

just in the last 10 years,

在过去的十年中，

computers have effectively developed the ability to see;

电脑快速地发展出了‘看’的能力；

10 years ago, they couldn't see, now they can see.

十年前，它们看不到， 但现在可以了。

You can imagine this has had a transformative effect

这在计算机运用上

on what we can do with computers.

具有变革性的影响。

让我们来看几个由这些功能促成的

So let's look at a couple of the great applications

了不起的实际应用。

enabled by these capabilities.

我们可以通过机器学习 更准确地预测洪水，

We can better predict flooding, keep everyone safe,

保证所有人的安全。

using machine learning.

我们可以翻译一百多种语言 以便更顺畅地交流，

We can translate over 100 languages so we all can communicate better,

还可以更精准地预测和诊断疾病，

and better predict and diagnose disease,

where everyone gets the treatment that they need.

让所有人得到所需的治疗。

So let's look at two key components

让我们来看看

that underlie the progress in AI systems today.

构成当代人工智能系统 基础的两个关键元素。

The first is neural networks,

首先是神经网络，

a breakthrough approach to solving some of these difficult problems

它是解决这些难题的一项重大突破，

that has really shone in the last 15 years.

它在过去的十五年已经大放异彩。

But they're not a new idea.

但它并不是一个新的点子。

And the second is computational power.

第二个是运算能力。

It actually takes a lot of computational power

驱动神经网络运作

to make neural networks able to really sing,

实际需要大量的运算能力，

and in the last 15 years, we’ve been able to halve that,

在过去的十五年， 我们做到了使其减半，

and that's partly what's enabled all this progress.

那也是整个人工智能 得以发展至此的原因之一。

But at the same time, I think we're doing several things wrong,

但同时，我认为我们做错了几件事，

and that's what I want to talk to you about

那也是我想在这次演讲的最后

at the end of the talk.

和大家谈论的。

First, a bit of a history lesson.

首先，给大家上点历史课。

So for decades,

数十年间，

almost since the very beginning of computing,

几乎从计算机科学最早出现，

people have wanted to be able to build computers

人们就想建造

that could see, understand language, understand speech.

可以识别语言及理解谈话的电脑。

The earliest approaches to this, generally,

最初的方法一般是

people were trying to hand-code all the algorithms

人们手动写下

that you need to accomplish those difficult tasks,

完成难题所需的算法。

and it just turned out to not work very well.

但成效一般。

But in the last 15 years, a single approach

但在过去的十五年间，

一个方法出其不意地 一次性解决了所有难题：

unexpectedly advanced all these different problem spaces all at once:

neural networks.

神经网络。

So neural networks are not a new idea.

神经网络并非一个新想法。

They're kind of loosely based

它们大致上基于

on some of the properties that are in real neural systems.

一些真实神经系统的特性。

And many of the ideas behind neural networks

大部分神经网络背后的理念

have been around since the 1960s and 70s.

出现于1960和70年代。

A neural network is what it sounds like,

神经网络如同其字面意思一样，

a series of interconnected artificial neurons

是一连串互相连接的神经元。

that loosely emulate the properties of your real neurons.

它们大致上效仿了 人体真正神经元的特性。

An individual neuron in one of these systems

这种系统中的一个独立神经元，

has a set of inputs,

拥有一组输入信息，

each with an associated weight,

每组输入信息有对应的比重，

and the output of a neuron

神经元的信息输出

is a function of those inputs multiplied by those weights.

就等于那些输入信息 乘以它们对应的比重。

So pretty simple,

所以其实挺简单的，

and lots and lots of these work together to learn complicated things.

无数神经元协同运作 就可以学习复杂的东西。

So how do we actually learn in a neural network?

所以我们到底是 如何在神经网络中学习的？

It turns out the learning process

其实，在学习过程中，

consists of repeatedly making tiny little adjustments

比重在不断被微调，

to the weight values,

strengthening the influence of some things,

增强一些东西的影响，

weakening the influence of others.

削弱其他的影响。

By driving the overall system towards desired behaviors,

在驱使整个系统去往 理想运作模式的过程中，

these systems can be trained to do really complicated things,

可以训练其完成 一些非常复杂的事情，

like translate from one language to another,

例如翻译语言，

detect what kind of objects are in a photo,

识别照片中的物体，

all kinds of complicated things.

任何类型的复杂工作。

I first got interested in neural networks

我对神经网络的兴趣，

when I took a class on them as an undergraduate in 1990.

始于1990年本科阶段时 学到的一门相关课程。

At that time,

那时，

neural networks showed impressive results on tiny problems,

神经网络在精细问题的解决上 取得了惊人的成果，

but they really couldn't scale to do real-world important tasks.

但还达不到完成真实世界中 重要工作的程度。

But I was super excited.

但我非常兴奋。

(Laughter)

（笑声）

I felt maybe we just needed more compute power.

我觉得我们可能只是 需要更强的运算能力。

And the University of Minnesota had a 32-processor machine.

明尼苏达大学当时 有一个32位处理器。

I thought, "With more compute power,

我想：“如果有更强的运算能力， 我们真能用神经网络干点大事。”

boy, we could really make neural networks really sing."

So I decided to do a senior thesis on parallel training of neural networks,

所以我决定以神经网络的并行训练 作为我毕业论文的课题，

the idea of using processors in a computer or in a computer system

理念是将电脑或电脑系统中 所有的处理器

to all work toward the same task,

运用到同一件任务上，

that of training neural networks.

用来训练神经网络。

32 processors, wow,

32位处理器，哇，

we’ve got to be able to do great things with this.

我们肯定能用它做点大事。

But I was wrong.

但我错了。

Turns out we needed about a million times as much computational power

我发现如果想用神经网络 做些引人注目的事情，

as we had in 1990

before we could actually get neural networks to do impressive things.

我们所需的算力大概是 90年代算力的一百万倍。

But starting around 2005,

但从大概2005年开始，

thanks to the computing progress of Moore's law,

多亏了摩尔定律，

we actually started to have that much computing power,

我们真的开始拥有这么多运算能力了，

and researchers in a few universities around the world started to see success

世界上一些大学里的研究员们

开始成功用神经网络完成各种任务。

in using neural networks for a wide variety of different kinds of tasks.

I and a few others at Google heard about some of these successes,

我和其他几个在谷歌的同事 听闻了这些成功事例，

and we decided to start a project to train very large neural networks.

于是决定启动一个项目， 训练大型神经网络。

One system that we trained,

其中一个系统，

we trained with 10 million randomly selected frames

我们用油管视频里随机截取的 一千万帧照片对其进行训练。

from YouTube videos.

The system developed the capability

这个系统发展出了能够识别 所有不同种类物体的能力。

to recognize all kinds of different objects.

And it being YouTube, of course,

然后因为是油管的关系，

it developed the ability to recognize cats.

所以它发展出了识别猫的能力。

YouTube is full of cats.

油管上全是猫。

(Laughter)

（笑声）

But what made that so remarkable

但让它如此引人注目的是

is that the system was never told what a cat was.

从未有人告诉过这个系统 猫到底是什么。

So using just patterns in data,

仅仅依靠数据的形态规律，

the system honed in on the concept of a cat all on its own.

它就能自己琢磨出来猫究竟是什么。

All of this occurred at the beginning of a decade-long string of successes,

所有的这一切标志着 一个长达十年的，

of using neural networks for a huge variety of tasks,

在谷歌及其它地方成功运用神经网络 完成繁杂任务的开始。

at Google and elsewhere.

Many of the things you use every day,

这些你日常使用的东西，

things like better speech recognition for your phone,

例如手机更准确的语音识别系统，

improved understanding of queries and documents

提升对问题和文本的理解力，

for better search quality,

优化搜索结果，

better understanding of geographic information to improve maps,

对地理信息更完善的理解等等 提升地图及其他方面。

and so on.

Around that time,

在那个时候，

we also got excited about how we could build hardware that was better tailored

我们还对如何打造一个更适合 神经网络运算所需的计算机硬件感兴趣。

to the kinds of computations neural networks wanted to do.

神经网络运算有两个特性。

Neural network computations have two special properties.

The first is they're very tolerant of reduced precision.

第一个是它们对精准度要求很低。

Couple of significant digits, you don't need six or seven.

几个有效位就够了， 不需要六七个那么多。

And the second is that all the algorithms are generally composed

第二个是所有算法都普遍由多个 不同的矩阵和向量的运算组成。

of different sequences of matrix and vector operations.

So if you can build a computer

所以如果你可以造一台 擅长低精准度矩阵及向量运算

that is really good at low-precision matrix and vector operations

but can't do much else,

但除此之外没啥用的电脑，

that's going to be great for neural-network computation,

它会非常适用于神经网络运算，

虽然你无法用它做太多别的事。

even though you can't use it for a lot of other things.

如果你造出类似的东西， 人们会发现它的妙用。

And if you build such things, people will find amazing uses for them.

这是我们制作的 第一个成品，TPU v1。

This is the first one we built, TPU v1.

"TPU" stands for Tensor Processing Unit.

“TPU”是张量处理器的意思。

These have been used for many years behind every Google search,

多年来，这一技术运用于谷歌搜索，

for translation,

翻译，

in the DeepMind AlphaGo matches,

以及AlphaGo围棋比赛，

so Lee Sedol and Ke Jie maybe didn't realize,

所以李世石和柯洁可能没意识到，

but they were competing against racks of TPU cards.

他们其实是在和TPU架构比赛。

And we've built a bunch of subsequent versions of TPUs

后来我们制作了很多个 TPU的后续版本，

that are even better and more exciting.

他们更强大也更振奋人心。

But despite all these successes,

但即使我们取得了这些成功，

I think we're still doing many things wrong,

我觉得我们仍然做错了很多事，

接下来我将讲三件我们做错的事情，

and I'll tell you about three key things we're doing wrong,

以及如何修正他们。

and how we'll fix them.

第一个是，现如今的大部分神经网络

The first is that most neural networks today

are trained to do one thing, and one thing only.

只被训练进行单一种类的任务。

You train it for a particular task that you might care deeply about,

你训练它去做一件你很关心的事情，

but it's a pretty heavyweight activity.

但这是一项非常繁重的工作。

You need to curate a data set,

你需要搜索数据组，

you need to decide what network architecture you'll use

选择这个问题所需的网络架构，

for this problem,

you need to initialize the weights with random values,

接着随机分配起始比重，

apply lots of computation to make adjustments to the weights.

然后为调整比重进行大量运算。

And at the end, if you’re lucky, you end up with a model

到最后，如果你幸运的话，

that is really good at that task you care about.

可以得到一个非常适用于 你关心的问题的模型。

But if you do this over and over,

但如果你一直这样做，

you end up with thousands of separate models,

到最后会得到几千个独立的模型，

each perhaps very capable,

每个可能都很有用，

but separate for all the different tasks you care about.

但都只针对某个单一类型的问题。

But think about how people learn.

想一想人类是怎样学习的。

In the last year, many of us have picked up a bunch of new skills.

去年，我们中的很多人 都掌握了一些新的技能。

I've been honing my gardening skills,

我沉浸于钻研园艺，

experimenting with vertical hydroponic gardening.

尝试垂直水培园艺。

To do that, I didn't need to relearn everything I already knew about plants.

我无需为此重新学习一遍 我已经掌握的有关植物的知识。

I was able to know how to put a plant in a hole,

我知道怎么把植物放进洞里，

how to pour water, that plants need sun,

怎么浇水，以及植物需要光照，

and leverage that in learning this new skill.

我只需要整合这些知识 用以学习新的技术。

Computers can work the same way, but they don’t today.

电脑也可以这样运作， 但现如今它们并没有。

If you train a neural network from scratch,

如果你从头开始训练一个神经网络，

it's effectively like forgetting your entire education

每次学习新的东西

都像是让你忘掉 之前学习的所有知识。

every time you try to do something new.

That’s crazy, right?

这太疯狂了对吧！

So instead, I think we can and should be training

所以比起这样，我认为我们可以，

multitask models that can do thousands or millions of different tasks.

也应该去训练能完成成千上万种 不同任务的多任务处理模型。

Each part of that model would specialize in different kinds of things.

此模型的每个部分都有自己的专长。

And then, if we have a model that can do a thousand things,

然后，假设我们有一个 可以完成一千种任务的模型，

and the thousand and first thing comes along,

这时候第一千零一种任务来了，

we can leverage the expertise we already have

我们就可以整合系统中已有的 和这个新任务相关的知识，

in the related kinds of things

so that we can more quickly be able to do this new task,

从而更快速的完成这项新任务，

just like you, if you're confronted with some new problem,

就像你本人一样， 如果你面临了一些新问题，

you quickly identify the 17 things you already know

你会快速识别17件自己已知 并能够帮助解决这些新问题的知识。

that are going to be helpful in solving that problem.

Second problem is that most of our models today

第二个问题是，

大部分现今的模型 只能应对一种形态的数据，

deal with only a single modality of data --

with images, or text or speech,

图片，文字或语音，

but not all of these all at once.

但无法做到一网打尽。

But think about how you go about the world.

但想一想你如何在这世上生活。

You're continuously using all your senses

你不断地动用你所有的感官

to learn from, react to,

去学习，去做出反应，

figure out what actions you want to take in the world.

去搞清楚现在应该做什么。

Makes a lot more sense to do that,

这样显然更加合理，

and we can build models in the same way.

我们也可以用同样的方式建造模型。

We can build models that take in these different modalities of input data,

我们可以建造一个可以接收 所有不同种类数据的模型，

文字，图像，语音，

text, images, speech,

but then fuse them together,

然后把它们融合在一起，

so that regardless of whether the model sees the word "leopard,"

这样无论这个模型看到文字“豹子”，

sees a video of a leopard or hears someone say the word "leopard,"

看到豹子的视频，还是 听到有人说出“豹子”这个词

the same response is triggered inside the model:

它都会触发同样的反应：

the concept of a leopard

一个豹子的概念

can deal with different kinds of input data,

可以应对很多种不同的数据输入项，

even nonhuman inputs, like genetic sequences,

甚至是非人工的输入项， 例如基因序列，

3D clouds of points, as well as images, text and video.

3D点云数据，当然也包括 图片，文字和影像。

The third problem is that today's models are dense.

第三个问题是现今的模型是稠密的。

There's a single model,

拿一个模型来举例，

the model is fully activated for every task,

进行任何一项任务时 这个模型都需要被完全激活，

for every example that we want to accomplish,

任何一个我们想要完成的事情，

whether that's a really simple or a really complicated thing.

无论困难与否。

This, too, is unlike how our own brains work.

这再一次有悖人脑运作的习惯。

Different parts of our brains are good at different things,

大脑不同的区块擅长不同的工作，

and we're continuously calling upon the pieces of them

我们会不断的运用 与需要完成的任务相关的区块。

that are relevant for the task at hand.

For example, nervously watching a garbage truck

比如，紧张地看着一辆垃圾车 朝自家车的方向倒车，

back up towards your car,

the part of your brain that thinks about Shakespearean sonnets

你大脑中负责思考莎士比亚的 十四行诗的区块应该不会被激活。

is probably inactive.

(Laughter)

（笑声）

AI models can work the same way.

人工智能模型也可以这样运作。

Instead of a dense model,

比起稠密模型，

we can have one that is sparsely activated.

我们可以制造一个 激活反应较稀松的模型。

So for particular different tasks, we call upon different parts of the model.

在做不同任务的时候， 我们运用模型中不同的区块。

During training, the model can also learn which parts are good at which things,

在训练时，模型也可以学习 哪个区块擅长哪个领域，

to continuously identify what parts it wants to call upon

从而不断地识别需要运用 哪些区块来完成新的任务。

in order to accomplish a new task.

The advantage of this is we can have a very high-capacity model,

此类模型的优势在于 虽然容量大但十分高效，

but it's very efficient,

because we're only calling upon the parts that we need

因为我们只会用到 完成某项任务所需的区块。

for any given task.

So fixing these three things, I think,

因此我认为，解决了这三个问题，

will lead to a more powerful AI system:

人工智能的能力将更上一层楼：

instead of thousands of separate models,

比起几千个独立的模型， [独立模型/通用模型]

train a handful of general-purpose models

训练几个能够应对 成千上万件事情的通用模型。

that can do thousands or millions of things.

Instead of dealing with single modalities,

比起接收单一的数据形态，

deal with all modalities,

让它处理所有的数据形态 并把它们整合到一起。

and be able to fuse them together.

And instead of dense models, use sparse, high-capacity models,

比起稠密模型， 运用稀松且高容量的模型，

where we call upon the relevant bits as we need them.

让我们能按需激活指定区块。

We've been building a system that enables these kinds of approaches,

我们已经在制造 符合以上条件的模型了，

and we’ve been calling the system “Pathways.”

我们给它取了个名字 叫“Pathways”。

So the idea is this model will be able to do

我们的理念是

thousands or millions of different tasks,

这个模型可以完成 成千上万种不同类型的任务，

and then, we can incrementally add new tasks,

然后我们可以逐步增加新的任务，

and it can deal with all modalities at once,

它也可以同时处理各种形态的数据，

and then incrementally learn new tasks as needed

然后逐步学习新技能，

and call upon the relevant bits of the model

并按需为不同任务启动不同区块。

for different examples or tasks.

And we're pretty excited about this,

我们对此感到非常兴奋，

we think this is going to be a step forward

我们认为这将是人工智能 系统建造迈出的重要一步。

in how we build AI systems.

But I also wanted to touch on responsible AI.

不过我也想浅谈一下 什么是可靠的人工智能。

We clearly need to make sure that this vision of powerful AI systems

我们一定要确保这样强大的 人工智能系统造福所有人。

benefits everyone.

These kinds of models raise important new questions

此类模型将一些重要的 新问题带进大众的视野，

about how do we build them with fairness,

我们应该如何确保系统于所有人而言

interpretability, privacy and security,

都具有公平性、可解释性、 私密性以及安全性。

for all users in mind.

For example, if we're going to train these models

例如，如果我们要训练 这些模型完成成千上万种任务，

on thousands or millions of tasks,

we'll need to be able to train them on large amounts of data.

我们需要极大量的数据。

And we need to make sure that data is thoughtfully collected

我们一定要确保 这些数据的采集经过深思熟虑，

and is representative of different communities and situations

它们需要代表世界上 不同的社群和情况。

all around the world.

And data concerns are only one aspect of responsible AI.

数据上的担忧仅仅是 可靠的人工智能这个议题的一部分。

We have a lot of work to do here.

我们要做的还有很多。

So in 2018, Google published this set of AI principles

2018年，谷歌发表了在开发 此类科技时应注意的人工智能守则。

by which we think about developing these kinds of technology.

And these have helped guide us in how we do research in this space,

这帮助指导了我们 如何在这个领域从事研究，

how we use AI in our products.

以及如何在产品中使用人工智能。

And I think it's a really helpful and important framing

我认为这对思考复杂且有深度的问题，

for how to think about these deep and complex questions

及在社会中运用人工智能 都有帮助且很重要。

about how we should be using AI in society.

We continue to update these as we learn more.

我们在学习过程中不断更新该守则，

Many of these kinds of principles are active areas of research --

很多此类准则都是 现在的热点研究领域，

super important area.

它们非常重要。

Moving from single-purpose systems that kind of recognize patterns in data

由只能识别数据中的 图案规律的单一用途系统，

to these kinds of general-purpose intelligent systems

到对世界有更深理解的 通用智能系统的转变，

that have a deeper understanding of the world

will really enable us to tackle

真的可以赋予我们能力 去解决人类面临的重大问题。

some of the greatest problems humanity faces.

For example,

比如，

we’ll be able to diagnose more disease;

我们可以诊断更多疾病，

we'll be able to engineer better medicines

我们可以通过给这些模型灌输化学 和物理知识从而设计出更好的药品，

by infusing these models with knowledge of chemistry and physics;

we'll be able to advance educational systems

我们可以通过个性化定制的补习服务 来优化现有的教育系统，

by providing more individualized tutoring

to help people learn in new and better ways;

让人们用更先进的方法来学习，

we’ll be able to tackle really complicated issues,

我们可以解决十分复杂的难题，

like climate change,

例如全球变暖，

and perhaps engineering of clean energy solutions.

或是帮助清洁能源的发展。

So really, all of these kinds of systems

所以其实，所有此类系统

are going to be requiring the multidisciplinary expertise

都需要世界各地的 多学科专家共同协作。

of people all over the world.

So connecting AI with whatever field you are in,

将人工智能和你所在的领域相结合，

in order to make progress.

从而推动产业的进程。

So I've seen a lot of advances in computing,

我看到了许多计算机科学的优势，

and how computing, over the past decades,

以及在过去的几十年中计算机科学 如何帮助几百万人更好地理解世界。

has really helped millions of people better understand the world around them.

And AI today has the potential to help billions of people.

今天的人工智能拥有 帮助数十亿人的潜力。

We truly live in exciting times.

我们真的生活在 一个振奋人心的时代。

Thank you.

谢谢。

(Applause)

（掌声）

Chris Anderson: Thank you so much.

克里斯·安德森：非常感谢。

I want to follow up on a couple things.

我还有几个问题。

This is what I heard.

据我所知，

Most people's traditional picture of AI

大部分人对于人工智能的印象

is that computers recognize a pattern of information,

是电脑识别出信息的模式，

and with a bit of machine learning,

运用一些机器学习的技术，

they can get really good at that, better than humans.

就可以比人类更擅长某项工作。

What you're saying is those patterns

你所说的模式已经不再是 人工智能所处理的原子层面了，

are no longer the atoms that AI is working with,

that it's much richer-layered concepts

而是一种层次更加丰富的概念 比如组成一只美洲豹的，

that can include all manners of types of things

that go to make up a leopard, for example.

各种形式的各种信息。

So what could that lead to?

那这会产生什么样的结果呢？

Give me an example of when that AI is working,

请举个例子，当人们使用人工智能

what do you picture happening in the world

你预想接下来五到十年中 世界上会发生什么有趣的事情？

in the next five or 10 years that excites you?

Jeff Dean: I think the grand challenge in AI

杰夫·迪恩：我认为人工智能 面临的最大的挑战是

is how do you generalize from a set of tasks

如何尽可能简单且高效地

you already know how to do

to new tasks,

运用已知的知识来完成新任务。

as easily and effortlessly as possible.

And the current approach of training separate models for everything

现在我们训练一个通用模型的方法

means you need lots of data about that particular problem,

需要大量的与该问题 相关的数据支持，

because you're effectively trying to learn everything

因为我们在试图从零开始高效地学习 世界上包括那个问题在内的所有知识。

about the world and that problem, from nothing.

But if you can build these systems

不过如果你可以建造出一个

that already are infused with how to do thousands and millions of tasks,

能完成成千上万种任务的系统，

then you can effectively teach them to do a new thing

你就只需要用较少的数据 去教会它一个新的东西。

with relatively few examples.

So I think that's the real hope,

我认为这是一个希望，

that you could then have a system where you just give it five examples

只需给这个系统五个 和你想完成的任务相关的事例，

of something you care about,

and it learns to do that new task.

它就能学会这个新技能。

CA: You can do a form of self-supervised learning

克里斯：你只需要一点点‘肥料’ 就能打造一个自我监督学习系统。

that is based on remarkably little seeding.

JD: Yeah, as opposed to needing 10,000 or 100,000 examples

杰夫：是的，比起用一万或者十万个 事例来搞明白世界上所有的事情。

to figure everything in the world out.

CA: Aren't there kind of terrifying unintended consequences

克里斯：它可能带来什么 非常恐怖的后果吗？

possible, from that?

JD: I think it depends on how you apply these systems.

杰夫：我认为这取决于 你如何使用此类系统。

It's very clear that AI can be a powerful system for good,

显然人工智能可以是 普惠大众的一大利器，

or if you apply it in ways that are not so great,

也可能因为不当的使用 而产生负面影响。

it can be a negative consequence.

So I think that's why it's important to have a set of principles

所以我认为这体现出了 行为准则的重要性，

by which you look at potential uses of AI

在你放眼人工智能的无限可能时，

and really are careful and thoughtful about how you consider applications.

谨慎且缜密地思考它的用途。

CA: One of the things people worry most about

克里斯：其中一个人们最担心的点是

is that, if AI is so good at learning from the world as it is,

如果人工智能如此擅长学习，

it's going to carry forward into the future

它可能对未来世界的格局产生影响，

aspects of the world as it is that actually aren't right, right now.

但这在现在看来还不太能被接受。

And there's obviously been a huge controversy about that

显然最近针对谷歌也有 一个非常大的争议。

recently at Google.

Some of those principles of AI development,

一些之前提到的人工智能开发准则

you've been challenged that you're not actually holding to them.

被质疑并未被遵守。

Not really interested to hear about comments on a specific case,

倒不用说一些具体事例，

but ... are you really committed?

不过……你们真的有遵守吗？

How do we know that you are committed to these principles?

我们如何知道谷歌 是否遵守这些条规？

Is that just PR, or is that real, at the heart of your day-to-day?

它只是一个公关行为， 还是它真的时刻存于你们心中。

JD: No, that is absolutely real.

杰夫：毋庸置疑是真的。

Like, we have literally hundreds of people

我们差不多有几百人

working on many of these related research issues,

在研究与此相关的议题，

because many of those things are research topics

因为很多这种准则 本身就是研究议题。

in their own right.

How do you take data from the real world,

你如何在现实世界中提取数据，

that is the world as it is, not as we would like it to be,

我们需要还原世界本来的样子， 而非我们想让它变成的样子，

and how do you then use that to train a machine-learning model

接着是如何用这些数据 训练一个机器学习模型，

and adapt the data bit of the scene

调整或在原有基础上增加一些数据，

or augment the data with additional data

so that it can better reflect the values we want the system to have,

从而使系统反映出我们想要的价值，

not the values that it sees in the world?

而非那些它在世界中观察到的。

CA: But you work for Google,

克里斯：但你在谷歌工作，

Google is funding the research.

谷歌在为这些研究提供资金。

How do we know that the main values that this AI will build

我们怎么知道开发这些人工智能 最大的价值在于全人类的福祉，

are for the world,

and not, for example, to maximize the profitability of an ad model?

而不是，比如说使一个 广告模型利益最大化？

When you know everything there is to know about human attention,

当你知道了有关 人类关注点的所有信息，

you're going to know so much

你对于我们内心的扭曲 和阴暗就了解的太多了。

about the little wriggly, weird, dark parts of us.

In your group, are there rules about how you hold off,

在你的团队中，有没有阻止 不当行为的规则，

在商业压力前，有没有 竖起一面政教之墙，

church-state wall between a sort of commercial push,

"You must do it for this purpose,"

要求“你必须这样做，”

so that you can inspire your engineers and so forth,

从而激励你的工程师们之类的，

to do this for the world, for all of us.

为世界，为我们考虑。

JD: Yeah, our research group does collaborate

杰夫：是的，我们的研究团队 会和谷歌的其它一些团队合作，

with a number of groups across Google,

包括广告部门， 搜索部门，地图部门，

including the Ads group, the Search group, the Maps group,

我们有这样的合作， 但也有许多纯粹的公开发表的文献。

so we do have some collaboration, but also a lot of basic research

that we publish openly.

我们去年发表了一千多份论文，

We've published more than 1,000 papers last year

涉及各种领域，也包含了 许多你之前提到的议题，

in different topics, including the ones you discussed,

例如公平性， 机器学习模型的可解释性，

about fairness, interpretability of the machine-learning models,

那些很重要的东西，

things that are super important,

and we need to advance the state of the art in this

且我们需要优化它的最高标准，

in order to continue to make progress

从而继续确保这些模型 在安全且可靠的情况下被开发。

to make sure these models are developed safely and responsibly.

CA: It feels like we're at a time when people are concerned

克里斯：我感觉我们处在一个人们 十分忌惮大型科技公司能力的时代，

about the power of the big tech companies,

and it's almost, if there was ever a moment to really show the world

如果能有机会真正向世界展示

that this is being done to make a better future,

这是为了建造更好的未来，

that is actually key to Google's future,

这是谷歌未来的关键，

as well as all of ours.

也是所有人的。

JD: Indeed.

杰夫：完全正确。

CA: It's very good to hear you come and say that, Jeff.

克里斯：非常高兴 听到你这么说，杰夫。

非常感谢你来到TED。

Thank you so much for coming here to TED.

杰夫：谢谢

JD: Thank you.

(掌声）

(Applause)