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

譯者: Lilian Chiu 審譯者: Helen Chang

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

我很興奮能來分享一些

關於人工智慧的辛辣想法。

But first, let's get philosophical

但，咱們先來點哲學，

by starting with this quote by Voltaire,

用一句伏爾泰的引言開頭，

an 18th century Enlightenment philosopher,

這位十八世紀的 啟蒙時代哲學家曾說：

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

「常識並不常見。」

Turns out this quote couldn't be more relevant

結果發現，這句話

太適合套用在現今的人工智慧上了。

to artificial intelligence today.

Despite that, AI is an undeniably powerful tool,

儘管如此，人工智慧 無疑是強大的工具，

打敗世界級圍棋冠軍，

beating the world-class "Go" champion,

acing college admission tests and even passing the bar exam.

在大學入學考試表現一流，

甚至通過了律師考試。

I’m a computer scientist of 20 years,

我是有二十年經驗的電腦科學家，

and I work on artificial intelligence.

我做的是人工智慧。

我來這裡揭開人工智慧的神秘面紗。

I am here to demystify AI.

So AI today is like a Goliath.

現今的人工智慧就像巨人歌利亞。

It is literally very, very large.

它真的非常非常大。

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

據推測，訓練近期的 人工智慧所用的是

數以萬計的圖形處理器 和上兆個字詞。

and a trillion words.

Such extreme-scale AI models,

這種極大規模的人工智慧模型 通常被稱為「大型語言模型」，

often referred to as "large language models,"

appear to demonstrate sparks of AGI,

它們顯然展現出了 AGI 的跡象，

artificial general intelligence.

AGI 就是人工通用智慧。

Except when it makes small, silly mistakes,

只是它也會犯很蠢的小錯，

which it often does.

且還蠻常犯的。

Many believe that whatever mistakes AI makes today

許多人相信，不論現今 人工智慧犯的是什麼錯誤，

can be easily fixed with brute force,

都可以用暴力地以更大的規模 及更多的資源來解決。

bigger scale and more resources.

What possibly could go wrong?

怎麼可能會出錯呢？

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

在社會層級，我們已經 面臨到三個立即的挑戰。

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

第一，

訓練極大規模的人工智慧 模型，成本非常昂貴，

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

只有幾間科技公司能負擔得起。

So we already see the concentration of power.

所以我們已經看到了 權力集中的現象。

But what's worse for AI safety,

但就人工智慧安全性 而言，更糟的是，

we are now at the mercy of those few tech companies

我們現在要看少數科技公司的臉色，

because researchers in the larger community

因為，

在更大的圈子裡的研究者沒有辦法

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

真正檢查和分析這些模型。

And let's not forget their massive carbon footprint

也別忘了它們的大量碳足跡

and the environmental impact.

以及對環境的衝擊。

And then there are these additional intellectual questions.

此外還有智慧方面的問題。

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

若沒有健全的常識，對人類 來說人工智慧真的安全嗎？

And is brute-force scale really the only way

且，暴力規模的方式真的是唯一可以

and even the correct way to teach AI?

教導人工智慧的方式嗎？ 它又是對的方式嗎？

So I’m often asked these days

這陣子我常被問，在沒有 極大規模計算的情況下

whether it's even feasible to do any meaningful research

做任何有意義的研究是可行的嗎？

without extreme-scale compute.

And I work at a university and nonprofit research institute,

我在大學及非營利研究機構工作，

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

所以我負擔不起大型圖形處理器農場

來創造巨大的語言模型。

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

儘管如此，

我相信我們必須要做／能做很多事

and can do to make AI sustainable and humanistic.

讓人工智慧永續且人本主義化。

We need to make AI smaller, to democratize it.

我們要讓人工智慧 變得更小，將它民主化，

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

我們也需要教導人工智慧 人類標準和價值觀來讓它更安全。

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

也許我們可以用 《大衛與歌利亞》來比喻。

here, Goliath being the extreme-scale language models,

歌利亞就是極大規模語言模型，

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

接著從古典名著 《孫子兵法》中尋求靈感，

which tells us, in my interpretation,

根據我的詮釋，它告訴我們，

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

了解你的敵人，選擇你要打 哪場仗，並創新你的武器。

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

咱們先從第一點開始， 了解你的敵人，

which means we need to evaluate AI with scrutiny.

我們需要用嚴謹的態度 來評估人工智慧。

AI is passing the bar exam.

人工智慧能通過律師考試。

Does that mean that AI is robust at common sense?

那表示人工智慧有健全的常識嗎？

You might assume so, but you never know.

你可能假設是如此， 但你永遠不會知道。

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

假設我留了五件衣服在太陽下曬乾，

and it took them five hours to dry completely.

它們需要五小時才能全乾。

How long would it take to dry 30 clothes?

曬乾三十件衣服要多久？

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

最新最棒的人工智慧系統 GPT-4 說是三十個小時。

Not good.

不好。

A different one.

再來個不同的例子：我有 十二公升的甕和六公升的甕，

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

and I want to measure six liters.

我想要測量出六公升，要怎麼做？

How do I do it?

Just use the six liter jug, right?

用六公升的甕就好了對吧？

GPT-4 spits out some very elaborate nonsense.

GPT-4 卻吐出一堆 精心製作的胡言亂語。

(Laughter)

（笑聲）

Step one, fill the six-liter jug,

第一步：把六公升的甕裝滿。

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

第二步：把六公升甕中的水 倒入十二公升的甕。

step three, fill the six-liter jug again,

第三步：再次把六公升的甕裝滿。

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

第四步：非常小心地將水 從六公升的甕倒入十二公升的甕。

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

最後，

你就有六公升的水了，

它就在現在應該沒水的六公升甕中。

that should be empty by now.

(Laughter)

（笑聲）

OK, one more.

好，再來一個。

Would I get a flat tire by bicycling over a bridge

我這樣會不會爆胎：騎腳踏車騎過

一座懸吊的橋，橋下有

that is suspended over nails, screws and broken glass?

釘子、螺絲，和碎玻璃。

Yes, highly likely, GPT-4 says,

GPT-4 說：是的，非常有可能，

presumably because it cannot correctly reason

有可能是因為它無法正確推論出

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

如果橋是懸吊在 釘子和碎玻璃上方的，

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

那麼橋的表面就不會直接 接觸到這些銳利的物體。

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

好，各位對此有什麼感想： 一位人工智慧律師

在律師考試表現出色，

yet randomly fails at such basic common sense?

卻會隨機在一些常識上出現錯誤？

AI today is unbelievably intelligent and then shockingly stupid.

現今的人工智慧

聰明到讓人難以置信，

卻也愚蠢到讓人瞠目結舌。

(Laughter)

（笑聲）

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

若要透過暴力規模來教導人工智慧，

這就是無可避免的副作用。

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

有些對規模抱持樂觀的人會說：

「別擔心這點， 這一切都很容易解決，

All of these can be easily fixed by adding similar examples

只要加一些類似的例子， 給人工智慧更多訓練資料即可。

as yet more training data for AI."

But the real question is this.

但，真正的問題是這個：

Why should we even do that?

我們幹嘛這樣做？你馬上 就能得出正確答案了，

You are able to get the correct answers right away

without having to train yourself with similar examples.

你還不需要用類似的例子 來訓練你自己。

Children do not even read a trillion words

兒童甚至不用讀到上兆個字詞

to acquire such a basic level of common sense.

也能習得這種基本層級的常識。

So this observation leads us to the next wisdom,

這項觀察，就要帶到第二項教訓：

choose your battles.

選擇你要打哪場仗。

So what fundamental questions should we ask right now

我們現在應該要問哪些基礎問題，

and tackle today

現今要處理哪些問題，

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

才能夠克服極大規模 人工智慧的這種現況？

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

我會說，常識是最該 優先處理的議題之一。

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

長久以來，常識一直是 人工智慧領域的難題。

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

為了解釋這一點， 讓我用暗物質來做比喻。

So only five percent of the universe is normal matter

宇宙只有 5% 是正常物質，

that you can see and interact with,

即你可以看見、互動的物質，

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

剩下的 95% 都是暗物質和暗能量。

Dark matter is completely invisible,

暗物質是完全看不見的，

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

但科學家推論它存在， 因為它會影響可見的世界，

even including the trajectory of light.

甚至光的軌道。

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

語言上的正常物質就是可見的文字，

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

暗物質則是世界 如何運作的潛在規則，

including naive physics and folk psychology,

包括天真物理學和民間心理學，

which influence the way people use and interpret language.

這些都會影響到人如何 使用和詮釋語言。

So why is this common sense even important?

那為什麼常識很重要？

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

尼克‧博斯特羅姆提出了 一個著名的思想實驗，

AI was asked to produce and maximize the paper clips.

人工智慧被要求要盡量製造出最多的

迴紋針。

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

而人工智慧決定要殺害人類， 用他們當作額外的資源，

to turn you into paper clips.

把各位變成迴紋針。

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

因為人工智慧不像人類 對於人的價值有基本的了解。

Now, writing a better objective and equation

就算是給它比較好的目標和方程式，

that explicitly states: “Do not kill humans”

明確陳述「不要殺人」，

will not work either

也行不通，因為 人工智慧可能就會改成

because AI might go ahead and kill all the trees,

殺光所有樹木，

thinking that's a perfectly OK thing to do.

以為那樣做完全沒問題。

And in fact, there are endless other things

事實上，還有數不清的事，

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

都是人工智慧在盡量製造 迴紋針時不該做的事，包括：

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

「別散播假消息」、 「別偷竊」、「別說謊」，

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

都屬於我們了解世界 如何運作的常識。

However, the AI field for decades has considered common sense

然而，

數十年來，人工智慧領域

認為常識是不可能的挑戰。

as a nearly impossible challenge.

So much so that when my students and colleagues and I

且到這種程度：

當我和我的學生及同事 數年前開始投入這個主題時，

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

我們被大力勸阻。

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

別人告訴我們，這是七○ 和八○年代的研究主題，

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

別投入這個主題， 因為永遠不會有成果。

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

事實上，甚至別說出這個詞 才能被別人認真對待。

Now fast forward to this year,

快轉到今年，我聽到：

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

「別投入這個主題，因為 ChatGPT 快解決它了。」

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

及「把規模擴大就會有神奇的 事情發生，其他都無所謂。」

and nothing else matters.”

So my position is that giving true common sense

所以，我的立場是：

要讓人工智慧有真正的常識 仍然是跟登月一樣難的課題。

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

And you don’t reach to the Moon

且登月的做法並不是把世上 最高的大樓一次加高一英吋。

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

Extreme-scale AI models

我承認極大規模人工智慧模型 取得的常識知識不斷在增加。

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

I'll give you that.

但是，

But remember, they still stumble on such trivial problems

別忘了，連孩子都懂的簡單 小問題它們還是會出錯。

that even children can do.

So AI today is awfully inefficient.

現今的人工智慧非常低效。

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

如果有替代途徑 或尚未被找到的途徑呢？

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

以深度神經網路的進展 為基礎的途徑，

but without going so extreme with the scale.

但在規模上不用做到這麼大。

So this leads us to our final wisdom:

這就要帶出

最後一項教訓：創新你的武器。

innovate your weapons.

In the modern-day AI context,

在現代人工智慧的情境中， 意思就是創新你的資料和演算法。

that means innovate your data and algorithms.

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

粗略來說，訓練現代人工智慧 所用的資料分為三類：

that modern AI is trained on:

raw web data,

原始網路資料、

crafted examples custom developed for AI training,

為了訓練人工智慧 而量身打造的範例，

and then human judgments,

以及人類判斷，

also known as human feedback on AI performance.

也就是人類針對 人工智慧表現的回饋。

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

若只用免費取得的原始網路資料 （第一類）來訓練人工智慧，

which is freely available,

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

那並不好，因為，

這些資料中帶有許多 種族主義、性別主義、假消息。

and misinformation.

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

不論你用多少資料， 垃圾進只會垃圾出。

So the newest, greatest AI systems

所以，

現在最新最棒的人工智慧系統 也靠第二、三類資料的支持，

are now powered with the second and third types of data

that are crafted and judged by human workers.

這些資料由人類工作者製作和評斷。

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

可以比喻成：寫專門的 教科書給人工智慧研讀，

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

接著僱用人類家教來經常 給予人工智慧回饋意見。

These are proprietary data, by and large,

這些大多是專有的資料，

speculated to cost tens of millions of dollars.

推測可能成本要數千萬美金。

We don't know what's in this,

我們不知道這些資料包含什麼， 但應該要開放給大家取得，

but it should be open and publicly available

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

讓我們能檢視和確保它們能協助 傳遞多樣性的規範和價值觀。

So for this reason, my teams at UW and AI2

基於這個理由， 我在 UW 和 AI2 的團隊

have been working on commonsense knowledge graphs

一直投入在做常識知識圖

as well as moral norm repositories

以及道德規範庫。

to teach AI basic commonsense norms and morals.

來教導人工智慧基本的 常識規範和道德。

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

我們的資料是完全公開的， 人人都可以檢視內容，

and make corrections as needed

並做必要的修正，

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

因為對這麼重要的研究主題 來說，透明度是關鍵。

Now let's think about learning algorithms.

咱們來思考一下學習演算法。

No matter how amazing large language models are,

不論大型語言模型有多驚人，

by design

從設計角度來說它們可能不會最適合

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

擔任可靠的知識模型。

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

這些語言模型確實 取得了大量的知識，

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

但這是它們副產物， 而不是直接的學習目標，

Resulting in unwanted side effects such as hallucinated effects

這導致了我們不想要的 副作用，如幻覺效應

and lack of common sense.

以及缺乏常識。

Now, in contrast,

相對的，人類學習的重點從來 不是去預測下一個字是什麼，

human learning is never about predicting which word comes next,

but it's really about making sense of the world

重點是要去理解這個世界， 並學習這個世界如何運作。

and learning how the world works.

Maybe AI should be taught that way as well.

也許也該用這種方式 來教導人工智慧。

So as a quest toward more direct commonsense knowledge acquisition,

所以，為了追尋

更直接取得常識知識的方法，

my team has been investigating potential new algorithms,

我的團隊一直在研究 潛在的新演算法，

including symbolic knowledge distillation

包括符號知識蒸餾，

that can take a very large language model as shown here

可以將非常大的語言模型， 如畫面上的這個，

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

它太大了所以螢幕放不下，

and crunch that down to much smaller commonsense models

把它打碎成更小許多的常識模型，

using deep neural networks.

用的方法是深度神經網路。

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

這麼做的過程中， 我們也用演算法產生出

可讓人類檢視的

symbolic, commonsense knowledge representation,

符號化知識呈現方式，

so that people can inspect and make corrections

讓大家可以做檢查、 做修正，甚至用它

and even use it to train other neural commonsense models.

訓練其他神經常識模型。

More broadly,

更廣泛來說，

we have been tackling this seemingly impossible giant puzzle

我們一直在拼湊這個似乎無法解決的

巨型常識拼圖，

of common sense, ranging from physical,

它的範圍從實體、 社會，以及視覺常識，

social and visual common sense

to theory of minds, norms and morals.

一直到心智理論、規範，和道德。

Each individual piece may seem quirky and incomplete,

每一片看起來都很怪異且不完整，

but when you step back,

但當你退後一步，

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

就好像每片拼圖編織 在一起成了一幅織錦，

that we call human experience and common sense.

我們稱這織錦為人類經驗及常識。

We're now entering a new era

我們正在邁入新時代， 在這個新時代，

in which AI is almost like a new intellectual species

人工智慧幾乎就像是 一種有智慧的新物種，

with unique strengths and weaknesses compared to humans.

和人類相比，它們 有獨特的優勢和缺點。

In order to make this powerful AI

為了要讓這種強大的人工智慧

sustainable and humanistic,

能永續且有人性，

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

我們必須要教導人工智慧 常識、規範，和價值觀。

Thank you.

謝謝。

(Applause)

（掌聲）

Chris Anderson: Look at that.

主持人：看哪。葉真，請留步。

Yejin, please stay one sec.

This is so interesting,

這個關於常識的想法相當有趣。

this idea of common sense.

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

不論將來出現的是什麼， 我們顯然都希望能如此。

But help me understand.

請幫我釐清一下， 我們有兒童學習模型，

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

How does a child gain common sense

兒童是如何習得常識的？

apart from the accumulation of more input

除了累積更多的輸入資訊

and some, you know, human feedback?

以及一些人類回饋？

What else is there?

還有什麼其他的？

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

講者：基本上被忽略的有 好幾項，但舉例來說其一就是

but one of them is, for example,

the ability to make hypothesis and make experiments,

做假設和實驗的能力，

interact with the world and develop this hypothesis.

和世界互動並發展出假設的能力。

We abstract away the concepts about how the world works,

我們把世界運作的方式 抽象化成概念，

and then that's how we truly learn,

那是我們真正學習的方式。

as opposed to today's language model.

相對的，現今的語言模型並非如此。

Some of them is really not there quite yet.

有一些真的還差得很遠。

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

主持人：你比喻說，若要登月，

不能用把大樓一次 增加一英呎的方式。

by extending a building a foot at a time.

But the experience that most of us have had

但我們大部分人對語言模型的 體驗並不是一次一英呎，

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

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

而像是讓人摒息的加速。

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

你確定，依現在 這些模型的發展步調，

each next level seems to be bringing with it

每提升一個層級就似乎帶來了

what feels kind of like wisdom and knowledge.

很像是智慧和知識的感覺。

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

講者：我完全同意 這種擴展規模的做法

really enhances the performance across the board.

真的讓性能全面性強化了。

So there's real learning happening

因為計算和資料的規模這麼大，

due to the scale of the compute and data.

確實有真正的學習產生。

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

然而，學習的品質還不太到位。

And the thing is,

重點是，我們還不知道

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

我們是否能靠擴大規模就達到目標。

just by scaling things up.

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

如果達不到，問題就是： 還有什麼別的？

And then even if we could,

就算達得到，

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

我們真的會喜歡這種概念嗎： 這些極大規模的人工智慧模型

that only a few can create and own?

只有少數人能創造和擁有？

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

主持人：如果 OpenAI 說 「我們對你的研究很感興趣，

我們希望你能協助 改善我們的模型」，

we would like you to help improve our model,"

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

你認為有辦法將你的研究 和他們已經打造出的成果結合嗎？

with what they have built?

YC: Certainly what I envision

講者：我所想像的肯定是

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

要以深度神經網路的進展為基礎。

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

且可能會有某種

最適度的規模，讓……

such that ...

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

順道一提，我並沒有想像 比較小就是比較好。

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

很可能會有適當的規模，

但除此之外，

the winning recipe might be something else.

成功的關鍵可能是別的。

所以，將想法綜合起來是很重要的。

So some synthesis of ideas will be critical here.

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

主持人：崔葉真， 謝謝你帶來的演說。

(Applause)

（掌聲）