Why AI Is Incredibly Smart and Shockingly Stupid | Yejin Choi | TED

翻译人员: Yip Yan Yeung 校对人员: Yanyan Hong

So I'm excited to share a few spicy thoughts on artificial intelligence.

我很高兴可以分享几个关于 人工智能（AI）的“真知灼见”。

But first, let's get philosophical

但首先，我们从哲学看起，

by starting with this quote by Voltaire,

引用一句来自 18 世纪 启蒙思想家伏尔泰的名言：

an 18th century Enlightenment philosopher,

who said, "Common sense is not so common."

“常识不平常。”

Turns out this quote couldn't be more relevant

结果这句名言和

to artificial intelligence today.

如今的人工智能息息相关。

Despite that, AI is an undeniably powerful tool,

虽然如此，AI 毋庸置疑 是个强大的工具，

beating the world-class "Go" champion,

它能赢得世界级围棋大赛，

acing college admission tests and even passing the bar exam.

顺利通过大学入学考试， 甚至通过律师资格考试。

I’m a computer scientist of 20 years,

我从事计算机科学家这一职业 已经 20 年了，

and I work on artificial intelligence.

我研究人工智能。

I am here to demystify AI.

我来到这里，是为了揭秘 AI。

So AI today is like a Goliath.

如今的 AI 就像是个 歌利亚（巨人）。

It is literally very, very large.

真的非常、非常大型。

It is speculated that the recent ones are trained on tens of thousands of GPUs

据推测，最新的 AI 由几万个 GPU （图形处理器）和

and a trillion words.

一万亿个词语训练而成。

Such extreme-scale AI models,

如此超巨型的 AI 模型，

often referred to as "large language models,"

通常被称为“大语言模型”，

appear to demonstrate sparks of AGI,

它们的出现是 AGI，

artificial general intelligence.

即通用人工智能的一簇火花。

Except when it makes small, silly mistakes,

虽然它会犯一些愚蠢的小错误，

which it often does.

而且总是会犯。

Many believe that whatever mistakes AI makes today

很多人认为，AI 现在犯的错误

can be easily fixed with brute force,

都可以强行依靠

bigger scale and more resources.

更大的规模和 更多的资源轻松解决。

What possibly could go wrong?

这有什么不对的呢？

So there are three immediate challenges we face already at the societal level.

我们如今在社会层面上 面临着三个亟待解决的问题。

First, extreme-scale AI models are so expensive to train,

首先，训练超大规模 AI 模型的成本非常高，

and only a few tech companies can afford to do so.

只有屈指可数的科技公司 具备负担的实力。

So we already see the concentration of power.

借此我们已经可以看出 权力的集中化了。

But what's worse for AI safety,

但就 AI 安全而言， 更不好的情况是

we are now at the mercy of those few tech companies

我们现在任凭这仅有的 几家科技公司的摆布，

because researchers in the larger community

因为业界的研究者们

do not have the means to truly inspect and dissect these models.

还没有找到真正检查、 剖析这些模型的方法。

And let's not forget their massive carbon footprint

我们也不能忽略它们大量的碳足迹

and the environmental impact.

和环境影响。

And then there are these additional intellectual questions.

还有几个智能方面的问题。

Can AI, without robust common sense, be truly safe for humanity?

如果 AI 没有可靠的常识， 它对人类来说真的是安全的吗？

And is brute-force scale really the only way

强行扩张真的是教授 AI 的

and even the correct way to teach AI?

唯一且正确的途径吗？

So I’m often asked these days

最近总是有人问我，

whether it's even feasible to do any meaningful research

如果没有超大规模计算， 还有没有可能做出一些有意义的研究。

without extreme-scale compute.

And I work at a university and nonprofit research institute,

我在一所大学和非营利研究机构工作，

so I cannot afford a massive GPU farm to create enormous language models.

所以我负担不起用大规模的 GPU 集群做出大语言模型。

Nevertheless, I believe that there's so much we need to do

但是，我相信还有很多 我们需要做、

and can do to make AI sustainable and humanistic.

可以做的事， 让 AI 可持续、以人为本。

We need to make AI smaller, to democratize it.

我们得缩小 AI、让它触手可及。

And we need to make AI safer by teaching human norms and values.

我们得通过传授人类的规范和价值观 让 AI 更安全。

Perhaps we can draw an analogy from "David and Goliath,"

也许我们可以引用 《大卫和歌利亚》的比喻，

here, Goliath being the extreme-scale language models,

在这个例子中，歌利亚就是 超大规模语言模型，

and seek inspiration from an old-time classic, "The Art of War,"

受到古代经典作品 《孙子兵法》的启发，

which tells us, in my interpretation,

根据我自己的解读，

know your enemy, choose your battles, and innovate your weapons.

我们需要了解对手、 选择战与不战、更新武器。

Let's start with the first, know your enemy,

我们从第一点了解对手开始，

which means we need to evaluate AI with scrutiny.

也就是说我们得对 AI 细细审视。

AI is passing the bar exam.

AI 通过了律师资格考试。

Does that mean that AI is robust at common sense?

这能说明 AI 有着可靠的常识吗？

You might assume so, but you never know.

你可以这么认为， 但你也没法验证。

So suppose I left five clothes to dry out in the sun,

假设我晒了 5 件衣服，

and it took them five hours to dry completely.

要花 5 个小时才能晒干。

How long would it take to dry 30 clothes?

那晒干 30 件衣服要多久？

GPT-4, the newest, greatest AI system says 30 hours.

最新、最厉害的 AI 系统 GPT-4 说 30 个小时。

Not good.

不咋地。

A different one.

换个问题。

I have 12-liter jug and six-liter jug,

我有一个 12 升的壶 和一个 6 升的壶，

and I want to measure six liters.

我想量出 6 升水。

How do I do it?

该怎么做？

Just use the six liter jug, right?

直接用 6 升的壶就行了，对吧？

GPT-4 spits out some very elaborate nonsense.

GPT-4 输出了一堆复杂的狗屁。

(Laughter)

（笑声）

Step one, fill the six-liter jug,

第一步，装满 6 升的壶。

step two, pour the water from six to 12-liter jug,

第二步，把水从 6 升的壶 倒进 12 升的壶里。

step three, fill the six-liter jug again,

第三步，再装满 6 升的壶。

step four, very carefully, pour the water from six to 12-liter jug.

第四步，小心翼翼地把水 从 6 升的壶倒进 12 升的壶里。

And finally you have six liters of water in the six-liter jug

最后，6 升的壶就能量出 6 升的水，

that should be empty by now.

而这个壶现在应该是空的。

(Laughter)

（笑声）

OK, one more.

再来一个。

Would I get a flat tire by bicycling over a bridge

如果我骑着自行车经过了一座

that is suspended over nails, screws and broken glass?

跨过钉子、螺丝和碎玻璃的桥， 我的轮胎会爆掉吗？

Yes, highly likely, GPT-4 says,

“会，非常有可能会。” GPT-4 是这么回答的，

presumably because it cannot correctly reason

可能是因为它无法正确地解读

that if a bridge is suspended over the broken nails and broken glass,

这座桥是架在碎钉子和碎玻璃之上的，

then the surface of the bridge doesn't touch the sharp objects directly.

桥面也不会直接接触到尖锐物体。

OK, so how would you feel about an AI lawyer that aced the bar exam

那你对这位通过了律师资格考试，

但偶尔会在这些基本常识上犯错的 AI 律师有何感想？

yet randomly fails at such basic common sense?

AI today is unbelievably intelligent and then shockingly stupid.

如今的 AI 聪明绝顶却又愚蠢不堪。

(Laughter)

（笑声）

It is an unavoidable side effect of teaching AI through brute-force scale.

如果要通过强行扩张教授 AI， 那就会产生不可避免的副作用。

Some scale optimists might say, “Don’t worry about this.

有些看好扩张的人可能会说： “别担心这个。

All of these can be easily fixed by adding similar examples

这些都可以通过给 AI 再加点类似的实例

as yet more training data for AI."

和训练数据轻松解决。”

But the real question is this.

但真正的问题是这个。

Why should we even do that?

我们干嘛要这么做呢？

You are able to get the correct answers right away

你甚至都不用自己拿着 近似实例去训练一遍，

without having to train yourself with similar examples.

就能立即得出正确答案。

Children do not even read a trillion words

要让儿童获取基本的常识，

to acquire such a basic level of common sense.

根本不需要阅读一万亿个单词。

So this observation leads us to the next wisdom,

这个现象将我们引向了 下一条大智慧：

choose your battles.

选择战与不战。

So what fundamental questions should we ask right now

我们现在该问、 该解决什么关键问题，

and tackle today

in order to overcome this status quo with extreme-scale AI?

才能应对超大规模 AI 的现状？

I'll say common sense is among the top priorities.

我想说，常识是重中之重。

So common sense has been a long-standing challenge in AI.

常识一直是 AI 长久以来的挑战。

To explain why, let me draw an analogy to dark matter.

让我引用暗物质的比喻 来解释一下这是为什么。

So only five percent of the universe is normal matter

宇宙中只有 5% 是正常物质，

that you can see and interact with,

是你可以看见、互动的，

and the remaining 95 percent is dark matter and dark energy.

剩下的 95% 都是 暗物质和暗能量。

Dark matter is completely invisible,

暗物质是完全不可见的，

but scientists speculate that it's there because it influences the visible world,

但科学家们推测出了它的存在， 是因为它影响着可见世界，

even including the trajectory of light.

甚至包括了光路。

So for language, the normal matter is the visible text,

对语言来说，正常物质 就是可见的文本，

and the dark matter is the unspoken rules about how the world works,

暗物质就是潜规则， 描述世界是如何运行的，

including naive physics and folk psychology,

包括朴素物理学和民间心理学，

which influence the way people use and interpret language.

它们影响着人们使用、 解读语言的方式。

So why is this common sense even important?

这种常识有什么重要的呢？

Well, in a famous thought experiment proposed by Nick Bostrom,

尼克·博斯特罗姆 （Nick Bostrom）

曾提出这样一个著名的思想实验，

AI was asked to produce and maximize the paper clips.

要求 AI 产生最大量的回形针。

And that AI decided to kill humans to utilize them as additional resources,

AI 最终决定杀死人类， 将人类当作额外的资源，

to turn you into paper clips.

把你们都做成回形针。

Because AI didn't have the basic human understanding about human values.

因为 AI 对于人类的价值 没有基本的人类认知。

Now, writing a better objective and equation

如果写了这么一个 更好的目标和等式，

that explicitly states: “Do not kill humans”

明确表示：“不要杀死人类。”

will not work either

也无济于事，

because AI might go ahead and kill all the trees,

因为 AI 有可能会杀死所有的树木，

thinking that's a perfectly OK thing to do.

认为这完全没问题。

And in fact, there are endless other things

其实还有无穷无尽的事，

that AI obviously shouldn’t do while maximizing paper clips,

都是 AI 在生产最多回形针的 同时显然不应该做的，

including: “Don’t spread the fake news,” “Don’t steal,” “Don’t lie,”

包括：“不要散布假消息”、 “不要盗窃”、“不要撒谎”，

which are all part of our common sense understanding about how the world works.

这些都是我们对这个世界 该如何运行的常识性理解。

However, the AI field for decades has considered common sense

但是，几十年以来， AI 领域一直将常识

as a nearly impossible challenge.

视为几乎不可能被征服的挑战。

So much so that when my students and colleagues and I

不可能到我和我的学生、同事

started working on it several years ago, we were very much discouraged.

多年前开始研究这个领域时， 都非常挫败。

We’ve been told that it’s a research topic of ’70s and ’80s;

有人告诉我们这个研究课题 该是上世纪 70、80 年代的；

shouldn’t work on it because it will never work;

不该研究这个， 因为永远得不到答案；

in fact, don't even say the word to be taken seriously.

这个词甚至都不该被摆到台面上。

Now fast forward to this year,

时间跳到今年，

I’m hearing: “Don’t work on it because ChatGPT has almost solved it.”

我听到了：“别研究这个了，因为 ChatGPT 几乎已经把它搞定了。”

And: “Just scale things up and magic will arise,

还有：“什么都扩张一下就行了， 会发生奇迹的，

and nothing else matters.”

别的都无所谓。”

So my position is that giving true common sense

我的观点是，给 AI 真正的常识，

human-like robots common sense to AI, is still moonshot.

类人的机器人常识， 依旧难如登天。

And you don’t reach to the Moon

你要登天，

by making the tallest building in the world one inch taller at a time.

也不可能一英尺一英尺地 拔高世界上最高的楼。

Extreme-scale AI models

超大规模的 AI 模型

do acquire an ever-more increasing amount of commonsense knowledge,

确实需要比以往更大量的常识，

I'll give you that.

我可以这么说。

But remember, they still stumble on such trivial problems

但记住，它们仍然会在 一些小朋友都会做的

that even children can do.

小问题上犯错误。

So AI today is awfully inefficient.

现在的 AI 极度低效。

And what if there is an alternative path or path yet to be found?

也许还有一条还没有 被发掘的道路呢？

A path that can build on the advancements of the deep neural networks,

一条基于深度神经网络进步的道路，

but without going so extreme with the scale.

也不用走向极端的规模。

So this leads us to our final wisdom:

这就说到了我们最后一条大智慧：

innovate your weapons.

更新你的武器。

In the modern-day AI context,

在当代的 AI 环境中，

that means innovate your data and algorithms.

指的就是在你的数据和算法上创新。

OK, so there are, roughly speaking, three types of data

现在的 AI 大概 由 3 类数据训练而成：

that modern AI is trained on:

raw web data,

原始网页数据、

crafted examples custom developed for AI training,

专为 AI 训练定制的人工实例

and then human judgments,

和人类判断，

also known as human feedback on AI performance.

也就是人类就 AI 的表现 提供的反馈。

If the AI is only trained on the first type, raw web data,

如果 AI 只由 第一种原始网页数据训练，

which is freely available,

此类数据唾手可得，

it's not good because this data is loaded with racism and sexism

这就会是个不好的选择，因为 这类数据充满了种族歧视、性别歧视

and misinformation.

和错误信息。

So no matter how much of it you use, garbage in and garbage out.

无论你用了多少此类数据， 就是输入了垃圾又输出了垃圾。

So the newest, greatest AI systems

最新最好的 AI 系统

are now powered with the second and third types of data

现已接入了 第二种和第三种数据，

that are crafted and judged by human workers.

由人类员工创建、评判。

It's analogous to writing specialized textbooks for AI to study from

这就类似于专为 AI 写了一本教科书，让它学，

and then hiring human tutors to give constant feedback to AI.

然后再请人类辅导老师 不断给 AI 提供反馈。

These are proprietary data, by and large,

这些都是专有数据，

speculated to cost tens of millions of dollars.

大约估算要花费上亿美元。

We don't know what's in this,

我们都不知道其中有什么，

but it should be open and publicly available

但这些数据得是公开的、 公众可以获取的，

so that we can inspect and ensure [it supports] diverse norms and values.

这样我们可以检视， 保证多种规范和价值观。

So for this reason, my teams at UW and AI2

因此，我在华盛顿大学和 艾伦人工智能研究所（AI2）的团队

have been working on commonsense knowledge graphs

一直在研究常识知识图谱

as well as moral norm repositories

和道德规范库，

to teach AI basic commonsense norms and morals.

借此将基本常识中的规范和道德 教授给 AI。

Our data is fully open so that anybody can inspect the content

我们的数据是完全公开的， 任何人都可以检查内容，

and make corrections as needed

必要时做出修改，

because transparency is the key for such an important research topic.

因为透明度是如此重要的 研究课题的关键。

Now let's think about learning algorithms.

我们来谈一谈学习算法。

No matter how amazing large language models are,

无论大语言模型有多牛，

by design

它们可能本来就不是 可靠的知识模型的最佳选择。

they may not be the best suited to serve as reliable knowledge models.

And these language models do acquire a vast amount of knowledge,

这些语言模型确实能获取海量知识，

but they do so as a byproduct as opposed to direct learning objective.

但这是与它直接的学习目标 相反的意外收获。

Resulting in unwanted side effects such as hallucinated effects

这会导致多余的副作用， 比如幻觉

and lack of common sense.

和缺乏常识。

Now, in contrast,

相比之下，

human learning is never about predicting which word comes next,

人类学习从来就不是 预测接下来该输出什么词，

but it's really about making sense of the world

而是理解世界，

and learning how the world works.

学习世界运作的方式。

Maybe AI should be taught that way as well.

也许也该这么教 AI。

So as a quest toward more direct commonsense knowledge acquisition,

为了探寻获取常识的 更直接的方式，

my team has been investigating potential new algorithms,

我的团队一直在研究 潜在的新算法，

including symbolic knowledge distillation

比如符号知识提炼，

that can take a very large language model as shown here

需要的巨型模型如图所示，

that I couldn't fit into the screen because it's too large,

这页都放不下， 因为实在是太大了，

and crunch that down to much smaller commonsense models

再通过深度神经网络把它 缩小成一个小得多的常识模型。

using deep neural networks.

And in doing so, we also generate, algorithmically, human-inspectable,

与此同时，我们通过算法 生成人类可以检视、

symbolic, commonsense knowledge representation,

以符号表达的常识知识表示，

so that people can inspect and make corrections

让人们可以检查、修正，

and even use it to train other neural commonsense models.

甚至用其训练其他神经常识模型。

More broadly,

更广泛地说，

we have been tackling this seemingly impossible giant puzzle

我们正在解开这个看似 不可能解开的巨幅常识拼图，

of common sense, ranging from physical,

从物理的、

social and visual common sense

社会的、视觉上的常识，

to theory of minds, norms and morals.

到心智理论、规范和道德。

Each individual piece may seem quirky and incomplete,

每一块都古怪又不完整，

but when you step back,

但如果退后一步看，

it's almost as if these pieces weave together into a tapestry

这些碎片好像交织在一起， 形成一幅我们称之为

that we call human experience and common sense.

人类经验和常识的画卷。

We're now entering a new era

我们现正迈入一个新时代，

in which AI is almost like a new intellectual species

AI 就像是一种 拥有知识的新物种，

with unique strengths and weaknesses compared to humans.

相较人类有着独特的优势和弱势。

In order to make this powerful AI

要让这强大的 AI

sustainable and humanistic,

可持续又以人为本，

we need to teach AI common sense, norms and values.

我们得把常识、规范和 价值观教给 AI。

Thank you.

谢谢。

(Applause)

（掌声）

Chris Anderson: Look at that.

克里斯·安德森 （Chris Anderson）：瞧瞧。

Yejin, please stay one sec.

艺珍（Yejin），请留步。

This is so interesting,

太有趣了，

this idea of common sense.

常识的话题。

We obviously all really want this from whatever's coming.

显然我们都很期待。

But help me understand.

但请你帮我理解一下。

Like, so we've had this model of a child learning.

我们有了这个类似 儿童学习的模型。

How does a child gain common sense

除了更多输入的积累和人类的反馈，

apart from the accumulation of more input

小孩子该如何获取常识呢？

and some, you know, human feedback?

What else is there?

还有什么？

Yejin Choi: So fundamentally, there are several things missing,

崔艺珍（Yejin Choi）： 从根本上来说，缺少了几样东西，

but one of them is, for example,

但以其中一样为例，

the ability to make hypothesis and make experiments,

即做出假设和尝试的能力，

interact with the world and develop this hypothesis.

与世界互动，形成假设。

We abstract away the concepts about how the world works,

我们不会去归纳总结 世界运作的方式，

and then that's how we truly learn,

这才是我们学习的真正方式，

as opposed to today's language model.

而不是如今语言模型采用的方式。

Some of them is really not there quite yet.

有些模型还达不到这种程度。

CA: You use the analogy that we can’t get to the Moon

CA: 你打了个比方， 说我们无法通过

by extending a building a foot at a time.

每次把楼拔高一英尺登天。

But the experience that most of us have had

但很多人在这些语言模型上的 体验可不是每次一英尺，

of these language models is not a foot at a time.

It's like, the sort of, breathtaking acceleration.

而是像猛地一脚油门。

Are you sure that given the pace at which those things are going,

你确定照现在发展的节奏，

each next level seems to be bringing with it

每到下一个阶段都会带来

what feels kind of like wisdom and knowledge.

某种智慧的心得和知识吗？

YC: I totally agree that it's remarkable how much this scaling things up

YC: 我完全同意扩大规模

really enhances the performance across the board.

真的可以总体提高性能 是一件了不起的事。

So there's real learning happening

计算和数据的规模 真的能让我们有所收获。

due to the scale of the compute and data.

However, there's a quality of learning that is still not quite there.

但是收获的质量不尽如人意。

And the thing is,

问题是，

we don't yet know whether we can fully get there or not

我们都不知道 到底能不能“如人意”，

just by scaling things up.

仅仅通过扩大规模这一途径。

And if we cannot, then there's this question of what else?

如果这样是达不到的， 问题就变成了：还有什么途径呢？

And then even if we could,

就算我们可以借此达到想要的效果，

do we like this idea of having very, very extreme-scale AI models

我们真的会喜欢使用这种 非常、非常大规模的 AI 模型，

that only a few can create and own?

只有屈指可数的人 可以创造、拥有的模型吗？

CA: I mean, if OpenAI said, you know, "We're interested in your work,

CA: 如果 OpenAI 说： “我们对你的工作很感兴趣，

we would like you to help improve our model,"

我希望你能帮我们改进我们的模型。”

can you see any way of combining what you're doing

你觉得有没有将你的研究内容

with what they have built?

与他们做的东西相结合的可能？

YC: Certainly what I envision

YC: 我的畅想当然

will need to build on the advancements of deep neural networks.

必须建立在深度神经网络的突破之上。

And it might be that there’s some scale Goldilocks Zone,

也许存在一个规模的“适居带”，

such that ...

这样……

I'm not imagining that the smaller is the better either, by the way.

我不是说越小越好，

It's likely that there's right amount of scale, but beyond that,

很有可能有一个合适的规模， 但除此之外，

the winning recipe might be something else.

取胜秘籍可能另有他物。

So some synthesis of ideas will be critical here.

各种想法的碰撞就是关键。

CA: Yejin Choi, thank you so much for your talk.

CA: 崔艺珍，感谢你的演讲。

(Applause)

（掌声）