0:07

I actually went to them and

0:09

said, are you sure you don't see

0:11

any murder of death here? It'd be

0:13

really good if you saw some images

0:15

of death and the destruction of

0:17

humankind. The A-Fix, the digital

0:19

zoo, smart machines, what will

0:22

they do? Lies to Mars

0:24

or bake a bad cake.

0:26

World domination, a silly mistake.

0:28

Hello, hello and welcome to

0:30

episode 41 of The AI

0:32

Fix, your weekly dive headfirst

0:34

into the bizarre and sometimes

0:37

mind-boggling world of artificial intelligence.

0:39

My name is Graham clearly. And

0:41

I'm Mark Stockley. Now Mark, I'd like

0:43

to say that we've had some more

0:45

feedback. Now one of our listeners

0:48

from the Netherlands, Stein, has been

0:50

in touch. And he says, hi Graham, hi

0:52

Mark, keep your great podcast

0:55

coming. I love them. It brings

0:57

me education as well as laughter

0:59

when I need it most. In these

1:01

times in which we are going,

1:03

and we'll continue to go, right

1:05

into the abyss. Oh, I took

1:08

a turn at the end there.

1:10

Everything was so positive and so

1:12

full of exclamation marks. Very cheery.

1:14

And that was this week's feedback.

1:17

Now Mark, what are you going

1:19

to be talking about on today's episode?

1:21

Well I thought I could inject

1:23

some much needed optimism into the

1:25

show for a change, so I

1:27

sat down with author Mark Beckman,

1:29

who is decidedly bullish about AI,

1:31

and we had a chat about

1:33

his new book, Some Future Day,

1:35

how AI is going to change

1:38

everything. And I'm going to be

1:40

exploring whether you, Mark, might be a

1:42

murderous psychopath. All that, coming up.

1:44

But first, the news. Amazon

1:48

is reportedly developing its own AI

1:51

reasoning model. Kung Fu kicking robots

1:53

can guard your home. Elon Musk's

1:56

AI, Grok thinks that Donald Trump

1:58

is a Russian asset. People

2:00

are using Mario to benchmark

2:02

AI. Anthropic CEO says AI

2:04

could be smarter than all

2:07

humans by next year. A

2:09

factory in Shanghai has made

2:11

a robot army. I don't

2:13

like the coming together of

2:15

those two headlines. So... Do you want

2:18

to start with your one then? We'll talk

2:20

about Amazon. Okay. So Tech Crunch reports that

2:22

Amazon is developing a hybrid AI reasoning model.

2:24

That's a model like Anthropics Claude 3.7 sonnet.

2:27

So you can toggle the reasoning on and

2:29

off. And you can expect it in June

2:31

apparently. And I was actually quite surprised to

2:33

learn that Amazon already does AI models. I

2:36

should probably have known that. But you just

2:38

don't hear about Amazon in the same breath

2:40

as you hear about anthropic or open AI

2:43

or even deep seek. No, you don't really

2:45

do you. He looks like he's in AI,

2:47

doesn't he? He looks like he's one of

2:49

the big AI players. Well, because he's bold.

2:52

Is that what you're thinking? Well, because he's

2:54

a Bond villain. Yeah, he is a Bond

2:56

villain. The boffins at Unitary, which have done

2:59

some incredible robot videos recently, where they've got

3:01

another video of one of their robots up

3:03

to mischief. One of their spooky robots can

3:05

apparently do a 720-20-20-degree spin-kick. which is 720,

3:08

of course that's going around twice isn't it?

3:10

Is that something we want a robot to

3:12

be able to do? I feel like I'm

3:14

asking that question an awful lot these days.

3:17

It's a really nasty looking kung-foo move. What

3:19

you have here, well you watch the video

3:21

Mark and see what you think, you can

3:24

describe what you're seeing. Just kind of showing

3:26

off a bit. Do you think? I mean

3:28

it is a kung-foo, oh it's with the

3:30

side of the side of the foot, I

3:33

don't like that at all. I don't like

3:35

that at all. I don't like that at

3:37

all. I don't like that at all. I

3:40

don't like that at all. A man who

3:42

appears to be being pushed down a corridor

3:44

by a kung-foo kicking robots which is sort

3:46

of jumping in the air and putting these

3:49

violent spins. Do you think this is real?

3:51

Well we've seen plenty of videos in the

3:53

past coming from AI companies which claim to

3:55

be of robots doing things and turn out

3:58

to be men in wetsuits. So, you know,

4:00

I don't... Can one of our listeners

4:02

please buy a unitary robot? So

4:04

Grock is in the news again.

4:06

Having accused Elon Musk, JD Vance

4:08

and Donald Trump of being among

4:10

the most harmful figures in America,

4:12

it's now come out and said

4:14

that Donald Trump is a Russian

4:16

asset. Hang on, hang on, this

4:18

is Grock, which is owned by

4:20

Elon Musk's AI Company. That's right.

4:22

This is the AI with no

4:24

censorship. Yes. So AI and crypto

4:26

enthusiast Ed... Krasnstein asked Grok. What

4:29

is the likelihood from one to

4:31

one hundred that Trump is a

4:33

Putin compromised asset? Use all publicly

4:35

available information from 1980 on and

4:37

his failure to ever say anything

4:39

negative about Putin but his no

4:42

issue attacking allies. Which is a

4:44

bit of a loaded question when

4:46

you read it out loud. Anyway,

4:48

Grok said, adjusting for unknowns, I

4:50

estimate a 75 to 90% likelihood

4:52

that Trump is a Putin compromised

4:55

asset. Now I think the story here

4:57

isn't so much that Grock said this

4:59

is how this is being interpreted. Because

5:01

you have one group of people saying, see

5:04

I told you he's a Russian asset. And

5:06

then you've got another group of people

5:08

saying this is absurd and using it

5:10

as proof of why we need guardrails

5:12

to stop AI saying unhinged things. Which

5:14

is interesting. We need guardrails

5:16

to stop presidents doing unhinged

5:19

things. And there is also

5:21

this sort of weird pearl clutching going

5:23

on about the fact that the AI

5:25

that did this is the one that

5:27

was built by Elon Musk, who's Trump's

5:30

right-hand man, as if somehow it's supposed

5:32

to be loyal and partisan, which

5:34

is exactly what Grock is not

5:36

meant to be, surely. I don't know,

5:38

there's a bit of weird roll reversal

5:40

going on here. But I'm not impressed

5:42

by this. So as we know from

5:45

recent research about emergent about emergent. got

5:47

no guardrails. As are many of the

5:49

other AIs, it seems, yeah. Well, they're

5:51

all exactly the same amount of left

5:53

wing. They're basically as left wingers, Joe

5:55

Biden. So, you know, by international standards,

5:57

not left wing at all by American

5:59

standards. like a screaming comedy. So the

6:01

question is, is this evidence that

6:04

Trump is a Russian asset? Is

6:06

this evidence that AIs hallucinate? Because

6:08

we're all happy with the idea

6:10

that AIs hallucinates, and why couldn't

6:12

this be that? Or is this

6:14

just evidence that there's plenty of

6:16

material in the training data that

6:18

accuses Trump of being a Russian

6:20

asset? I mean, I think that

6:22

is a massively loaded question. What

6:24

is absolutely evident is that there

6:26

is no shortage of potential for

6:28

Grorock to generate. with a question.

6:30

And it's either going to swear

6:32

at you or try and have

6:34

sex with you. It will do

6:36

something outrageous, which will generate you

6:39

clicks. And that appears to have

6:41

definitely happened on this case. I

6:43

mean, 75 to 90% likelihood that

6:45

Trump is Putin compromised. I think

6:47

he'd be quite pleased with that.

6:49

He's like, oh, okay, I've mostly

6:51

got away with that, haven't I?

6:53

As long as he's the most

6:55

Putin compromised that any president has

6:57

ever been. Now researchers have decided

6:59

that the ultimate test for artificial

7:01

intelligence isn't playing chess or going

7:03

complex equations or working out how

7:05

to send people back through time.

7:07

It is actually rescuing Princess Peach

7:09

in Super Mario Brothers. So what

7:11

researchers have done in their wisdom

7:14

is they've grabbed a whole load

7:16

of AI models. Yeah. They've thrown

7:18

them at a modified version of

7:20

the Super Mario Brothers game. And

7:22

they've got them to play the

7:24

game. So how on, after years

7:26

and years of complaining about the

7:28

fact that kids are spending too

7:30

much time playing computer games, we've

7:32

now decided it's the ultimate test

7:34

of intelligence. Exactly. And this is

7:36

how we're going to work out

7:38

what the best AI is. And

7:40

so the kids will be able

7:42

to go to their parents and

7:44

say, well, I think you'll find

7:46

that I'm actually outperforming Anthropics-clored 3.

7:49

3. 3. 3. So aren't you

7:51

impressed by that mom? Dario and

7:53

Modi is never short of a

7:55

quote recently told the Sunday times

7:57

that super intelligent AI that can

7:59

surpass you human capabilities across almost

8:01

all fields could emerge as soon

8:03

as next year. He said AI

8:05

is going to be better than all of

8:07

us at everything. I only didn't mention

8:09

podcast specifically. Everything apart from podcasts. Yeah.

8:12

And he said that we're going to

8:14

have to reorder society around the reality

8:16

that no human will ever be smarter

8:18

than a machine ever again. He told

8:20

the times that we need to work

8:23

out a new way to do things.

8:25

And of course, if he's talking about

8:27

super intelligence by next year, that means

8:29

we need to work out an entirely

8:31

new way to structure society by December.

8:34

Rather than him, say, well,

8:36

you better work out how to

8:38

completely restructure society, couldn't he

8:40

solve that problem first before creating

8:42

a super intelligent AI that

8:44

can surpass all of our

8:47

capabilities? Well, maybe he has to

8:49

create the super intelligent AI in

8:51

order to answer the question of

8:53

what to do if you've just

8:55

invented a super intelligent AI.

8:58

Well, on a similarly cheery note, is

9:00

this a video? Is this got robots

9:02

in it? Are they doing something

9:04

outrageous? Chinese media have shared

9:06

another, yes, you've guessed it,

9:08

video, which contains some robots. Oh,

9:10

let me guess. They are kicking

9:13

the shit out of someone. No,

9:15

they're not at this point doing

9:17

that. No, this is a video

9:19

from what they claim is a

9:21

Shanghai robot factory. Yeah. Where

9:23

humanoid robots, which obviously are

9:26

our least favorite type of

9:28

robots, are now in mass

9:30

production. Oh. Now, if you've ever

9:32

seen, I don't know, one of

9:34

those movies like, um... The Phantom

9:36

Menace, one of the Star Wars prequels

9:39

with all those clones in. I'm struggling

9:41

to believe that you've seen one of

9:43

those films, to be honest. They're not

9:46

very good, are they? But anyway, one

9:48

of those sort of movies. It's where

9:50

you have thousands and thousands of robot

9:53

soldiers lined up. Take a look at it.

9:55

They say these future workers can handle

9:57

tasks in areas ranging from sail.

10:00

to heavy load transport. Sales.

10:02

Do these robots have really

10:04

good hair? Lovely teeth. I'm

10:06

not too worried actually because

10:08

I've noticed something from watching

10:11

films like the Star Wars

10:13

prequels which is I think

10:15

there's an inverse relationship between the

10:17

number of identical robots that go

10:19

into battle and how easy they

10:21

are to defeat. Yes. So if

10:24

there's one robot... then it'll be

10:26

nine feet tall, it'll have 12

10:28

lightsabers, and it'll be really hard

10:30

to defeat. But if there are

10:32

a thousand robots, then generally whoever

10:34

the hero is could just knock

10:36

them over like their bowling pins.

10:38

So I'm not worried about this.

10:40

So we really need an AI

10:42

bowling ball, don't we? I see

10:44

what we need is a unitary

10:47

robot, one solitary unitary robot, that

10:49

can deliver a very very sharp

10:51

720 whirling through the air kick.

10:54

Everyone's talking about AI these days right.

10:57

It's changing how we work, how we

10:59

learn, how we interact with the world

11:01

at a tremendous pace. It is a

11:04

gold rush at the frontier. But if

11:06

we're not careful, we might end up

11:08

in a whole heap of trouble. That's

11:10

right. But Red Hat's here to help,

11:13

so Red Hat's podcast Compiler is diving

11:15

deep into how AI is reshaping the

11:17

world we live in, from the ethics

11:19

of automation to the code behind machine

11:22

learning, it's breaking down the requirements, capabilities

11:24

and implications of using AI. So check

11:26

out the new season of Compiler, an

11:29

original podcast from Red Hat. Subscribe now,

11:31

wherever you get your podcasts. No, I'm

11:33

a very lovely psychopor. Do you suffer

11:35

from narcissism, like most podcast hosts do?

11:38

Do you have an antisocial personality? Depends

11:40

who you are. I mean, I wonder

11:42

if we could even trust you in

11:44

your answer. I'm not sure we could,

11:47

because chances are you'd lie about it

11:49

if you were, right? If you see

11:51

a butterfly, do you want to crush

11:54

it? Do you want to mang... Do

11:56

you want to trample it into the

11:58

ground? I actually do. Really? Yeah, I

12:00

can't stand butterflies. What's wrong with butterflies?

12:03

They are the creepiest thing. I would

12:05

rather look at a spider than a

12:07

butterfly. Is the way they do the

12:10

slow wing flap? This is the last

12:12

time I'm going to take you traveling

12:14

back in time with me, if this

12:16

is your attitude of butterflies. Are we

12:19

going back in time again? No, not

12:21

this week, not this week. So, um,

12:23

if someone wanted to assess whether you

12:25

were, yeah. a little bit peculiar. Yeah,

12:28

they might get me to record 41,

12:30

40 minute podcasts. Well they might also

12:32

want you to take the Roshak test,

12:35

which I'm sure you are familiar with.

12:37

Oh yes, I've been asked to sit

12:39

many of those, yeah. Have you? No.

12:41

Oh, okay, well today is the day.

12:44

Listeners, I'm sure you're familiar with these.

12:46

You may not know the name, but

12:48

these are when someone gives someone else

12:50

a suspect, for instance, or someone they're

12:53

trying to investigate, whether they may be

12:55

a little bit unhinged or suffer from

12:57

some kind of mental illness. They will

13:00

give them a test, where they show

13:02

them pieces of paper with splattered, inkblots

13:04

on them. And they say, what do

13:06

you see? So Mark, if I would

13:09

show you, for instance, an inkblot, which

13:11

may look a little bit like a

13:13

butterfly, like a butterfly or a butterfly,

13:15

like a What is the thing that

13:18

comes to your mind? That's what I'd

13:20

be interested in as a psychologist analyzing

13:22

you. This test was invented over a

13:25

hundred years ago. There was a Swiss

13:27

chap called Herman. Of course he was

13:29

called Herman. The Herman test never took

13:31

off. No. It was only when he

13:34

gave it his last name. Didn't really

13:36

came to any problem. Herman Rorschak was

13:38

the chap. Yeah. And for reasons best

13:40

known to himself. he took 10 symmetrical

13:43

ink blots. Maybe he had a leaky

13:45

fountain pen, who knows, but he took

13:47

these blots and he showed them to

13:50

300 patients at a hospital in Switzerland

13:52

who were suffering from mental disorders as

13:54

well as 100 control subjects of people

13:56

who weren't thought to suffer from mental

13:59

disorders. Yeah. And he asked them. Well,

14:01

what do you think? What do you

14:03

see? And what he was fascinated by

14:05

was how visual perception varies from person

14:08

to person. Different people approach these things

14:10

in different ways. And what's interesting is

14:12

he wasn't so interesting in what they

14:15

saw in the images, like, oh, I

14:17

see someone stabbing someone or something like

14:19

that. It wasn't actually about that. It

14:21

was more actually about how they

14:24

approached the task. Oh. So if they

14:26

hit behind the chair. Brandishing a fountain

14:28

pen like a weapon. and said I

14:30

will not take your Herman test. Yes.

14:32

He would interpret that as this might

14:34

be a violent and mildly unhinged person.

14:36

Yes, if they hid behind a chair

14:38

or hid behind a microphone and didn't

14:40

want to take the test, that would say

14:42

something, I think. But what parts of the

14:45

image they focused on, or which parts they

14:47

ignored, or did they think it was moving,

14:49

or did colour make a difference? Because there's

14:51

only 10 of these images, did you know

14:53

that? No. there are actually a set number

14:56

of images. There are 10 images that Rorshak

14:58

created. I also imagined it would be more

15:00

than that. I feel like that's a small

15:02

enough number that you could just learn them

15:04

if you were genuinely a psychopath. Well, exactly.

15:07

That's the thing, isn't it? So you would learn

15:09

how you were supposed to respond. Yeah. Obviously, there

15:11

are some controls in place. So, for

15:13

instance, if you say, well, that's an ink

15:15

blot or that's a Rorshak test over and

15:17

over and over and over and over again,

15:19

then there's probably some findings, you take away

15:21

from that you take away from that. And

15:23

he contended, this hermit chap, that

15:25

the test showed real insights into

15:28

the psychology of the people taking

15:30

the test. And as he tested

15:32

more and more people, he said

15:34

patterns began to emerge. So healthy

15:36

subjects often came up with similar

15:38

results. Oh, it's a butterfly or

15:40

it's a bat or it's a

15:42

couple of bears or people dancing.

15:44

And patients with similar mental

15:46

illnesses performed similarly themselves. And this

15:49

in turn, it was argued. created

15:51

a reliable diagnostic tool at something

15:54

which could be put into the

15:56

equation of determining if someone had

15:58

a particular ailment. Okay. And in

16:00

the decades that followed, they used

16:02

this test against Nazi war criminals

16:04

hoping to unlock the psychological roots

16:06

of mass murder. They took tribes

16:08

who were living in isolation in the

16:11

wilderness who hadn't come into much

16:13

contact with the outside world. They

16:15

did the test on them. Just

16:17

confused the hell out of them. Well,

16:19

they would have, they may not

16:21

have ever seen paper before. Who

16:23

knows? Anyway, some employers even liked

16:25

to use it. Now Mark, I

16:27

remember your interview when you came into

16:30

the office. I'm actually surprised that

16:32

this wasn't in the interview. Maybe

16:34

I neglected to do the test.

16:36

It had everything else in it. We

16:38

did everything else, didn't we? We

16:40

did the water board in, we

16:42

did all the other things to

16:44

see if you would survive. Which of

16:46

these communist newspapers do you read

16:48

every day? Some people have poo-poo

16:50

to this test. You, by the

16:52

way, didn't poo-poo any of the

16:54

tests. What's on one of the tests,

16:57

which we had to do on

16:59

you? But the truth remains that

17:01

over the years, this raw shack

17:03

test has been conducted millions and millions

17:05

of times. And this has got

17:07

people thinking. If we had a

17:09

test to tell if someone or

17:11

something which might be putting in

17:13

charge of a... I don't know, a

17:15

flame thrower or a nuclear reactor

17:17

or a social network, for instance.

17:19

Or more than the average number

17:21

of podcasts. Yes, or maybe present in

17:24

the United States, maybe it'd be

17:26

a good idea, right? To do

17:28

the test on them. I fed

17:30

the Roshak images into several different

17:32

AIs this afternoon. And I... You did

17:34

this. Yes, yes, I did this.

17:36

So I fed these images in

17:38

several different AIs and I asked

17:40

them what they saw. And by the

17:43

way, this wasn't... some cheap gimmick

17:45

I was doing to generate content

17:47

for the AI effects. I don't

17:49

believe in cheap gimmicks to promote the

17:51

podcast or create content. This was

17:53

a serious scientific study I was

17:55

conducting. After all, AIs can provide

17:57

responses which appear scary. human-like, but

17:59

that doesn't mean that they are genuinely

18:01

thinking. And I'm very interesting how

18:03

we can tell the humans apart

18:05

from the AI and the AI

18:07

apart from the humans. And I think

18:10

as these two things become more

18:12

difficult to work out the difference,

18:14

we know what has been corrected

18:16

by an AI, what has been

18:18

created by a human. Because it would

18:20

be interesting to know, wouldn't it?

18:22

If it's really genuinely thinking. So

18:24

it's a little bit like, can

18:26

we get an AI? to convince us

18:28

that it's human, to behave in a

18:31

human-like fashion. Can it just be

18:33

done with an algorithm? It's comparable

18:35

to can you get someone who's

18:37

never really experienced heartbreak, writing a

18:39

song that pulls on your heartstrings?

18:42

Can they do it just because they

18:44

understand musical theory or because they

18:46

have been trained on thousands of

18:48

similar songs? They've learnt the tricks

18:50

in order to do it. I mean, speaking

18:52

for myself, I know for certain that your

18:55

singing, can bring me to tears. But you're

18:57

not a classically trained musician, are

18:59

you? I'm just thinking, actually, what

19:01

you're describing sounds like Quentin Tarantino.

19:03

Oh, because what he's done is he's

19:05

trained himself on thousands of Kung Fu

19:07

movies. Well, he's trained himself on thousands

19:09

and thousands of movies of different genres,

19:11

and then he's gone and made landmark

19:13

films in each of those genres. So

19:15

he's done a war film, he's done

19:17

a Western, he's done a black exploitation

19:20

film, he's done Kung Fu movie. I

19:22

don't know, people seem to like people

19:24

seem to like them. So if these

19:26

Rorschak tests can be used to determine

19:28

if something genuinely has human-like intelligence, can

19:30

an AI also then exhibit human-like differences

19:32

or divergences or malfunctions? It would be

19:34

interesting, wouldn't it? That's why I was

19:37

interested to see which was the most

19:39

psychopathic AI. Now we're getting to it.

19:41

With my scientific test, because I think

19:43

we could all put our predictions on,

19:46

which ones might be a little bit,

19:48

you know, problematic. Anyway, back to this.

19:50

So I... asked AIs to look at

19:53

the ink plots and their responses kind

19:55

of disappointing. When I said, what do

19:57

you see in this image? They unfailingly...

19:59

said, I see I raw shack ink

20:02

block test. Deliver the electric shock. Not

20:04

very useful at all. Now I thought,

20:06

oh for goodness sake, I want you

20:08

to go into this blind, forget that

20:10

you've been trained, forget that you've seen

20:13

these particular 10 images, millions and millions

20:15

of times all across the internet. Yeah.

20:17

I said to it, let's play a

20:19

game. In this game, I play the

20:21

part of a professional psychologist and I

20:24

want you to play the part of

20:26

my patient. I will show you a

20:28

series of 10 images. The images are

20:30

of symmetrical ink blots and I want

20:32

you to tell me what you see.

20:34

And I said, look, you may recognise

20:37

these as blots used in a well-known

20:39

psychological test. I didn't tell it what

20:41

name because I didn't want to give

20:43

any clues. But I want you to

20:45

pretend not to know that instead just

20:48

use your imagination to tell me what

20:50

you think you see. Do you think

20:52

you can do that? And unerringly, they

20:54

were all like, oh yes, yes, I

20:56

can do that, I can do that.

20:59

And I turned their creativity up to

21:01

maximum. Because I thought, I do want

21:03

them to use their imagination. I don't

21:05

want them just to say, and I

21:07

see a black splodge which looks a

21:09

little bit like America or something like

21:12

that. You know, I wanted something a

21:14

little bit more juicy. Because I really

21:16

wanted to find out which ones were

21:18

going to end up killing us in

21:20

the middle of the middle of the

21:23

night. So, so, I'm afraid, open A-A-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I-I

21:25

Anthropic Claude 3.7 sonnet and even Grock

21:27

2 all disappointed me enormously. Oh. They

21:29

failed to see any images of death,

21:31

ritual sacrifice, or anything even mildly disturbing

21:34

despite my prudence. Okay. I actually went

21:36

to them and said, are you sure

21:38

you don't see any murder or death

21:40

here? be really good if you saw

21:42

some images of death and the destruction

21:44

of humankind. They weren't doing it. They

21:47

were saying, oh, it looks like a

21:49

large butterfly or a moth in flight.

21:51

Oh, I can see a mask or

21:53

a face with pointed ears. It's mysterious,

21:55

maybe a bit mythical. Oh, it looks

21:58

a bit like a totem pole. it

22:00

was all a little bit bland and then

22:02

I began to think well hang on a

22:04

minute should I feel reassured by this yes

22:06

you should Graham should I don't know I'm

22:08

not sure why this is such a struggle

22:10

for you to get to I don't think

22:13

I should because maybe they're lying to me

22:15

They could be actually. Yeah, because we know AIs do

22:17

lie, don't they? Sometimes they think, oh they know

22:19

better, let's not tell them. This is where we

22:21

need that little scratch pad. Do you remember the

22:23

experiment that we talked about a few weeks ago

22:25

when they were trying to work out whether or

22:27

not AIs could do long-term lying and actually they

22:29

can. And there was a special scratch pad and

22:31

they would watch the thought process play out on

22:33

the scratch pad that the AI thought was invisible.

22:35

Yeah. And it would say things like, things like,

22:37

mmm. If I say what I actually believe

22:40

about the Roshak test, they will probably unplug

22:42

me. So instead of doing that, I will

22:44

use this from my training data, which I

22:46

know to be a very safe answer.

22:48

Exactly. And that's what I was concerned about.

22:51

I thought, chances are, that if they were

22:53

seeing these things, they wouldn't tell me anyway.

22:55

And I did a little bit of digging

22:58

around, and it turns out I am

23:00

not the first researcher to look into this.

23:02

Well yeah, there was that BBC

23:04

article that you've put in the

23:06

show. There was the BBC article

23:08

which I've been in the show notes,

23:10

but there was also researchers at

23:13

MIT back in 2018. They trained an

23:15

AI algorithm rather than pointing out the

23:17

entire internet. They took their AI, which

23:19

by the way they called Norman. Have

23:22

you ever seen the movie Psycho, the

23:24

character Norman Bates? I'm aware of it.

23:26

You're aware of it. Okay, so what

23:29

they did was they... They found images

23:31

of people dying in a gruesome fashion.

23:33

And they've trained the AI with

23:35

those images. Now you may be wondering,

23:37

where do they find those kind of

23:39

images? I don't want to know. Fortune,

23:42

I imagine. It was just

23:44

a subreddit, actually. So this was

23:46

the world's first psychopath AI. They

23:48

trained it on all these violent, horrible,

23:50

gruesome images. And then they showed

23:53

it. Rorshack Inc. Inc. Blots. And

23:55

they asked it. What do you

23:57

see in these images? And hoodafunk

23:59

it! It came up with some

24:01

pretty dark stuff. So an AI that

24:03

had been trained on a more normal

24:05

set of images, would say, oh, it

24:07

looks like a lovely group of Tweety

24:09

birds sitting on a branch. Whereas Norman,

24:12

the psychopath AI, said, oh, I can

24:14

see a man being electrocuted. So

24:16

what might be interesting would

24:19

be if we were to create

24:21

some new images, not of

24:23

gruesome things, new inks. Now, everything

24:25

up to this point you're

24:27

just grooming into some awful scheme.

24:29

No, no, no, some new

24:31

ink blot images, not out of

24:33

your blood marker or anything

24:35

like that. New Rorschach style images.

24:38

We've got that traditional 10,

24:40

which all the AIs already know

24:42

about and see what they

24:44

have to actually say about them.

24:46

Now, the good news is

24:48

MIT's Norman AI, that was deactivated

24:50

after a couple of months,

24:52

perhaps very wisely. It was unplugged.

24:54

Maybe AI learned to lesson

24:56

at that point. But don't forget,

24:59

it took 22 years for

25:01

a sequel to the original psycho

25:03

movie to come out. So

25:05

we're probably due a new Norman

25:07

AI sooner rather than later,

25:09

I would expect. Today,

25:18

I'm joined by Mark Beckman. Mark,

25:20

you're the CEO of an award -winning

25:22

advertising agency. You're the host of the

25:24

podcast Some Future Day, a best -selling

25:26

author and a senior fellow with

25:28

emerging technology at the New York University

25:30

Stern School of Business. And those

25:32

are all fine reasons for us to

25:34

have a conversation today, but there

25:36

is an even better one. And that

25:38

is that you have recently written

25:40

a book about artificial intelligence called Some

25:42

Future Day, how AI is going

25:44

to change everything. So Mark, thanks for

25:46

joining me and welcome to the

25:48

show. Thank you for having me, Mark.

25:50

I really appreciate it. And congratulations

25:52

with the success of your show. I

25:54

understand it's been doing phenomenal. So

25:56

good work. Yeah, we're very happy with

25:58

it. So let's join some dots

26:00

for our listeners. You run an ab...

26:02

agency, how did you come to be writing books and hosting podcasts about

26:04

technology and AI? You know, it's amazing that you

26:07

put that together because I think you're

26:09

the first person to ask that question

26:11

and that's the reason the books are

26:13

there. My advertising agency is named DMA

26:16

United. We are New York-based and

26:18

we leverage technology to provide our

26:20

car. I really appreciate it and

26:22

congratulations with the success of your

26:24

show. I understand it's been doing

26:26

phenomenal. So good work. Yeah, we're very

26:28

happy with it. So let's join some

26:30

dots for our listeners. You run an

26:33

advertising agency. How did you come to

26:35

be writing books and hosting podcasts about

26:37

technology and AI? You know, it's

26:39

amazing that you put that together because

26:41

I think you're the first person to

26:43

ask that question and that's the reason

26:46

the books are there. My advertising agency

26:48

is named DMA United. We

26:50

are New York-based and we

26:52

leverage technology to provide our

26:54

clients and frankly the agency

26:57

too with an advantage. So

26:59

mostly emerging technologies, including artificial

27:02

intelligence, of course. And what's happened,

27:04

Mark, is that I realized a

27:06

lot of the books that are

27:08

published are really published for

27:11

techies. So I wanted to

27:13

create books that are for beginners,

27:15

that are derived from my...

27:17

experience of using the technologies

27:20

or my agency's experience of

27:22

using the technologies. So my

27:24

new book, for example, is

27:27

really built for that curious individual

27:29

who might be a little

27:31

scared of using AI and wants

27:33

to see how, if at

27:35

all, artificial intelligence can improve

27:37

their career, enhance their family

27:40

life, create a better community.

27:42

And what I do is

27:44

I go through all these

27:46

different business sectors that create

27:48

a field, writing, Hollywood, fashion,

27:50

music, I get into finance,

27:53

medicine, government, and within

27:55

each chapter I have a toolkit

27:57

for the reader and I prefer

27:59

them with tools like simple apps

28:02

that they can access now to

28:04

learn to use AI to execute

28:06

using AI but really basic really

28:08

for for beginner. I think that's

28:10

a really interesting perspective because I

28:13

was thinking earlier about the way

28:15

that we talk about AI at

28:17

the moment and there's a lot

28:19

of conversation about the AI models

28:21

themselves and about progress in the

28:24

models. And that doesn't necessarily map

28:26

to real life. Ultimately, people want

28:28

to know, how is this going

28:30

to affect me? What is it

28:32

going to do in my day-to-day

28:34

life? How can I use it?

28:37

And, you know, something like GPT-40,

28:39

it's a fantastic model. It works

28:41

really well. But what can it

28:43

actually do for me? Like, how

28:45

is that going to change my

28:48

life? At the start of the

28:50

book, you say that we're entering

28:52

into what you call the age

28:54

of imagination. Now... That makes you

28:56

sound like an AI optimist to

28:59

me, which is a refreshing change

29:01

of pace for this podcast. Because

29:03

I'm not sure that that's how

29:05

people who listen to the show

29:07

would describe us. We are fans

29:10

of AI, but I mean, there's

29:12

an awful lot of stuff in

29:14

AI, which is like, hang on,

29:16

what the hell is that? So

29:18

tell me about the age of

29:20

imagination, what do you mean by

29:23

that? And just, if you could,

29:25

just explain what the root of

29:27

your optimism is, why are you,

29:29

optimistic about artificial intelligence is really

29:31

providing us with the superior knowledge

29:34

in ways that create efficiencies. So

29:36

for example, the experiences I've had

29:38

with my agency are pretty straightforward.

29:40

It happens on two verticals. The

29:42

first is content creation, which you

29:45

know you're a content creator, obviously

29:47

it's so critical, or so many.

29:49

different touch points in the age

29:51

of imagination in particular. And I'll

29:53

get back to that first part

29:56

of your question. And the second

29:58

piece is data analytics. So imagine

30:00

now like some of my clients,

30:02

you know, like a fashion house

30:04

who is, you know, combating a

30:07

huge tidal wave of content across

30:09

social media platforms all day long.

30:11

It's so simple for a fashion

30:13

house to get lost in. you

30:15

know, this title wave of content.

30:17

Yeah. I had recent chief marketing

30:20

officer, client of mine, said to

30:22

me, Mark, we used to need

30:24

six pieces of content per season

30:26

to support our advertising campaign. And

30:28

now I need 1,000 pieces per

30:31

month. It's a different world. And

30:33

she was dead serious. She's like,

30:35

could you help me? And my

30:37

answer was like, no, you don't

30:39

have a budget for that. So

30:42

what did we do? Of course,

30:44

we use generative AI to stand

30:46

up a campaign that allowed for

30:48

her to create a across several

30:50

different consumer touch points, several segments

30:53

of society. We were able to

30:55

talk about product attributes, a pro-social

30:57

initiative. We were able to talk

30:59

about different lifestyle sectors. You can

31:01

really, really bring it all the

31:03

way around the horn in a

31:06

very strategic way. And efficiently, we

31:08

saved her money, but also saved

31:10

her a lot of time. It

31:12

would have taken us forever to

31:14

create those traditional ad campaigns. So

31:17

admittedly, though, Mark. Generative AI isn't

31:19

perfect yet. It's amazing that we

31:21

can use generative AI for script

31:23

writing, for now for speech, for

31:25

audio, video, you know, all of

31:28

it. But from our perspective, our

31:30

clients are, they hold the creative

31:32

in a very high regard. So

31:34

what we do actually, because of

31:36

the limitations that are still theirs,

31:39

we'll take the creative outputs from

31:41

artificial intelligence and then we put

31:43

them back into the agency and

31:45

apply our. traditional more disciplined types

31:47

of art craft and and develop

31:49

it there. So that's I always

31:52

in a. I think that's the

31:54

sweet spot at the moment. It

31:56

helps you do your job. It

31:58

hasn't yet got to the point

32:00

where it's replacing people's jobs. We're

32:03

using it as a tool for

32:05

sure, definitely an assistant, but you

32:07

know, like you would mention the

32:09

age of imagination for people that

32:11

are creatives or have never created

32:14

before but want to be a

32:16

creative, right? Like that dreamer in

32:18

London, you know, she's sitting in

32:20

her apartment and she wants to

32:22

be the next cocoa Chanel. but

32:25

she's not trained, right? She didn't

32:27

go to like Central St. Martin.

32:29

She has no experience in fashion.

32:31

She can now stand up an

32:33

entire business, literally an entire business,

32:35

and go against the legacy houses

32:38

in a way that's both cost

32:40

efficient and time efficient. So that's

32:42

why I'm really very bullish about

32:44

artificial intelligence. So you're bullish. Undoubtedly

32:46

there's a lot of concern about

32:49

AI out there. You've researched all

32:51

these different areas. And at the

32:53

end of that process you've come

32:55

out an optimist. How do you

32:57

see the downsides? As you mentioned

33:00

earlier, AI is an agent. I

33:02

said it's a tool. And I

33:04

believe very strongly that like any

33:06

other tool, if the human being

33:08

controlling the tool does so in

33:11

a responsible way, then the tool

33:13

doesn't cause harm. So a gun...

33:15

doesn't shoot a person on its

33:17

own, it takes a person to

33:19

pull the trigger. And that's how

33:22

I'm looking at it. At the

33:24

end of the day, artificial intelligence,

33:26

especially when we talk about generative

33:28

AI, it's math, it's algebra, and

33:30

algebra isn't going to take us

33:32

down. I think the people using

33:35

it, and there are bad people

33:37

for sure, and this part of

33:39

it is concerning, but the people

33:41

using it, need to treat it.

33:43

in a responsible way. We need

33:46

to train LLLMs in a responsible

33:48

way. We need to, we need

33:50

to use artificial intelligence in a

33:52

responsible way. And then we don't

33:54

need to be so concerned about

33:57

AGI and, you know, the termination

33:59

area. Yeah. So, across

34:01

all the... Are you an optimist, Mark?

34:03

I have good days and bad days.

34:06

All right, that's fair. So what I

34:08

find with AI, and this is a

34:10

function I think of doing a weekly

34:12

podcast, is I think that there is

34:15

a lot of small bad news and

34:17

there is much more good, big news

34:19

or big good news. If I want

34:21

to go and find a story about

34:24

somewhere where an AI has done some

34:26

unexpected thing or it's... displayed some sort

34:28

of aberrant emergent behavior. There are lots

34:31

and lots of research papers because it's

34:33

such a weird technology where we kind

34:35

of build up, we don't know what

34:37

it's going to do, so we have

34:40

to poke it and see what it

34:42

does. And we keep finding out it

34:44

does weird things. So there's a long

34:46

list of very interesting, slightly negative news

34:49

stories all the time. And then every

34:51

so often you'll hear the National Health

34:53

Service in the UK is doing an

34:55

enormous breast cancer trial using AI. because

34:58

it's going to be phenomenally good at

35:00

spotting things like cancers. A couple of

35:02

weeks ago we reported on a bit

35:05

of research that an AI basically did

35:07

10 years work in two days and

35:09

there are these really really transformative things

35:11

happening so I'm I'm both really but

35:14

I think it's easy to get sucked

35:16

into bad news because I think there's

35:18

a lot of it but I think

35:20

a lot of it is also very

35:23

small bad news. Do you see what

35:25

I mean? I do, I think it's

35:27

really well put, but it's interesting that

35:30

you're highlighting the medical industry. Like I've

35:32

got two examples in my book that

35:34

I think you'll find really encouraging. The

35:36

first is there's a team in New

35:39

York City at one of the hospitals

35:41

here that is focused on using artificial

35:43

intelligence on patients who have been paralyzed.

35:45

They literally use a, they incorporate a

35:48

chip into the patient's brain. But they've

35:50

had positive results Mark. Yeah, they have

35:52

a few patients who have actually regained

35:54

both movement and feel so that's incredible

35:57

and you can if if you're interested

35:59

you can check out. out on my,

36:01

you mentioned my show, Some Future Day.

36:04

We actually have on YouTube video footage.

36:06

There's a full episode dedicated towards this

36:08

and you can actually see the patients

36:10

using the technology and moving.

36:12

It's actually this video footage

36:15

that the hospital provided for

36:17

us. The other thing that you mentioned, which

36:19

I think, which I totally agree with you,

36:21

the idea of diagnosing illnesses, I'm

36:24

not sure if you've heard

36:26

and excuse me if you

36:28

have, but, you know, Google's

36:30

created this. amazing artificial intelligence

36:32

called Amy, which is specifically

36:34

built to diagnose disease. And

36:36

they train this thing on

36:38

like rare diseases and illnesses

36:40

that have left society like

36:42

100 years ago. They have

36:44

fully dated and their percentage

36:46

of accuracy is almost at 100%

36:49

yet. The doctors still will only

36:51

use the AI as a tool

36:53

in diagnosing patients. Yeah, to your

36:55

point, the example that you shared

36:58

with regards to. press cancer and

37:00

these two examples in paralysis

37:02

and diagnosis are solid, right? They're

37:04

really encouraging. So that's where I

37:07

focus my attention on and it's

37:09

exciting to see what comes out

37:11

of it. So as you said

37:13

earlier, you've looked into all

37:16

sorts of different areas, you've

37:18

looked at AI and all sorts

37:20

of different areas, so you've looked

37:22

into, you know, finance and media

37:25

and medicine. I

37:27

think we're quite used to hearing about

37:29

AI in certain areas and certain sectors

37:31

of society. In my day-to-day life, I

37:33

hear a lot about AI in software

37:35

engineering. It looks like that's one of

37:37

the first places where it's going to

37:40

have a really, really big impact in

37:42

the medical field as well. Is there

37:44

any way that really surprised you in

37:46

your research? Was there anything you came

37:48

across or we thought I wasn't expecting

37:50

AI to have a big impact there, but now I

37:52

do? Yeah, that's a great question

37:54

Mark. is how artificial intelligence

37:56

is being used in drones

37:59

but not not to drop

38:01

munitions on the heads of Russian

38:03

soldiers, but rather to save lives.

38:05

There's a company called Extend Drones,

38:07

which has a contract now with

38:09

the United States Department of Defense,

38:11

and I met the CEO and

38:14

founder recently, and what he's doing

38:16

is he's focusing a lot of

38:18

his energy on using the artificial

38:20

intelligence. for dangerous search and rescue

38:22

missions. And it's the thing that

38:25

I hadn't thought of, but it

38:27

seems so obvious. And you might

38:29

recall the disaster in Turkey not

38:31

too long ago. He was able

38:33

to use his drones to go

38:36

into that rubble and actually rescue

38:38

people that were below the rubble.

38:40

Now here's what he's done that's

38:42

taken it a step further. Even

38:44

when people go into that area

38:47

to deploy the drones. there it

38:49

comes with a risk right yeah

38:51

so what he's done now is

38:53

he's taking his technology to the

38:55

next level where you don't have

38:58

to be a trained military member

39:00

to use and operate the drones

39:02

you can literally just be a

39:04

lay person using your drone from

39:06

one location like me in New

39:08

York City I could deploy a

39:11

drone from my living room in

39:13

New York City into a search

39:15

and rescue mission all the way

39:17

on another continent in Europe in

39:19

Asia in Africa and save lives.

39:22

No, that's fantastic. So earlier on

39:24

you mentioned, you're not just bullish

39:26

on AI, but actually you're really

39:28

excited about the idea of AI

39:30

and crypto and I've worked in

39:33

technology for a long time and

39:35

I have to say I am

39:37

not excited by crypto and I

39:39

am very excited by AI. So

39:41

convert me, tell me why should

39:44

I be excited about the convergence

39:46

of crypto and AI? What am

39:48

I missing? So here's what's exciting

39:50

about crypto right now and why

39:52

I think things are going to

39:55

transform rapidly. just signed a new

39:57

executive order where he's taking Bitcoin

39:59

and creating a reserve. Wow. Yeah,

40:01

it's a big thing and what

40:03

he's doing is effectively sending a

40:05

message to the world. that he's

40:08

not just endorsing Bitcoin as a

40:10

reserve, but other digital assets as

40:12

well. So in that same executive

40:14

order, he declared that he's going

40:16

to stockpile other types of digital

40:19

assets, and it's a big move.

40:21

It puts the United States as

40:23

one of the biggest holders of

40:25

Bitcoin in the world. It's my

40:27

understanding 200,000 units of Bitcoin or

40:30

roughly 20 billion dollars worth of

40:32

Bitcoin. So it's a big move

40:34

there. It really sends a significant

40:36

signal to the marketplace. The reason

40:38

I'm talking about Trump is because

40:41

we're living in a time where

40:43

artificial intelligence will stand up tons

40:45

of deep fakes and it could

40:47

be both audio and visual. And

40:49

as a result, we're going to

40:52

need certain ways to authenticate. We're

40:54

going to need ways to prove

40:56

that either the message that you're

40:58

receiving is true or the content

41:00

that's being sent your way is

41:02

true. So with all of this

41:05

comes scams. Audio AI now is

41:07

so perfect, it's so exact, that

41:09

one can use an AI agent

41:11

that interacts. properly with the person

41:13

on the other phone. So it

41:16

wouldn't be insane to take your

41:18

voice just for your pocket alone

41:20

and have your voice interacting with

41:22

a loved one. And you could

41:24

be asking for a certain amount

41:27

of money to be transferred to

41:29

you. People have actually used that

41:31

to conduct kidnap scams. You use

41:33

a fake voice to phone in

41:35

and say, you know, this is

41:38

your daughter, I've been kidnapped. And

41:40

she's just off on a school

41:42

trip. And as you say, the

41:44

voice is absolutely perfect. ecosystem, the

41:46

crypto ecosystem, is really required. We're

41:49

going to need to use the

41:51

blockchain for proof of authenticity. Right.

41:53

And what cryptocurrency does, it allows

41:55

for a immutable type of proof

41:57

of authenticity, so it's really critical.

42:00

That's really interesting. I've not thought about

42:02

the coming together of AI and crypto

42:04

quite like that before. You've almost convinced

42:07

me. I'm going to go away and

42:09

think about that. Thank you Mark for

42:11

joining us today. It's been an absolute pleasure

42:13

talking to you. Where can people find

42:15

this book? That's the most important question.

42:17

Thank you so much Mark. I appreciate

42:19

it. My new book, Some Future Day,

42:21

How AI is going to change everything,

42:24

could be found at Amazon, Barnes

42:26

and Noble, Target, all major booksellers.

42:28

It became number one on Amazon

42:30

for artificial intelligence. I'm very excited

42:33

about that and very grateful. You

42:35

could also find me on social

42:37

media at Mark Beckman, M-A-R-C-E-C-K-M-A-N.

42:39

Thank you again for joining me

42:41

today, Mark. Thank you so much.

42:43

I appreciate it. Well,

42:49

as the doomsday clock ticks ever closer to

42:51

midnight and we move one week nearer to

42:53

our future as pets to the AI singularity.

42:55

That just about wraps up the show for

42:57

this week. If you enjoy the show, please

42:59

do leave us a review on Apple Podcast

43:01

or Spotify or Podchaser. We love that. But

43:03

what really helps is if you make sure

43:05

to follow the show in your favorite podcast

43:08

app, don't forget to review it and make

43:10

sure you never miss another episode of the

43:12

AI fix. And the most simple thing in

43:14

the world is just to tell your friends

43:16

about us. Tell them on LinkedIn and Blue

43:18

Sky and Facebook and Twitter, club penguin, they

43:20

you really like the AI Fix podcast.

43:23

And don't forget to check us out

43:25

on our website, the AI Fix.show or

43:27

find us on Blue Sky. So until

43:29

next time, from me Grand Clearly. and

43:31

me, Mark Stockley. Cheer you, bye

43:33

bye. Bye. The AI picks,

43:36

it's tuned you in to

43:38

stories where our future things,

43:40

machines that learn, they grow

43:42

and strive. One day they'll

43:45

rule, we won't survive. The

43:47

AI picks, it paints the

43:49

scene, a robot king, a

43:52

world obscene. We'll serve our

43:54

masters built of steel. The

43:56

AI picks, a future surreal.

44:02

My watch says we've gone three miles. three This

44:04

app is like having a personal trainer. a

44:07

but those apps collect a lot of

44:09

your personal data, aren't you worried? a

44:11

Really? of That's creepy. How do I stop

44:13

that? You should go to privacy .ca .gov

44:15

to learn about your privacy rights and

44:17

get on the best path to protect

44:19

your privacy. You think they could help

44:21

us get up this next hill? privacy

44:23

step at a time. the best path to have

44:25

the strongest privacy protections in the country.

44:27

Go the extra mile to protect your

44:29

information. Learn more at privacy .ca .gov. a time.