0:02

This is an ABC podcast. I

0:08

remember the first time I sat down to

0:10

use OpenAI's Chat GPT.

0:14

I logged on to the chatbot's website and

0:17

was presented with a blinking cursor and

0:19

that now familiar prompt, send a message.

0:22

I had the knowledge of the internet at

0:25

my fingertips. I could ask it

0:27

anything and its powerful brain would

0:29

sort through all that knowledge and give

0:31

me an answer in a lovely

0:33

digestible form. So what

0:35

did I type? Write

0:37

a short rhyming poem from my

0:40

dog to me. And

0:42

then I added please and they didn't want to sound

0:44

rude.

0:50

Dear James, With you I long depart

0:52

to that wondrous open park. Birds

0:55

to chase and grass to tread, thoughts

0:57

of sausages fill my head. So

0:59

please, dear friend, with a loving heart,

1:02

let's go and let my adventures start.

1:06

I think for an AI, that's pretty good.

1:09

I showed it to my dog, but I think

1:11

she prefers the more avant-garde stuff.

1:24

This is Hello AI Overlords,

1:26

a science fiction series about how artificial

1:28

intelligence has burst into our lives

1:31

in just a few short years. I'm

1:33

James Pertill. At the start of the

1:35

year, it was Chat GPT and its

1:37

amazing ability to communicate that

1:40

was the thing that broke AI

1:42

into the mainstream. But like everything

1:45

in AI, it didn't come from nothing. Its

1:47

rise has given one company, OpenAI,

1:50

and one man, its CEO, Sam Altman,

1:53

a huge amount of

1:54

power and a lot of say over

1:56

where AI is heading. It's

1:58

created an arms race. as the world's

2:01

massive tech companies scramble to

2:03

control the most important technology

2:05

in the world. This episode, what

2:08

happens if one man controls the

2:10

future of intelligent machines?

2:20

If modern AI and chat GPT

2:23

in particular had a face, it's

2:25

this guy. Sam Altman. What

2:27

was Sam Altman's pitch to you?

2:29

Sam Altman, everybody! Sam

2:32

Altman is the CEO of OpenAI,

2:35

the creator of chat GPT, arguably

2:38

the world's most advanced AI tool.

2:40

And depressingly, Sam is young. He's

2:43

six months younger than me. I'm 39. The

2:46

whole world wants this technology. The whole

2:48

world needs the benefits of this. For me personally,

2:51

the thing I'm most excited about is using

2:53

this technology to

2:54

increase the scientific progress.

2:56

Like other tech CEOs before him, Sam

2:58

Altman's rise to power has followed a similar

3:00

pattern. He built a cool app, dropped

3:03

out of Stanford. By 30 was

3:06

leading Y Combinator, one of the world's best

3:08

startup investment companies. And for

3:10

a brief period, eight days to be exact,

3:13

he was the CEO of Reddit. And

3:16

by the way, he's a prepper. He

3:19

said he has guns, gold, antibiotics

3:21

and gas masks stored on a rural

3:23

property somewhere. Just in case. But

3:26

at the moment, he's optimistic about

3:29

the future. I am a huge believer that the

3:31

only sustainable way that our

3:33

lives all get better is scientific and technological

3:36

progress. And as young tech CEOs do,

3:39

Altman made a lot of nice sounding statements

3:41

about his mission and how he's going to change

3:43

the world for good. We started this company

3:45

because we thought AI could destroy the whole world and we wanted to figure

3:48

out how to prevent that. For those that know

3:50

him, like reOne Child who used to work

3:52

at OpenAI, Sam Altman is

3:54

a straight shooter. He'd often have

3:56

kind of asked me anything like we'd have CEO as a

4:00

anything, he would ask them hard questions,

4:02

he'd always answer them directly, usually pretty

4:04

succinctly, and when he didn't know stuff he

4:06

would just say, I'm not sure. I

4:08

still do feel a lot of admiration and respect for him. I think

4:10

he's a very plain spoken but

4:13

really focused on the important

4:15

elements kind of guy. But the interesting

4:18

thing about Sam Altman is he has this

4:20

crazy ambitious dream. He

4:22

wants to create an artificial super

4:25

intelligence, a machine smarter than

4:27

any human. You can't just

4:30

conjure that up from nothing. You need

4:32

to take a series of steps and make a bunch

4:34

of technological advances to get

4:36

there. And that's where chat GPT

4:39

comes into it. Because up until 2017

4:42

computers were terrible at reading

4:44

and writing. Sure they could be taught to

4:46

recognize certain words, but

4:49

they struggled with sentences and paragraphs.

4:52

In the computing world, reading and writing is

4:55

known as natural language processing. It's

4:57

an incredibly complex trick and

5:00

computers were getting straight fails. It

5:02

was just garbage, you know, like

5:05

every time somebody tried to do anything

5:07

with computers and natural language, it was an embarrassment.

5:10

Jeremy Howard is a world leading machine

5:13

learning expert and yep, that's

5:15

an Australian accent. And

5:18

in 2017, Jeremy was living in

5:20

tech central San Francisco

5:23

and he believed teaching AI to read and

5:25

write was the next big step. The

5:27

vast body of knowledge of

5:30

what humans have written down and the huge ability

5:32

to communicate with humans through text

5:35

was outside of the purview of computers.

5:38

It was the biggest thing holding

5:41

back in my opinion, computers

5:43

from being, you know, as

5:45

useful a tool as they might be. Jeremy

5:48

had an idea. So he downloaded

5:51

all of Wikipedia. That's

5:53

three billion words. This idea

5:56

that if you train a big enough language model

5:59

for a long enough time, time on enough

6:02

general text, you know Wikipedia

6:04

covers a lot of different territory, you

6:07

end up with something

6:09

with a huge amount of kind of latent capabilities.

6:12

Jeremy hoped that his AI could

6:14

do more than just be an encyclopedia.

6:17

It would hopefully understand the relationship

6:20

between words. So if you fed

6:22

it the sentence the day after Wednesday

6:25

is, it would know the answer was not

6:27

potato or the battle of Waterloo

6:29

but Thursday and if it understood

6:32

enough of these relationships maybe

6:34

that would add up to intelligence. So

6:37

Jeremy used his fancy new AI to

6:39

work out if a movie was hot or

6:41

not. So I actually picked the hardest

6:44

and most well studied task which is

6:47

to read an entire multi-thousand

6:49

word movie review and

6:51

say whether it's a positive or

6:53

a negative sentiment. The experiment worked.

6:57

The model was better at working out if

6:59

a reviewer liked a movie than any AI ever.

7:02

For sure I had goosebumps. It was definitely

7:05

beating the best ever result

7:08

and so I just set to work trying to figure

7:11

out is this real because if it is real

7:13

this is the thing I've been wondering about for 30 years. Jeremy

7:17

had figured out a way to teach AI to read

7:19

and he clicked publish on his research. What

7:22

happened next shocked him. OpenAI

7:25

picked up the idea and ran with

7:27

it. This is what it had been looking

7:29

for, a way to teach AI to

7:32

read and write. A

7:34

few months later OpenAI released its

7:36

own large language model and it

7:38

said it was partly based on Jeremy's

7:41

ideas. It was an early version

7:43

of the now world famous chat GPT.

7:46

It was basic but it worked. AI

7:49

could now digest human knowledge in

7:51

written form. Sam Altman had taken

7:53

another step towards his goal

7:56

of super intelligence but

7:58

the success would put him on a collision. course

8:00

with his own principles and

8:03

he'd because

8:10

Altman didn't want to just develop super

8:13

intelligent AI. He wanted to do

8:15

this safely. Now

8:17

Altman knew the history of Silicon Valley. Small

8:19

company invents technology, dominates

8:22

the market and becomes a tech

8:24

giant. Think Google with search

8:27

or Apple with smartphones and Silicon

8:30

Valley historian Margaret O'Mara says

8:32

Altman had an idea about how

8:35

to avoid this happening to AI.

8:37

You

8:37

know he is very much committed to open source

8:40

stuff. He's an open source guy. That's

8:43

his philosophy.

8:44

And Altman was so dedicated to

8:47

this principle it's there in the company's

8:49

name. Open AI. The

8:52

open means open source. Sharing

8:54

the source code of software so it can't be

8:57

controlled by one company. But

8:59

there's a problem with this.

9:01

Money. They have money.

9:04

They've

9:04

already acquired the top research,

9:06

many of the top researchers on this in the world

9:08

work for one of these companies. Now

9:11

they're all racing against one another.

9:13

If you're a growing AI company you're

9:15

going to need people and

9:17

you can't match the salaries of the big players

9:20

without a lot of money. So

9:23

Altman was in a bind and

9:25

he knew from his time growing startups at Y

9:27

Combinator you've got to think big

9:30

because you also need money for computers.

9:33

Training AI ain't done in your laptop. Like

9:35

basically imagine a huge

9:39

warehouse like maybe 10,000 square

9:42

foot warehouse which is

9:44

just filled with racks of machines. RiiOne

9:46

Child actually went inside the data

9:48

centers that trained open AI's early

9:51

GPT models. Each of which cost

9:53

like

9:54

I don't know 20 or 40 thousand dollars.

9:56

Or when you start running your model just the sound

9:59

of all of the sound,

9:59

turning on is like

10:02

you know a jet turbine or something.

10:05

Altman had reached the crossroads. Taking

10:08

the money would allow investment growth

10:10

and speed. Not taking the money

10:13

would mean giving up on his dream of

10:15

super intelligent AI. And

10:17

so in 2019 he announces

10:20

his decision. He's taking

10:22

the money. OpenAI signs

10:24

a one billion dollar deal with

10:26

Microsoft and its technology

10:29

will no longer be shared freely

10:31

with the world. Some employees

10:33

even quit and protest at OpenAI

10:36

becoming too commercial and

10:38

they start their own AI company, Anthropic.

10:41

But with Microsoft's money Altman

10:44

can pursue his ambition and

10:46

his company begins training an enormous

10:48

AI model much much larger

10:51

than anything that's been seen before.

10:53

Personal money terms it's absolutely ridiculous

10:56

like oh that you spend like millions

10:59

of dollars more money than most people see in their lifetime

11:02

on a language model where it's

11:04

like you're not even sure if it'll be useful for anything.

11:07

So Sam Altman has to compromise

11:09

on one of his big ideals. He

11:11

believes he has to do so to

11:13

invent the technology for a better

11:15

future. But a better future

11:18

for who? Yes it was a dream

11:21

job. This is Richard Methenge and these

11:23

days he's the head of the African Content

11:26

Moderators Union and he's been listed

11:28

among the top people in AI. But

11:31

back in 2021 he was just starting a new job

11:35

with an outfit called Summer. This

11:37

was a company OpenAI had contracted.

11:40

Summer is based in Richard's hometown of

11:42

Nairobi, Kenya. Richard's

11:44

job and that of dozens of other workers

11:47

was to sift through the output of

11:49

chat GPT as it was being developed.

11:52

Because the model that's trained from all the text

11:54

from the internet ends up spitting

11:56

out a lot of terrible stuff. So

11:59

humans had to to tell AI what was

12:01

acceptable and what was not. We

12:03

had two factions. One was dealing

12:06

with violent content and

12:08

the other one, which was myself

12:10

and my team, we were actually working

12:13

on sexual content. We were

12:15

training the chatbots how to

12:18

work with toxic pieces of

12:20

text that was very disturbing, very

12:23

grotesque, very traumatizing.

12:26

For 10 hours a day, five

12:28

days a week. He and his team worked

12:31

in a secluded room. Often

12:33

the material they had to read was

12:35

chat GPT's written depictions of

12:38

child sex abuse and bestiality.

12:42

So chat GPT would churn out content

12:45

and Richard would tick a box saying

12:47

if it was acceptable or not. But

12:50

as the months wore on, the flow

12:52

of disturbing content didn't slow

12:54

down and the team grew

12:56

more and more traumatized.

12:59

And some of them, by the way, as we speak

13:01

right now, it was so harsh to

13:04

the point that they even separated

13:06

or they even divorced with their partners

13:09

just because of the experience

13:11

that they went through in terms

13:14

of content moderation. And the kicker, they

13:16

got paid peanuts. It

13:19

was less than less than a dollar

13:21

an hour. Less than a dollar.

13:24

Less than a dollar an hour.

13:26

Now, OpenAI says the story is more

13:28

complicated. It believed Summer

13:30

was paying its contractors more

13:32

than $1 an hour. Summer says

13:34

it didn't do anything wrong and

13:37

that it offered workers living wages.

13:40

Either way, Richard says he and

13:42

his team haven't had an apology from OpenAI

13:44

and they've set up a union to

13:46

help African content moderators negotiate

13:49

with big tech. You don't just work

13:52

with someone and throw them

13:54

away, you know, just

13:56

because you have maximized the process. OpenAI

13:59

was now working. billions and

14:01

it was no longer nonprofit nor

14:04

very committed to open source. Instead

14:06

it was allied to Microsoft. Sam

14:09

Altman had done what it took to chase

14:11

his dream. He may have had to compromise

14:14

on his open source ideals and

14:16

take billions in funding but

14:18

he was about to change the world. ChatGPT

14:22

was unleashed and everyone

14:24

lost their mind. This is reOne

14:27

Child who worked at OpenAI. I think

14:29

the public reaction is something I had completely

14:32

just like completely gobsmacked

14:34

by like I had no ability to predict. ChatGPT

14:37

was the fastest selling consumer

14:39

app ever. It made OpenAI

14:42

and Sam Altman famous. It

14:44

wasn't super intelligent but it was

14:47

freakishly good. It could pass

14:49

uni exams and it looked like

14:52

it could replace a lot of workers. People

14:55

woke up to the potential of AI. People

14:57

like leaders of countries, politicians

15:00

and they were nervous about this new power

15:03

that Sam Altman wielded. What

15:05

future was he creating? Somehow

15:13

Sam Altman had found himself as

15:15

the spokesperson for modern AI

15:18

and he was happy to use his unofficial

15:21

status. In May 2023 he

15:23

hopped on a plane for a 22 country, 25 city world tour. It's

15:29

basically a very nerdy conteci

15:31

tour. Instead of getting drunk and

15:34

passing up beneath a historical statue

15:36

he addressed the United States Congress.

15:39

My name is Sam Altman. I'm the chief executive

15:41

officer of OpenAI. He met Francis president

15:44

Emmanuel Macron. Thanks for the opportunity to speak to

15:46

you today about large neural networks. It's

15:48

really an honor to be here. Then Britain's prime minister

15:50

Rishi Sunak. We think it can be

15:53

a printing press moment. We have to work together

15:55

to make it so. Then India's PM Narendra

15:57

Modi. We believe that the benefits of the tools

15:59

we have to deployed so far vastly outweigh the

16:01

risks, but ensuring their safety is

16:03

vital to our work. Margaret O'Mara

16:06

says Altman presented himself to world leaders

16:08

as the inevitable new tech superpower.

16:11

Sam Altman's on a world tour as

16:14

if he were a leader

16:17

of a nation state, a head of state, is kind

16:19

of in a way underscoring that

16:21

even if that was the intent.

16:23

It wasn't just a simple PR

16:25

trip. Altman's message was more

16:27

complicated than the arguments of previous

16:30

tech CEOs talking to governments. He

16:33

said AI was great, but

16:35

it could be dangerous, and it

16:38

needed to be regulated. We think that regulatory

16:40

intervention by governments will be critical

16:42

to mitigate the risks of increasingly powerful

16:45

models. Altman was like an evangelist

16:47

preaching the coming of the higher power.

16:50

We are working to build tools that one day could help

16:52

us make new discoveries and address some of humanity's

16:54

biggest challenges, like climate change

16:57

and curing cancer. He said super intelligent

16:59

AI was right around the

17:01

corner, and it was going to be everywhere

17:04

in our lives. Like a really

17:06

clever Google Assistant. So

17:08

when I gave the current generation of AI go,

17:11

it would seem this future is

17:13

not here yet. Hey Google, are

17:15

you there? Of course I am. I'm always here

17:17

for you whenever you need me. I'm trapped

17:20

in the bathroom. I need your help

17:22

to get me out. Sorry,

17:24

I didn't understand that.

17:25

Just tell me if you need

17:27

help with anything or is this a good time

17:29

to have fun? I need help getting

17:32

out of the bathroom. It's an emergency. Bummer,

17:34

I still don't understand. Okay. But

17:37

I'll be here if you need me.

17:38

I'm actually quite hungry and I

17:41

need help. I've been in here for some

17:43

time now Google. Is there anything

17:45

you can do? Hey

17:48

Google? Anyone? Hello? But

17:55

Altman says super intelligence is

17:57

coming. And it's not all. Good

18:00

news. My worst fears are that we cause

18:02

significant, we the field, the technology,

18:05

the industry cause significant harm to

18:07

the world. For example, AI could

18:09

be used to hack elections through

18:12

creating videos of politicians doing

18:14

things they didn't actually do or using

18:16

their voice to frame them for scandals.

18:20

Whatever the invent of use, whoever

18:22

has this power will steamroll

18:24

those who don't. It is essential that powerful

18:27

AI is developed with democratic values in mind.

18:29

Now Altman prides himself on being

18:31

unlike the other tech CEOs. He's

18:34

a cool CEO. And now after

18:36

years of hard work, he's the

18:39

guy who's found himself as the

18:41

voice for AI. The multimillionaire

18:44

from Y Combinator who's behind disruptive

18:46

companies like Airbnb is

18:49

the guy government seems to be listening to. They

18:52

ask him, what should be the rules

18:54

around this technology? And the future

18:57

that he sketches up is kind of what you'd

18:59

expect from a guy out of Silicon Valley.

19:02

It's a future where a bunch of CEOs make

19:04

a lot of money. Altman

19:07

says, open AI and other tech companies

19:09

should continue to own and operate the

19:12

most powerful AI models. And

19:14

yeah, they should be regulated, but

19:17

not too much.

19:18

Well, at the end of the day, these are all

19:20

private sector companies and their purpose

19:23

is to make money. It's capitalism.

19:26

So when you scrape away the hype and

19:28

Altman's warnings, his vision of the

19:30

future is a familiar one. It's

19:33

the same Silicon Valley dream we've heard

19:35

for years from Google, Facebook

19:37

and others. Technology

19:40

will make the world better. It

19:42

will solve everything. And

19:44

yeah, it will be very profitable.

19:46

The Sam Altman's and others in this whole

19:48

debate, they are very,

19:51

very privileged. They are by and large,

19:53

extremely wealthy, extremely wealthy

19:56

and great wealth can create a bubble.

19:58

Even if you have

19:59

every desire to stay in touch

20:02

with the pulse of the world and

20:04

what people are thinking, great

20:06

power and wealth is very isolating.

20:08

You think he's being a little naive?

20:10

I fear so. Yes, I do. I

20:12

do. I wish it were otherwise.

20:14

I wish it were otherwise.

20:15

Altman's world tour feels a bit like

20:17

a global victory lap for his vision

20:20

of AI and his message is

20:22

that he's the good guy. He wants to

20:24

share AI with the world, not lock

20:26

it up but others grow worried that

20:28

actually he's tightening his control.

20:31

OpenAI is now believed to be valued

20:33

at $100 million. It's

20:35

gone up 3x in a few months. I

20:38

don't see why it won't go up another 3x in another few

20:40

months.

20:44

That's Jeremy Howard, the Australian who developed

20:47

the language model that helped OpenAI

20:49

make chat GPT. You

20:51

know, these AI companies, I suspect

20:53

are on track to being the most

20:57

powerful organizations in the world and will

20:59

continue to grow. But what's stopping

21:01

others from making their own powerful

21:04

AI? Well, Jeremy

21:06

says there's a couple reasons. First,

21:09

there's a big cost barrier. It costs

21:11

over a billion dollars to train a cutting

21:13

edge AI and that cost is

21:16

going up. And then there's

21:18

a data barrier. GPT-3

21:20

was trained on the entire internet. GPT-4

21:24

was trained on the internet plus data

21:26

created by GPT-3. GPT-5

21:29

will be trained on data from GPT-4. And

21:33

so we're at the point now where you need an

21:35

existing very large language model to

21:37

train a new one. This positive

21:40

feedback will be so great that

21:43

that company becomes

21:46

the biggest monopolist in history.

21:50

So Jeremy says only a few companies

21:52

will have the resources to build the

21:55

best AI. And that means

21:57

whoever is in front now or in

21:59

the next. years will dominate

22:01

AI for at least the next

22:04

few decades. They'll make the British

22:06

East India companies seem like peanuts. They'll

22:09

be vastly, vastly powerful. And

22:12

so yeah, what happens to the rest of us in that situation?

22:14

Well, we don't really have much to add, do we?

22:18

We then become, most of the world then becomes

22:20

actors with little ability

22:23

to generate surplus economic value.

22:26

Even that already right now, the humans

22:28

in the world who we feel that cannot

22:31

generate huge amounts of surplus economic value

22:33

we treat like absolute crap, you should

22:36

assume that when most of us are like

22:38

that, we'll be treated exactly the way that

22:41

we treat those people now.

22:45

Now we don't know how the future

22:48

will work out, but it looks

22:50

like the early idealistic

22:52

phase of AI research has

22:54

passed. We're now entering

22:56

a new era of profit. We've

22:59

seen this before. Google's first

23:02

motto way back in 2000 was don't

23:05

be evil. I quietly got

23:07

rid of this five years ago. And

23:10

this is why we're talking about Sam Altman and

23:12

this scramble for control. There's

23:14

two sides to Altman. One a

23:16

techno optimist excited about

23:19

future technology. And the

23:21

other side, well, he's

23:23

just another Silicon Valley CEO.

23:26

I know how power works. I've seen it enough times

23:28

that when people start making

23:30

bucket loads of money, they get

23:32

surrounded by people who want them to keep making

23:35

bucket loads of money so that they can also make bucket loads

23:37

of money and they will convince themselves, oh,

23:40

you know, the more important thing for us is to make bucket

23:42

loads of money. Is

23:44

it a case of power corrupt? I

23:47

think it drags in the wrong kinds of people.

23:50

I think you can convince yourself very

23:52

easily that you being

23:55

in control of vast power and resources

23:57

is actually a net point to the world.

23:59

convince themselves that. I

24:02

think that's very

24:06

unlikely to be true.

24:08

This is Hello AI Overlords, a

24:10

science fiction series. I'm James

24:12

Pertill.

24:17

Our show is made on the lands of the Wajak

24:20

Nunga, Wurundjeri and Palawa, with

24:24

production by Jordan Fennell, Erica Volles

24:26

and Will Ockenden. Our sound engineer was Marcus

24:29

Hobbs. Next

24:31

episode, we've talked about the history

24:33

of AI and the companies that control it. Now

24:35

it's time to talk impacts.

24:38

I was brought on as a

24:40

background actor in a very

24:42

large superhero movie.

24:44

And hear from people whose lives were disrupted

24:46

by the sudden arrival of this powerful

24:49

new technology. There were weird vibes. There were a lot

24:51

of people from

24:52

the studio there.

24:55

One of the PAs

24:55

was coming through and kind of like hand selecting people

24:59

from a list. Actors who went on strike

25:01

saying studios wanted to digitally clone them and

25:03

do the amount of acting work. They said, you know, we got

25:05

this really

25:08

cool opportunities and we're going to

25:10

scan you with the full body scan.

25:12

It's like, you know, video games. It's cool. It's like, it's the thing

25:15

that we

25:15

do. This is just like a new way

25:17

of doing it. That's the story of 2023, the

25:19

year the world

25:22

woke up to AI. You can find our previous

25:25

episodes right now on ABC Listen.

25:28

Search for science friction. Don't forget

25:30

to tell a friend about the show. See

25:32

you soon.