0:00

You've probably heard about the job

0:02

of an intimacy coordinator. But

0:04

do you know what they actually do? In

0:06

every intimacy coordinator's kit is going to be some

0:09

form of mint. Really? I think most

0:11

of the time. Wait. Oh, 100%. Huh.

0:14

Then we add in the full retro Listerine breath strip, which

0:17

is

0:17

crucial. Several years into

0:19

the era of the intimacy coordinator,

0:22

we ask what they've changed for the

0:24

better in Hollywood. And

0:26

what still needs work? This week

0:29

on Intuit, Vulture's Pop

0:31

Culture Podcast.

0:38

It's some of the most

0:39

expensive meat in the world. But

0:42

is it even meat? This summer

0:44

for the first time, Americans are going to be able

0:46

to try actual chicken meat that didn't

0:49

involve killing a chicken. This episode of Gastropod,

0:52

we are among the very first people

0:54

to taste our way through these brand new lab

0:56

grown offerings. Chicken, hamburger,

0:58

bacon, salmon, bluefin tuna. We tasted

1:01

it all. We wanted to know whether it matches up

1:03

to the real thing, but we also

1:05

wanted to know if it can ever really

1:07

replace meat from animals, not

1:09

to mention keep our planet from going up

1:11

in smoke. Find Gastropod wherever you get your

1:14

podcasts and taste the future.

1:18

I went to see the latest Mission Impossible

1:20

movie this weekend, and it had a bad

1:23

guy that felt very 2023. The

1:26

entity has since become sentient.

1:29

An AI becoming super intelligent and

1:31

turning on us. You're telling me this thing has

1:33

a mind of its own?

1:34

And it's just the latest entry in a

1:36

long line of super smart AI

1:39

villains. Open the pod bay doors, Hal.

1:42

Like in 2001, a Space Odyssey. I'm

1:44

sorry, Dave. I'm afraid I can't

1:46

do that. Or Ex Machina. Ava,

1:50

go back to your room. Or maybe the

1:52

most famous example, Terminator.

1:55

They say it got smart. A new

1:57

order of intelligence decided

1:59

our fate.

1:59

microseconds. But

2:02

AI doesn't need to be super intelligent in

2:05

order to pose some pretty major risks.

2:08

Last week on the first episode of our Black Box

2:10

series, we talked about the unknowns at the

2:12

center of modern AI. How even the

2:15

experts often don't understand how these

2:17

systems work,

2:18

or what they might be able to do. And

2:20

it's true that understanding isn't necessary

2:23

for technology. Engineers don't always

2:25

understand exactly how their inventions

2:28

work when they first design them. But

2:30

the difference here is that researchers using

2:32

AI often can't predict what

2:34

outcome they're going to get. They can't

2:37

really steer these systems all that well. And

2:40

that's what keeps a lot of researchers up at night.

2:42

It's not Terminator. It's

2:44

a much likelier, and maybe even

2:46

stranger scenario. It's the story

2:49

of a little boat. Specifically

2:52

a boat in this retro-looking online

2:54

video game. It's called Coast

2:57

Runners, and it's a pretty straightforward racing

2:59

game. There are these power-ups that give

3:01

you points if your boat hits them. There are

3:03

obstacles to dodge. There are these kind of lagoons

3:06

where your boat can get all turned around. And

3:09

a couple years ago, the research company

3:11

OpenAI wanted to see if they could get an

3:13

AI to teach itself how to get a high

3:16

score on the game without being

3:18

explicitly told how. We are

3:20

supposed to

3:20

train a boat to complete a

3:22

course from start to finish. This

3:25

is Dario Amade. He used to be a researcher

3:27

at OpenAI. Now he's the CEO of another

3:29

AI company called Anthropic. And

3:32

he gave a talk about this boat at a think tank called

3:34

the Center for a New American Security.

3:37

I remember setting it running one day, just

3:39

telling it to teach itself. And I

3:41

figured that it would learn to complete the course.

3:44

Dario had the AI run tons of

3:46

simulated races over and over. But

3:49

when he came back to check

3:50

on it, the boat hadn't even come

3:53

close to the end of the track. What it does

3:55

instead, this thing that's been looping, is it

3:57

finds this isolated lagoon.

3:59

and it goes backwards in the course. The

4:03

boat wasn't just going backwards in this

4:05

lagoon. It was on fire,

4:07

covered in pixelated flames, crashing

4:10

into docks and other boats, and

4:12

just spinning around in circles. But

4:17

somehow the AI's score was going

4:19

up. Turns out that by

4:21

spinning around in this isolated lagoon in

4:24

exactly the right way, it can

4:26

get more points than it could possibly ever

4:28

have gotten by completing the race

4:30

in the most straightforward way. When he looked

4:32

into it, Dario realized that the game didn't

4:34

award points for finishing first. For

4:37

some reason, it gave them out for picking up

4:39

power-ups. Every time you get

4:41

one, you increase your score, and they're kind of laid

4:44

out mostly linearly along the course. But

4:46

this one lagoon was just full of these power-ups,

4:49

and the power-ups would regenerate after

4:52

a couple seconds. So the AI

4:54

learned to time its movement to get these

4:56

power-ups over and over by

4:58

spinning around and

4:59

exploiting this weird game design.

5:03

There's nothing wrong with this in the sense

5:05

that we asked it to find a solution to a mathematical

5:07

problem. How do you get the most points? And this

5:10

is how it did it. But if this was

5:12

a passenger ferry or something, you wouldn't

5:14

want it spinning around, setting itself on fire,

5:16

crashing into everything.

5:19

This boat game might seem like a small,

5:22

glitchy example, but it

5:24

illustrates one of the most concerning aspects

5:26

of AI. It's called the alignment problem.

5:29

Essentially, an AI's solution to a problem

5:32

isn't always aligned with its designers'

5:34

values, how they might want

5:36

it to solve the problem.

5:38

And like this game, our world isn't

5:40

perfectly designed. So if scientists don't

5:43

account for every small detail in our society

5:45

when they train in AI, it can solve

5:47

problems in unexpected ways,

5:50

sometimes even harmful ways. Something

5:52

like this can happen without us even knowing that it's happening,

5:55

where our system has found a way to do the thing

5:57

we think we want in a way that we really don't want.

5:59

The problem here isn't with the AI itself.

6:02

It's with our expectations of it.

6:05

Given what AIs can do, it's tempting

6:07

to just give them a task and assume the whole

6:09

thing won't end up in flames. But

6:12

despite this risk, more and more institutions,

6:15

companies, and even militaries are

6:17

considering how AI might be useful to

6:19

make important real-world decisions.

6:22

Hiring. Self-driving cars. Even

6:25

battlefield judgment calls.

6:28

Using AI like this can almost feel like making

6:30

a wish with a super annoying, super

6:33

literal genie. You

6:35

got real potential for a wish, but

6:37

you need to be extremely careful. This

6:40

reminds me of the tale of the man who

6:42

wished to be the richest man in the world. That

6:44

was then crushed at the mountain of corgoines.

6:49

I'm Noam Hasenfeld, and this is the second episode

6:52

of The Black Box, unexplainable series

6:54

on the unknowns at the heart of AI. If

6:57

there's so much we still don't understand about

6:59

AI, how can we make sure it does

7:01

what we want, the way we want? And

7:04

what happens if we can't?

7:07

Thinking intelligent thoughts is a mysterious

7:10

activity. The future of the computer

7:12

is just heartwarming.

7:13

I just have to admit I don't

7:15

really know. You're confused, Doctor. How do you

7:17

think I'd feel? Activity. Intelligence.

7:19

Can

7:21

the computer think? No!

7:29

So given the risks here that AI

7:31

can solve problems in ways its designers

7:33

don't intend, it's easy to wonder

7:35

why anyone would want to use AI to

7:37

make decisions in the first place. It's

7:40

because of all this promise, the positive

7:43

side of this potential genie. Here's

7:45

just a couple examples. Last year,

7:47

an AI built by Google predicted almost

7:50

all known protein structures. It

7:52

was a problem that had frustrated scientists for

7:54

decades, and this development has already

7:57

started accelerating drug discovery. AI

8:00

has helped astronomers detect undiscovered stars,

8:03

it's allowed scientists to make progress on decoding

8:05

animal communication, and

8:07

like we talked about last week, it was able

8:09

to beat humans at Go, arguably

8:12

the most complicated game ever made.

8:15

In all of these situations, AI has

8:17

given humans access to knowledge we

8:19

just didn't have before. So

8:22

the powerful and compelling thing about AI

8:24

when it's playing Go is sometimes

8:26

it will tell you a brilliant Go move that you would

8:28

never have thought of, that no Go master would ever have

8:30

thought of, that does advance your

8:33

goal of winning the game.

8:34

This is Kelsey Piper. She's a reporter for

8:37

Vox who we heard from last episode, and

8:39

she says this kind of innovation is really useful,

8:41

at

8:42

least in the context of a game. But

8:44

when you're operating in a very complicated context

8:47

like the world, then those brilliant

8:49

moves that advance your goals might

8:52

do it by having a bunch of side effects

8:54

or inviting a bunch of risks that you don't

8:56

know, don't understand, and aren't evaluating.

8:59

Essentially, there's always that risk of

9:02

the boat on fire. We've

9:04

already seen this kind of thing happen outside of

9:06

video game. Just take the example of

9:09

Amazon back in 2014. So

9:11

Amazon tried to use an AI hiring

9:13

algorithm, looked at candidates and then recommended

9:15

which ones would proceed in the interview process. Amazon

9:19

fed this hiring AI 10 years

9:21

worth of submitted resumes, and they told

9:23

it to find patterns that were associated with

9:25

stronger candidates.

9:26

And then an analysis came

9:28

out, finding that the AI was biased. It had

9:30

learned that Amazon generally preferred

9:32

to hire men, so it was happily more likely to recommend

9:35

Amazon men.

9:36

Amazon never actually used this AI

9:38

in the real world. They only tested it. But

9:41

a report by Reuters found exactly which

9:43

patterns the AI might have internalized. The

9:46

technology thought, oh, Amazon doesn't

9:48

like any resume that has the word women's

9:51

in it. So this is a woman's university, captain

9:53

of a women's chess club, captain of women's

9:56

soccer team.

9:57

Essentially, when they were training their AI, Amazon

9:59

was not a Amazon hadn't accounted for

10:01

their own flaws in how they'd been measuring

10:04

success internally. Kind of like

10:06

how OpenAI hadn't accounted for the way the

10:08

boat game gave out points based on power-ups,

10:10

not based on who finished first.

10:12

And of course, when Amazon realized

10:14

that, they took the AI out of

10:17

their process. But it seems like they

10:19

might be getting back in the AI hiring

10:21

game. According to an internal document

10:23

obtained by former Vox reporter Jason

10:25

Del Rey, Amazon's been working on a new

10:28

AI system for recruitment. At the same

10:30

time, they've been extending buyout offers

10:32

to hundreds of human recruiters. And

10:35

these flaws aren't unique to hiring AIs.

10:38

The way AIs are trained has led to all kinds

10:40

of problems. Take what happened with Uber

10:42

in 2018, when they didn't include jaywalkers

10:45

in the training data for their self-driving cars,

10:48

and then a car killed a pedestrian.

10:50

Tempe, Arizona police say 49-year-old

10:53

Elaine Herzberg was walking a bicycle

10:55

across a busy thoroughfare frequented by pedestrians

10:57

Sunday night. She was not in a crosswalk.

11:01

And a similar thing happened a few years ago with

11:03

the self-training AI Google used in its

11:05

photos app.

11:06

The company's automatic image recognition

11:08

feature in its photo application

11:10

identified two black persons as gorillas,

11:13

and in fact even tagged them as soap.

11:16

According to some former Google employees,

11:18

this may have happened because Google had a biased

11:20

data set. They may just not have

11:22

included enough black people.

11:24

The worrying thing is if you're using AIs to

11:26

make decisions, and the data they have

11:29

reflects our own biased processes,

11:31

like a biased justice system that sends

11:34

some people to prison for crimes where it lets

11:36

other people off with a slap on the wrist, or

11:38

a biased hiring process, then

11:40

the AI is going to learn the same thing.

11:44

But despite these risks, more companies

11:46

are using AI to guide them in making

11:48

important decisions. This is changing

11:51

very fast. Like there are a lot more companies

11:53

doing this now than there were even a year ago,

11:56

and there will be a lot more in a couple

11:59

more years.

11:59

Companies see a lot of benefits here.

12:02

First, on a simple level, AI is

12:04

cheap. Systems like chat GPT

12:06

are currently being heavily subsidized by investors,

12:09

but at least for now, AI is way cheaper

12:12

than hiring real people.

12:13

If you want to look over thousands of job

12:15

applicants, AI is cheaper than having humans

12:18

screen those thousands of job applicants. If

12:20

you want to make salary decisions, AI is

12:22

cheaper than having a human whose job is to

12:24

think about and make those salary decisions. If

12:26

you want to make firing decisions, those get done

12:28

by algorithm because it's easier

12:30

to fire who the algorithm spits out than to

12:32

have human judgment and

12:35

human analysis in the picture. And

12:37

even

12:37

if companies know that AI decision-making

12:39

can lead to boat-on-fire situations,

12:42

Kelsey says they might be OK with that

12:44

risk. It's so much cheaper that that's

12:47

a good business trade-off. And so we hand

12:49

off more and more decision-making to AI

12:51

systems for

12:53

financial reasons.

12:55

The second reason behind this push to use AI

12:57

to make decisions is because it could

12:59

offer a competitive advantage.

13:01

Companies that are employing AI

13:04

in a very winner-take-all capitalist

13:06

market, they might outperform the companies

13:08

that are still relying on expensive human

13:11

labor. And the companies that aren't

13:13

are much more expensive, so fewer people want

13:15

to work with them, and they're a smaller share of the economy.

13:18

And you might have huge

13:20

economic behemoths that are making

13:22

decisions almost entirely with AI systems.

13:24

But it's not just companies.

13:27

Kelsey says competitive pressure is even leading

13:29

the military to look into using AI to make

13:31

decisions.

13:32

I think there is a lot of fear

13:34

that the first country to successfully

13:37

integrate AI into its decision-making

13:39

will have a major battlefield advantage over

13:41

anyone still relying on slow humans.

13:44

And that's the driver of a lot in the military,

13:46

right? If we don't do it, somebody else

13:48

will, and maybe it will be a huge advantage.

13:51

This kind of thing may have already happened

13:54

in actual battlefields. In 2021,

13:56

a UN panel determined that an autonomous

13:59

Turkish drone

13:59

may have killed Libyan soldiers

14:02

without a human controlling it or even

14:04

ordering it to fire.

14:06

And lots of other countries, including the US, are

14:08

actively researching AI-controlled weapons.

14:10

You don't want to be the people,

14:13

you know, still fighting on horses when

14:15

someone else has invented fighting with guns,

14:17

and you don't want to be the people who don't have AI when

14:19

the other side has AI. So I think there's

14:22

this very powerful pressure not just

14:24

to figure this out, but to have

14:26

it ready to go.

14:27

And finally, the third reason behind the push toward AI

14:29

decision making is because of the promise we talked

14:32

about at the top. AI can provide

14:34

novel solutions for problems humans

14:36

might not be able to solve on their own. Just

14:39

look at the Department of Defense. They're

14:41

hoping to build AI systems that, quote,

14:43

function more as colleagues than as tools.

14:46

And they're studying how to use AI to help soldiers

14:49

make extremely difficult battlefield

14:51

decisions, specifically when it comes to medical

14:53

triage. I'm going to talk about

14:55

how we can build AI-based systems

14:57

that we would be willing to bet our lives with

15:00

and not be foolish to do so.

15:02

AI has already shown an ability to beat

15:04

humans in war game scenarios, like with

15:06

the board game diplomacy. And researchers

15:08

think this ability could be used to advise

15:11

militaries on bigger decisions, like strategic

15:13

planning.

15:14

Cybersecurity expert Matt Davos talked

15:16

about this on a recent episode of On the Media.

15:19

I think it'll probably get really good at threat

15:21

assessment. I think analysts might also

15:24

use it to help them through their thinking, right?

15:26

They might come up with an assessment and

15:28

say, tell me how I'm wrong. So I think there'll be

15:30

a lot of unique ways in which the technology

15:33

is used in the intelligence community.

15:35

But this whole time, that boat

15:37

on fire possibility is just lurking.

15:40

One of the things

15:42

that makes AI so promising, the

15:45

novelty of its solutions, it's

15:47

also the thing that makes it so hard to predict. Kelsey

15:50

imagines a situation where AI recommendations

15:53

are initially successful, which leads

15:55

humans to start relying on them uncritically,

15:58

even when the recommendations seem counter- are intuitive.

16:01

Humans might just assume the AI sees something

16:03

they don't,

16:04

so they follow the recommendation anyway. We've

16:07

already seen something like this happen in a game context

16:09

with AlphaGo, like we talked about last week. So

16:12

the next step is just imagining it happening

16:14

in the world.

16:16

And we know that AI can have fundamental

16:19

flaws. Things like bias training

16:21

data or strange loopholes engineers haven't

16:23

noticed.

16:24

But powerful actors relying on AI

16:26

for decision-making might not notice

16:29

these faults until it's too late.

16:31

And this is before we get into the AI like

16:34

being deliberately adversarial. This

16:36

isn't the terminator scenario with AI

16:38

becoming super intelligent and wanting to kill us.

16:41

The problem is more about humans and

16:44

our temptation to rely on AI uncritically.

16:46

This isn't the AI trying to trick

16:49

you. It's just the AI exploring

16:52

options that no one

16:54

would have thought of that get us into weird territory

16:57

that no one has been in before. And

16:59

since they're so untransparent, we can't

17:01

even ask the AI, hey, what are the risks of

17:04

doing this?

17:08

So if it's hard to make sure that AI operates

17:10

in the way its users intend, and

17:13

more institutions feel like the benefits

17:15

of using AI to make decisions might outweigh

17:17

the risks,

17:19

what do we do? What can

17:21

we do? There's a lot that we don't

17:23

know, but I think we should be changing

17:26

the policy and regulatory incentives so

17:28

that we don't have to learn from a horrible

17:30

disaster. And so that we like understand

17:33

the problem better and can start making progress

17:35

on solving it.

17:36

How to start solving a problem that

17:39

you don't understand

17:41

after the break. 100 years

17:46

ago, Louis Armstrong walked into a

17:49

tiny studio in Richmond,

17:51

Indiana, and made his first recording.

17:54

A century later, we're still living

17:56

in the musical world that this extraordinary trumpeter

17:58

and vocalist helped create. Listen to

18:01

virtually any pop song and whether you know

18:03

it or not, you're hearing the legacy of Louis Armstrong.

18:05

If you think of Armstrong today, you might think of a

18:07

funny voiced slightly corny entertainer

18:10

whose music serves as the soundtrack for cruise

18:12

ship commercials and comic impressions. But

18:15

there's a lot to learn about this iconic

18:18

musician. For one, his name. He

18:20

preferred Louis, not Louie.

18:22

His success. He was the oldest artist

18:25

to ever score a number one billboard

18:27

hit and he knocked the Beatles off the

18:29

charts

18:29

to do it. His influence, he

18:32

made a new mold for the

18:34

modern pop star that everyone from

18:36

the Wu Tang Clan to Harry Styles has

18:38

followed. I'm Nate Sloan, co-host

18:41

of the Vulture Music Podcast, Switched on Pop,

18:44

and this week we're discussing how Louis

18:46

Armstrong continues to shape the sound

18:48

of popular music 100 years later

18:51

and why his music resonates today more

18:53

than ever. Listen to Switched on Pop

18:56

anywhere you get podcasts.

19:07

So here's what we know. Number

19:09

one, engineers often struggle to

19:12

account for all the details in the world when they

19:14

program an AI.

19:15

They might want it to complete a boat race and

19:17

end up with a boat on fire.

19:19

A company might want to use it to recommend a set of

19:21

layoffs only to realize that the AI

19:23

has built in biases. Number

19:26

two, like we talked about in the first episode of this

19:28

series, it

19:29

isn't always possible to explain

19:31

why modern AI makes the decisions it

19:33

does, which makes it difficult

19:36

to predict what it'll do.

19:38

And finally, number three,

19:40

we've got more and more companies, financial

19:42

institutions, even the military,

19:44

considering how to integrate these AIs into

19:46

their decision making.

19:48

There's essentially a race to deploy this

19:50

tech into important situations, which

19:53

only makes the potential risks here more

19:56

unpredictable. on

20:00

unknowns. So what

20:02

do we do? I would say at this point

20:04

it's sort of unclear. Seagal

20:07

Samuel writes about AI and ethics for

20:09

Vox, and she's about as confused

20:11

as the rest of us here. But she says

20:13

there's a few different things we can work

20:15

on. The first one is interpretability,

20:18

just trying to understand how these AIs

20:20

work. But like we talked about last

20:22

week, interpreting modern AI systems

20:25

is a huge challenge.

20:26

Part of how they're so

20:28

powerful and they're able to give us info that we can't just

20:30

drum up easily ourselves is that they're so

20:33

complex. So there might be something

20:35

almost inherent about lack of interpretability

20:38

being an important feature of

20:40

AI systems that are going to be much more

20:42

powerful than my human brain.

20:44

So interpretability may not be

20:47

an easy way forward, but some

20:49

researchers have put forward another idea. Monitoring

20:52

AIs by using more AIs. At

20:54

the very least, just to alert users if AIs

20:57

seem to be behaving kind of erratically.

20:59

But it's a little bit circular

21:01

because then you have to ask, well, how would we be

21:03

sure that our helper AI is

21:06

not tricking us in the same way that we're

21:08

worried our original AI is doing?

21:10

So if these kind of tech-centric solutions

21:12

aren't the way forward, the best path

21:15

could be political, just trying to reduce

21:17

the power and ubiquity of certain kinds

21:19

of AI.

21:20

A great model for this is the EU, which

21:22

recently put forward some promising AI

21:24

regulation. The European Union is

21:27

now trying to put forward these regulations

21:29

that would basically require companies that

21:32

are offering AI products

21:35

in especially high-risk areas

21:38

to prove that these

21:41

products are safe.

21:42

This could mean doing assessments for bias,

21:44

requiring humans to be involved in the process of

21:46

creating and monitoring these systems, or

21:49

even just trying to reasonably demonstrate that

21:51

the AI won't cause harm.

21:53

We've unwittingly bought this premise

21:55

that they can just bring anything to market

21:57

when we would never do that for other similar

21:59

impactful technologies like think about

22:02

medication. You gotta get your FDA

22:04

approval. You gotta jump through these hoops. Why

22:06

not with AI?

22:09

Why not with AI? Well, there's

22:12

a couple reasons regulation might be pretty hard

22:14

here.

22:15

First, AI is different from something

22:17

like a medication that the FDA would approve.

22:19

The FDA has clear agreed upon hoops

22:21

to jump through, clinical testing. That's

22:24

how they assess the dangers of a medicine before

22:26

it goes out into the world. But with

22:28

AI, researchers often don't know what

22:30

it can do until it's been made public. And

22:33

if even the experts are often in the dark, it

22:35

may not be possible to prove to regulators

22:37

that AI is safe. The second

22:40

problem here is that even aside from

22:42

AI, big tech regulation doesn't

22:44

exactly have the greatest track record of

22:46

really holding companies accountable,

22:49

which might explain why some of the biggest AI companies

22:51

like OpenAI have actually been publicly

22:54

calling for more regulation.

22:56

The cynical read is that this is

22:58

very much a repeat of what we saw with a company

23:00

like Facebook, now Meta, where

23:02

people like Mark Zuckerberg were going to Washington,

23:05

D.C. and saying, oh, yes,

23:07

we're all in favor of regulation. We'll help you.

23:09

We wanna regulate too.

23:11

When they heard this, a lot of politicians said

23:13

they thought Zuckerberg's proposed changes were

23:15

vague and essentially self-serving,

23:18

that he just wanted to be seen supporting the rules,

23:21

rules which he never really thought

23:23

would hold them accountable.

23:24

Allowing them to

23:27

regulate in certain ways, but where really

23:29

they maintain control of their data sets,

23:31

they're not being super transparent and having

23:33

external auditors. So really they're

23:36

getting to continue to drive the ship and

23:38

make profits while

23:40

creating the semblance that society

23:43

or politicians are really driving the ship.

23:45

Regulation with real teeth seems

23:47

like such a huge challenge that one

23:49

major AI researcher even wrote an op-ed

23:51

in Time magazine calling for an indefinite

23:54

ban on AI research, just

23:56

shutting it all down. But Seagal

23:58

isn't sure that. That's such a good idea. I

24:01

mean, I think we would lose all the potential

24:03

benefits it stands to bring. So drug

24:06

discovery, you know, cures for certain

24:08

diseases, potentially

24:10

huge economic growth that

24:13

if it's managed wisely, big if,

24:15

could help alleviate some kinds of poverty.

24:18

I mean, at least potentially, it could

24:20

do a lot of good. And so you

24:23

don't necessarily want to throw that baby out with the bathwater.

24:25

At the very least, Seagal does want to turn

24:27

down the faucet. I think the problem

24:30

is we are

24:31

rushing at breakneck speed

24:33

towards more and more advanced forms of

24:35

AI. When the AIs that

24:38

we already currently have, we don't even know

24:40

how they're working.

24:41

When chatGPT launched, it was the fastest

24:43

publicly deployed technology in history.

24:47

Twitter took two years to reach a million users.

24:49

Instagram took two and a half months. ChatGPT

24:52

took five days.

24:54

And there are so many things researchers learned ChatGPT

24:57

could do only after it was

24:59

released to the public. There's so much we still

25:01

don't understand about them. So what

25:04

I would argue for is just slowing

25:06

down. Slowing down AI could

25:08

happen in a whole bunch of different ways. So

25:10

you could say, you know, we're going to stop

25:13

working on making AI more powerful

25:15

for the next few years, right? We're just not

25:17

going to try to develop AI that's got even

25:20

more capabilities than it already has.

25:22

AI isn't just software. It

25:24

runs on huge, powerful computers.

25:27

It requires lots of human labor. It

25:29

costs tons of money to make

25:32

and operate, even if those costs

25:34

are currently being subsidized by investors.

25:37

So the government could make it harder to get

25:39

the types of computer chips necessary for huge

25:41

processing power. Or it could

25:44

give more resources to researchers in

25:46

academia who don't have the same profit incentive

25:48

as researchers in industry.

25:50

You could also say, all right, we

25:52

understand researchers are going to keep doing the development

25:55

and try to make these systems more powerful, but

25:57

we're going to really halt or slow down deployment.

25:59

and like release to commercial

26:02

actors or whoever. Slowing down the development

26:04

of a transformative technology like

26:06

this, it's a pretty big ask, especially

26:09

when there's so much money to be made. It

26:11

would mean major cooperation, major regulation,

26:14

major complicated discussions with stakeholders

26:17

that definitely don't all agree. But

26:19

Segal isn't hopeless. I'm

26:21

actually reasonably

26:23

optimistic. I'm

26:26

very worried about the direction AI is

26:28

going in. I think it's going way

26:31

too fast. But

26:33

I also try to look at things with

26:36

a bit of a historical perspective. Segal

26:38

says that even though tech progress can seem

26:40

inevitable, there is precedent for

26:42

real global cooperation.

26:44

We know historically there

26:46

are a lot of technological innovations

26:49

that we could be doing, but we're

26:51

not because society just seems like a bad

26:53

idea. Human cloning or like

26:55

certain kinds of genetic experiments, like

26:58

humanity has shown that we are capable

27:00

of putting a stop or at least a slowdown

27:03

on things that we think are dangerous.

27:05

But even if guardrails are possible,

27:08

our society hasn't always been good about

27:10

building them when we should. The

27:12

fear is that sometimes society

27:15

is not prepared to design

27:17

those guardrails until there's been some huge

27:19

catastrophe, like Hiroshima,

27:21

Nagasaki, it's just horrific things that

27:23

happen. And then we pause and we say, hmm,

27:26

okay, maybe we need to go to the drawing board, right?

27:29

That's what I don't want to have happen with AI.

27:32

We've seen this story play out before.

27:35

Tech companies or technologists essentially

27:38

run mass experiments on society.

27:41

We're now prepared, huge harms happen,

27:44

and then afterwards we start to catch up and we

27:46

say, oh, we shouldn't let that catastrophe happen again. I

27:49

want us to get out in front of the catastrophe.

27:52

Hopefully that will be by slowing down the

27:54

whole AI race. If

27:56

people are not willing to slow down,

27:59

at least...

27:59

At least let's get in front by trying

28:02

to think really hard about what

28:04

the possible harms are and how we

28:06

can use regulation to

28:09

really prevent harm as much as we possibly

28:11

can.

28:15

Right now, the likeliest potential

28:17

catastrophe might have a lot less to

28:19

do with the sci-fi terminator scenario than

28:21

it does with us and how we could end up using

28:24

AI, relying on it in more and

28:26

more ways.

28:27

Because it's easy to look at AI and just

28:30

see all the new things it can let us do. AIs

28:33

are already helping enable new technologies,

28:35

they've shown potential to help companies and militaries

28:38

with strategy, they're even helping advance

28:40

scientific and medical research. But

28:43

we know they still have these blind spots that

28:45

we might not be able to predict. So

28:48

despite how tempting it can seem to rely

28:50

on AI, we should be honest

28:52

about what we don't know here. So

28:55

hopefully the powerful actors who are actually shaping

28:57

this future, companies, research

28:59

institutions, governments, will

29:02

at the very least stay skeptical of

29:04

all of this potential. Because if

29:06

we're really open about how little we know, we

29:09

can start to wrestle with the biggest question here.

29:12

Are all of these risks

29:14

worth it?

29:25

That's it for our Black Box series. This

29:28

episode was reported and produced by me,

29:30

Noam Hasenfeld. We had editing

29:32

from Brian Resnick and Catherine Wells, with

29:35

help from Meredith Hadnott, who also manages our team.

29:38

Mixing and sound design from Vince Fairchild,

29:40

with help from Christian Ayala. Music

29:42

from me, fact-checking from Tien Nguyen.

29:45

Mandy Nguyen is a potential werewolf, we're

29:47

not sure. And Bird Pinkerton sat

29:50

in the dark room at the Octopus Hospital, listening

29:53

to this prophecy.

29:55

and

30:01

that only a bird would be able to ensure

30:03

the survival of our species. You

30:06

are that bird, Pinkerton.

30:10

Special thanks this week to Pawan Jain, Jose

30:13

Hernandez-Orreo, Samir Rawashte,

30:15

and Eric Aldridge. If you have

30:17

thoughts about the show, email us at unexplainable

30:20

at vox.com, or you could leave

30:22

us a review or a rating, which we'd also love.

30:25

Unexplainable is part of the Vox Media Podcast

30:27

Network, and we'll be back in your feed next

30:29

week.