0:00

Everyone deserves access to clean, affordable

0:02

energy. Everyone. Millions

0:04

of Americans rely on propane for dependable

0:07

energy that is independent from the electric

0:09

grid. Propane is a reliable, versatile energy

0:11

produced in the United States, the power

0:13

school buses, hospitals, and our homes, with

0:16

lower carbon emissions than conventional fuels. The

0:19

path to a low-carbon future includes cleaner

0:21

energy like propane. And the production of

0:23

renewable propane, which is made from used

0:25

cooking oils and plant oils, is growing

0:27

rapidly. Propane is the

0:30

energy for everyone. Learn more at

0:32

propane.com. Hey

0:34

everybody, it's Sabrina. I'm

0:36

here to talk to you one last time this

0:38

week about the New York Times audio subscription. Sorry.

0:42

We just know that change is hard

0:44

and we want to make absolutely certain

0:46

that all of you, our lovely, incredible

0:48

audience, understand what's going on.

0:51

So, two things. First,

0:53

today's show is not as long as

0:55

it looks. This week, we're doing this

0:57

kind of unusual thing where we attach

1:00

episodes of other New York Times podcasts

1:02

right to the end of our show. So

1:04

today, that's the fabulous podcast, Hard

1:06

Fork, where hosts Kevin Roose and

1:09

Casey Newton break down the latest

1:11

tech news with smarts, humor, and

1:13

expertise. We're doing that because,

1:15

and here's the second part of all of this,

1:17

the New York Times has launched an audio subscription

1:19

this week. That subscription gets

1:21

you full access to shows like us

1:24

and Hard Fork, Ezra Klein, The

1:27

Runup, The Interview, and Headlines. So,

1:30

this is included for people who subscribe

1:32

to all of the New York Times,

1:35

the big bundle with news, cooking, and

1:37

games. But you can

1:39

also sign up for audio separately.

1:42

Either kind of these subscriptions will

1:44

allow you to log on, Apple

1:46

Podcasts, Spotify, and you'll have access

1:48

to all past shows. Bonus content,

1:50

early access, stuff like that. Reminder,

1:53

you don't have to subscribe to

1:55

keep listening to the daily. Recent

1:58

episodes will still be... free. But

2:01

we hope you'll see it as a way to support

2:03

the show and the work we do. Okay.

2:05

Thank you for bearing with us. And with

2:08

all of these announcements, all this week, TGIF,

2:11

we just want everyone to know the deal. And

2:14

as always, thank you for listening. From

2:18

the New York times, I'm Sabrina

2:20

Taverducee, and this is the Daily. We

2:28

are following a major breaking story out

2:31

of the Middle East. Israel says the

2:33

leader of Hamas is dead. Yahya

2:35

Sinwar, the leader of Hamas and

2:38

the architect of the October 7th

2:40

attack, was killed by Israeli forces

2:42

in Gaza. Today, the images of

2:44

Sinwar's body lying in rubble, surrounded

2:46

by Israeli troops, sent shockwaves through

2:48

the region. Sinwar's assassination,

2:50

dealing a major blow to Hamas

2:53

amid the threat of wider escalation

2:55

in this region. It was a

2:57

major victory for Israel and prompted

2:59

calls from Israeli leaders for Hamas

3:02

to surrender. This war can

3:04

end tomorrow. It can end if

3:07

Hamas lays down its arms and returns

3:09

our hostages. But

3:11

what actually happens next is unclear.

3:14

Today, my colleague Ronan

3:17

Bergman on how Israel

3:19

got its most wanted man and

3:22

what his killing means for the future of

3:24

the war. It's

3:39

Friday, October 18th. So

3:46

Ronan, we're talking to you around 3 p.m. New York

3:48

time, 10 p.m. your time. Just

3:51

a few hours ago, we started getting hints

3:54

that Israel possibly killed the leader of Hamas,

3:56

Yahya Sinwar, and just a little while ago,

3:58

we started getting hints that Israel was

4:00

announced that he was in fact killed. What

4:04

was your first thought when you heard

4:06

the news? So candidly,

4:08

I thought, Oh, here

4:10

we go again, because on the 25th of

4:12

August, I got a call

4:14

from a source who said, we believe Sinoire

4:17

is dead. And then the source called

4:19

again, said, they thought it's

4:22

Sinoire, but it was not. So

4:25

I thought that the beginning, maybe it's the

4:27

same thing going again. But then we

4:29

got the first picture. And

4:33

when you look at the picture that was

4:35

just taken from the site,

4:39

though, you know, I'm not a

4:41

forensic expert, but it

4:43

looks like

4:46

the body of the

4:49

leader of Hamas, when

4:52

an hour later, my hutch was

4:54

that it was him, though

4:56

not yet confirmed, I

4:59

thought this is a

5:01

watershed moment where the

5:03

war can end. And

5:06

maybe, maybe the hostages could come

5:08

back. This is a

5:10

critical moment where things maybe

5:13

can go for the

5:15

first time in so long for

5:17

the better. Okay, we'll get

5:19

to that. But remind us

5:22

first who Sinoire was. Sinoire

5:26

was one of

5:28

the most important people in the history of

5:31

Hamas. Yekha Sinoire was

5:33

born in a refugee camp in

5:36

Hanyones, south of the

5:38

Gaza Strip. And

5:40

he was one of the young

5:43

first followers of Sheikh Ahmed Yasin,

5:46

the founder and the

5:49

leader and the spiritual compass

5:52

for Hamas, the

5:55

jihadist religious movements

5:58

he founded in the 80s. the

10:00

circles of secrecy, the intelligence community

10:02

and the leadership of the military

10:04

and later the Prime Minister office

10:07

heard that there is a strong chance that he

10:09

is dead. It's all by

10:11

coincidence. So

10:15

basically, after a year

10:17

of trying to kill him, using

10:20

all of the technology and

10:22

intelligence that Israel has at

10:25

its fingertips, they

10:27

killed him kind of by accident

10:30

is what you're saying. Like almost, you know,

10:33

by mistake. And it was a

10:35

bunch of trainees who did it. And he wasn't even

10:37

in a tunnel like everyone assumed. He was just kind

10:39

of walking around out there. So

10:42

one of the things that we

10:44

already identified this phenomena which

10:47

was discovered by Israeli intelligence

10:49

only in the latest stages of the

10:52

war is that it

10:54

is very hard to spend constant

10:57

days and weeks in the tunnels. You

11:01

know, our embed with

11:03

the IDF tours in Gaza, I have been in

11:05

many of these tunnels. And

11:07

let me tell you Sabine, it's even

11:09

in the tunnels that are built

11:13

for Hamas leadership, we have been

11:15

in few. It's very

11:17

hard to stay human,

11:21

claustrophobic, small,

11:26

narrow, and

11:28

everybody needed to go

11:30

out from time to time. Sinwar

11:32

thought that the

11:35

area is free from

11:37

enemy hostiles. He was

11:39

wrong. He was killed. Now

11:43

Ronan, I understand from your reporting

11:45

recently that this was not the

11:48

first time that the IDF was at

11:50

least close to killing him. Tell

11:52

me that story. Yes,

11:54

okay. So in January of 2024,

11:57

this year, fantastic

14:00

sci-fi, inspiring motivation, and more. It's all

14:02

there in the Audible app. There are

14:05

also thousands of included titles with more

14:07

added every week, so you've always got

14:09

something new to try. There's more to

14:11

imagine when you listen on Audible. Find

14:14

out for yourself. Sign up

14:16

for a free 30-day trial at

14:18

audible.com/the daily. How can

14:21

a microchip manufacturer keep track of

14:23

250 million control points at once?

14:26

How can technology behind animated movies

14:28

help enterprises reimagine their future? Built

14:30

for Change listeners know those answers,

14:32

and more. I'm Elise Hugh. And

14:34

I'm Josh Klein. We're the hosts

14:36

of Built for Change, a podcast

14:38

from Accenture. We talk to leaders

14:40

of the world's biggest companies to

14:42

hear how they've reinvented their business

14:44

to create industry shifting impact. And

14:46

how you can too. New episodes

14:49

are coming soon, so check out

14:51

Built for Change wherever you get

14:53

your podcasts. So what are these

14:55

important documents on this computer that

14:59

they found? And what's the story that

15:02

the documents reveal? So

15:05

what is there is

15:07

the first understanding of

15:09

the decision-making process, the

15:11

preparations, the

15:14

deception, everything that Hamas

15:17

leaders were doing throughout

15:19

the last two years before the war. There

15:23

are the minutes of these meetings, 10

15:25

meetings from July 2021 until the 7th

15:27

of August, 2023. So

15:34

exactly two months before the attack,

15:37

where Hamas leaders,

15:41

Sinoar, and five other

15:44

military leaders were

15:46

talking freely because

15:49

they were sure that Israeli

15:51

intelligence doesn't listen

15:53

it to that room. So

15:56

these documents are minutes of

15:58

high-level meetings. peace

24:00

to the north with Hezbollah

24:05

and at the

24:07

end form something that could turn

24:09

the page into something better, a

24:11

better day for the

24:14

Middle East. Unfortunately, I'm

24:16

not sure that this

24:18

will happen because

24:22

Israel is now gearing

24:24

towards a

24:26

massive attack on Iran. But

24:29

Ronan, I want to understand that because Israel

24:32

has always said since October 7

24:34

that they want to eliminate Hamas

24:36

and a huge part of that

24:38

was killing the head of Hamas,

24:40

Sid Israel

24:57

has gained some successes

24:59

in its war against Hezbollah. Israeli

25:03

defense establishment is regaining its

25:05

pride for these successes

25:09

and they feel that this can go on. And

25:13

even without a clear exit

25:15

strategy, they believe they can,

25:18

some of them believe they can just win again and again.

25:21

The second is Netanyahu

25:24

is coerced by parts

25:26

of his extreme parts of his

25:28

coalition to continue, not sign a

25:30

ceasefire, and

25:33

start some kind of

25:35

implementation of military rule

25:37

in Gaza. So, Israeli

25:40

coming back to Gaza and we hear

25:42

parts of the coalition talking about re-establishment

25:44

of settlements in Gaza that were taken

25:47

down when Israel disengaged from the

25:49

Strip. And also,

25:52

Prime Minister Netanyahu knows that when the

25:54

war is over, officially over, so

25:56

there is some kind of an agreement, there will

25:58

be a new war. Teen

30:01

accounts are private by default, helping

30:04

to protect them from unwanted contact. And

30:07

teenagers under 16 require

30:09

parental approval to change these protections. Instagram

30:12

Teen Accounts. New

30:14

built-in protections for parents' peace

30:17

of mind. Learn

30:19

more at instagram.com/teenaccounts. Did

30:25

you know that there are everyday actions you

30:27

can take to help keep energy reliable and

30:29

electricity costs down? Try PGE's new home energy

30:32

analysis tool and get personalized tips to lower

30:34

your bill. Visit

30:36

portlandgeneral.com/take action and

30:38

save. Here's

30:43

what else you should know today. On

30:46

Thursday, an independent panel

30:48

reviewing the failures that led

30:50

to the attempted assassination of

30:52

former President Donald Trump in

30:54

Butler, Pennsylvania, said that agents

30:56

involved in the security planning did

30:58

not take responsibility in the lead-up to

31:01

the event, nor did they

31:03

own the failures in the aftermath. The

31:06

panel, which included former Department of

31:09

Homeland Security Secretary Janet Napolitano, called

31:12

on the Secret Service to replace

31:14

its leadership with people from

31:16

the private sector and shed

31:19

much of its historic role

31:21

investigating financial crimes to focus

31:23

almost exclusively on its protective

31:25

mission. And

31:28

federal prosecutors have charged a

31:30

man they identified as an

31:32

Indian intelligence officer with trying

31:35

to orchestrate from abroad an

31:37

assassination on U.S. soil. An

31:40

indictment unsealed in Manhattan on

31:42

Thursday said that the man,

31:44

Vikash Yadav, whom authorities

31:46

believe is currently in India, directed

31:49

the plot that targeted a New

31:51

York-based critic of the Indian government,

31:54

a Sikh lawyer and political activist. The

31:57

charges are part of an escalating response from

32:00

the U.S. and Canada to

32:02

what those governments see as

32:04

brazenly illegal conduct by India,

32:07

a long-time partner. Today's

32:13

episode was produced by Mooj

32:16

Zaidi, Rob Zipko, Diana Nguyen,

32:18

and Eric Krupke. It

32:21

was edited by Paige Cowett and

32:23

MJ Davis-Lynn, with help

32:25

from Chris Haxl. It

32:27

contains original music by Rowan

32:29

Nemestow and Diane Wong, and

32:32

was engineered by Chris Wood. Our

32:35

theme music is by Jim Brunberg and Ben

32:37

Landsberg of Wonderly. Special

32:39

thanks to Patrick Kingsley. Remember

32:50

to catch a new episode of

32:52

The Interview right here tomorrow. David

32:55

Marchese talks to adult film

32:57

star turned influencer Mia Khalifa.

33:00

I am so ashamed of the things that

33:03

I've said and thought about myself and allowed

33:05

others to say and jokes

33:07

that I went along with and contributed to

33:09

about myself or about other women or anything

33:12

like that. I'm extremely ashamed of that.

33:24

That's it for The Daily. I'm

33:27

Sabrina Taverneesi. See you on

33:29

Monday. We

33:38

just got another email from somebody who said they thought I

33:40

was bald. I have

33:43

an apparently crazy bald energy that I bring

33:45

to this podcast. What do you think is

33:48

bald seeming about you? I

33:50

think for me, they think of me as

33:52

a wacky sidekick, which is a bald energy.

33:56

I think so. I don't think of...

33:58

I don't associate wacky and bald. Because

34:00

I'm thinking Jeff Bezos. I'm like, I'm

34:02

like, I know a lot of like

34:04

very hardcore ball interesting. So do you

34:06

think that maybe people think that I

34:08

sound like a sort of Titan of

34:10

industry plutocrat? I would not

34:13

say that's the energy you're giving is plutocrat

34:15

energy, but really because I

34:17

just fired 6,000 people to show

34:19

that I could you did

34:21

order me to come to the office today. I

34:24

did. I said there's a return to office and effect immediately.

34:27

No questions. I'm

34:34

Kevin is a tech columnist from the

34:36

New York Times. I'm Casey noon from

34:38

platformer and this is hard for this

34:40

week. Are we reaching the AI endgame

34:42

a new essay from the CEO of

34:44

Antropic has Silicon Valley talking then uber

34:46

CEO Dara Koshra Shaw. He joins us

34:48

to discuss his company's new partnership with

34:50

Waymo and the future of autonomous vehicles

34:52

and finally internal TikTok documents tell us

34:54

exactly how many videos you need to

34:56

watch to get hooked and so I

34:58

did very brave. God help me. Well,

35:10

Kevin the AI race continues to accelerate

35:12

and this week the news is coming

35:14

from an anthropic now last year you

35:16

actually spent some time inside this company

35:18

and you called it the white hot

35:21

center of AI doomerism. Yes. Well the

35:23

headline of my piece called it the

35:25

white hot center of AI doomerism. Just

35:27

want to clarify blame the headline. Well,

35:29

you know reporters don't often write our

35:32

own headlines. So I just feel the

35:34

need to clarify that fair enough. But

35:36

the story does talk about how many

35:38

of the people you met inside this

35:40

company seemed strangely pessimistic

35:42

about what they were building. Yeah, it

35:45

was a very interesting reporting experience because

35:47

I got invited to spend you know

35:49

several weeks just basically embedded at anthropic

35:51

as they were gearing up to launch

35:54

an update of their chatbot Claude and

35:57

I sort of expected it you know, they would

35:59

go in and try. to impress me with how

36:01

great Claude was and talk about all the amazing

36:03

things it would allow people to do. And then

36:05

I got there and it was like, all they

36:07

wanted to do was talk about how scared they

36:09

were of AI and of releasing these systems into

36:11

the wild. I compared it in the piece to

36:13

like being a restaurant critic who shows up at

36:15

like a buzzy new restaurant and all

36:17

anyone wants to talk about is food poisoning. And

36:20

so for this reason, I was very

36:22

interested to see over the past

36:25

week, the CEO of Anthropic, Mario

36:27

Amadei, write a 13,000 word

36:30

essay about his vision of the

36:32

future. And in

36:35

this essay, he says that he is

36:37

not an AI doomer does not think

36:39

of himself as one but actually thinks

36:42

that the future is quite bright and

36:44

might be arriving very quickly. And

36:46

then shortly after that, Kevin, the company

36:49

put out a new policy, which they

36:51

call a responsible scaling policy that I

36:53

thought had some interesting things to say

36:55

about ways to safely build AI systems.

36:58

So we want to talk about this today for

37:00

a couple reasons. One is

37:02

that AI CEOs have kept telling

37:04

us recently that major changes are

37:06

right around the corner. Sam Altman

37:08

recently had a blog post where

37:10

he said that an artificial super

37:12

intelligence could be just a few

37:15

thousand days away. And now

37:17

here Amadei is saying that AGI could

37:19

arrive in 2026, which

37:22

check your calendar, Kevin, that is in 14 months. Certainly

37:25

there is a case that this is just hype.

37:28

But even so, there are some very wild

37:30

claims in here that I do think deserve

37:32

broader attention. The second reason

37:34

that we want to talk about this

37:36

today is that anthropic is really the

37:38

flip side to a story that we've

37:40

been talking about for the past year

37:42

here, which is what happened to open

37:45

AI during and after Sam Altman's temporary

37:47

firing as CEO and proper was started

37:49

by a group of people who left

37:51

open AI primarily over safety concerns. And

37:53

recently, several more members of open AI's

37:55

founding team and their safety research teams

37:57

have gone over to anthropic. And so

37:59

in a way, Kevin, anthropic is an

38:02

answer to the question of what would

38:04

have happened if OpenAI's executive team hadn't

38:06

spent the past few years falling apart?

38:09

And while they are still the underdog compared to

38:11

OpenAI, is there a chance

38:13

that anthropic is the team that builds

38:15

AGI first? So that's what we want

38:17

to talk about today, but I want

38:19

to start by just talking about this

38:21

essay. Kevin, what did Dario Amadei have

38:24

to say in his essay, Machines of

38:26

Loving Grace? Yeah, so the first thing

38:28

that struck me is he's clearly reacting

38:30

to this perception, which I

38:32

may have helped create through my story last

38:34

year that sort of he and anthropic are

38:36

just doomers, right? That they are just a

38:39

company that goes around warning about how badly

38:41

AI could go if we're not careful. And

38:44

what he says in this essay that I thought was

38:46

really interesting and important is, we're

38:49

going to keep talking about the risks of

38:51

AI. This is not him saying, I don't

38:53

think this stuff is risky. I've been taken

38:56

out of context and I'm actually an AI

38:58

optimist. What he says is it's important to

39:00

have both, right? You can't just be going

39:02

around warning about the doom all the time.

39:04

You also have to have a positive vision

39:06

for the future of AI, because

39:10

that's not only what inspires and motivates

39:12

people, but it matters what

39:14

we do. I thought that

39:16

was actually the most important thing that he did

39:18

in this essay was he basically said, look,

39:21

this could go well or it could go

39:23

badly. And whether it

39:25

goes well or badly is up to us.

39:27

This is not some inevitable force. Sometimes people

39:29

in the AI industry, they have a habit

39:31

of talking about AI as if it's just

39:33

kind of this disembodied force that is just

39:35

going to happen to us. Inevitably.

39:37

Yes, and we either have to sort of like

39:39

get on the train or get run over by

39:41

the train. And what Dario says is

39:43

actually different. He says, here's a

39:45

vision for how this could go well, but

39:47

it's going to take some work to get there. It

39:50

also made me realize that for the past

39:52

couple of years, I have heard much more

39:54

about how AI could go wrong than how

39:56

it could go right from the AI CEOs.

39:59

As much as these guys get knocked for

40:01

endlessly hyping up their products, they also have,

40:03

I think to their credit, spent a lot

40:05

of time trying to explain to people that

40:08

this stuff is risky. And so there was

40:10

something almost counterintuitive about Dario coming

40:12

out and saying, wait, let's get really specific

40:14

about how this could go well. Totally. So

40:16

I think the first thing that's worth pulling

40:19

out from this essay is the timelines. Because

40:21

as you said, Dario Amade is claiming that

40:23

powerful AI, which is sort of his term,

40:25

he doesn't like AGI, he thinks it sounds

40:28

like too sci-fi, but powerful AI,

40:30

which he sort of defines as like an

40:33

AI that would be smarter than a Nobel

40:35

Prize winner in basically any field and

40:37

that it could basically control tools, go

40:40

do a bunch of tasks simultaneously. He

40:42

calls this sort of a country of

40:44

geniuses in a data center. That's sort

40:46

of his definition of powerful AI. And

40:49

he thinks that it could arrive as soon as 2026. I

40:52

think there's a tendency sometimes to be cynical

40:54

about people with short timelines like these, like,

40:57

oh, these guys are just saying this stuff

40:59

is going to arrive so soon because they

41:01

need to raise a bunch of money for

41:03

their AI companies. And, you know, maybe that

41:05

is a factor. But I

41:07

truly believe that at least Dario

41:09

Amade is sincere and serious about

41:12

this. This is not a

41:14

drill to him. And Anthropic is actually

41:16

making plans, scaling teams, and building products

41:18

as if we are headed into a

41:21

radically different world very soon, like within the

41:23

next presidential term. Yeah. And look, Anthropic is

41:25

raising money right now. And that does give

41:27

Dario motivation to get out there in the

41:30

market and start talking about curing cancer and

41:32

all these amazing things that he thinks that

41:34

that AI can do. At the

41:36

same time, you know, I think

41:39

that we're in a world where the discourse

41:41

has been a little bit poisoned by folks

41:43

like Elon Musk, who are constantly going

41:46

out into public, making bold claims about things that

41:48

they say are going to happen, you know, within

41:50

six months or a year, and then truly just

41:52

never happen. And our understanding of Dario based on

41:54

our own conversations with them and of people who

41:56

work with them is like, he is not that

41:59

kind of person. And this is not somebody who

42:01

lets his mouth run away with him. When

42:03

he says that he thinks this stuff could

42:06

start to arrive in 14 months, I actually

42:08

do give some credibility. Yeah. And and, you

42:10

know, you can argue with the time scales

42:12

and plenty of smart people disagree about this.

42:14

But I think it's worth taking this seriously, because

42:17

this is the head of one of the leading

42:19

AI labs, sort of giving you his thoughts on

42:21

not just what AI is going to change about

42:23

the world, but when that's going to happen. And

42:26

what I liked about this essay was that it

42:28

wasn't trying to sell me a vision of a

42:30

glorious AI future. Right. Dario says, you know, all

42:32

or some or none of this might come to

42:35

pass. But it was basically a thought

42:37

experiment. He has this idea in the

42:39

essay about what he calls the compressed

42:41

21st century. He

42:44

basically says, what if all AI

42:46

does is allow

42:48

us to make 100 years worth

42:50

of progress in the next 10

42:53

years in fields like biology? What

42:55

would that change about the world? And I thought that

42:57

was a really interesting way to frame it. Give

43:00

us some examples, Kevin, of what Dario says

43:02

might happen in this compressed 21st century. So

43:05

what he says in this essay

43:07

is that if we do get

43:09

what he calls powerful AI relatively

43:11

soon, that in the sort of

43:13

decade that follows that, we would

43:15

expect things like the prevention and

43:17

treatment of basically all natural infectious

43:20

disease, the elimination of most

43:22

types of cancer, very sort of

43:24

good embryo screening for genetic diseases

43:26

that would make it so that

43:28

more people didn't die of these

43:30

sort of hereditary things. He

43:33

talks about there being improved treatment for

43:35

mental health and other ailments. Yeah, I

43:37

mean, and a lot of this comes

43:39

down to just understanding the human brain,

43:41

which is an area where we still

43:43

have a lot to learn. And the

43:45

idea is if you have what he

43:47

calls this country of geniuses that's just

43:49

operating on a server somewhere, and they

43:52

are able to talk to each other,

43:54

to dream, to suggest ideas, to give

43:56

guidance to human scientists in labs to

43:58

run experiments, then you have this massive

44:00

compression. effect and all of a sudden

44:02

you get all of these benefits really

44:04

soon. And you know, obviously the headline

44:06

grabbing stuff is like, you know, Dario

44:08

thinks we're going to cure all cancer

44:10

and we're going to cure Alzheimer's disease.

44:13

Obviously those are huge, but there's also

44:15

kind of the more mundane stuff like

44:17

do you struggle with anxiety? Do you

44:19

have other mental health issues? Like are

44:21

you like mildly depressed? It's possible that

44:23

we will understand the the neural circuitry

44:25

there and be able to develop treatments

44:27

that would just lead to a general

44:29

rise in happiness. And that really struck

44:31

me. Yeah. And it sounds when you

44:34

just describe it that way, it sounds

44:36

sort of utopian and crazy, but what

44:38

he points out and what I actually

44:40

find compelling is like most scientific progress

44:42

does not happen in a straight line,

44:44

right? You have these kind of moments

44:46

where there's a breakthrough that enables a

44:48

bunch of other breakthroughs. And

44:50

we've seen stuff like this already happen

44:52

with AI, like with AlphaFold, which won

44:54

the freaking Nobel Prize this year in

44:57

chemistry, where you don't just have a

44:59

cure for one specific disease, but you

45:01

have a way of potentially discovering cures

45:03

for many kinds of diseases all at

45:05

once. There's a part of an essay

45:07

that I really liked where he points

45:09

out that CRISPR was

45:11

something that we could have invented long

45:13

before we actually did. But essentially, no

45:15

one had noticed the things they needed

45:17

to notice in order to make it

45:19

a reality. And he posits that there

45:21

are probably hundreds of other things like

45:23

this right now that just no one

45:25

has noticed yet. And if you had

45:27

a bunch of AI agents

45:29

working together in a room and they were

45:31

sufficiently intelligent, they would just notice those things

45:34

and we'd be off to the races. Right.

45:36

And what I liked about this section of

45:38

the essay was that it didn't try to

45:40

claim that there was some, you know,

45:43

completely novel thing that would

45:45

be required to result

45:47

in the changed world that he envisions. Right.

45:49

All that would need to happen for

45:52

society to look radically different 10

45:54

or 15 years from now in Dario's

45:57

mind is for that sort of base

45:59

rate. of discovery to accelerate rapidly due

46:01

to AI. Yeah. Now let's take a

46:03

moment to acknowledge folks in the audience

46:06

who might be saying, oh my gosh,

46:08

will these guys stop it with the

46:10

AI hype? They're accepting every premise that

46:12

these AI CEOs will just shovel it.

46:15

They can't get enough and it's irresponsible.

46:17

These are just stochastic parrots, Kevin. They

46:19

don't know anything. It's not intelligence and

46:21

it's never going to get any better

46:23

than it is today. And I just

46:26

want to say I hear you and

46:28

I see you and our email address

46:30

is Ezra Klein show and playtime suck

46:32

up. But here's

46:35

the thing. You can look at

46:37

the state of the art right now.

46:39

And if you just extrapolate what is

46:41

happening in 2024 and you assume some

46:44

rate of progress beyond where we currently

46:46

are, it seems likely to

46:48

me that we do get into a world where

46:50

you do have these sort of simulated PhD students

46:52

or maybe simulated super geniuses and they are able

46:55

to realize a lot of these kinds of statistics.

46:57

Now maybe it doesn't happen in five, 10 years.

46:59

Maybe it takes a lot longer than that. But

47:01

I just wanted to underline like we are not

47:04

truly living in the realm of fantasy. We are

47:06

just trying to get a few

47:08

years and a few levels of

47:10

advancement beyond where we are right

47:12

now. Yeah. And Dario does in

47:14

his essay make some caveats about

47:16

things that might constrain the rate

47:18

of progress in AI like regulation

47:21

or clinical trials taking a long

47:23

time. You know, he also

47:25

talks about the fact that some people may just opt

47:28

out of this whole thing

47:30

like they just may not want anything to

47:32

do with AI. There might be some political

47:34

or cultural backlash that sort of slows down

47:36

the rate of progress. And

47:38

he says you know like that could actually

47:41

constrain this and we need to think about

47:43

some ways to address that. Yeah. So

47:45

that is sort of the suite of things

47:47

that Dario thinks will benefit our lives. You

47:49

know there's a bunch more in there. You

47:51

know he thinks it will help with climate

47:54

change other issues. But

47:56

the essay has five parts and there was

47:58

another part of the essay. that really caught

48:00

my attention, Kevin. And it is

48:03

a part that looks a little bit

48:05

more seriously at the risks of this

48:07

stuff, because any super genius that was

48:09

sufficiently intelligent to cure cancer could otherwise

48:11

wreak havoc in the world. So

48:13

what is his idea for ensuring that AI

48:16

always remains in good hands? So he admits

48:18

that he's not like a geopolitics expert. This

48:20

is not his forte. Unlike the two of

48:22

us. Right. And

48:25

there have been, look, a lot of

48:27

people theorizing what the

48:30

politics of advanced AI are going

48:32

to look like. Dario says that

48:34

his best guess currently about how

48:37

to prevent AI from empowering autocrats

48:39

and dictators is through what he

48:41

calls an Entente strategy. Basically, you

48:44

want a bunch of democracies to

48:46

come together to secure their supply

48:48

chain, to block adversaries from getting

48:51

access to things like GPUs and

48:53

semiconductors, and that

48:55

you could basically bring countries

48:57

into this democratic alliance and

49:00

ice out the more authoritarian regimes from

49:02

getting access to this stuff. But I

49:04

think this was not the

49:06

most fleshed out part of the argument. Yeah,

49:09

well, and I appreciate that

49:11

he is at least making an

49:13

effort to come up with ideas

49:16

for how would you prevent AI

49:18

from being misused. But as I

49:20

was reading the discussion around the

49:22

blog post, I found this interesting

49:24

response from a guy named Max

49:26

Tegmark. Max is a professor at

49:28

MIT who studies machine learning. And

49:31

he's also the president of something called the

49:33

Future of Life Institute, which is a sort

49:35

of nonprofit focused on AI safety. And

49:38

he really doesn't like this idea

49:40

of what Dario calls the Entente,

49:42

the group of these democracies working

49:44

together. And he says he

49:47

doesn't like it because it essentially sets

49:49

up and accelerates a race. It

49:51

says to the world that

49:53

essentially whoever invents super powerful

49:55

AI first will win forever.

49:58

Because in this view, AI is essentially the

50:00

final technology that you ever need to invent, because

50:02

after that it'll just invent anything else it

50:04

needs. And he calls that

50:07

a suicide race. And the reason is this,

50:09

and he has a great quote, horny

50:11

couples know that it is easier to make a

50:14

human level intelligence than to raise and align it.

50:16

And it is also easier to make an AGI

50:18

than to figure out how to align or control

50:20

it. Wow. I

50:23

never thought about it like that. Yeah, you probably never thought I

50:25

would say horny couple on the show, but I just did. So

50:29

Kevin, what do you make of

50:31

this sort of feedback? Is there

50:33

a risk there that this effectively

50:35

serves as a starter pistol that

50:38

leads maybe our adversaries to

50:40

start investing more in AI and sort

50:42

of racing against us and triggering

50:44

some sort of doom spiral? Yeah, I mean, look,

50:46

I don't have a problem with China racing us

50:49

to cure cancer using AI, right? If

50:51

they get there first, like more power to them. But

50:54

I think the more serious risk is that

50:56

they start building the kind of AI that

50:58

serves Chinese interests, right? That it becomes a

51:01

tool for surveillance and control

51:03

of people rather than some of these more

51:05

sort of democratic ideals. And this is actually

51:07

something that I asked Dario about back last

51:09

year when I was spending all that time

51:12

at Anthropik, because this is

51:14

the most common criticism of Anthropik is like, well,

51:16

if you're so worried about AI and

51:19

all the risks that it could pose, like why are

51:21

you building it? And I asked him about this, and

51:23

his response was, he basically said, look, there's this problem

51:25

of, in AI

51:27

research of kind of intertwining, right? Of the

51:30

same technology that sort of advances the

51:32

state of the art in AI also

51:34

allows you to advance the state of

51:37

the art in AI safety, right? The

51:39

same tools that make the language models

51:41

more capable also make it possible to

51:43

control the behavior of the language models.

51:46

And so these things kind of go

51:48

hand in hand. And if

51:50

you want to compete on the frontier of

51:52

AI safety, you also have to compete on

51:54

the frontier of AI capabilities. Yeah, and I

51:56

think it's an idea worth considering. To me,

51:59

it just sounds like. like, wow, you are

52:01

really standing on a knife's edge there. If

52:03

you're saying in order to

52:05

have any influence over the future, we

52:07

have to be right at the frontier

52:10

and maybe even gently advance the

52:12

frontier and yet somehow not

52:14

accidentally trigger a race where all of

52:16

a sudden everything gets out of control.

52:19

But I do accept and respect that that

52:21

is Dario's viewpoint. But isn't that kind of

52:23

what we observed from the last couple of

52:26

years of AI progress, right? Like OpenAI, it

52:28

got out there with Chat GPT before

52:31

any of the other labs had released

52:33

anything similar. And Chat GPT

52:35

kind of set the tone for all of

52:37

the products that followed it. And so I

52:39

think the argument from Anthropik would be like,

52:43

yes, we could sort of be way behind the state

52:45

of the art. That would probably make us safer than

52:47

someone who was actually advancing the state of the art.

52:50

But then we missed the chance to kind of

52:52

set the terms of what future AI products from

52:54

other companies will look like. So it's sort of

52:56

like using a soft power in an effort to

52:58

influence others. Yeah, and the way they

53:00

put this to me last year was that they wanted

53:02

instead of there to be just

53:04

a race for raw capabilities of AI

53:06

systems, they wanted there to be a

53:09

safety race, right? Where companies would start

53:11

competing about whose models were the safest

53:13

rather than whose models could do

53:15

your math homework better. So let's talk

53:17

about the safety race and the other

53:19

thing that Anthropik did this week to

53:22

lay out a future vision for AI. And

53:24

that was with something that has, I'll say

53:26

it, kind of a boring name, the responsible

53:29

scaling policy. I understand. This

53:31

maybe wasn't going to come up over drinks

53:33

at the club this weekend. Yeah, but I

53:36

think this is something that people should pay attention

53:38

to because it's an example of what you just

53:40

said, Kevin. It is Anthropik trying to use some

53:42

soft power in the world to say, hey, if

53:44

we went a little bit more like this, we

53:47

might be safer. All right, so talk about what's

53:49

in the responsible scaling policy that Anthropik released this

53:51

week. Well, let's talk about what it is. subjected

54:00

to more scrutiny and they should have more

54:02

safeguards added to them. They put this out

54:04

a year ago and it

54:06

was actually a huge success in

54:09

this sense. Kevin open AI went

54:11

on to release its own version

54:13

of it. And then Google deep

54:15

mine released a similar scaling

54:18

policy as well this spring. So

54:21

now when Tropic is coming back just over

54:23

a year later and they say, we're going

54:25

to make some refinements. And

54:27

the most important thing that they say is

54:29

essentially we have identified two

54:32

capabilities that we think would be

54:34

particularly dangerous. And so if anything

54:36

that we make displays these capabilities,

54:39

we are going to add a

54:41

bunch of new safeguards. The

54:43

first one of those is if a

54:46

model can do its own AI research

54:48

and development, that is going to

54:50

start ringing a lot of alarm bells

54:52

and they're going to put many

54:54

more safeguards on it. And second, if

54:57

one of these models can

54:59

meaningfully assist someone who has

55:02

a basic technical background in

55:04

creating a chemical, biological, radiological

55:06

or nuclear weapon, then

55:08

they would add these new safeguards. What are

55:10

these safeguards? Well, they have a super long

55:12

blog post about it. You can look it

55:14

up, but it includes basic things like taking

55:17

extra steps to make sure that a foreign

55:19

adversary can't steal the model weights, for example,

55:21

or otherwise hack into the systems and run

55:23

away with it. Right. And this is some

55:25

of it is similar to things that were

55:27

proposed by the Biden White House in its

55:29

executive order on AI last year. This

55:32

is also, these are some of the steps that came

55:34

up in SB 1047, the

55:37

AI regulation that was

55:39

vetoed by Governor Newsom in California recently. So

55:41

these are ideas that have been floating out

55:43

there in the sort of

55:46

AI safety world for a while. But Anthropic

55:48

is basically saying, we are going to proactively

55:50

commit to doing this stuff even

55:52

before a government requires us to. Yeah. There's

55:54

a second thing I like about this, and

55:56

it relates to this SB 1047 that we

55:58

talked about on the show. Something that a

56:01

lot of folks in Silicon Valley didn't like

56:03

about it was the way that it tried

56:05

to identify danger. And it was not because

56:07

of a specific harm that a model could

56:09

cause. It was by saying,

56:11

well, if a model costs a certain

56:13

amount of money to train, right? Or

56:16

if it is trained with a certain amount

56:18

of compute, those were the proxies that

56:20

the government was trying to use to understand why

56:23

this would be dangerous. And

56:25

a lot of folks in Silicon Valley said,

56:27

we hate that because that has nothing to

56:29

do with whether these things could cause harm

56:31

or not. So what anthropic is doing here

56:33

is saying, well, why don't we try to

56:35

regulate based on the anticipated harm? Obviously, it

56:37

would be bad if you could log

56:40

on to Claude, anthropic's rival to chat GBT, and

56:42

said, hey, help me build a radiological weapon, which

56:44

is something that I might type into Claude because

56:46

I don't even know the difference between a radiological

56:48

weapon and a nuclear weapon. Do you? I hope

56:51

you never learn. I

56:53

hope I don't either. Because sometimes I have bad days, Kevin,

56:56

and I get to scheming. So

56:58

for this reason, I think that governments, regulators around

57:00

the world might want to look at this approach

57:02

and say, hey, instead of trying to regulate this

57:04

based on how much money AI labs are spending

57:06

or like how much compute is involved, why

57:08

don't we look at the harms we're trying

57:11

to address and say, hey, if you build

57:13

something that could cause this kind of harm,

57:15

you have to do X, Y, and Z.

57:17

Yeah, that makes sense to me. So I

57:19

think the biggest impact that both the sort

57:21

of essay that Dario wrote and this responsible

57:23

scaling policy had on me was not about

57:25

any of the actual specifics of the idea.

57:27

It was purely about the timescales and the

57:29

urgency. It is one thing to hear a

57:31

bunch of people telling you that

57:33

AI is coming and that it's going to be

57:35

more powerful than you can imagine, sooner than you

57:37

can imagine. But if you

57:40

actually start to internalize that and

57:42

plan for it, it just

57:44

feels very different. If

57:47

we are going to get powerful AI

57:49

sometime in the next, let's call it

57:51

two to 10 years, you

57:53

just start making different choices. Yeah, I

57:56

think it becomes sort of the calculus.

57:58

I can imagine it affecting. what

58:00

you might want to study in college if

58:02

you are going to school right now. I

58:05

have friends who are thinking

58:07

about leaving their jobs because they think the place

58:10

where they are working right now will

58:12

not be able to compete in

58:14

a world where AI is very

58:16

widespread. So yes, you're absolutely starting

58:19

to see it creep into the

58:21

calculus. I don't

58:23

know kind of what else it could do. There's

58:26

no real call to action here

58:28

because you can't really do very

58:30

much until this world begins to

58:32

arrive. But I do think psychologically,

58:34

we want people to at least imagine,

58:37

as you say, what it would be

58:39

like to live in this world because

58:41

I have been surprised

58:43

at how little discussion this has been

58:45

getting. Yeah, I totally agree. I mean,

58:47

to me, it feels like we

58:50

are entering, I wouldn't call it like

58:52

an AI end game because I think

58:55

we're closer to the start than the

58:57

end of this transformation. But it does

58:59

feel like something is happening. I'm starting

59:01

to notice AI's effects in my life

59:03

more. I'm starting to feel more dependent

59:06

on it. And I'm also like, I'm

59:08

kind of having an existential crisis. Really?

59:10

Like not a full blown one, but

59:12

like typically I'm a guy who likes

59:14

to plan. I like to strategize. I

59:16

like to have like a five year

59:18

and a 10 year plan. And I've

59:20

just found that my own certainty about

59:22

the future and my ability to plan

59:24

long-term is just way lower

59:27

than it has been for any time

59:29

that I can remember. That's interesting. I mean, for myself,

59:31

I feel like that has always been true. In

59:34

1990, I did not know what things were gonna look like in

59:36

2040. And I would be

59:38

really surprised by a lot of things that have happened along the

59:40

way. But yeah, there's a

59:42

lot of uncertainty out there. It's scary, but I also

59:44

like, do

59:47

you not feel a little bit excited about it?

59:50

Of course. Look,

59:52

I love software. I love tools. I wanna live

59:54

in the future. And it's already happening to me.

59:57

There is a lot of that uncertainty and like that stuff for

59:59

you. It just freaks me out. But

1:00:02

if we could cure cancer, if we

1:00:04

could cure depression, if we could cure

1:00:06

anxiety, you'd be talking about the greatest

1:00:09

advancement to human well-being, certainly

1:00:11

in decades, maybe that we've ever seen. Yeah. I

1:00:14

mean, I have some priors

1:00:17

on this because my dad died of a

1:00:19

very rare form of cancer that was, he

1:00:24

was a sub 1% type of cancer.

1:00:27

And when he got sick, it was like,

1:00:30

I read all the clinical trials and

1:00:32

it was just like, there hadn't been

1:00:34

enough people thinking about this specific type

1:00:36

of cancer and how to cure

1:00:38

it because it was not breast cancer, it

1:00:40

was not lung cancer, it was not something

1:00:42

that millions of Americans get. And

1:00:45

so there just wasn't the kind of

1:00:47

brain power devoted to trying to solve

1:00:49

this. Now it has subsequently, it hasn't

1:00:51

been solved, but there are now treatments

1:00:53

that are in the pipeline that didn't

1:00:55

exist when he was sick. And I

1:00:58

just constantly am wondering like,

1:01:01

if he had gotten sick now instead

1:01:03

of when he did, like maybe

1:01:05

he would have lived. And I think

1:01:08

that is one of the things that

1:01:10

makes me really optimistic about AI is

1:01:12

just like, maybe

1:01:14

we just do have the brain power or

1:01:16

we will soon have the brain power to

1:01:18

devote, world-class research teams

1:01:21

to these things that might not affect

1:01:23

millions of people, but that do affect

1:01:25

some number of people. Absolutely. I just,

1:01:27

I don't know, it really, I

1:01:30

got kind of emotional reading

1:01:32

this essay because it was just like, obviously it's,

1:01:35

I'm not someone who believes all the

1:01:37

hype, but I'm like, I assign some

1:01:40

non-zero probability to the possibility

1:01:42

that he's right, that all this stuff could happen.

1:01:44

And I just find that so much more

1:01:48

interesting and fun to think about than like

1:01:50

a world where everything goes off the rails.

1:01:52

Well, it's just the first time that we've

1:01:55

had a truly positive, transformative

1:01:57

vision for the world. world coming

1:01:59

out of Silicon Valley in a

1:02:01

really long time. In fact, this

1:02:03

vision, it's more positive and optimistic

1:02:05

than anything that has been like

1:02:08

in the presidential campaign. It's

1:02:10

like when the presidential candidates talk about the future of

1:02:12

this country, it's like, well, we'll

1:02:14

give you this tax break, right? Or we'll

1:02:16

make this other policy change. Nobody's

1:02:19

talking about how they're gonna fricking cure cancer, right?

1:02:21

So I think, of course we're drawn

1:02:23

to this kind of discussion because it

1:02:25

feels like there are some people in

1:02:28

the world who are taking really, really

1:02:30

big swings, and if they connect, then

1:02:32

we're all gonna benefit. Yeah, yeah. Yeah.

1:02:41

When we come back, why Uber has way more

1:02:44

autonomous vehicles on the road than it used to.

1:02:58

OK, so one of the biggest developments

1:03:00

over the past few months in tech

1:03:02

is that self-driving cars now are actually

1:03:04

working. Yeah, but this is no longer

1:03:06

in the realm of sci-fi. Yes, so

1:03:09

we've talked, obviously, about the self-driving cars

1:03:12

that you can get in San Francisco now. It used

1:03:15

to be two companies, Waymo, and Waverly. And it's a

1:03:17

big deal. And it's a big deal. And it's a

1:03:19

big deal. And it's a big deal. And it's a

1:03:21

big deal. And it's a big deal. And it's a

1:03:23

big deal. And it's a big deal. And it's a

1:03:26

big deal. And it's a big deal. It's gonna be

1:03:28

two companies, Waymo and Cruise. Now it's just Waymo. And

1:03:31

there have also been a bunch of different

1:03:33

autonomous vehicle updates from other companies that are

1:03:35

involved in this space. And the

1:03:37

one that I found most interesting recently was

1:03:39

about Uber. Now, as you

1:03:42

will remember, Uber used to try

1:03:44

to build its own robotaxes. They

1:03:46

gave that back in 2020. That

1:03:49

was the year they sort of sold off their autonomous

1:03:52

driving division to a startup called

1:03:54

Aurora after losing just an

1:03:56

absolute ton of money on it. But

1:03:59

now they are back. back in

1:04:01

the game and they just recently

1:04:03

announced a multi-year partnership with Cruise,

1:04:05

the self-driving car company. They also

1:04:08

announced an expanded partnership with Waymo,

1:04:10

which is going to allow Uber

1:04:12

riders to get AVs in Austin,

1:04:14

Texas and Atlanta, Georgia. They've

1:04:16

been operating this service in Phoenix since

1:04:19

last year and that's going to keep

1:04:21

expanding. They also announced that self-driving Ubers

1:04:23

will be available in Abu Dhabi through

1:04:26

a partnership with the Chinese AV company

1:04:28

WeRide. And they've also

1:04:30

made a long-term investment in Wave,

1:04:32

which is a London-based autonomous driving

1:04:34

company. So they are investing

1:04:37

really heavily in this and they're doing it

1:04:39

in a different way than they did back

1:04:41

when they were trying to build their own

1:04:43

self-driving cars. Now they are essentially saying, we

1:04:45

want to partner with every company that we

1:04:47

can that is making self-driving cars. Yeah, so

1:04:49

this is a company that many people take

1:04:51

several times a week, Uber, and

1:04:54

yet I feel like it sometimes is

1:04:56

a bit taken for granted. While

1:04:59

we might just focus on the cars you

1:05:01

can get today, they are thinking very long-term

1:05:03

about what transportation is going to look like

1:05:05

in five or 10 years. And increasingly for

1:05:07

them, it seems like autonomous vehicles are a

1:05:09

big part of that answer. Yeah, and what

1:05:11

I found really interesting, so Tesla had this

1:05:13

RoboTaxi event last week where Elon Musk talked

1:05:15

about how you'll soon be able to hail

1:05:17

a self-driving Tesla. And what

1:05:19

I found really interesting is that Tesla's

1:05:21

share price plummeted after that event, but

1:05:23

Uber's stock price rose to an all-time

1:05:25

high. So clearly people think that, or

1:05:27

at least some investors think that Uber's

1:05:29

approach is better here than Tesla's. It's

1:05:32

the sort of thing, Kevin, that makes

1:05:34

me want to talk to the CEO

1:05:36

of Uber. And lucky for you, he's

1:05:38

here. Oh, thank goodness. So today we're

1:05:40

going to talk with Uber CEO Dara

1:05:42

Khazurshawi. He took over at Uber in

1:05:44

2017 after a bunch

1:05:46

of scandals led the founder of Uber,

1:05:48

Travis Kalanick, to step down. He

1:05:51

has made the company profitable for the first time in its

1:05:53

history, and I think a lot of people think he's

1:05:55

been doing a pretty good job over there. And

1:05:58

he is leading this charge. into autonomous

1:06:00

vehicles, and I'm really curious to hear

1:06:02

what he makes, not just of Uber's

1:06:04

partnership with Waymo, but of sort of

1:06:06

the whole self-driving car landscape. Let's bring

1:06:08

him in. Let's do it. ["The

1:06:11

Star-Spangled Banner"] Dara

1:06:18

Kazurshawi, welcome to Hard Fork. Thank you for

1:06:20

having me. So you were previously on the

1:06:22

board of the New York Times Company until

1:06:24

2017, when you stepped down right

1:06:27

after taking over at Uber. I assume

1:06:30

you still have some pull with our bosses, though,

1:06:32

because of your years of service. So can you

1:06:34

get them to build us a nicer studio? I

1:06:37

didn't have pull when I was on the board,

1:06:39

and I certainly have zero pull now. I've got

1:06:41

negative pull, I think. They're taking revenge on me.

1:06:45

Well, since you left the board, they're making

1:06:47

all kinds of crazy decisions, like letting us

1:06:49

start a podcast and stuff. Yeah. Oh, my

1:06:52

god. But all right, so we are going

1:06:54

to talk today about your new partnership with

1:06:56

Waymo and the autonomous driving future.

1:06:58

I would love to hear the story of

1:07:01

how this came together, because I think for

1:07:03

people who've been following this space for a

1:07:05

number of years, this was surprising. Uber and

1:07:07

Waymo have not historically had a great relationship.

1:07:10

The two companies were embroiled in litigation and

1:07:12

lawsuits and

1:07:14

trade secret theft and things like that. There was

1:07:17

a big deal. And

1:07:19

so how did they approach

1:07:21

you? Did you approach them? How did this

1:07:23

partnership come together? I guess it's time healing.

1:07:25

Right? When

1:07:27

I came on board, we thought

1:07:29

that we wanted to establish a better

1:07:31

relationship with Google generally, Waymo generally. And

1:07:34

even though we were working on our

1:07:36

own self-driving technology, it was

1:07:38

always within the context of we were

1:07:41

developing our own, but we want to work with third parties as well.

1:07:44

One of the disadvantages of developing our own

1:07:46

technology was that some

1:07:48

of the other players of Waymo's of the world, et cetera, hurt

1:07:51

us, but didn't necessarily believe

1:07:53

us. It's difficult to work

1:07:55

with players that you compete. So

1:07:58

one of the first decisions that we made

1:08:00

that we made was we can't be

1:08:02

in between here. Either you have to

1:08:04

go vertical or you

1:08:06

have to go platform strategy. You can't achieve

1:08:10

both and we have to make it back. We either

1:08:12

have to do our own thing or we have to

1:08:14

do it with partners. Yeah, absolutely. And

1:08:16

so that strategic kind of

1:08:18

fork became quite apparent to me.

1:08:21

And then the second was just, what are we good at? Listen,

1:08:24

I'll be blunt, we sucked at

1:08:26

hardware, right? We try to

1:08:29

apply software principles to hardware. It

1:08:31

doesn't work. Hardware is a different

1:08:33

place, different demand in terms

1:08:35

of perfection, et cetera. And

1:08:38

ultimately that fork, do we go vertical?

1:08:40

And there are very few companies that

1:08:42

can do software and hardware. Well, Apple,

1:08:44

Tesla are arguably one of the few

1:08:46

in the world. And

1:08:49

we decided to make a bet on the platform. And

1:08:52

so once we made that bet, we

1:08:54

went out and identified who are the leaders.

1:08:56

And so I think that's why we got

1:08:58

Google to be a bigger shareholder. And then

1:09:00

over a period of time, we built relationships.

1:09:02

And I do think there's a synergy between

1:09:05

the two. So it just makes sense, the relationship,

1:09:09

and we're very, very excited to, on

1:09:12

a forward basis, expand it pretty significantly. So

1:09:14

this was, I feel like maybe your

1:09:17

most consequential decision to date as

1:09:19

the CEO of this company. If

1:09:21

you believe that Google is a big part

1:09:23

of this company, if

1:09:26

you believe that AVs are gonna become the norm

1:09:28

for many people hailing a ride in 10 or

1:09:30

15 years, it's

1:09:32

conceivable that they might open up the Waymo

1:09:34

app, and not the Uber app. Waymo has

1:09:37

an app to order cars. I use it

1:09:39

fairly regularly. So what gave

1:09:41

you the confidence that in that world,

1:09:43

it will still be Uber that is

1:09:45

the app that people are turning to

1:09:47

and not Waymo or whatever other apps

1:09:49

might have arisen for other AV companies?

1:09:52

The first is that it's not a

1:09:54

binary outcome. Okay, I think that a

1:09:56

Waymo app and an Uber app can

1:09:58

coexist. We saw it. in my

1:10:00

old job in the travel business, right? I ran

1:10:02

Expedia and there's this dramatic,

1:10:04

is Expedia going to put the hotel chains

1:10:06

out of business or the hotel chains going

1:10:09

to put Expedia out of business? The fact

1:10:11

is both thrived. And there's a

1:10:13

set of customers who have booked through Expedia. There's

1:10:15

a set of customers who books hotel direct and

1:10:18

both businesses have grown and interactivity

1:10:20

in general has grown. Same

1:10:22

thing if you look in food, right?

1:10:24

McDonald's has its own app. It's a

1:10:26

really good app. It has a loyalty

1:10:28

program. Starbucks has its own app, has

1:10:30

a loyalty program, yet both are engaging

1:10:32

with us through the Uber Eats marketplace.

1:10:35

So my conclusion was that there isn't an

1:10:37

either or. I do believe there will be

1:10:39

other companies. There'll be cruises and there'll be

1:10:41

wee rides and waves, et cetera. There'll be

1:10:44

other companies and self-driving choices. And

1:10:46

the person who wants utility, speed,

1:10:48

ease, familiarity will choose Uber and

1:10:50

both can coexist and both can

1:10:52

thrive and both are really going

1:10:54

to grow because autonomous will be

1:10:56

the future eventually. So tell us

1:10:58

more about the partnership with Waymo

1:11:00

that is going to take place

1:11:03

in Austin and Atlanta. Who is

1:11:05

actually paying for the

1:11:07

maintenance of the cars? Does Uber have to

1:11:09

sort of make sure that there's no trash

1:11:12

left behind in the cars? What is Uber

1:11:14

actually doing in addition to just making these

1:11:16

rides available through the app? Sure. So I

1:11:18

don't want to talk about the economics because

1:11:20

they're confidential in terms of

1:11:22

the deal, but in those

1:11:24

two cities, Waymo will

1:11:26

be available exclusively through the

1:11:28

Uber app and we will

1:11:30

also be running the fleet

1:11:32

operations as well. So depots, recharging,

1:11:34

cleaning, if something gets lost, making

1:11:36

sure that it gets back to

1:11:38

its owner, et cetera. And

1:11:41

Waymo will provide the software driver, will

1:11:44

obviously provide the hardware, repair the hardware,

1:11:46

et cetera. And then we will be

1:11:48

doing the upkeep and operating the networks.

1:11:51

For riders, if you want to get in

1:11:53

a Waymo in one of those cities through

1:11:55

Uber, is there an option to specifically request

1:11:57

a self-driving Waymo or is

1:11:59

it? just kind of chance. Like if

1:12:02

the car that's closest to you happens to be

1:12:04

a Waymo, that's the one you get. Right now,

1:12:06

the experience, for example, in Phoenix, is that it's

1:12:08

by chance. I think you got one by chance,

1:12:10

and you can say, yes, I'll do it or

1:12:12

not. And I think that's what we're going

1:12:14

to start with. But there may be some people who only

1:12:16

want Waymos, and there are some people who may not want

1:12:18

Waymos. And we'll solve for that over a period of time.

1:12:20

It could be personalizing preferences,

1:12:23

or it could be what you're talking about, which is

1:12:25

I only want a Waymo. Do the passengers get rated

1:12:27

by the self-driving car the way that they would in

1:12:30

a human-driven Uber? Not yet,

1:12:32

but that's not a bad idea. What

1:12:34

about tipping? If I get out of a self-driven Uber,

1:12:37

is there an option to tip the car if it did a good job? I'm

1:12:39

sure we could build that. Why not? I

1:12:42

don't know. I do wonder if people are going

1:12:44

to tip machines. I don't think

1:12:47

it's likely, but you never know. It sounds crazy,

1:12:49

but at some point, someone is going to start

1:12:51

asking because they're going to realize it's just free

1:12:53

margin. It's like even if only 100 customers do

1:12:55

it in a whole year. I don't know. It's

1:12:57

just free money. Totally. The good news is tipping

1:12:59

100% of tips go to drivers now, and we

1:13:01

definitely want to keep that. So we like the

1:13:03

tipping habits, but whether people tip machines is TBD.

1:13:06

Yeah. And how big are these fleets? I think

1:13:08

I read somewhere recently that Waymo has about 700

1:13:10

self-driving cars operating

1:13:12

nationwide. How many AVs are we talking

1:13:14

about in these cities? We're starting in

1:13:16

the hundreds, and then we'll expand

1:13:19

from there. I

1:13:21

know you don't want to discuss the economics, even though

1:13:23

I would love to learn what the split is there.

1:13:25

I'm not going to tell you. But

1:13:27

you did recently talk about the margins on

1:13:30

autonomous rides being lower than the

1:13:32

margins on regular Uber rides for

1:13:34

at least a few more years.

1:13:37

That's not intuitive to me because in an autonomous

1:13:39

ride, you don't have to pay the driver. So

1:13:41

you would think the margin would be way higher

1:13:44

for Uber. But why would you make less money

1:13:46

if you don't have to pay a driver? So

1:13:48

generally, our design spec in terms of how we

1:13:50

build businesses is any newer business, we're

1:13:52

going to operate at a lower margin while we're

1:13:54

growing that business. You don't want

1:13:56

it to be profitable day one. And that's my

1:13:59

attitude with autonomous, which is a good one. again,

1:14:01

get it out there, introduce it to as many

1:14:03

people as possible. At a maturity

1:14:05

level, generally, if you look at our take

1:14:07

rate around the world, it's about 20%,

1:14:09

we get 20%, the driver gets 80%. We

1:14:13

think that's a model that makes sense for any

1:14:15

autonomous partner going forward. And that's, that's what we

1:14:17

expect. I kind of don't care, honestly, what the

1:14:19

margins are for the next five years. The question

1:14:22

is, can I get lots of supply? Can

1:14:24

it be absolutely safe? And,

1:14:27

you know, does that 2080 split

1:14:29

look reasonable going forward? And I think it does.

1:14:32

Yeah. I want to ask

1:14:34

about Tesla. You mentioned them a little earlier.

1:14:37

They held an event recently

1:14:39

where they unveiled their plans

1:14:41

for a robotaxi service. Do

1:14:44

you consider Tesla a competitor? Well,

1:14:47

they certainly could be right if they develop

1:14:49

their own AV vehicle and

1:14:51

they decide to go

1:14:54

direct only through the Tesla

1:14:56

app, they would be a competitor. And

1:14:58

if they decide to work with us,

1:15:01

then we would be a partner as

1:15:03

well. And to some extent, again, both can be

1:15:05

true. So I don't think it's going to be

1:15:07

an either or I think

1:15:09

Elon's vision is pretty compelling, especially

1:15:11

like you might have these, uh,

1:15:13

cyber shepherds or these, these owners

1:15:16

of these fleets, et cetera. Those

1:15:18

owners, if they want to

1:15:20

have maximum earnings on those fleets, will

1:15:23

want to pull those fleets on

1:15:25

Uber, but at this point it's unknown what

1:15:27

his intentions are. There's this

1:15:29

big debate that's playing out right

1:15:31

now about who has the better,

1:15:34

uh, AV strategy between Waymo and Tesla

1:15:36

in the sense that the Waymos have

1:15:39

many, many sensors on them. The vehicles

1:15:41

are much more expensive to produce. Uh,

1:15:44

Tesla is trying to get

1:15:46

to full autonomy using only

1:15:48

its cameras, um, and software

1:15:51

and, uh, Andre Carpathi, the AI researcher recently

1:15:53

said that Tesla was going to be in

1:15:55

a better position in the long run because

1:15:57

it ultimately just had a software problem. Whereas

1:15:59

Waymo has a hardware problem, and those are

1:16:02

typically harder to solve. I'm curious

1:16:04

if you have a view on this,

1:16:06

whether you think one company is likely

1:16:08

to get to a better scale based

1:16:10

on the approach that they're taking with

1:16:12

their hardware and software. I mean,

1:16:14

I think that hardware costs scale

1:16:16

down over a period of time.

1:16:18

So sure, Waymo has a hardware

1:16:21

problem, but they can solve it.

1:16:23

I mean, the history of

1:16:25

of compute and hardware is like the

1:16:27

costs come down very, very significantly. The

1:16:30

Waymo solution is working right now, so it's

1:16:32

not theory, right? And I think the

1:16:34

differences are bigger, which is Waymo has

1:16:37

more sensors, has cameras, has LIDAR, so

1:16:39

there's a certain redundancy there. Waymo

1:16:42

generally has more compute, so

1:16:44

to speak. So the inference

1:16:46

of that computer is going to

1:16:49

be better. Right. And Waymo

1:16:51

also has a high definition of maps

1:16:54

that essentially makes the problem

1:16:56

of recognizing what's happening in the real

1:16:58

world a much simpler problem. So

1:17:01

under Elon's model, the weight that

1:17:03

the software has to carry is

1:17:05

very, very heavy versus

1:17:07

the Waymo and most other

1:17:09

player model where you

1:17:11

don't have to kind of weigh as much on training

1:17:13

and you make the problem much simpler as

1:17:16

a compute problem to understand. I

1:17:19

think eventually both will get there. But

1:17:21

if you had to guess, who's going to get to

1:17:24

sort of a viable scale first? Listen, I think

1:17:26

I think Elon eventually will get to a viable

1:17:28

scale. But for the next five years, I bet

1:17:30

on Waymo and we are betting on Waymo. I'll

1:17:33

say this. I don't want to get into an autonomous Tesla

1:17:35

in the next five years. I'm going to let somebody else

1:17:37

can test that out. I'm not going to

1:17:39

be an early adopter. FSD is getting pretty good. Have

1:17:41

you used it recently? I have not used it recently.

1:17:43

Yeah. All right. Yeah, it's really good. Now, again, it's

1:17:46

the first example, the cost of a solid state LIDAR now

1:17:49

is 500, 600. Right.

1:17:51

So why wouldn't you put that

1:17:54

into your sensor stack? It's not

1:17:56

that expensive. And for a

1:17:58

fully self drive. I think that makes a

1:18:00

lot of sense to me. Now, Elon has

1:18:02

accomplished the unimaginable many, many, many times, so

1:18:07

I wouldn't bet against him. Yeah, I

1:18:09

don't know. This is always, you know, my

1:18:11

secret dream for you, you know, obviously you should stay at

1:18:13

Uber as long as you want. When you're done with that,

1:18:15

I actually do think you should run Tesla because I think

1:18:17

you would be, just as you've done Uber, you'd be

1:18:20

willing to make some of the sort of easy compromises like, just

1:18:23

put a $500 freaking LiDAR on the thing and we'd go much faster.

1:18:25

So anyway, what do you think about that? I have a full-time job

1:18:27

and I'm very happy with it. Thank you. Well,

1:18:30

the Tesla board is listening. I don't

1:18:32

know if the Tesla board listens to

1:18:34

you two. Good point. I

1:18:38

made too many Kennedy Chokes. We're opening up

1:18:40

the board meeting with an episode of Fart

1:18:42

Fork, everybody. You think you learned a lot

1:18:44

from this show. What's

1:18:46

your best guess for when, say, 50% of Uber

1:18:48

rides in the U.S. will

1:18:50

be autonomous? I'd say close to

1:18:53

eight to ten years is my best guess, but I

1:18:55

am sure that'll be wrong. Probably

1:18:58

closer to ten. Close

1:19:01

to ten? Okay, interesting. Most

1:19:03

people have overestimated, you

1:19:06

know, again, it's a wild guess. The probabilities

1:19:08

of your being a rider are just as

1:19:10

much as mine. I'm curious if

1:19:12

we can sort of get into a

1:19:14

future imagining mode here. Like, in

1:19:17

the year, whether it's ten years or fifteen

1:19:19

years or twenty years from now when maybe

1:19:21

a majority of rides in at least big

1:19:23

cities in the U.S. will be autonomous. Do

1:19:27

you think that changes the city at all?

1:19:29

Like, do the roads look different? Are there

1:19:31

more cars on the road? Are there fewer

1:19:33

cars on the road? What does that even

1:19:35

look like? So I think that the cities

1:19:38

will have much, much more space

1:19:41

to use. Parking often

1:19:43

takes up 20, 30%

1:19:45

of the square miles in a city,

1:19:48

for example, and that parking space will

1:19:50

be open for living, parks, et cetera.

1:19:52

So there's no doubt that it will

1:19:54

be a better world. You

1:19:56

will have greener, cleaner cities, and you'll

1:19:59

never have to go there. Park again, which I think is pretty cool.

1:20:02

I'm very curious what you think

1:20:05

about the politics of autonomy in

1:20:07

transportation. In the early days of

1:20:09

Uber, there was a lot of backlash and

1:20:11

resistance from taxi drivers. And,

1:20:13

you know, they saw Uber as a threat

1:20:16

to their livelihoods. There were some, you know,

1:20:18

well-publicized cases of sort of sabotage and big

1:20:20

protests. Do you anticipate there

1:20:22

will be a backlash from

1:20:24

either drivers or the public to the

1:20:27

spread of AVs as they start to

1:20:29

appear in more cities? I think there

1:20:31

could be. And what I'm hoping is

1:20:34

that we avoid the backlash by having

1:20:36

the proper conversations. Now, historically, society as

1:20:39

a whole, we've been able to adjust to

1:20:41

job displacement because it does happen gradually. And

1:20:45

even in a world where there's greater automation now

1:20:47

than ever before, employment rates,

1:20:49

etc., are at historically great

1:20:51

levels. But the fact is

1:20:53

that AI is going to displace jobs.

1:20:55

What does that mean? How quickly should

1:20:57

we go? How do we think about

1:21:00

that? Those are discussions that we're going to have. And

1:21:02

if we don't have the discussions, sure, there will be

1:21:04

backlash. There's always backlash against societal

1:21:06

change that's significant. Now,

1:21:09

we now work with taxis in San Francisco and

1:21:12

taxi drivers who use Uber make more than 20

1:21:14

percent more than the ones who don't. So there

1:21:17

is a kind of solution space where

1:21:20

new technology and established

1:21:22

players can win. But

1:21:25

I don't know exactly what that looks like. But

1:21:27

that calculus does not apply to self-driving. You know,

1:21:29

it's not like the Uber driver who's been driving

1:21:31

an Uber for 10 years and that's their main

1:21:33

source of income can just start driving a self-driving

1:21:35

way. But you don't need a driver. No, you

1:21:37

don't need a driver. It's not just that they

1:21:39

have to switch the app they're using, it's that

1:21:41

it threatens to put them out of a job.

1:21:43

Well, listen, could they be part

1:21:46

of fleet management, cleaning, charging, etc.?

1:21:49

That's a possibility. We are

1:21:51

now working with some of our

1:21:53

drivers. They're doing AI map labeling

1:21:55

and training of AI models, etc.

1:21:57

So we're expanding the solution. set

1:21:59

of work on demand work that

1:22:01

we're offering our drivers because there

1:22:03

is part of that work, which

1:22:06

is driving maybe going away or the growth

1:22:08

in that work is going to slow down

1:22:10

at least over the next 10 years. And

1:22:13

then we'll look to adjust. But listen, these are

1:22:15

issues that are real and I

1:22:17

don't have a clean answer for them at this

1:22:19

point. Yeah. You brought

1:22:21

up shared rides earlier and back

1:22:24

in the day, I think when Uber X first rolled

1:22:26

out shared rides, I did that a couple of times

1:22:28

and then I don't know, I got a raise at

1:22:31

my job and I thought from here on out, I think it's just

1:22:33

going to be me in the car. How

1:22:35

popular do you think you can make shared

1:22:37

rides and is there anything that you can

1:22:40

do to make that more appealing?

1:22:42

Well, I think the way that we have to make

1:22:44

it more appealing is to reduce the penalty, so

1:22:47

to speak, of the shared rides. I think the number one

1:22:49

reason why people use Uber is they want to save time,

1:22:51

they want to have their time back. And

1:22:53

a shared ride would, you

1:22:55

would get about a 30% decrease in

1:22:58

price historically, but there could be a 50 to 100%

1:23:01

time penalty. We're working now. Well, you might

1:23:03

end up sitting next to Casey Newton. That

1:23:06

would be cool. That would be amazing.

1:23:08

Although I would feel very short. Otherwise,

1:23:10

I would have no complaints. People so

1:23:13

far we've heard don't have a problem

1:23:15

with company. It really is time and

1:23:17

they don't mind riding with other people.

1:23:19

There's a certain sense of satisfaction with

1:23:21

riding with other people, but we're now

1:23:23

working with both algorithmically and

1:23:25

I think also fixing the product. Previously,

1:23:28

you would choose a shared ride

1:23:30

and you get an upfront discount. So your

1:23:32

incentive as a customer is to get the discount,

1:23:35

but not to get a shared ride. So we

1:23:37

would have customers gaming the system. They get a

1:23:39

shared ride at 2 a.m. when they know they're

1:23:41

not going to be matched up, et cetera. Now

1:23:43

you get a smaller discount and you

1:23:45

get a reward, which is a higher discount

1:23:48

if you're matched. So part of it is

1:23:50

we're not customers aren't working against us and

1:23:52

we're not working against customers, but we're working

1:23:54

on tech. We are reducing the time penalty,

1:23:56

which is we avoid these weird

1:23:59

routes, et cetera. that's going to cost you 50% of

1:24:01

your time or 100% of your time. Now,

1:24:04

in Autonomous, if

1:24:06

we are the only player that then has the liquidity

1:24:09

to introduce shared Autonomous into

1:24:11

cities, that lowers congestion, lowers the price,

1:24:14

that's another way in which our marketplace can add

1:24:16

value to the ecosystem. Speaking

1:24:19

of shared rides, Uber just

1:24:21

released a new airport shuttle service

1:24:23

in New York City. It

1:24:25

costs $18 a person. You book a seat.

1:24:28

It goes on a designated route

1:24:30

on a set schedule. I

1:24:33

don't have a question. I just wanted to congratulate you on

1:24:35

inventing a bus. It's a

1:24:37

better bus. You know exactly when it's coming, picking

1:24:40

up. Just knowing exactly where your

1:24:42

bus is, pick up, knowing what

1:24:44

your path is, real time, it just gives

1:24:46

a sense of comfort. We think this can

1:24:48

be a pretty cool product. And again, is

1:24:51

bus going to be hugely profitable for us long

1:24:53

term? I don't know, but it will introduce us

1:24:55

to a bigger audience to

1:24:57

come into the Uber ecosystem. And

1:25:00

we think it can be good for cities as well. If

1:25:02

you're in Miami, by the way, over the weekend, we

1:25:05

got buses to the Taylor Swift concert as well. So

1:25:07

I'm just saying. Well, I mean, look, it should not

1:25:09

be hard to improve on the experience of a city

1:25:11

bus. Yeah. Like, do you know what I mean? So

1:25:14

actually, I love city buses. When was the last time

1:25:16

you were on a city bus? Well, I took the

1:25:18

train here. So it wasn't a bus, but it was

1:25:20

transit. He doesn't take shared. He doesn't take bus. This

1:25:23

guy is like, I like to ride public

1:25:25

transit. You're an elitist. No, I would love to see a

1:25:27

picture of you on a bus sometimes in the past five

1:25:29

years, because I'm pretty sure that's never happened. Let me ask

1:25:31

you this. I think we can make the experience better. So

1:25:34

far, I've resisted giving you any product feedback, Dara.

1:25:36

But I had this one thing that I have

1:25:39

always wanted to know the explanation for, and it

1:25:41

says, at some point in the past couple years, you

1:25:43

all, when I ordered an Uber, started sending me a

1:25:46

push notification saying that the driver

1:25:48

was nearby. And I'm

1:25:50

the sort of person, when I've ordered an Uber, Dara, I'm

1:25:52

going to be there when the driver pulls. I'm not making

1:25:54

this person wait. OK, I'm going to respect their time. And

1:25:57

what I've learned is when you tell me the driver is What

1:26:00

that means is they're at least three minutes away and they

1:26:02

might be two miles away. And what

1:26:05

I want to know is why do you send me that notification? We

1:26:08

want you to be prepared to not keep

1:26:10

the driver waiting. Maybe we should personalize it.

1:26:12

I would love that. I think that's a

1:26:14

good question, which is depending on whether or

1:26:16

not you keep the driver waiting. I think

1:26:18

that is one of the cool things with

1:26:20

AI algos that we can do. At

1:26:22

this point, you're right. The experience

1:26:24

is not quite optimized. But

1:26:27

it's for the driver. It's for the driver. No, I get it.

1:26:29

And if I were a driver, I would be happy that

1:26:31

you were sending that. But you also send me this

1:26:33

notification that says the driver's arriving. And that's when I'm like,

1:26:36

OK, it's time to go downstairs. But it sounds like

1:26:38

we're making progress on this. I think the

1:26:40

algorithm just likes you. It just wants to have a conversation with

1:26:42

you. They know that I love my rides. Well,

1:26:45

Casey has previously talked about how he doesn't like his Uber

1:26:47

drivers to talk to him. This is

1:26:49

a man who doesn't share. Listen,

1:26:52

this man likes to coast through life in

1:26:54

a corset and bubble. Here's

1:26:57

what I'm saying. If you're on your way to the

1:26:59

airport at 6.30 in the morning, do you truly want

1:27:01

a person you've never met before asking you who you're

1:27:03

going to vote for in the election? Is that an

1:27:05

experience that anyone enjoys? By the way, I drive. I

1:27:07

drove. And reading the rider

1:27:10

as to whether they want to have

1:27:12

a conversation or not, I was not

1:27:14

good at the art of conversation

1:27:17

as a driver. Were you two talking to him? No,

1:27:19

no. Hey, how's it going? Are you having a good

1:27:21

day? Going to work. And then I just shut up.

1:27:24

Yeah. And have a nice day. To me, that's ideal.

1:27:27

I don't know if that's how I feel. No, that's perfect. That's

1:27:29

going to give you all the information that you need. I'll be your

1:27:32

driver any day. This is Casey's real attraction

1:27:34

to self-driving cars, is that he never has to talk

1:27:36

to another human. Look, you can make fun of me

1:27:38

all you want. I am not the only person who

1:27:40

feels this way. Let me tell you. When I check

1:27:42

into a hotel, same thing. Did you have a nice

1:27:44

day? Yeah, but where are

1:27:46

you coming in from? Let's not get into it. I

1:27:49

would love to see you checking into a hotel. So did you have

1:27:51

a nice day? And you're like, well, let me tell you about this

1:27:53

board meeting I just went to. Because the pressure I'm under, you don't

1:27:56

want to hear about it. All

1:27:58

right, well, I think we're at time. Thank

1:28:01

you so much for coming. Really appreciate it. It

1:28:03

was fun. When

1:28:05

we come back, well, AI is driving progress

1:28:07

and it's driving cars. Now we're going to

1:28:09

find out if it can drive Casey insane.

1:28:13

He watched 260 TikTok

1:28:15

videos and he'll tell you all about

1:28:19

it. Well,

1:28:25

Casey, aside from all the drama in

1:28:27

AI and self-driving cars this week, we

1:28:29

also had some news about TikTok. One

1:28:32

of the other most powerful AI forces on earth.

1:28:34

No, truly. Yes. I

1:28:36

ironically believe that. Yeah, that was not a joke. So

1:28:39

this week we learned about some

1:28:41

documents that came to light as

1:28:43

part of a lawsuit that is

1:28:45

moving through the courts right now.

1:28:48

As people will remember, the federal government

1:28:50

is still trying to force ByteDance to

1:28:52

sell TikTok. But last week,

1:28:54

13 states and the District of

1:28:56

Columbia sued TikTok, accusing the company

1:28:58

of creating an intentionally addictive app

1:29:00

that harmed children. And

1:29:02

Kevin, and this is my favorite part

1:29:04

of this story, is that Kentucky Public

1:29:06

Radio got ahold of these court documents

1:29:08

and they had many redactions. You know,

1:29:10

often in these cases, the most interesting

1:29:12

sort of facts and figures will just

1:29:14

be redacted for who knows what reason.

1:29:16

But the geniuses over at Kentucky Public

1:29:18

Radio just copy and pasted everything in

1:29:20

the document. And when they pasted it,

1:29:23

everything was totally visible. This keeps

1:29:25

happening. I feel like every year

1:29:27

or two we get a story about

1:29:29

some failed redaction. Like is it that hard

1:29:31

to redact a document? I'll say this. I hope

1:29:33

it always remains this hard to redact a document

1:29:35

because I read stuff like this,

1:29:37

Kevin, and I'm in heaven. Yes. So

1:29:40

they got ahold of these documents. They copied

1:29:42

and pasted. They figured out what was behind

1:29:44

sort of the black boxes in the redacted

1:29:46

materials. And it was pretty

1:29:48

juicy. These documents included details like

1:29:51

TikTok's knowledge of a high number

1:29:53

of underage kids who were

1:29:55

stripping for adults on the platform. The

1:29:57

adults who were paying them in digital.

1:30:01

These documents also claimed that TikTok

1:30:03

had adjusted its algorithm to prioritize

1:30:05

people they deemed beautiful. And

1:30:08

then there was this stat that I

1:30:10

know you honed in on, which was

1:30:12

that these documents said, based on internal

1:30:14

conversations, that TikTok had figured out exactly

1:30:16

how many videos it needed to show

1:30:18

someone in order to get them hooked

1:30:20

on the platform. And

1:30:22

that number is 260. 260

1:30:25

is what it takes. You

1:30:27

know, it reminds me, this is sort of ancient, but do you remember the

1:30:29

commercial in the 80s where they would say, like, how

1:30:31

many licks does it take to get to the center of a

1:30:33

Tootsie Pop? Yes. To me, this is

1:30:35

the sort of 2020's equivalent. How

1:30:39

many TikToks do you have to watch

1:30:41

until you can't look away ever again?

1:30:43

Yes. So this is,

1:30:45

according to the company's own research, this is

1:30:47

about the tipping point where people start to

1:30:49

develop a habit or an addiction of going

1:30:51

back to the platform, and

1:30:54

they sort of become sticky in the parlance of

1:30:56

social media apps. In the disgusting

1:30:58

parlance of social media apps, it becomes

1:31:00

sticky. So, Casey,

1:31:02

when we heard about this

1:31:04

magic number of 260 TikTok videos,

1:31:06

you had what I thought was

1:31:08

an insane idea. Tell us

1:31:11

about it. Well, Kevin, I thought if 260

1:31:13

videos is all it takes, maybe I should

1:31:15

watch 260 TikToks, and here's why. I

1:31:19

am an infrequent user of TikTok.

1:31:21

I would say once

1:31:23

a week, once every two weeks, I'll check in,

1:31:25

I'll watch a few videos, and I would say

1:31:28

I generally enjoy my experience, but not to the

1:31:30

point that I come back every day. And

1:31:33

so I've always wondered what I'm missing

1:31:35

because I know so many folks that

1:31:37

can't even have TikTok on their phone

1:31:40

because it holds such a power over

1:31:42

them, and they feel like

1:31:44

the algorithm gets to know them so quickly

1:31:46

and so intimately that it can only be

1:31:49

explained by message. So

1:31:51

I thought if I've not been able to

1:31:53

have this experience just sort of normally using

1:31:56

TikTok, what if I tried

1:31:59

to consume two... 260

1:32:01

tiktoks as quickly as I possibly could

1:32:03

and just saw what would happen after

1:32:05

that not all heroes wear capes Okay,

1:32:08

so Casey you watched 260

1:32:11

tiktok videos last night. Yeah, tell

1:32:13

me about it. So I did create

1:32:15

a new account So I started fresh.

1:32:17

I didn't just reset my algorithm Although

1:32:19

that is something that you can do

1:32:21

in tiktok and I decided

1:32:23

a couple of things one is I was

1:32:26

not going to follow Anyone like no friends,

1:32:28

but also no influencers no enemies No enemies

1:32:30

and I also was not going to do

1:32:32

any searches right a lot of the ways

1:32:34

that tiktok will get to know you Is

1:32:36

if you do a search and

1:32:38

I thought I want to get

1:32:41

the sort of broadest most mainstreamy

1:32:43

experience of tiktok that I can so

1:32:46

that I can develop a better sense

1:32:48

of how does it sort of Walk

1:32:51

me down this funnel toward my eventual interest

1:32:53

Whereas if I just follow ten friends and

1:32:55

did like three searches for my favorite subjects

1:32:57

Like I probably could have gone there faster

1:33:00

And so do you know the very first thing that tick

1:33:03

tock showed me Kevin? What's that? It showed me a 19

1:33:05

year old boy flirting with an 18 year old girl trying

1:33:07

to get her phone number And

1:33:09

when I tell you I could not have been any less interested

1:33:11

in this content. It was aggressively

1:33:14

straight Yes, and it was

1:33:16

very young and it had nothing to

1:33:18

do And

1:33:23

so over the next several hours this total

1:33:25

process I did About

1:33:27

two and a half hours last

1:33:30

night, and I did another 30

1:33:32

minutes this morning And I would

1:33:34

like to share you know maybe

1:33:36

the first Nine

1:33:38

or ten things that tick tock showed me

1:33:40

again You know that the assumption is it

1:33:42

knows basically nothing about me Yes, and I

1:33:44

do think there is something quite revealing about

1:33:47

an algorithm that knows nothing Throwing

1:33:49

spaghetti at you seeing what will stick and

1:33:51

then just picking up the spaghetti afterwards and

1:33:53

saying well What is it you know that

1:33:55

I thought was interesting so here's what it

1:33:57

showed me second video a disturbing

1:34:00

911 call, like a very upsetting sort

1:34:02

of domestic violence situation, skip. Three,

1:34:05

two people doing trivia on a diving board and

1:34:07

like the person who loses has to jump off

1:34:09

the diving board. Okay, fine. Four,

1:34:11

just free booted clip of audition

1:34:13

for America's Got Talent. Five,

1:34:17

vegetable mukbang. So just a guy

1:34:19

who had like rose and rose

1:34:21

of beautiful multicolored vegetables in front

1:34:23

of them who was just eating

1:34:25

them. Six, a comedy

1:34:28

skit, but it was like running on

1:34:30

top of a Minecraft video. So

1:34:32

one of my key takeaways after

1:34:34

my first six or seven TikTok videos was

1:34:37

that it does actually assume that you're quite

1:34:39

young, right? That's why it started out by

1:34:41

showing me teenagers. And as I would go

1:34:43

through this process, I found that over and

1:34:45

over again, instead of just showing me a

1:34:48

video, it would show me a video that

1:34:50

had been chopped in half and on top

1:34:52

was whatever the sort of core content was.

1:34:54

And below would be someone is playing Subway

1:34:57

Surfers, someone is playing Minecraft or someone is

1:34:59

doing those sort of oddly satisfying

1:35:01

things. This is a growth hack. I'm

1:35:04

combing through a rug or whatever. And

1:35:06

it's like, it's literally people trying to

1:35:08

hypnotize you, right? It's like, if you

1:35:10

just see the, oh,

1:35:12

someone is trying to smooth something out

1:35:14

or someone is playing with like slime.

1:35:17

Have you seen the soap cutting? Soap

1:35:19

cutting is huge. Again, there

1:35:21

is no content to it. It is

1:35:23

just trying to stimulate you on some

1:35:25

sort of like lizard brain level. It

1:35:27

feels vaguely narcotic. Absolutely. It is like,

1:35:29

yes. It is just purely a drug.

1:35:32

Video number seven, an ad. Video

1:35:34

number eight, a dad

1:35:36

who was speaking in Spanish and dancing, I mean,

1:35:38

it was very cute. Now, can I ask you

1:35:40

a question? Are you doing

1:35:43

anything other than just swiping from one video to

1:35:45

the next? Are you liking anything? Are you saving

1:35:47

anything? Are you sharing anything? Because all of that

1:35:49

gets interpreted by the algorithm as like a signal

1:35:52

to keep showing you more of that kind of

1:35:54

thing. Absolutely. So for the first 25 or so

1:35:56

videos, I did not like anything, but because I

1:35:58

truly didn't. like anything, like nothing was really doing

1:36:00

it for me. But my intention was always like,

1:36:03

yes, when I see something I like, I'm gonna

1:36:05

try to reward the algorithm, give it a like,

1:36:07

and I will maybe get more like that. So

1:36:10

the process goes on and

1:36:12

on. And I'm

1:36:14

just struck by the absolute

1:36:16

weirdness and

1:36:19

disconnection of everything in the feed. At

1:36:22

first, truly nothing has any relation to

1:36:24

anything else. And it sort of feels

1:36:26

like you've put your brain into like

1:36:28

a Vitamix, you know? Where it's like,

1:36:30

swipe, here's a clip from friends. Swipe,

1:36:32

kids complaining about school. Swipe, Mickey Mouse

1:36:34

has a gun and he's in a

1:36:36

video game. Those are three videos that

1:36:39

I saw in a row. And

1:36:41

the effect of it is just like

1:36:43

disorienting, right? And I've had this experience

1:36:45

when you like go onto YouTube but

1:36:47

you're not logged in, you know,

1:36:49

on like a new account. And it's sort of just, it's

1:36:51

just showing you sort of a random assortment of things that

1:36:54

are popular on YouTube. It does feel very much like they're

1:36:56

just firing in a bunch of

1:36:58

different directions, hoping that something will stick. And

1:37:01

then it can sort of, it can then

1:37:03

sort of zoom in on that thing. Yes, absolutely.

1:37:05

Now I will add that in the

1:37:07

first 30 or so videos,

1:37:10

I saw two things that I thought were

1:37:12

like actually disturbing and bad. What were they?

1:37:14

Like things that should never have been shown

1:37:16

to me. Was it a clip from the

1:37:18

All In podcast? Yes, no.

1:37:21

Fortunately, it didn't get that bad. But

1:37:23

one, there was a clip of a

1:37:25

great in like a busy city and

1:37:27

there was air blowing up from the

1:37:29

great. And the TikTok was just women

1:37:31

walking over the great and their skirts

1:37:33

blowing up. That seems bad. That's horrible.

1:37:35

That was in the first 20 videos

1:37:37

that I saw. Wow. This

1:37:39

video, okay. I guess if you like that video, it says a lot

1:37:41

about you, right? But it's like that. The

1:37:44

second one, and I truly, I do not even know if

1:37:46

we are, we'll want

1:37:48

to include this on our podcast because I

1:37:50

can't even believe that I'm saying that I

1:37:52

saw this, but it is true. It

1:37:55

was an AI voice of someone

1:37:57

telling an erotic story, which

1:37:59

involves involved incest and it was

1:38:02

shown over a video of someone

1:38:04

making soap. Wow. Like,

1:38:07

what? This is dark

1:38:10

stuff. This is dark stuff. Now, at

1:38:12

what point did you start to wonder if

1:38:14

the algorithm has started to pick up on

1:38:16

your clues that you were giving it? Well,

1:38:18

so I was desperate to find out this

1:38:20

question because I am gay and I wondered

1:38:22

when I was going to see the first

1:38:24

gay content, like when it was actually just

1:38:27

going to show me two gay men who

1:38:29

were talking about gay concerns and it

1:38:32

did not happen. Ever?

1:38:34

No. It never quite got

1:38:36

there. On this morning... In 260 videos. It's

1:38:38

over 260 videos. Now, it did show me queer people. Actually,

1:38:42

do you know the first queer person,

1:38:44

identifiably queer person that the TikTok algorithm

1:38:46

showed me? Are you familiar

1:38:48

with the very popular TikTok meme from this

1:38:51

year, very Jameer, very mindful? a

1:38:58

piece of sponsored content, and she was trying to sell me

1:39:01

a Lenovo laptop. And that

1:39:03

was the queer experience that I

1:39:05

got in my romp through the

1:39:08

TikTok algorithm. Now, it did

1:39:10

eventually show me a couple of queer people. It

1:39:12

showed me one. And

1:39:19

then it showed me a video by

1:39:21

Billie Eilish, a queer pop star. And

1:39:24

I did like that video. And now Billie Eilish was

1:39:26

one of the most famous pop stars in the entire

1:39:28

world. I mean, like, truly, like on the Mount Rushmore

1:39:30

of famous pop stars right now. So it makes a

1:39:33

lot of sense to me that TikTok would show me

1:39:35

also incredibly popular with teenagers. And so

1:39:37

I liked one Billie Eilish video and then

1:39:39

that was when the floodgates opened and it

1:39:41

was like, okay, here's a lot of that.

1:39:44

Just from like sort of scrolling it. No,

1:39:46

we did not get to the gay zone.

1:39:51

Now, I did notice the algorithm adapting to me.

1:39:53

So something about me was because again, I was

1:39:55

trying to get through a lot of videos in

1:39:57

a relatively short amount of time. And TikTok now

1:39:59

will often show you three, four, five minute long

1:40:01

videos, I frankly did not have the time for

1:40:03

that. The longer I scrolled, the shorter the videos

1:40:05

were that I got. And I do feel like

1:40:07

the content aged up a little bit. You know,

1:40:09

it started showing me a category

1:40:11

of content that I call people being weird

1:40:13

little freaks, you know, is like somewhat.

1:40:17

These are some real examples. A man dressed

1:40:19

as the cat in the hat dancing to

1:40:22

Sierra's song Goodies. OK. There was

1:40:24

a man in a horse costume

1:40:26

playing the Addams Family theme song

1:40:28

on an accordion using a toilet

1:40:30

lid for percussion. This

1:40:34

is the most important media platform in

1:40:36

the world. Yes. Hours a

1:40:38

day, teenagers are staring at this. And

1:40:41

this is one of the. We

1:40:44

are so screwed. Yeah, you know,

1:40:47

it it figured out that I was more likely

1:40:49

to like content about animals than other things. So

1:40:51

there started to become a lot of dogs doing

1:40:53

cute things, cats doing cute things or, you know,

1:40:55

other other things like that. But,

1:40:57

you know, there was also just a lot of like, here's

1:41:00

a guy going to a store and showing you objects

1:41:02

from the store or like here is a guy telling

1:41:04

you a long story. Can

1:41:06

I ask you a question? Like, was it any any in these

1:41:08

260 videos? Were there any that you

1:41:10

thought like that is a great video? I

1:41:15

don't know if I saw anything truly great. I

1:41:17

definitely saw some animal videos that if I showed

1:41:19

them to you, you would laugh. Or you would

1:41:21

say that was cute. There was stuff that that

1:41:23

gave me an emotional response. And I would say

1:41:25

particularly as I got to the end of this

1:41:27

process, I was seeing stuff that I enjoyed a

1:41:29

bit more. But there I

1:41:32

did this morning. I decided to

1:41:34

do something, Kevin, because I got so frustrated

1:41:36

with the algorithm. I thought it is time

1:41:38

to give the algorithm a piece of data about me. So do

1:41:40

you know what I did? What did you do? I searched the

1:41:42

word gay. Very subtle.

1:41:45

Which like in fairness is an insane

1:41:48

search query. Because what is TikTok supposed to show me

1:41:50

in response? You can show me all sorts of things.

1:41:52

But on my like real TikTok account, it just shows

1:41:54

me your creators all the time. And they're doing all

1:41:56

sorts of things. They're singing, they're dancing, they're telling jokes.

1:41:58

They're telling stories. So I was like, I would like

1:42:01

to see a little bit of stuff like that. Do

1:42:03

you know the first clip that

1:42:06

came up for me when

1:42:08

I searched gay on TikTok to train my algorithm?

1:42:10

What was it? It was a clip from an

1:42:12

adult film. Now, like

1:42:14

explicit, unblurred, it was

1:42:17

from, and I don't know this,

1:42:19

I've only read about this, but apparently at the

1:42:21

start of some adult films, before the explicit stuff,

1:42:23

there'll be some sort of story content, that sort

1:42:25

of establishes the premise of the scene. And

1:42:27

this was sort of in that vein. But

1:42:30

I thought it, if I just

1:42:32

sort of said off-handed, oh,

1:42:35

TikTok, yeah, I bet if you just search gay,

1:42:37

they'll just show you porn. People

1:42:39

would say, it sounds like you're being insane. Why

1:42:42

would you say that? That's being insane. Obviously, they're

1:42:44

probably showing you their most famous

1:42:46

queer creator, something like that. No, they

1:42:48

literally just showed me porn. So

1:42:52

it was like, again, so much of this process for

1:42:54

me was hearing the

1:42:56

things that people say about TikTok, assuming

1:42:58

that people were sort of exaggerating or being

1:43:01

too hard on it, and then having the

1:43:03

experience myself and saying like, oh

1:43:05

no, it's actually like that. That was interesting. An

1:43:07

alternative explanation is that the algorithm is actually really,

1:43:09

really good, and the reason it show you all

1:43:11

the videos of people being weird little freaks is

1:43:13

because you are actually a weird little freak. That's

1:43:16

true, I will accept those allegations. I will not

1:43:18

fight those allegations. So,

1:43:20

okay, you watched 260 videos, you

1:43:22

reached this magic number that is supposed to

1:43:24

get people addicted to TikTok. Are

1:43:27

you addicted to TikTok? Kevin,

1:43:29

I'm surprised and frankly

1:43:32

delighted to tell you, I have never

1:43:34

been less addicted to TikTok than

1:43:36

I have been after going through this experience.

1:43:38

Do you remember back when people would

1:43:41

smoke cigarettes a lot, and if a parent caught a

1:43:43

child smoking, the thing that they would do is they

1:43:45

say, you know what? You're gonna smoke this whole pack,

1:43:47

and I'm gonna sit in front of you, and you're

1:43:49

gonna smoke this whole pack of cigarettes, and the accumulated

1:43:51

effect of all that stuff that you're breathing into your

1:43:53

lungs, by the end of that, the teenager says, Dad,

1:43:56

I'm never gonna smoke again. This

1:43:59

is how I feel. It

1:44:01

cured your addiction. After watching hundreds

1:44:03

of these TikToks. So, okay, you are not

1:44:05

a TikTok addict. In fact, it seems like

1:44:07

you are less likely to become a TikTok

1:44:09

power user than you were before this experiment.

1:44:11

I think that's right. Did this experiment change

1:44:14

your attitudes about whether TikToks should be banned

1:44:16

in the United States? I

1:44:19

feel so bad saying it, but I think the answer

1:44:21

is yes. Like, not ban it,

1:44:23

right? Like, you know, my feelings about

1:44:25

that still have much more to

1:44:27

do with like free speech. Freedom

1:44:30

of expression. And I think that a

1:44:32

ban raises a lot of questions that the United

1:44:34

States approach to this issue, it just makes me

1:44:36

super uncomfortable with. You can go back through our

1:44:38

archive to hear a much longer discussion about that.

1:44:41

But if I

1:44:44

were a parent of a teen who

1:44:46

had just been given their first smartphone,

1:44:48

hopefully not any younger than like 14,

1:44:52

it would change the way that I talk with them

1:44:54

about what TikTok is. And it would change the way

1:44:56

that I would check in with them about what they

1:44:59

were seeing, right? Like I would say, you are

1:45:01

about to see something that is going to make you

1:45:03

feel like your mind is in a blender and it

1:45:06

is going to try to addict you. And here's how

1:45:08

it is gonna try to addict you. And

1:45:10

I might sit with my child and might

1:45:12

do some early searches to try to precede

1:45:14

that feed with stuff that was good and

1:45:17

would give my child a greater chance of

1:45:19

going down some positive rabbit holes and seeing

1:45:21

less of, you know, some of the more

1:45:23

disturbing stuff that I saw there. So if

1:45:25

nothing else, like I think it was a

1:45:27

good educational exercise for me to go through.

1:45:29

And if there is someone in your life,

1:45:32

particularly a young person who is spending a

1:45:34

lot of time on TikTok, I

1:45:36

would encourage that you go through this process yourself

1:45:38

because these algorithms are changing all the time. And

1:45:40

I think you do wanna have a sense of

1:45:43

what is it like this very week if you

1:45:45

really wanna know what it's gonna be showing your

1:45:47

kid. Yeah, I mean, I will say, you know,

1:45:49

I spent a lot of time on TikTok. I

1:45:53

don't recall ever getting

1:45:56

done with TikTok and

1:45:58

being sort of... happy and

1:46:00

fulfilled with how I spent the time. There's

1:46:03

a vague sense of shame about it. There's

1:46:06

a vague sense that sometimes it helps me turn

1:46:08

my brain off at the end of a stressful

1:46:10

day. It has this sort of narcotic

1:46:12

effect on me. And

1:46:16

sometimes it's calming, and sometimes I find things

1:46:18

that are funny, but rarely do I come

1:46:20

away saying that was the best possible use

1:46:22

of my time. There is something

1:46:24

that happens when you adopt this

1:46:27

sort of algorithm first, vertical

1:46:30

video, mostly short form, infinite

1:46:32

scroll. You put all of

1:46:34

those ingredients into a bag,

1:46:36

and what comes out does

1:46:38

have this narcotic effect, as

1:46:40

you say. Well, Casey,

1:46:42

thank you for exposing your brain

1:46:44

to the TikTok algorithm for the

1:46:46

sake of journalism. I appreciate you.

1:46:49

And I will be donating it to

1:46:51

science when my life ends. People

1:46:54

will be studying your brain after you die. I

1:46:56

feel fairly confident. I don't know why they'll be

1:46:58

studying your brain, but there will be research teams

1:47:00

looking at it. Can't wait to hear which I'll

1:47:03

find out. I'm Casey. I'm

1:47:05

the director of the The

1:47:20

Hard Fork is produced by Whitney Jones and Rachel

1:47:22

Cohn, where edited by Jen Poient. Today's

1:47:25

show was engineered by Alyssa Moxley. Original

1:47:27

music by Mary Lozano, Sophia

1:47:30

Landman, Diane Wong, Rowan Nemestow,

1:47:32

and Dan Powell. Our

1:47:34

audience editor is Nuggler Loegely. Video production by

1:47:36

Ryan Manning and Chris Schott. As

1:47:39

always, you can watch this full episode

1:47:41

on YouTube at youtube.com/hard fork. Special

1:47:44

thanks to Paula Schumann, Hui Wing Tam, Dalia

1:47:47

Hadad, and Jeffrey Miranda. You can email

1:47:49

us at hardfork at nytimes.com. Thanks

1:47:52

for watching. I'll see you next time. Bye. And

1:48:10

now, a next-level moment from AT&T Business.

1:48:12

Say you've sent out a gigantic shipment

1:48:15

of pillows, and they need to be

1:48:17

there in time for International Sleep Day.

1:48:19

You've got AT&T 5G, so you're fully

1:48:21

confident. But the vendor isn't responding. And

1:48:23

International Sleep Day is tomorrow! Luckily, AT&T

1:48:25

5G lets you deal with any issues

1:48:27

with ease, so the pillows will get

1:48:29

delivered and everyone can sleep soundly. Especially

1:48:32

you. AT&T 5G requires a compatible plan

1:48:34

and device. 5G is not available everywhere.

1:48:36

See att.com/5G for you for details.