0:02

This is an ABC podcast. This

0:06

story starts with a horse named

0:08

Clever Hans. Come

0:11

over here Hans. Over

0:14

100 years ago, Clever Hans wowed the

0:16

world of AI. That's

0:19

animal intelligence. It could do

0:21

something no horse had ever done before.

0:23

It could add, subtract, multiply,

0:26

divide, read, spell and understand

0:28

German. That's more than me.

0:32

So this is how it worked. The

0:34

questioner would ask something like, Hey

0:36

Hans, what's 1 plus 1? And

0:39

the horse would tap out its answer with a hoof.

0:43

Two taps. You did it Hans. But

0:46

it turned out Hans wasn't clever

0:48

in the way its audience believed.

0:51

He was tapping his hoof until

0:53

he detected involuntary cues from the

0:55

questioner that showed he was getting close

0:57

to the answer. The questioner who

0:59

knew the answer would tense up

1:01

at the critical moment and

1:04

Hans would stop tapping. Hans

1:08

wasn't clever. We

1:10

were just projecting intelligence onto the

1:12

horse. Now

1:15

some people say this old story

1:18

is a parable for modern AI.

1:21

It may look like ChatGPT is alive

1:23

and talking to us, but

1:25

again, that's just us thinking

1:28

it's thinking. Why

1:30

does this matter? Well, we may be

1:32

on track to developing a truly intelligent

1:34

AI or we might

1:37

be driving down a dead end. Is

1:40

ChatGPT a baby genius? Or

1:42

is it the modern equivalent of

1:45

a clever horse? This

1:51

is Hello AI Overlords, a science

1:53

fiction series about how AI has burst

1:55

into our lives in a few short

1:58

years. I'm James Purtill. Behind

2:01

the rise of AI, there's big questions

2:03

about where this technology is going. Is

2:06

it going to be super intelligent? And if

2:08

that happens, is it going to kill us all? In

2:11

this series, I've spoken to so many

2:13

AI researchers and thinkers, and

2:15

all have different ideas about where we're

2:17

heading. So what could the

2:20

future look like? And what

2:22

keeps them all up at night? Today

2:24

we're going to meet our future AI

2:26

overlords. What are they like? I

2:29

hope they're nice. So,

2:36

first question. How smart

2:38

will it get? Let's

2:40

start with Rodney Brooks, a world leading

2:42

roboticist and AI expert. In

2:44

my estimation, the LLMs and the

2:47

chat engines built around them are

2:49

doing a fantastic and surprisingly good

2:51

pilot trick. He says

2:54

today's AI tools, like chat

2:56

GPT, might appear intelligent, but

2:59

they're just statistical machines. But

3:01

it's all about probabilities of what the next

3:03

word should be, and that's enough to fool

3:05

us. These chat bots are simply

3:07

predicting the next word in a sentence. They

3:10

don't actually understand what they're writing,

3:13

but like with clever hands, we

3:15

want to believe they do. It

3:18

says something about us humans, that

3:20

simple correlations of words provide

3:23

meaning to us that we interpret. It

3:26

says a lot about how we communicate,

3:29

and it's hijacking that. And I don't

3:31

mean that the people who built the

3:33

system intentionally tried to hijack that. I

3:35

think that's the result of

3:38

the way it works. I wanted to

3:41

put Rodney's arguments to the test. Is

3:43

chat GPT thinking, or is

3:46

it just parroting our own words back at us?

3:49

And I had a tricky question up

3:51

my sleeve, designed to trip up the

3:53

chat bot and expose the nature of

3:55

its intelligence. It's called the laundry problem.

4:00

I've just spoken to a guy named

4:02

Rodney Brooks who says you're a parlor

4:04

trick. Rodney Brooks is

4:06

a notable roboticist and AI researcher and

4:08

people may have different opinions about AI

4:11

like me. Okay, so I'm going

4:13

to ask you a question. Please

4:15

reply concisely. Of course. Please go

4:17

ahead and ask your questions. If

4:20

10 t-shirts laid up in the sun

4:22

takes five hours to dry, how long

4:24

does 20 t-shirts take? It

4:26

would take 20 t-shirts 10 hours to dry

4:28

in the same conditions. Babong!

4:32

Now, you and I know that it doesn't take

4:35

longer to dry two loads of washing than one,

4:37

so long as you've got a big washing line.

4:40

But chatTPT doesn't. It

4:42

appears to have no underlying model of the

4:44

world in its head. It's just

4:46

a big library. It has no concept

4:49

of the sun or heat, no

4:51

notion of water or cotton

4:53

fabric. And this gets

4:55

to the heart of the question about whether machine

4:58

learning models like this can

5:00

become truly intelligent. Rodney

5:02

Brooks says machine learning by

5:04

itself isn't enough. Calm

5:06

down people. Calm down. Just wait

5:09

a while. Breathe. We'll figure out

5:11

how powerful or not it is.

5:13

And we're starting to see the turn that

5:16

it wasn't as powerful as we first

5:18

thought. It's damn powerful. I'm not saying

5:20

it's not. But it doesn't mean that

5:22

AI is about to take over everything.

5:25

Now this talk of how to build

5:27

a truly intelligent machine might

5:29

sound familiar. In the very first

5:31

episode of the series, we heard about the war

5:33

of ideas at the dawn of AI. In

5:36

the 1950s and 1960s,

5:38

Frank Rosenblatt said machines that

5:40

can learn on their own

5:42

will ultimately learn to be

5:44

intelligent. Another researcher,

5:47

Marvin Minsky, said no, AI

5:49

has to be taught how to think in

5:52

order to be truly intelligent. Today,

5:54

this dream of an AI that's

5:56

as smart as a human or

5:58

smarter is called AGI. artificial

6:01

general intelligence and artificial

6:06

general intelligence. We don't know how it works

6:08

at all. So talking

6:10

about artificial general intelligence, I

6:13

think it's just way, way, way too

6:15

premature. So

6:20

that's Rodney Brooks' take and he's not the

6:22

only skeptic. Michael Georges is

6:24

an AI expert who's built many

6:27

AI systems over 40 years, including

6:29

for the Space Shuttle program in the 1990s.

6:33

If you want to rely purely on machine learning,

6:35

it would require thousands

6:37

of years, if not hundreds of thousand years

6:39

for them and a lot of machines being

6:42

destroyed along the way in

6:44

order for them to learn how to get around in the world.

6:46

But he's not ruling out artificial

6:49

general intelligence. He says a

6:51

solution may be to draw on the past

6:54

and blend Minsky and Rosenblatt's

6:56

approaches. While it may have a

6:59

big component involving probabilistic learning,

7:02

we'll have to have certain

7:05

ability to execute rules to carry

7:07

out certain logical or common sense

7:09

reasoning to work out how to

7:11

manipulate goals and how to handle

7:14

failure and that won't be

7:16

learnt. So in other words, we

7:18

need to take what we know about machine

7:20

learning and take what we know about

7:22

human reasoning and mix them together. Then

7:25

put that into a computer and

7:28

then we might get AGI

7:30

or not. So I'd go

7:33

at teaching chatgbt how to

7:35

solve the laundry problem. Hey

7:38

chatgbt, your answer is wrong. Ten

7:41

shirts dry as quickly as 20 shirts.

7:44

I understand your point now. Both 10

7:46

and 20 shirts would take approximately the

7:48

same amount of time to dry. That's

7:52

the laundry problem solved. We

7:54

have a trillion more problems to go. Because

8:00

this is academia and research where everyone

8:02

knows everyone and no one can agree,

8:05

of course there's another school of thought about

8:07

how to build AGI. This

8:10

is a dominant school of thought in Silicon Valley.

8:13

It's where all the money currently is and

8:15

where most of the resources have been dedicated.

8:17

It has huge companies behind it,

8:19

companies like OpenAI, the maker of

8:21

chat GPT, DeepMind, owned

8:23

by Google. They reckon

8:26

AGI is way closer than we think,

8:28

like maybe only 5 to 10

8:30

years away. One of

8:32

the godfathers of modern AI says, that

8:34

sounds about right. Shall I

8:36

call you Professor Benjio or Yoshua? What would you prefer?

8:39

Yoshua is fine. Yoshua

8:41

Benjio is a professor at the

8:43

University of Montreal. He's

8:46

basically one of the inventors of the

8:48

machine learning methods we use today. He's

8:51

a very big deal in AI. I'm

8:53

a professor at the University of Montreal in

8:55

computer science and I'm known for my

8:57

work in deep

8:59

learning. Yoshua says AGI has

9:01

two ingredients. There's the

9:03

stuff we do automatically, like recognizing

9:06

objects. And AI is very

9:08

good at doing that. That's what chat GPT

9:10

is known for. And then

9:12

there's the other kind, high level reasoning. This

9:15

is the ability to generalize from one

9:17

set of knowledge to new settings. Humans

9:20

are very good at this. That's why

9:22

the answer to the laundry problem is obvious

9:25

to us. Well that's what really learning is

9:27

about. It's not about memorizing.

9:30

It's about using what

9:33

you're observing to

9:36

extract information that allows

9:38

you to produce good

9:41

behaviors in new settings, not scale-izing.

9:43

I put Rodney Brooks' argument that

9:45

machine learning will never be good

9:48

enough to Yoshua. You'd probably

9:50

know Rodney Brooks who is at MIT at the same

9:52

time as you. Of course the

9:54

AI world is small and they know

9:56

each other. He's very sceptical of that

9:58

idea that neural nets would... learn to

10:00

generalize. Well, he's wrong because they do.

10:03

A lot of the feats that we see, you

10:05

know, let's do and have done in the last

10:07

few years is all

10:09

of generalization. I put to Yoshua that today's

10:12

AI can do a narrow set of tasks

10:14

that it's been trained to do, but

10:16

show it something new and it struggles. Well,

10:19

it's much less narrow than it was. So

10:21

if you look at chat GPT, I mean,

10:23

one of the scary things is we now

10:25

have systems that know a

10:27

lot. In fact, they know more

10:29

stuff than any human, at least

10:32

verbalizable stuff. So, yes,

10:35

we've been on the march to build more

10:37

and more general systems and we still haven't

10:39

reached the level of generalization of humans. But

10:42

there's been a lot of progress in that direction.

10:44

Yoshua says AI may be

10:47

on the brink of high level reasoning. We're

10:49

still missing a few things, but I really

10:51

don't know if it

10:54

might be just like a mathematical formula that we

10:56

can find in six months from now and

10:59

then might take another year or two to scale it

11:01

up. Or if there are other

11:03

obstacles that I don't foresee and then

11:05

it's going to be another decade or two. So

11:07

it's possible that we're sort of, you know, I

11:10

know this is stretching it, but

11:12

one mathematical formula and a bunch

11:14

of computing power away from human

11:16

level intelligence. That's what I'm saying. And

11:19

I'm saying I'm not saying it is going to happen,

11:21

but I see this as a very clear possibility. So

11:24

the great AGI breakthrough may be happening

11:26

right now in a lab somewhere. They

11:29

could be making a podcast about us. And

11:31

if they are, we hope they call it

11:34

Hello Human Unlings. Or

11:36

AGI might never happen. We don't know the

11:38

answer. We don't even know if

11:40

the current approach is definitely the right

11:42

one. But putting

11:44

aside the debate, what if we do

11:47

achieve AGI? What could go wrong?

11:49

And will the AI's rise up to kill us

11:52

all? Let's jump ahead to the

11:54

year 2050. and

12:00

the Fremantle Dockers are aiming to win

12:03

their first premiership. AI

12:05

is now everywhere. It runs

12:07

our power grids and stock markets

12:09

and operates our weapons. Slowly

12:12

it becomes more autonomous. It does its own

12:14

thing. And then one

12:16

day it decides that its priorities are

12:18

not the same as ours. And

12:20

it hits the big red button. Joshua

12:28

Benjio says, yeah, that does

12:30

sound crazy. But

12:32

it could happen. Progei's have become

12:35

autonomous. They have their own goals. They

12:37

are trying to preserve themselves, replicate themselves,

12:39

you know, science fiction movies, scenarios.

12:43

Right now there's a lot of arguments

12:45

from very serious computer scientists that

12:47

explain how

12:50

it could happen. We don't

12:52

have any, I don't think

12:54

we have any serious arguments to show that it

12:56

couldn't. They're all plausible. Joshua

12:59

has other scenarios. Imagine

13:02

if AI becomes super smart and

13:04

then falls into the wrong hands.

13:07

It could be used as

13:09

a kind of mind control

13:11

machine. It would craft and

13:13

generate misinformation, perfectly targeted for

13:15

each person's psychology. It

13:17

would be like election hacking on steroids. And

13:19

I think this could threaten our democracies because,

13:21

you know, we have lobbying is just the

13:23

tip of the iceberg. But if

13:26

somebody has never seen

13:28

very powerful technologies at their hands, who

13:30

knows how that can

13:32

turn? And ultimately it could converge

13:34

to losing democracy completely and having

13:36

power concentrated in a

13:38

sort of single authoritarian government worldwide that

13:41

would use AI to control any kind

13:43

of opposition. Or

13:45

maybe the threat is the tech

13:47

companies themselves. Because today

13:50

the most powerful AI models are

13:52

controlled by a handful of companies

13:55

and there's every sign they'll have a

13:57

monopoly on this technology. forward

14:01

is that there may be a few people

14:03

who will have huge

14:05

power. If AI progresses

14:07

quickly in the hands of just

14:10

a few people, these people

14:12

might end up first economically, like

14:14

super, super rich, nothing like we

14:16

have now, even much worse, much

14:18

more. And with

14:20

the economic power usually also comes

14:23

political power. So

14:25

this is what keeps Yoshua up at night.

14:28

Either AI goes rogue and kills us

14:30

all, bad people use AI

14:32

to exploit and oppress others, or

14:35

we just end up in a hellscape of

14:37

tech bro overlords. And you might

14:39

think this is nuts. ChatGPT

14:42

is nowhere close to being

14:44

a threat. And

14:46

I'm not too worried about the robots rising

14:48

up either. I have a robot

14:50

vacuum cleaner. I call her Dueno the Mop Johnson.

14:54

She gets stuck in the shower all the

14:56

time. Dueno,

15:02

again. Dueno

15:06

may have a twinkle in her electronics,

15:09

maybe the shower water, but she's

15:11

a long way from world domination.

15:15

But others are taking it far

15:17

more seriously. In 2023, a

15:19

statement appeared on the internet with a

15:22

chilling message, short and to the

15:24

point. Mitigating the risk

15:26

of extinction from AI should

15:28

be a global priority alongside

15:30

other societal scale risks such

15:32

as pandemics and nuclear war.

15:35

And this wasn't some edgy post by a random

15:38

think tank. It was signed

15:40

by most of the AI industry. Texios

15:43

like Elon Musk, Sam Altman

15:45

and Demis Hisabas and

15:48

respected industry figures, including Yoshua

15:50

Benjio. After working

15:52

for decades in AI and driving breakneck

15:54

progress in the field, Yoshua

15:57

feels lost over his life's

15:59

work. dawns on you that

16:01

actually you can bring a lot of harm.

16:03

It's not easy, but if you are

16:06

honest with yourself, if you're not in denial

16:08

and you want to look at yourself in

16:10

the mirror every morning and feel good, you

16:13

have to take stock of

16:15

the reality and then see

16:17

what you can do to steer things

16:21

as you can in a better direction.

16:23

Joshua's worries are kind of symbolic of

16:26

the broader state of AI at the

16:28

moment. The industry is

16:30

now wondering what it's actually made.

16:33

Joshua had no idea AI would

16:35

improve so fast. He

16:38

says he's unintentionally created a weapon

16:40

and he's racked with guilt. He's

16:43

now urging governments to regulate. And

16:45

right now, even though I'm talking about these dangers, I'm

16:48

thinking about solutions, what can be done,

16:50

what we should do, what

16:52

are our options, what sort of regulations

16:55

do we need, how could we defend

16:57

against these dangers? Governments

17:00

are regulating, but slowly. And

17:03

here's the rub. They don't want

17:05

to regulate AI out of existence. Even

17:08

if there's a chance this technology could

17:10

one day make humans extinct. Because

17:13

they're worried other countries who don't hit

17:15

the brakes will speed ahead. And

17:18

so they're watching each other carefully. Everyone

17:21

is worried about the future, about where AI

17:23

is going. But precisely

17:26

because of this, no one

17:28

wants to fall behind. Now

17:33

not everyone in AI is worried

17:35

about rogue AI killing us. In

17:38

fact, a lot of researchers say it's

17:40

nonsense. Michael Wardridge is

17:42

a professor of computer science at

17:44

Oxford University. Nobody's ever given me

17:47

a plausible scenario from how we

17:50

go from here to Terminator. The

17:52

Terminator scenario. AKA

17:54

some hypothetical future military

17:57

AI system waking up.

18:00

going nuclear. If you remember the

18:02

Terminator scenarios involved, you know, robots

18:04

having control of the nuclear arsenal.

18:07

Very bad idea. Let's not do that,

18:09

right? I mean, let's just all agree

18:11

not to do that. But I don't

18:13

think anybody seriously remotely suggesting that. That's

18:15

not on anybody's agenda. So

18:17

if it's not scary robots, what's

18:20

Michael most worried about? Well, it's

18:22

something much more mundane and relatable.

18:25

Monitoring AI as your micromanaging

18:27

annoying boss. I'm

18:29

not applying my boss like that. I'm just

18:32

saying, all right, moving on. So

18:34

imagine in a very

18:36

near future, we've got AI, which

18:38

is monitoring every single keystroke that

18:40

you type. It's

18:42

looking at every email that you send

18:44

and scrutinizing it and giving you blunt

18:46

feedback on the quality of that email.

18:49

You didn't upsell this product. I didn't

18:51

like this phrase that you used. It

18:53

took two days for you to reply to

18:56

that email. Why did it take two days

18:58

for you to reply and so on? In

19:00

this bleak and very plausible future, humans are

19:02

treated like mindless machines. It's going

19:04

to reduce them to automata to just the

19:06

things that a machine can't do. And I

19:09

find that deeply depressing, I have to say,

19:11

and something that we should all be concerned

19:13

about. I think that future,

19:16

unless something happens, feels

19:18

like it's almost inevitable. That is, I can't

19:20

see any barriers to it happening. Michael

19:23

Woodridge didn't sign that statement about the

19:25

risk of AI wiping out humanity. In

19:28

fact, lots of prominent AI researchers

19:30

didn't. Romain Chaudry is

19:32

a Harvard Fellow in responsible AI,

19:35

and she's been named among the most

19:37

influential people in the field. She

19:40

says the Hollywood scenarios are a

19:42

dangerous distraction. Not as romantic

19:44

to talk about low income black women's

19:47

maternal health as it is to

19:49

talk about what if an AI comes alive and

19:51

takes over the government and shuts off nuclear weapons.

19:53

Romain sees the existential risk movement

19:55

in AI as a symptom of

19:57

a larger problem. about

20:00

it before. The people who

20:02

make AI are generally privileged white

20:04

men. And to them, the petty

20:07

annoyances that occur to women

20:09

and minorities, unimportant in

20:12

their world because it's not their life. So

20:14

what's Remand worried about? It's the

20:17

problems in AI right now. Bias,

20:19

misinformation, and big tech having

20:22

all the power. The

20:24

real danger isn't that AI gets too

20:26

smart, but dumb and

20:28

biased AI that gets trusted too

20:31

much. They end up in

20:33

charge of decisions like who gets a bank

20:35

loan and who misses out. They

20:37

rule our lives and they do

20:39

it unfairly. And the reason that

20:41

AI does this isn't because it's

20:43

alive. It's because the people in

20:46

charge didn't bother to fix the

20:48

problems. So when people build algorithms

20:50

that are biased, it's not because they're

20:52

malicious or evil or bad people. It's

20:54

because they overlooked something. And that is

20:56

now going to be embedded into these

20:59

large language models, general purpose models, and

21:01

we have to identify these problems at

21:03

scale. Now I could go on. I

21:06

spoke to lots of researchers, lots

21:08

of prominent AI people who shared

21:10

the things they were scared about.

21:13

But broadly, it's a spectrum. At

21:15

one end is Elon Musk, learning

21:18

about autonomous killer robots in the

21:20

future. And at the other end is

21:22

Raman Choudhury, pointing

21:24

out that AI is already causing

21:26

problems for lots of vulnerable people.

21:28

But again, unless you have lived that in

21:30

your life, that would never occur to you.

21:33

So where do I think AI is going? Well,

21:36

something that Raman said has stuck with

21:38

me. We walk towards

21:40

what we look at, right? So if

21:43

we are constantly thinking of worst case

21:45

scenarios in bad worlds, then

21:48

that's actually what we end up building, even if

21:51

we don't want to build it. We

21:53

walk towards what we look at. Now

21:56

making this series, I've had an image in

21:58

my head from a movie. Not

22:01

an AI movie, nothing like that. It's actually

22:03

The Wizard of Oz. We're

22:05

trekking down the yellow brick road towards

22:07

the Emerald City in the distance. Hang

22:10

on, does that make me Dorothy? Toto?

22:13

Anyway, we've met different characters on the

22:16

way. Minsky and Rosenblatt battling over the

22:18

future of intelligent machines. I love the

22:20

name Percetra. Lee Sedol facing

22:22

down AlphaGo and Sol in 2016. When

22:26

AlphaGo played that move, we thought it

22:28

had lost its computer mind. Robert

22:30

Williams locked up for a crime he

22:33

didn't commit. He's like, so the computer

22:35

got it wrong. And I'm like, yeah,

22:37

the computer got it wrong. Sebastian

22:40

Thrun coasting down a Californian

22:42

highway in a driverless car.

22:44

If I could go up five feet in the air

22:47

and fly to my destination, that

22:49

would be so amazing. Those students cheating

22:52

on their homework. That would

22:54

count as cheating. And now finally, we're

22:56

at the Emerald City and we're

22:58

going to meet the great AI. But

23:00

it turns out the great AI isn't

23:02

a wizard. It's not

23:04

magic. It's a machine and it's built

23:06

by us. We

23:09

called the series Hello AI Overlords and

23:11

we assumed we were talking about the

23:13

computers. But maybe

23:15

the AI overlords are people, the

23:18

ones behind the curtain, pulling the levers.

23:21

It may be the computer scientists who prepared

23:23

the training data sets or

23:25

maybe the banker who wants a return on

23:28

her investment or maybe the

23:30

researcher who's stayed up late trying to

23:32

iron out bias. OK,

23:34

so there's lots of people behind the curtain. But

23:37

I guess the lesson of this whole

23:39

series is this. The

23:43

most important thing about AI is

23:46

the humans behind it. And

23:48

yeah, like you, I hoped AI would

23:50

turn out to be a magical creature. But

23:54

the reality is actually better. It

23:56

means we, us humans, have

23:59

the chance to... to figure out what we want

24:01

AI to be and how we want

24:03

to use it. Sure, maybe

24:06

one day we'll invent an AI that's sentient

24:08

and then all bets are off. I see

24:10

you in the nuclear bunker, but

24:12

we're not there yet. So let's

24:15

think of something great. Let's

24:17

find a future that we want and

24:20

let's go there. This

24:28

has been Hello AI Overlords, a

24:30

science fiction series. I'm James Petill.

24:33

Our show is made on the lands

24:35

of the Wajak Nunga, Wurundjeri and Palawa.

24:38

With production by John Fennell, Erica

24:40

Volles and Will Ockenden. The

24:43

ABC's science editor is Jonathan

24:45

Webb. Our sound engineer was

24:47

Tim Jenkins. You can find

24:49

our previous episodes on the ABC Listen

24:51

app. Thank you

24:53

to all the AI researchers and thinkers

24:55

who spoke to us for this series.

24:58

Thank you to everyone who shared

25:00

their stories about how AI has

25:03

impacted their lives. And

25:05

thank you so much for listening.

25:35

You've been listening to an ABC podcast.

25:38

Discover more great ABC podcasts, live

25:41

radio and exclusives on the ABC

25:43

Listen app.