How to pronounce "annotated"

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

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

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

It is for me.

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

as long as they're funny paper clips.

(Laughter)

And the fact that it makes stuff up, hallucinates,

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

My entire career was making stuff up.

They're called cartoons.

This is probably the most famous one I hallucinated.

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

There's the superiority theory.

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

The incongruity theory.

There's a mismatch between the politeness of the language

and the rudeness of the message.

And the benign violation theory of humor,

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

See, I got in that term

(Laughter)

Which says, for something to be funny,

it has to have just the right amount of wrong.

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

And I once said at a talk at Google,

there is no algorithm for humor.

But now with the rapid pace of AI,

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

(Laughter)

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

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

I'm not above being helped by it.

Steve Jobs famously said

that computers are a bicycle for the mind.

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

And at my age, you know what?

I'd settle for a walker.

(Laughter)

The fears of machines replacing humans are not new.

This cartoon anticipates by five decades

what's now called the alignment problem,

when the goals of machines and humans go horribly awry,

at least for one of the parties.

Cartoons don't happen in a cultural vacuum.

They're part of the zeitgeist.

Here's a contemporaneous article

about the guy who invented cybernetics,

Norbert Wiener,

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

Now sadly and tragically,

Norbert Wiener died,

but not by a thinking machine,

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

(Laughter)

That's not true.

I made that up.

(Laughter)

I hallucinated it because it's funny.

(Laughter)

Here's a great update of that older cartoon.

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

which indeed it did.

So these fears are not new, not novel.

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

they go to 11.

They're cranked to the max.

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

Elon Musk, saying,

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

But certainly not as big as Elon Musk.

(Laughter)

(Applause and cheers)

People like Elon have a p(doom) number.

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

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

(Laughter)

I'm interested in p(funny),

and I’ve been using the New Yorker caption contest

to look into the probability of that.

Every week since 2005,

the New Yorker has presented a cartoon without a caption,

and challenged its readers to come up with the winning caption

in the caption contest.

And for that,

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

a huge amount of money,

house in the Bahamas that Sam Bankman-Fried --

Actually, it's just the glory.

On the page of the New Yorker,

there's a contest you enter,

the finalists from a few weeks before,

three finalists,

and a winning caption.

So it's staggered in that way.

Each one of these images are funny.

They're incongruous.

You'd certainly think they're humorous,

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

They're not mentally funny.

To make it that,

of course, you need the right caption.

"Any happily married people here tonight?"

(Laughter)

OK, but with up to 10,000 captions

every week,

how do you select that?

Now from 2005 to early 2016,

that burden fell on me and my assistants,

but mainly my assistants, to cull,

to try to cull the good captions

from what we uncharitably call the “craptions.”

(Laughter)

But then, in early 2016,

for the benefit of all humanity,

but mainly for me and my assistants,

we switched to crowdsourcing.

So now for every contest, you vote online,

and a funniness score from over a million judgments

is given for all the captions.

Now overall, I’m against mob rule,

but actually in this case,

the mob does a pretty good job.

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

the finalists come from the top 200 captions.

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

but it's caught the eye of data scientists,

creativity researchers,

cognitive scientists,

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

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

Oh, it's up there, thanks.

I wasn't really surprised when Vincent Vanhoucke,

then the chief data scientist for Google's DeepMind,

now their head of robotics,

sent me this email

indicating that winning the caption contest,

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

and intelligence.

And I was also flattered by that, of course,

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

And it turned out that was the case.

All of the AI juju circa 2016

wasn't up to the task.

It really couldn't even decode the image.

So for the sine qua non of the human mind,

DeepMind was non compos mentis and out of its depth.

But time and AI marched on,

AI marching quadruple time.

Vincent gets back to me and says,

while human creativity might still be out of reach,

we think we have understanding well in hand.

He sends me this ridiculous, uber-nerd example

of explaining humor.

And I said, you know what?

Let me give you a cartoon I did in 1997

of this other watershed moment

when IBM's DeepMind defeated Garry Kasparov,

the world chess champion.

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

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

I just want you to reheat the lasagna."

(Laughter)

I rate this explanation a solid B-minus.

But so what if it was an A?

Is there ever going to be a beautiful

New Yorker cartoon anthology of explanations?

I don't think so.

But the idea that understanding humor

could be a stepping stone to creating it

sort of made sense.

This paper I was involved in

tried to look at compared to smart humans,

what would the best AIs do on three tasks?

Could they,

from winning captions from different contests

match to the right image?

Could they, between two captions,

one that won and the one that was pretty good,

pick the right one?

And could they explain the humor?

Now for all of them, you know what?

Yeah, humans were still ahead,

but AI is closing the gap.

The most interesting thing about this paper for me

was it showed a pathway

for which you could create cartoon humor.

And that was how we trained the contest.

For 660 --

653 contests,

the AI was trained, fine-tuned,

on these examples which humans annotated.

A description of the cartoon, explanation of the humor.

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

Put a number of examples like this.

Put it in the prompt window.

Rinse and repeat and you get new cartoons.

Well Jack Hessel, the chief author of the paper,

did something more sophisticated.

And what he did was create 50 synthetic new cartoons

generated from this old data

in which there were five options for captions.

I picked four of them,

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

Here are the results.

Now Shannon said, well, these are weird.

They don't really seem like --

it's sort of an uncanny valley of cartoons.

They're not quite there.

But it is interesting.

All of these are new cartoons that never appeared anywhere

that are an idea of human --

of computer creativity.

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

it's weirdness in, weirdness out.

The caption contest cartoons are weird.

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

in that we played this completely straight.

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

or the caption.

Had he done that, it could have been better.

Also, we could have asked it to make more.

We could have put in the rankings for the humor.

We could do all this to improve it.

So quality comes out of quantity.

You can get an awful lot of quantity here.

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

But ...

I would not go so far

to give AI a true human sense of humor.

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

It's rooted in our vulnerability.

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

Mark Twain said the true source of humor

is not joy but sorrow.

If we gave AI the thousand natural shocks

that flesh is heir to,

that would be cruel.

If we did that,

it might very well want to wipe us out.

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

(Laughter)

Thank you.

(Applause)

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

we had a caption contest here, right?

It's on the next slide

that I had to decide from U of M people.

And this was the image we created for the contest.

It's that image.

And the one I picked up

by Kavya Davuluri was this one.

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

(Laughter)

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

(Laughter)

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

is the difference between the lightning and the lightning bug.

That caption should be

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

(Laughter)

Now you have a joke.

Thank you very much.

(Applause)