Can AI Master the Art of Humor? | Bob Mankoff | TED

翻译人员: Yip Yan Yeung 校对人员: Sue Lu

Whether you're thrilled by what AI can do for us

无论你对 AI 能为我们带来的成就而兴奋，

or terrified by what AI is going to do to us,

还是对 AI 将对我们造成的影响感到恐惧，

whether it can be funny, is probably not top of mind for you.

反正它是不是有趣， 大概不会是你优先考虑的事。

It is for me.

但却是我要优先考虑的。

I don't care if it turns all of us into paperclips,

我不在乎 AI 是否 会把我们都变成回形针，

as long as they're funny paper clips.

只要是好玩的回形针就行。

(Laughter)

（笑声）

And the fact that it makes stuff up, hallucinates,

AI 胡编乱造、出现幻觉的事实，

for me, that's not a bug, that's a feature.

对我来说， 不是故障，而是它的功能。

My entire career was making stuff up.

我的整个职业生涯都是在编造东西。

They're called cartoons.

我编造的，被称为“漫画”。

This is probably the most famous one I hallucinated.

这大概是我“幻想”出来 最出名的一幅。

There are a number of theories of humor that could explain this cartoon.

有许多幽默理论可以解释这幅漫画。

There's the superiority theory.

其中有优越论。

You're the guy on the phone, not on the other end.

你是打电话的人，不是接电话的。

The incongruity theory.

有乖讹论。

There's a mismatch between the politeness of the language

用语的礼貌 和消息内容的粗鲁不相配。

and the rudeness of the message.

And the benign violation theory of humor,

还有幽默的良性违规理论，

which is sort of a golden ratio theory of humor, if you will.

你可以把它当成 幽默的黄金比例理论。

See, I got in that term

看，我用了那个词。

(Laughter)

（笑声）

Which says, for something to be funny,

也就是说，要让什么东西好笑的话，

it has to have just the right amount of wrong.

就必须有恰到好处的不合时宜。

Now I didn’t use any theories to create that cartoon.

不过我在创作这幅漫画的时候 并没有使用任何理论。

And I once said at a talk at Google,

我曾经在谷歌的一次演讲中说过，

there is no algorithm for humor.

幽默没有算法。

But now with the rapid pace of AI,

但是现在随着 AI 的飞速发展，

I have to wonder, could there be a bot Mankoff?

我不禁想，会不会出现 一个机器人版曼科夫？

(Laughter)

（笑声）

You might think my reflexive answer to this would be, how about never?

你可能以为我对这个问题的 反应会是，永远不会吧？

But while I don't want to be replaced by a bot,

虽然我不想被机器人取代，

I'm not above being helped by it.

但我也并不排斥机器人的帮助。

Steve Jobs famously said

史蒂夫·乔布斯曾说过一句名言，

that computers are a bicycle for the mind.

计算机是心灵的自行车。

If that's the case, what's AI, a rocket ship?

照这个道理， AI 是什么，火箭吗？

And at my age, you know what?

我这个年纪，你猜怎么着？

I'd settle for a walker.

我会觉得是个助行器吧。

(Laughter)

（笑声）

The fears of machines replacing humans are not new.

对机器取代人类的恐惧早就有了。

This cartoon anticipates by five decades

这幅漫画五十年前就预测了

what's now called the alignment problem,

现在的“对齐”问题，

when the goals of machines and humans go horribly awry,

即机器和人类的目标大相径庭，

at least for one of the parties.

至少对其中一方来说是如此。

Cartoons don't happen in a cultural vacuum.

漫画不是在文化真空中产生的。

They're part of the zeitgeist.

它们是时代精神的一部分。

Here's a contemporaneous article

这是一篇同时代的文章，

about the guy who invented cybernetics,

讲述的是发明控制论的人

Norbert Wiener,

诺伯特·维纳（Norbert Wiener），

who said thinking machines were putting us on the eve of destruction.

他说会思考的机器 会让我们危在旦夕。

Now sadly and tragically,

可悲又不幸的是，

Norbert Wiener died,

诺伯特·维纳死了，

but not by a thinking machine,

但不是死于会思考的机器，

but by an unthinking one, he was run over by a bus.

而是死于一台不会思考的机器， 他被一辆巴士撞死了。

(Laughter)

（笑声）

That's not true.

这不是真的。

I made that up.

是我编的。

(Laughter)

（笑声）

I hallucinated it because it's funny.

我幻想出这个故事， 是因为它很好玩。

(Laughter)

（笑声）

Here's a great update of that older cartoon.

这是之前那幅漫画的重大更新。

"To think this all began with letting Autocomplete finish our sentence,"

“想想这一切竟然始于 让自动补全完成我们的句子”，

which indeed it did.

确实如此。

So these fears are not new, not novel.

所以这些恐惧不是什么新鲜事。

But now, in the immortal words of Nigel Tufnel of "Spinal Tap,"

但现在用 Spinal Tap 乐队的

奈杰尔·塔弗纳（Nigel Tufnel）的 经典名言来说，

they go to 11.

这种恐惧已经“顶破天”，

They're cranked to the max.

达到最高极限了。

And here is one of the maximum cranksters of all time,

有史以来最极端的一个怪人，

Elon Musk, saying,

埃隆·马斯克，曾说道：

"AI is one of the biggest threats to humanity."

“AI 是人类最大的威胁之一。”

But certainly not as big as Elon Musk.

但肯定没有埃隆·马斯克的威胁大。

(Laughter)

（笑声）

(Applause and cheers)

（掌声和欢呼）

People like Elon have a p(doom) number.

像埃隆这样的人 有一个 P(doom) （末日）值，

That's the probability AI is going to wipe us out.

是 AI 消灭我们的概率。

I think p(doom) is p(dumb).

我觉得 （末日）值 应该是 P(dumb) （傻子）值。

(Laughter)

（笑声）

I'm interested in p(funny),

我对 P(funny) （搞笑）值 很感兴趣，

and I’ve been using the New Yorker caption contest

我一直在通过 《纽约客》配文竞赛

to look into the probability of that.

研究出现搞笑配文的概率。

Every week since 2005,

2005 年以来的每一周，

the New Yorker has presented a cartoon without a caption,

《纽约客》都会刊登一张 没有配文的漫画，

and challenged its readers to come up with the winning caption

要求读者在配文竞赛中 提出最佳配文。

in the caption contest.

And for that,

获胜奖励是荣登《纽约客》杂志、

they get the glory of being in the New Yorker magazine,

a huge amount of money,

得到一大笔钱、

house in the Bahamas that Sam Bankman-Fried --

还有巴哈马的房子，

是山姆·班克曼-弗里德曾拥有的 （美国加密亿万富翁，因欺诈罪入狱）。

Actually, it's just the glory.

其实，奖励只有那份刊登的荣耀。

On the page of the New Yorker,

在《纽约客》杂志上，

there's a contest you enter,

定期登出一个你可以参与的配文竞赛，

the finalists from a few weeks before,

几周后选出决赛选手，

three finalists,

共三名，

and a winning caption.

从中再选出一名获胜者。

So it's staggered in that way.

竞赛就是这么分阶段进行的。

Each one of these images are funny.

竞赛的每一张图片都很好玩。

They're incongruous.

它们是违和的。

You'd certainly think they're humorous,

你肯定会觉得它们很幽默，

but they're not funny in a way that you get.

但它们的好笑并非是你理解的那样，

They're not mentally funny.

它们并不是需要动脑的幽默。

To make it that,

要达到这种幽默的效果，

of course, you need the right caption.

当然，你得有合适的配文。

"Any happily married people here tonight?"

“今晚在座的有婚姻幸福的吗？”

(Laughter)

（笑声）

OK, but with up to 10,000 captions

好吧，但是每周 都有多达 10,000 条配文，

every week,

how do you select that?

你该如何选择呢？

Now from 2005 to early 2016,

从 2005 年到 2016 年初，

that burden fell on me and my assistants,

这个负担落在了我和我的助理身上，

but mainly my assistants, to cull,

但主要是我的助理身上，挑选、

to try to cull the good captions

刨去我们不客气地 称为“垃圾”的配文，找出一些好的来。

from what we uncharitably call the “craptions.”

(Laughter)

（笑声）

But then, in early 2016,

但是，2016 年初，

for the benefit of all humanity,

为了全人类的利益，

but mainly for me and my assistants,

但主要是为了我和我的助理，

we switched to crowdsourcing.

我们转向了众包。

So now for every contest, you vote online,

现在每次竞赛，你都在网上投票，

and a funniness score from over a million judgments

所有配文都会从超过一百万次评判中

is given for all the captions.

得到一个有趣度评分。

Now overall, I’m against mob rule,

总体来说，我反对多数暴政，

but actually in this case,

但其实在这种竞赛中，

the mob does a pretty good job.

大众的表现相当不错。

Usually, the finalist almost certainly almost all the time, really,

一直以来决赛选手通常

the finalists come from the top 200 captions.

都选自前 200 名配文。

Well, this is popular, not only with the New Yorker,

嗯，这不仅在《纽约客》很受欢迎，

but it's caught the eye of data scientists,

还引起了各行广泛关注， 包括数据科学家、

creativity researchers,

创意研究人员、

cognitive scientists,

认知科学家，

and AI, of course, and everything adjacent to AI.

当然还有 AI 及与其有关的一切。

So that's why I wasn't really surprised when --

这就是为什么我不太感到意外……

Oh, it's up there, thanks.

哦，放出来了，谢谢。

I wasn't really surprised when Vincent Vanhoucke,

我不太意外文森特·范霍克——

then the chief data scientist for Google's DeepMind,

当时是谷歌 DeepMind 首席数据科学家，

now their head of robotics,

现为机器人负责人，

sent me this email

给我发了这封电子邮件，

indicating that winning the caption contest,

表明要在配文竞赛中获胜，

which was actually somewhat of the sine qua non of human creativity

其实必须有人类创造力

and intelligence.

和智力作为必要条件。

And I was also flattered by that, of course,

当然，他的话也让我很受宠若惊，

but I didn't think they had any chance at all of doing it.

不过我认为他们的机器人根本没戏。

And it turned out that was the case.

事实证明情况确实如此。

All of the AI juju circa 2016

2016 年左右的 AI 技术

wasn't up to the task.

还不足以胜任这项任务。

It really couldn't even decode the image.

它甚至无法解码竞赛图片。

So for the sine qua non of the human mind,

因此，对于人类思维这个必要条件来说，

DeepMind was non compos mentis and out of its depth.

DeepMind 既无判断能力， 也力不从心。

But time and AI marched on,

但是随着时间流逝和 AI 进步，

AI marching quadruple time.

AI 以四倍的速度向前迈进。

Vincent gets back to me and says,

文森特回复我说，

while human creativity might still be out of reach,

尽管人类创造力可能仍遥不可及，

we think we have understanding well in hand.

我们认为我们已有很好的理解。

He sends me this ridiculous, uber-nerd example

他发给我这个荒谬的、 超级书呆子的例子

of explaining humor.

用以解释幽默。

And I said, you know what?

我就说，这样吧，

Let me give you a cartoon I did in 1997

给你看看我 1997 年画的漫画，

of this other watershed moment

是关于另一个分水岭时刻，

when IBM's DeepMind defeated Garry Kasparov,

描绘了 IBM 的 DeepMind 击败 加里·卡斯帕罗夫（Garry Kasparov），

the world chess champion.

他是世界象棋冠军。

And here's the cartoon I did then, and it says,

这是我当时画的漫画，配文是：

"No, I don't want to play chess.

“不，我不想下棋。

I just want you to reheat the lasagna."

我只想让你去热一热千层面。”

(Laughter)

（笑声）

I rate this explanation a solid B-minus.

我只给这条配文注解打分 B-。

But so what if it was an A?

但如果它其实是 A 的水平呢？

Is there ever going to be a beautiful

会有一套精美的

New Yorker cartoon anthology of explanations?

《纽约客》漫画配文注解合集吗？

I don't think so.

我觉得不会。

But the idea that understanding humor

但是，理解幽默

could be a stepping stone to creating it

是创造幽默的基石，

sort of made sense.

这一想法是有道理的。

This paper I was involved in

我参与的这篇论文

tried to look at compared to smart humans,

研究与聪明的人类相比，

what would the best AIs do on three tasks?

最好的 AI 在三项任务上表现如何？

Could they,

它们能否把不同竞赛的获胜配文

from winning captions from different contests

match to the right image?

与正确的图片匹配？

Could they, between two captions,

它们能否面对两条配文——

one that won and the one that was pretty good,

一条获胜配文和一条很不错的配文，

pick the right one?

从中选出正确的那条？

And could they explain the humor?

而且，它们能解释幽默吗？

Now for all of them, you know what?

所有这些 AI 的表现证实，

Yeah, humans were still ahead,

没错，人类仍然处于领先地位，

but AI is closing the gap.

但 AI 正在缩小差距。

The most interesting thing about this paper for me

对我来说，这篇论文最有意思的地方

was it showed a pathway

是它展示了一条途径，

for which you could create cartoon humor.

可以帮助人们创作漫画幽默。

And that was how we trained the contest.

我们就是这样训练（AI）比赛的。

For 660 --

在 660……

653 contests,

653 场比赛中，

the AI was trained, fine-tuned,

AI 根据这些由人类注释的示例 进行了训练和微调，

on these examples which humans annotated.

A description of the cartoon, explanation of the humor.

比如漫画的描述，对幽默的解释。

OK, if you've used ChatGPT, you sort of get the idea now.

如果你用过 ChatGPT， 现在应该有点明白了吧。

Put a number of examples like this.

输入一些这样的例子，

Put it in the prompt window.

放进提示词窗口，

Rinse and repeat and you get new cartoons.

反复操作，就能得到新的漫画。

Well Jack Hessel, the chief author of the paper,

该论文的主要作者 杰克·赫塞尔（Jack Hessel）

did something more sophisticated.

做了更复杂的事情。

And what he did was create 50 synthetic new cartoons

他根据这些旧数据 创作了 50 幅合成的漫画，

generated from this old data

in which there were five options for captions.

每幅漫画有五个配文选项。

I picked four of them,

我挑了四个配文，

and I gave them to cartoonist Shannon Wheeler to draw up.

然后交给漫画家香农·惠勒 （Shannon Wheeler）画出来。

Here are the results.

以下是香农的画作。

Now Shannon said, well, these are weird.

香农说，嗯，这些画很奇怪。

They don't really seem like --

它们看起来不像……

it's sort of an uncanny valley of cartoons.

有点像漫画的恐怖谷效应。

They're not quite there.

显然 AI 的水平还不到位，

But it is interesting.

但很有意思。

All of these are new cartoons that never appeared anywhere

所有这些都是 从未在任何地方出现过的新漫画，

that are an idea of human --

一点不像来自人类的创作，

of computer creativity.

因为它们是计算机的创意。

And also when you look at it, what it is,

当你仔细看时，它们究竟是什么？

it's weirdness in, weirdness out.

因为怎么看都很奇怪。

The caption contest cartoons are weird.

配文竞赛的漫画很奇怪。

But I do see this now as a tool for brainstorming for cartoonists,

但是我确实把它看作一个 漫画家头脑风暴的工具，

in that we played this completely straight.

我们是很认真的。

Shannon wasn't able to manipulate the description of the picture

香农当时不能改动图片描述或配文。

or the caption.

Had he done that, it could have been better.

如果他改动了，画作可能会好些。

Also, we could have asked it to make more.

而且，我们可以让 AI 创作更多。

We could have put in the rankings for the humor.

我们还可以输入幽默排名。

We could do all this to improve it.

我们可以做所有这些来改善 AI。

So quality comes out of quantity.

也就是说，量变会引起质变。

You can get an awful lot of quantity here.

你可以得到非常大的数量，

You can have a human being in the loop to do this.

而且，你可以让一个人 介入流程来完成这件事。

But ...

但是……

I would not go so far

我不会做这么多来帮助

to give AI a true human sense of humor.

AI 拥有真正人类的幽默感。

A human sense of humor is not about making a joke or getting it.

人类的幽默感 并不只是开玩笑或理解玩笑。

It's rooted in our vulnerability.

它源于我们的脆弱性。

It's the blessing we get for the curse of mortality.

这是我们用死亡诅咒换来的祝福。

Mark Twain said the true source of humor

马克·吐温说，幽默的真正来源

is not joy but sorrow.

不是喜悦，而是悲伤。

If we gave AI the thousand natural shocks

如果我们要让 AI 经受上千种

that flesh is heir to,

肉体凡胎必经的自然冲击，

that would be cruel.

那会很残酷。

If we did that,

如果我们那样做，

it might very well want to wipe us out.

AI 很可能想要消灭我们。

And if they did, all I ask is that they take Elon first.

如果它们真要那么做， 就先消灭埃隆（Elon）吧。

(Laughter)

（笑声）

Thank you.

谢谢。

(Applause)

（掌声）

Now, I think, if I'm not mistaken,

如果我没记错的话，

we had a caption contest here, right?

我们这里也有过一场配文竞赛，对吧？

It's on the next slide

是在下一张幻灯片上，

that I had to decide from U of M people.

我不得不从密歇根大学的人中挑选。

And this was the image we created for the contest.

这就是我们为竞赛创作的图片，

It's that image.

这一张。

And the one I picked up

我选中的配文

by Kavya Davuluri was this one.

是卡维亚·达武鲁里 （Kavya Davuluri）创作的。

"Above all, I asked the class of 3000 not to forget your humanity."

“最重要的是，我请求 3000 届的学生 不要忘记你们的人性。”

(Laughter)

（笑声）

Now what I will say is, that is really close to a good caption.

我想说，这确实非常接近 一个好的配文了。

(Laughter)

（笑声）

Mark Twain said the difference between the right word and the wrong word

马克·吐温说， 正确用词和错误用词的区别

is the difference between the lightning and the lightning bug.

就是闪电和萤火虫的区别。

That caption should be

那条配文应该是：

"Above all, I ask that the class of 3000 to not forget my humanity."

“最重要的是，我请求 3000 届的学生 不要忘记我的人性。”

(Laughter)

（笑声）

Now you have a joke.

现在这就是个笑话了。

Thank you very much.

非常感谢。

(Applause)

（掌声）