0:00

City of Dreams is an incredible new

0:02

movie from executive producers of Sound of

0:04

Freedom and executive producer Vivek Ramaswamy, opening

0:06

in theaters everywhere on Labor Day weekend.

0:08

Inspired by a true story, it's about

0:10

a young boy from Mexico whose dream

0:12

of becoming a soccer star is shattered

0:14

when he's trafficked into America to work

0:16

in a sweatshop where he becomes one

0:18

of the 12 million children who are,

0:20

right now, victims of modern slavery. It's

0:22

the story of how he finds the

0:25

hope and courage to fight for his

0:27

freedom. See this important story that sheds

0:29

a light on the real-life issue of

0:31

child trafficking and join the mission to stop

0:33

this injustice. Tony Robbins calls City of Dreams

0:35

a prayer to the universe, and Grammy Award

0:37

winner Luis Fonsi says it's a rallying cry

0:39

for justice, and it will keep you on

0:41

the edge of your seat. Critics

0:43

are calling City of Dreams inspirational, Oscar-worthy, powerful,

0:46

and eye-opening. You don't want to miss this

0:48

film. Make sure you see it in a

0:50

theater. City of Dreams,

0:53

in theaters August 30th, rated

0:55

R, under 17 not admitted

0:57

without parent. Buy tickets at

0:59

cityofdreamsmovie.com. And

1:02

now, a Blaze Media Podcast. My

1:05

next guest is sounding the

1:07

alarm on the catastrophic risks

1:09

posed by AI, from

1:12

totalitarianism to bioengineered pandemics

1:14

to a total takeover

1:16

of mankind. When

1:20

you think about the things that we could

1:22

be facing, it doesn't look real good for

1:24

the human race, but it's not too late

1:26

to turn the ship around and harness the

1:29

power of AI to

1:31

serve our interests. But

1:33

if we don't, well, I'll

1:36

let him tell you what happens. Welcome

1:38

to the podcast, the Executive Director

1:40

at the Center for AI Safety

1:43

and an advisor for Elon Musk's

1:46

ex-AI, Dan

1:48

Hendricks. But

1:51

first, let me tell you

1:53

about Preborn, our sponsor. We're

1:55

going to be talking about life and

1:58

what is life. life, the

2:02

age of spiritual machines, if you will.

2:05

We know what life is now. Maybe

2:08

a quarter of the country doesn't

2:10

know what life is, but

2:13

it is worth living on

2:15

both ends of the scale in the womb

2:18

and towards the end. We

2:21

need to bring an end to

2:23

abortion and define

2:25

life and really appreciate

2:27

life, or AI

2:30

will change everything for us. It

2:32

will take our programming of, eh,

2:34

that one's not worth that much.

2:37

And God only knows where it will take us. The

2:40

Ministry of Preborn is working every

2:42

single day to stop abortion, and

2:44

they do it by introducing an

2:46

expecting mom to her unborn baby

2:49

through a free ultrasound that you and

2:51

I will pitch in and pay for.

2:54

They have rescued about 200

2:56

babies every day, and 280,000 babies have

3:03

been rescued so far just from the

3:05

ultrasound. And then also when

3:08

mom says, I don't have any support

3:10

system, they're there to offer assistance to

3:12

the mom and a support system for

3:14

up to two years after the baby

3:16

is born. Please help

3:19

out, if you will, make a

3:21

donation now. All you have to

3:23

do is just hit pound 250

3:25

and say the keyword baby that's

3:27

pound 250 keyword baby,

3:30

or you can go to

3:32

preborn.com/Glenn. This

3:34

episode is brought to you by Shopify. Forget

3:38

the frustration of picking commerce platforms

3:40

when you switch your business to

3:42

Shopify, the global commerce platform that

3:44

supercharges your selling wherever you sell.

3:46

With Shopify, you'll harness the same

3:48

intuitive features, trusted apps and powerful

3:50

analytics used by the world's leading

3:52

brands. Sign up today for your

3:55

$1 per month trial

3:57

period at shopify.com/tech all lower.

4:00

That's shopify.com/tech.

4:05

Hey Dan, welcome.

4:19

Hey, nice to meet you. Nice to

4:21

meet you. I am thrilled that you're on.

4:25

I have been thinking

4:27

about AI since I read The

4:30

Age of Spiritual Machines by Ray

4:32

Kurzweil. And that

4:34

so fascinated me. And

4:38

later I had a chance to talk to Ray. And

4:41

he's fascinating and terrifying, I think, at

4:43

the same time. Because,

4:46

you know, I don't see a lot of

4:48

people in your role. Can

4:50

you explain what you do

4:53

within the found, you

4:55

know, what you founded and what you do? Yes.

4:57

So I'm the director of the Center for AI Safety.

5:00

We focus on research and trying to

5:02

get other people to research and think

5:04

about risks from AI. And

5:07

we also help with policy

5:09

to try and suggest, suggest

5:12

policy interventions that will help reduce risks from AI.

5:16

Outside of that, I also

5:18

advise Elon Musk's AGI company,

5:20

XAI, as their sole safety

5:22

advisor. So we

5:24

have a variety of hats. There's a lot to do in AI

5:26

risk. So

5:29

research and policy advising are the main things I work

5:31

on. So how

5:34

many, how many heads

5:36

of AI projects are

5:40

concerned and are

5:45

not lost in, I'm going to

5:47

speak to God, this

5:50

drive that a lot of them

5:52

have to create something and

5:54

be the first to create it. How

5:57

many of them can balance that with, Maybe

6:00

we shouldn't do X, Y, and Z. I

6:05

think that a lot of the people who got

6:07

into this were concerned

6:09

about risks from AI, but

6:11

they also have another

6:14

constraint, which is that they want to make sure

6:16

that they're at the forefront and competitive. Because

6:19

if they take something like safety much

6:21

more seriously or slow down or proceed

6:23

more cautiously, they'll end up falling behind.

6:25

So although they would all

6:28

like there to be more safety and

6:31

for this to slow down, or

6:33

most of them, it's not

6:35

an actual possibility for them. So

6:38

I think that overall,

6:40

even though they have good

6:43

intentions, it doesn't matter,

6:45

unfortunately. Right. So

6:47

let me play that out a bit. Putin

6:52

has said whoever gets AI first will

6:54

control the world. I believe

6:56

that to be true. So

7:00

the United States can't slow down because

7:03

China is going to be, they're pursuing

7:05

it as fast as they can. And

7:08

I'm not sure. I don't want them

7:10

to be the first one with AI. It

7:12

might be a little spookier. So

7:16

is there any way to actually slow

7:18

down? Well,

7:21

we could possibly slow down

7:24

if we had

7:26

more control over the chips that

7:28

these AI systems run on. So

7:31

basically right now there are export controls to

7:33

make sure that the high end chips that

7:35

these AI's run on don't go to China,

7:37

but they end up going to China anyway.

7:39

They're smuggled left and right. And

7:42

if there were actually better constraint and we

7:44

had better export controls, then that

7:46

would make China substantially less competitive. Then we

7:48

would be out of this pernicious dynamic of

7:50

we all want safety, but you got to

7:52

do what you got to do and we

7:54

got to be competitive and keep racing forward.

7:56

So I think chips might be a way

7:58

of making a situation. not be in that

8:01

desperate situation. Are those chips

8:03

made in Taiwan or here? The

8:06

chips are made in Taiwan. However, most

8:08

of the ingredients that go

8:10

into those chips are

8:14

made in the US and made among

8:16

NATO allies. So about 90% of

8:18

those are in the US and NATO allies.

8:20

So we have a lot of influence over

8:22

the chips, fortunately. OK. But

8:25

if Taiwan is taken by China, we

8:29

lose all the, I

8:31

mean, we can't make those chips.

8:33

That's the highest end chip manufacturers,

8:35

right? And China will have

8:37

that. So what does that mean for

8:39

us? It seems

8:42

plausible that actually if China

8:44

were invading Taiwan, that the

8:46

place that makes those chips would actually just

8:48

be destroyed before they would fully take it.

8:51

So that would put us on more of an even playing

8:53

field. OK. So I've

8:58

been talking about this for 25, 30 years. And

9:03

it's always been over the horizon.

9:05

And I could never get people

9:07

to understand, no, you've got to

9:09

think about ethical questions right now.

9:12

Like, what is life? What is

9:14

personhood? All of these things. And

9:17

now it's just kind of like

9:19

the iPhone. It just happened.

9:21

And it's going to change

9:23

us. And it

9:25

hasn't even started yet. And it's amazing.

9:28

I go online now. I

9:30

don't know what's real or not.

9:33

I mean, I found myself this

9:35

week being on X or on

9:37

Instagram and looking and saying, is

9:40

that a real person? Is that a real

9:42

video? Is that

9:44

a real photo? You have no idea. Yeah.

9:48

And we've just begun. Yeah,

9:51

yeah, yeah. I

9:53

think that's a concern where

9:55

we don't have really great

9:58

ways to reliably detect. whether

10:00

something is fake or not. And this could

10:02

end up affecting our collective understanding of things.

10:05

I think another concern are AI companies

10:08

biasing their outputs. So people are wanting

10:10

to do things about safety, but it

10:12

creates a vacuum. If we got to

10:15

do something about it and what

10:17

takes its place is some culture war type of things.

10:19

As I think we saw with Google

10:22

Gemini, when you'd ask it to generate an

10:24

image of George Washington, then it'll

10:26

output, it'll make

10:28

him look black. To

10:30

make it, because its image outputs need to be

10:32

diverse. So that I

10:34

think is one reason

10:37

why Elon Musk's through

10:39

his company XAI is getting into

10:41

the arena and now

10:44

has a pretty competitive

10:46

AI system. So as to try

10:48

and change the norms so that

10:50

other big tech companies, when they're

10:52

sort of biasing their outputs, there

10:55

are alternatives so that we're

10:57

not all locked into whatever some random people

10:59

in San Francisco decide are the values of

11:01

AI systems. Yeah, it's

11:03

really difficult because you can see the

11:06

bias. It's quite

11:08

clear, the bias, especially if you know

11:10

history or you follow the news as

11:12

closely as I do. But

11:15

the average person won't see that. I

11:19

look at AI as a, like any

11:23

technology, a

11:25

tremendous blessing and

11:28

a horrible curse. But

11:31

this one has the potential of

11:34

enslaving all of us. Doesn't

11:41

it? I

11:43

think at least, I want at least

11:46

distinguish between the systems right now. The systems

11:48

right now have- Yeah, yeah, yeah. I mean

11:50

in potential, yeah, what's coming? Sure,

11:52

I mean, when it's as capable

11:54

as humans and when they have robotic

11:56

bodies and things like that, I mean,

11:58

there's basically no limits. to what

12:00

they could do. And it really matters how people

12:03

are using them, what instructions are given. Are

12:05

they given to cement a

12:07

particular government's power?

12:10

Are they used by non-state actors

12:12

for terrorism? All

12:14

of these things could lead to societal

12:16

scale risks, which

12:20

could include some sort of unshakable

12:22

totalitarian regime enabled by AI, or

12:26

unseen acts of terror.

12:28

So I think we're

12:31

at the same time, silver lining is

12:33

maybe if it all goes well, we

12:35

get automation

12:37

of things and we

12:39

don't have to work as much or at all.

12:43

So it's really divergent paths. Right.

12:47

Which do you think is more likely? I

12:52

think overall it's more likely that we end

12:54

up seeding

12:59

more and more control to AI systems and we can't

13:02

really make decisions without

13:04

them becoming extremely dependent on them. I

13:07

would also guess that some people would

13:10

give them various

13:12

rights in the farther future. And this will make

13:14

it be the case that we

13:17

don't control them or all of

13:19

them. So I am

13:22

not too optimistic for us overall.

13:25

There's still a lot of ways this could go. If

13:27

we said we're on team human, we

13:29

need to come together as a species and handle it, we're in

13:31

a different situation. But for instance,

13:33

if there were a

13:36

catastrophe, then we might actually

13:38

take this much more seriously. Otherwise we might

13:40

just sleepwalk into something and have the frog boil.

13:43

What would be a catastrophe that could

13:45

happen in the relative near future

13:47

that would wake us up that wouldn't destroy us?

13:51

Yeah. So I think one

13:54

possibility maybe say two to three years from

13:56

now is somebody instructing the dollar things

36:00

like that. So if we deleted

36:02

that knowledge from

36:04

the AI systems or had them just

36:06

refuse questions about reverse genetics or made

36:08

them use information about reverse genetics, then

36:11

we could still have

36:13

brain cancer research,

36:15

all these sorts of things, but

36:18

we're just bracketing off

36:21

virology, advanced expert level

36:23

virology. And maybe some

36:25

people could access that, like if they'd have

36:27

a clearance right now, we

36:29

have BSL-4 facilities. Like if you want to study

36:32

Ebola, you got to go to a BSL-4 facility.

36:34

So people can still do some research for it,

36:36

but it shouldn't be necessarily, everybody in the public

36:38

can ask questions about advanced

36:40

virology to how to increase the transmissibility of

36:42

a virus. So I think we can

36:45

partly decouple some of the

36:47

good from the bad with

36:50

biological capabilities. But as

36:54

it stands, the AI

36:56

systems keep learning more and more. There

36:58

aren't really guardrails to make sure that

37:00

they aren't answering those sorts of

37:02

questions. There aren't clear laws

37:04

about this. For instance, the US Bioterrorism Act

37:07

does not necessarily apply to AIs because

37:09

it requires that they are knowingly aiding

37:12

terrorism and AIs don't necessarily knowingly do

37:14

anything. We can't describe intent to them.

37:17

So it doesn't necessarily apply. A

37:19

lot of our laws on these don't necessarily

37:21

apply to AIs, unfortunately. So

37:24

yeah, I think

37:27

if we get expert level virologist AIs

37:29

and if they're ubiquitous and

37:31

it's easy to break

37:34

their guardrails, then that's

37:37

also walking into quite a potential

37:40

disaster. Right now, the

37:42

AI systems can't particularly help with making

37:44

bioweapons. They are better than Google, but

37:46

not that much better than Google. So

37:48

that's a source of

37:50

comfort. But I'm

37:54

currently measuring this with some

37:56

Harvard MIT virology PhD

37:59

students, where... where we're taking a

38:01

picture of virologists in the lab

38:04

and asking the AI, what should the virologists do next?

38:06

Like here's a picture of their Petri dish, here's their

38:08

lab conditions. And can it fill in the steps? And

38:10

right now it looks like it can fill in like

38:12

20% or so of the steps.

38:14

If that gets to 90%, then

38:17

we're in a very dangerous situation where

38:19

non-state actors are just randomly- How long

38:21

will that take? Yeah, so

38:23

I think progress is very surprising

38:25

in this space. Just last year,

38:27

the AIs could barely do basic

38:30

arithmetic where you're

38:32

adding two-digit numbers together. They would fail

38:34

at that. And then just last month,

38:36

or just this month, excuse me, now

38:38

they're getting a silver medal at the

38:40

International Mathematical Olympiad, which is the greatest

38:42

math competition. So it

38:46

could go from basically ineffective to

38:48

expert level, possibly within a year.

38:51

There's a bit of uncertainty about it, but it

38:54

wouldn't surprise me. So people,

38:56

it was a big debate whether

38:59

AGI and ASI could ever

39:01

happen. And

39:05

the point of singularity, I've

39:10

always felt like it, and I know nothing

39:12

about it, but I've always felt that it's

39:15

a little arrogant. We're

39:18

building something and we

39:20

look at it as a tool, but

39:23

it's not a tool. It's like

39:25

an alien, it's

39:27

like an alien coming down. We think they're gonna

39:30

think like us. Well, they won't think like

39:32

us. They have completely different

39:34

experiences. We don't know how this

39:36

will think. Do

39:40

you believe in the singularity that we

39:42

will hit ASI at some

39:44

point? I

39:46

think we'll eventually build a superintelligence if

39:49

we don't have substantial disruption

39:51

along the way, such as a huge

39:53

bio weapon that like harm

39:56

civilization or like TSMC gets blown.

39:58

Those might be things that would. really

40:00

extend the timeline. So by default

40:02

seems pretty plausible to me, like more likely

40:05

than not that we'd have a super intelligence

40:07

this decade. And

40:09

most people in the AI industry

40:11

think this as well. Like Elon thinks maybe

40:14

it's a few years away, Sam Mullen and

40:16

Jeddo does, Daria the head of entropic does,

40:19

one of the co-founders of Google DeepMind, I

40:22

think AGI is in 2026. So

40:25

yeah. But actually I

40:28

remember- Can you explain AGI? Can you

40:30

explain that to somebody who doesn't understand

40:32

what that means? Yeah,

40:35

so AGI has a constantly shifting

40:37

definition for many people, it used

40:39

to mean, an

40:41

AI that could basically talk like a human

40:44

and pass like a human. That

40:47

was the Turing test as it was called, but looks

40:49

like they're already able to do that. It

40:51

also was in contrast to narrow AI.

40:53

AI's just a few years ago could

40:55

only do a specific task. And

40:58

if you slightly change the specification of tasks,

41:00

they just fall apart. But now they can

41:02

do arbitrary tests, they can write poetry, they

41:04

can do calculus, they can generate images, whatever

41:07

you want. So

41:10

by some definitions, we have AGI. And

41:12

so there's been a moving goalpost where

41:14

people are now using it to mean

41:17

something like expert level in all domains.

41:20

Right. And able

41:22

to automate basically anything. So people

41:24

are like, we'll know when there's

41:26

AGI, when the AI lab stop

41:28

hiring people. Which

41:31

some of them have in their forecasts for

41:35

spending on labor, some of them are

41:37

expecting to stop hiring in a few

41:39

years because assuming there's automation. Wow. So

41:44

it varies quite a bit,

41:46

but you don't need AGI for a lot

41:48

of these malicious use

41:51

risks. It just needs to be very good at doing

41:53

like a cyber attack or it just needs to have

41:55

some expert level of virology knowledge and skills to

41:58

cause a lot of damage. So. big

56:01

AI upgrade maybe in the next six

56:03

months, late this year, early next year.

56:06

And that'll make the public go, what's going on?

56:10

And start having some demands of something to be

56:12

done about AI. How's that

56:14

gonna manifest itself in the next six months? So,

56:18

I make

56:20

this prediction largely just based on the

56:22

fact that it took them a long

56:24

time to build their 10X larger supercomputer

56:26

to train these AI systems. And now

56:28

they're basically built. And so now

56:30

they're training them. And they'll finish training

56:33

around the end of this year or early next year

56:35

and be released then. So, the

56:37

exact skills of them are unclear

56:39

each time, each 10X

56:43

in the amount of power and data that we throw

56:46

into these systems. We

56:48

can't really anticipate their capabilities

56:51

because AI systems are not really designed

56:54

like old traditional computer programs, they're more

56:56

grown. We just let them stew for

56:58

some months and then we see what

57:00

comes out. Wow, and it's like magic.

57:04

Kind of, we have extremely huge sources of energy, or

57:11

we have substantial sources of energy just

57:13

flowing directly into them for months. Right,

57:16

right. To create them. It's alive. Yeah,

57:19

yeah. So,

57:21

I think they should probably get a

57:26

lot more expert level reasoning, whereas right

57:28

now they're a bit shakier and this

57:30

could potentially improve their reliability for doing

57:32

a lot of these agent tasks. Right

57:34

now, they are closer to tools than

57:36

they are agents. But-

57:39

What's the difference between an agent and a tool?

57:43

Yeah, so a tool, it's like a

57:45

tool being like a hammer. Meanwhile,

57:47

an agent would be like an executive

57:50

assistant, a

57:52

secretary. You say, go do this for me, go book this for me,

57:55

arrange these sorts of plans, make me a PowerPoint,

57:58

write up this document. and

58:00

submitted, email it, and then handle the back and

58:02

forth in the email. I

58:05

think those capabilities could potentially turn on

58:07

with this next generation of AI systems.

58:09

We're already seeing signs of it, but

58:12

I think there could be

58:14

a substantial jump when we have

58:16

these 10X larger models. Wow.

58:18

I mean, think of it this way. Just in terms

58:20

of brain size, imagine like a

58:22

10X larger brain. Yeah. Shouldn't expect

58:24

that to be a lot more

58:26

capable. Well, at some point, it's

58:30

going to snap the neck, that

58:33

larger brain. So I hope that doesn't happen

58:35

soon. Dan,

58:37

you're fascinating. Thank you. 15

58:42

years ago, I was looking

58:44

for the people who had some

58:47

ethics that were saying, wait,

58:49

let's slow down. We should ask

58:51

these questions first. And I didn't

58:54

find a lot

58:56

of philosophy behind the progress

58:58

seekers. And

59:05

it really frightened me

59:08

because at some point,

59:12

they're going to say, you can live forever, but

59:15

it's just a downloaded you. And if

59:17

we haven't decided what life is, we

59:22

can easily be taught that, no,

59:24

that's grandma. And

59:27

then what value does

59:30

the actual human have, the body,

59:33

if it's just downloadable? So

59:38

I appreciate your look at safety

59:40

and what you're trying to do.

59:43

Thank you. And thank

59:46

you for bringing this topic to your

59:48

audience because it's important and it still

59:50

isn't discussed. Oh, yeah. Yeah.

59:52

Thank you. We'd love to have you back. Thank you.

59:55

Yeah. Have a good day. Bye. Thanks. Thank

59:58

you. Bye-bye. Just

1:00:04

a reminder, I'd love you

1:00:06

to rate and subscribe to the podcast and pass

1:00:08

this on to

1:00:16

a friend so it

1:00:19

can be

1:00:22

discovered by other people.