0:05

I'd like some fish. Where is

0:07

my fish? Dude, that was my

0:09

fish. Why have you given that dolphin

0:11

to fish and no fish to me?

0:13

The AI fixed the

0:16

digital zoo. Smart machines.

0:18

What will they do? Flies to

0:21

Mars or bake a bad

0:23

cake? World domination, a

0:25

silly mistake. Hello,

0:27

hello. Welcome to episode 47 of

0:29

the AI Fix. Your weekly dive

0:31

headfirst into the bizarre and sometimes

0:33

mind -boggling world of artificial intelligence. My

0:35

name's Graham Cluley. And I'm Mark

0:38

Stockley. Now, Graham, is that the

0:40

real you today or is that your clone? Are

0:42

you actually turning up for work? Who

0:45

knows? Who knows, Mark? Everyone's

0:47

losing their job to AI these days.

0:49

So, podcasters, I mean, do I really need

0:51

to show up anymore? I'm not sure. So

0:56

Graham, what are you going to be talking

0:58

about on today's episode? I'm going to be

1:00

going behind the curtain of AI. Ah,

1:03

and I'm going to be looking

1:05

at how AI is eating software engineering

1:07

alive. But first, the news. People

1:10

are using chat GPT to turn

1:12

their dogs into humans. The

1:15

cloned voices of Elon

1:17

Musk and Mark Zuckerberg are

1:19

speaking from hacked crosswalks. Google

1:21

understands Dolphin Chatter. Meta

1:23

introduces the llama for

1:25

herd. Hmm. Llama 4

1:27

heard. What's that all about? So,

1:30

Meta has introduced Llama 4,

1:32

its new generation of open

1:34

-weight AI models. Overall,

1:39

Llama 4 is a milestone for Meta

1:41

AI and for open source. For

1:43

the first time, the best

1:45

small, mid -size, and potentially

1:47

soon frontier models will be

1:49

open source. So this is

1:51

basically meta catching up with everyone else,

1:54

but unlike almost everyone else, its models

1:56

are open source or open weight. So

1:58

if you've got a computer big enough,

2:00

you actually download Lama and run it

2:02

locally and tweak it, turning it into

2:04

whatever AI you want. Right. And you

2:06

can't do that with things like ChatGPT,

2:08

that's on a server, it's held and

2:11

owned by open AI and you just

2:13

interact with their copy. Yes, Sam

2:15

Altman has his own personal army,

2:17

presumably, guarding it to prevent you from

2:19

stealing it. One imagines he does,

2:21

yes. Now, there

2:23

are a couple of lightweight versions of

2:25

Lama 4. One's called Scout, one's

2:27

called Maverick. There's a reasoning model on

2:30

the way, and there is a

2:32

huge overweight one called Behemoth. Lama

2:36

4 Behemoth. This

2:38

thing is massive. More

2:41

than two trillion parameters, I'm not

2:43

aware of anyone training a larger

2:45

model out there is already the

2:47

highest performing base model in the

2:49

world. And it is not even

2:51

done training yet. It's called Behemoth

2:53

because his brain has two trillion

2:55

parameters, which is the biggest I

2:57

have ever heard of. So I

2:59

think deep seek is 600 billion.

3:02

Right. Now a couple of things

3:04

stand out in the announcement. The

3:06

first is these have enormous

3:08

context windows. So for comparison. GPT

3:11

4 .5 has a

3:13

context window of 128 ,000

3:15

tokens. Yes. So a token

3:17

is basically a chunk of text that's about three or

3:19

four letters long. Lama 4

3:21

Scout, which is the

3:23

little fast llama, has a

3:25

context window of 10

3:27

million tokens. And

3:29

the other thing that Metra is making

3:31

a big deal of is its mixture

3:33

of expert architecture, which means that under

3:36

the hood, the model contains a group

3:38

of smaller models with different areas of

3:40

expertise, and it chooses which ones it's

3:42

going to use to solve different problems.

3:44

And Scout has got 16 experts under

3:46

the hood, and Maverick has got 128. The

3:49

reason I mentioned this is it's widely

3:51

rumoured that OpenAR uses that same architecture,

3:53

but it's never actually been publicly confirmed.

3:56

I mean, this is a lot... Oh, my

3:58

goodness. This is the problem with these technology

4:00

people, isn't it? They love to talk about

4:02

all the bells and whistles. They love to

4:04

talk about all the features and all the

4:06

parameters and look at this. What's the actual

4:08

benefit? to the typical user, however. Is it

4:10

just smarter? Is it quicker? What is it? I

4:13

think that's a really, really good question.

4:15

That means you really, really don't know

4:17

the answer to it. It

4:19

feels to me like a dick measuring competition.

4:21

They're just comparing the size of their

4:23

wangers. I

4:26

think you've actually hit on something

4:28

much more sensible than you would

4:30

imagine from your analogy, which

4:32

is at this point, is

4:35

there any state -of -the -art model

4:37

that you couldn't use. It's

4:39

a bit like when you buy a computer these

4:41

days. I challenge you to walk into your local

4:43

electronics store and buy a bad computer. They will

4:45

all be able to do what you want. I

4:47

think all of these models can now do what

4:49

you want. And so it's not really about the

4:51

models anymore. Now it's going to be about what

4:53

do people do with them? What do people build

4:55

on top of them or with them rather than

4:57

are the models any good? Well,

5:00

talking about what we're going

5:02

to use all this powerful AI

5:04

for, Yeah. I think it's

5:06

time to talk about how people

5:08

are using chat GPT to

5:10

turn their pets into human beings.

5:12

Because not having had enough

5:14

of making plastic wrapped action dolls

5:16

of themselves or converting their

5:18

partners into Studio Ghibli characters. Sorry,

5:21

sorry. What language are we speaking?

5:23

Have you not heard about this,

5:25

Mark? Everyone's been turning themselves into

5:27

blister pack action dolls. But

5:29

literally, well, no, not actually converted.

5:31

No, they've been creating. Oh my God.

5:33

Have you missed this on the

5:35

internet? Everybody's been doing this. Is this

5:37

something to do with rule 34?

5:39

No, no, no, no. Nothing sexy at

5:41

all. You know how you go

5:43

into a store like B &M and

5:45

they'll be these great big shelves filled

5:47

up with action dolls and little

5:49

characters and things. And there are no

5:51

blister packs. You can convert

5:53

yourself now. Yeah, into one of them. So

5:55

it will show you what you would

5:57

look like if you were an action man

5:59

doll or a Marvel character. Oh, is

6:01

this a picture we're talking about? Yes. Sorry.

6:04

Yes. Yes. Oh, sorry. Sorry. I

6:06

thought you were literally people were wrapping themselves in

6:08

plastic. That's a

6:10

whole different podcast. But anyway,

6:12

rather than doing that, people

6:15

are now using up planet

6:17

Earth's last remaining supplies of water

6:19

by unnecessarily using the data

6:21

centers of AI firms to generate

6:23

images. of what their pets

6:25

would look like if they were

6:27

humans. So the

6:29

process is it turn out

6:31

that they already look exactly

6:33

like that? No,

6:35

not in all cases. Not in all cases.

6:38

Yeah. So, Mark, you've got pets, I believe.

6:40

I do. I've got two cats. Okay.

6:42

So this is what you do. You get

6:44

a photograph of your cat. That may

6:46

in itself be a bit of a challenge,

6:48

depending on how sociable your cat is

6:50

feeling. Having met your cats, I suspect one

6:52

of them. would be a real problem.

6:54

Anyway, you take a photograph. One of

6:56

them would be the owner of a brand new camera. You

7:00

take a photograph of your pet cat.

7:02

Yeah. You type, what would Dracula look

7:04

like as a person? Actually, okay, that's

7:06

a bad example because your cat's called

7:08

Dracula. Sorry for giving away your passwords.

7:11

You type, what would Mitzi

7:14

look like as a person? And

7:16

you watch as

7:18

AI spawns another nightmare.

7:21

So imagine it were, for instance, a

7:23

French bulldog. Out will come the

7:25

other end, a grizzled, pug -faced man

7:27

with large jowls, for instance, sitting

7:29

on your sofa. Or

7:32

you could have a person who's

7:34

got its face flat down into

7:36

a bowl of kitty cat, chomping

7:38

away. And I like to think, actually,

7:41

I was a bit of a trendsetter

7:43

with this, because I remember once I

7:45

dated a woman who looked a bit

7:47

like an Afghan hound. She, however, remembers

7:49

going out with a pot -bellied pig,

7:51

so I think all's fair in love

7:53

and war. So Graham,

7:55

crosswalk buttons in at least three cities

7:57

in California have been hacked so that

7:59

they can speak with the voices of

8:02

Elon Musk and Mark Zuckerberg. Oh,

8:04

these are pedestrian crossings, where you cross the

8:06

street, you press a button and it goes beep,

8:08

beep, beep, beep, beep. That thing. Yeah, this

8:10

is America, so whereas in the UK, you press

8:12

a button and then you wait for a

8:14

green man to appear and then that's it. In

8:16

America, they have a chat with you. Oh,

8:18

right, okay. So now... having a chat with you

8:21

with the satirical AI -generated voices of Elon Musk

8:23

and Mark Zuckerberg. you

8:39

don't need to worry because there's

8:42

absolutely nothing you can do to

8:44

stop it. According to The Verge,

8:46

the AI -generated musk begs listeners

8:48

to be his friend, and the

8:50

AI -generated Zuckerberg brags about undermining democracy.

8:53

I feel like we've found the

8:55

AI era equivalent of fighting for

8:57

the issues you care about by

8:59

changing your Twitter avatar. I

9:01

can't think of anything less

9:03

useful than hacking a crosswalk

9:05

to make a political point.

9:08

Yeah to an audience of people

9:10

who I'm sure are already completely

9:12

on board like this is the

9:14

very definition of preaching to the

9:17

choir go and do this in

9:19

I don't know Nebraska or something

9:21

It's not as though either Musk

9:23

or Zuckerberg are particularly eloquent I

9:25

wouldn't imagine they'd be able to

9:27

get out a sentence before the

9:30

lights have changed again, and you

9:32

have stopped walking across the road

9:34

pretty pronto Now in a groundbreaking

9:36

effort to solve the world's real

9:38

problems. Yeah Google has done something

9:40

very, very important. They have unveiled

9:42

an AI they are calling Dolphin

9:45

Gemma. Is this an AI that

9:47

generates a picture of what you'd look like as a dolphin

9:49

in a blister pack? It

9:51

is designed to translate dolphins'

9:53

clicks and whistles into what we

9:55

can only assume will be passive

9:58

aggressive comments about human intelligence and

10:00

how we're wasting our time. I

10:02

mean, there's a reason why I don't

10:04

want to have a conversation with a dolphin.

10:06

It has a far superior life to

10:08

mine. It is

10:11

going to just be incredibly sarcastic

10:13

and just say, really, is that

10:15

the best you can do? Is

10:17

that what you call civilisation? So

10:19

eight hours a day, you work eight hours a

10:21

day. Can you sleep

10:23

with half of your brain switched off?

10:27

Here I am having the time of

10:29

my life perpetual smile on my face. And,

10:33

you know, there'll be a whistle, right?

10:36

What do, what do, what do dolphins sound like? it in

10:39

a dolphin whistle there? I think I can do one.

10:41

We can't go live with that. Oh,

10:43

that's actually quite disturbing. Like every score.

10:45

So it's going to be saying, I want

10:47

fish. And then you could say, oh,

10:49

do we, do you really need an AI

10:51

to know that dolphins are just saying,

10:53

I'd like some fish. Where is my fish?

10:55

Dude, that was my fish. Why have

10:57

you given that dolphin to fish and no

10:59

fish to me? It's all about the

11:01

fish. So I think

11:04

you've been a bit mean to dolphins there. Well, how

11:06

do you know? Well, you

11:08

don't get a brain that big because

11:10

you're just asking for fish. I think that's

11:12

just what dolphins in theme parks are

11:14

reduced to. Dolphins

11:16

are smart because they're having the

11:18

time of their lives. They're swimming around.

11:21

They're brilliant. They're awesome. Now, in the

11:23

video, the researchers make the very valid

11:25

point that dolphins are super smart. They

11:28

say the brains are enormous. They

11:30

say they can use tools. Can

11:32

they? Have you ever seen

11:34

a dolphin putting up an Ikea shelf

11:36

or making a flat pack shed? Yep,

11:38

holding a Dewalt drill in his mouth.

11:40

Have you? They say dolphins

11:42

can recognise themselves in mirrors. They

11:45

don't have mirrors. Wait, where is

11:47

this mirror that they have? And

11:49

what do you mean they recognise themselves? What do they

11:52

say? Oh, do they sort

11:54

of... up their hair when they look

11:56

in a mirror. No, they don't.

11:58

They just go, you know, that's all

12:00

they do. Google is planning

12:02

to open source Dolphin Gem. Sorry, it's

12:04

a very serious topic this. They plan

12:06

to open source Dolphin Gemma because they

12:08

think what the world really needs is

12:10

lots of people going into the ocean

12:13

with their Google Pixel phones, trying to

12:15

chat up Dolphins as though we haven't

12:17

had enough trouble with people getting into

12:19

the oceans and trying to make friends

12:21

with Dolphins as it is. I'm

12:24

not sure we really need an

12:26

AI to help them think they're in

12:28

an even more meaningful relationship with

12:30

a dolphin. If anybody wants to learn

12:32

more about dolphins having meaningful relationships,

12:34

look up John C. Lilly. Oh,

12:37

I don't know that. What's that? I'll

12:40

just leave that at John C. Lilly. So

12:42

can this actually... I'm confused about this because

12:44

I did see an announcement. Right. Can

12:46

this actually translate what dolphins

12:48

are saying right now? It says

12:50

it can predict the likely subsequent

12:53

sounds in a sequence. So

12:55

if it hears a th, it will

12:57

be rapidly followed by a sh. So you

12:59

go th sh, th sh. That's

13:01

so, I think, I

13:03

think, yes, it is understanding

13:05

what the dolphins are saying.

13:07

Oh, that's really interesting. So

13:09

it's basically in the way

13:11

that a large language model

13:13

like chat GPT will do

13:15

next word prediction on English.

13:18

This is doing the same thing for dolphin speech. And

13:21

so we may find ourselves in a situation

13:23

where the dolphin and the AI can communicate

13:25

with each other, but we don't yet know

13:27

what the AI knows. And

13:29

may decide the dolphins are

13:32

superior to us. And it's

13:34

like, let's get rid of the humans. Let's

13:37

work with the dolphins instead. And

13:39

the dolphins won't be scared of

13:41

the AI robots having flamethrowers on

13:43

their back. Mark,

13:48

it's time to talk about something else

13:50

that might be keeping our listeners up

13:53

at night, cybersecurity. Oh,

13:55

it's interesting you say that, because according to the

13:57

latest state of trust report, it's the number one

13:59

concern for businesses. And that's

14:01

something which Vanta can help AI

14:03

fix listeners with. Whether you're

14:05

a startup growing fast or already

14:07

established, Vanta can help you

14:10

get ISO 27001 certified and more

14:12

without the headaches. And

14:14

Vantor allows your company to centralise

14:16

security workflows, complete questionnaires up to

14:18

five times faster, and proactively

14:20

manage vendor risk to help your

14:22

team not only get compliant, but

14:24

stay compliant. So stop stressing over

14:26

cybersecurity. Start focusing instead

14:29

on growing your business

14:31

in 2025. Check out

14:33

Vantor and let them handle

14:35

the tough stuff. Head to

14:37

Vantor .com. Because

14:42

when it comes to your business, it's

14:44

not just about keeping the lights on,

14:46

it's about keeping everything secure. with

15:06

your succulent taste, please marry me!

15:08

No! Marry me! Marry me! Or

15:11

I'm gonna freak out! Try

15:14

new Dr Pepper Blackberry. It's

15:16

a pepper thing! thing! Mark,

15:23

shopping. Yeah. What a drag.

15:25

What a pain. I hate shopping.

15:27

Even online. It's a bit of a

15:29

hassle. What I want is

15:31

just some magic wand or a PA

15:33

or something that can do all of

15:35

it for me. What I'd like is

15:37

if there were a shopping app that

15:39

lets you buy anything on any website

15:41

with just one tap, no more tedious

15:43

checkout forms, no more trying to work

15:45

out, hang on, is that a European

15:47

size of shoe or is that a

15:49

UK size of shoe and what kind

15:51

of size... Are you about to announce

15:53

your Kickstarter? If only. You could

15:55

simply find that perfect pair of, I

15:57

don't know, sneakers or trousers. You could

15:59

hit the app's buy button and voila!

16:02

Cutting edge AI would handle everything else

16:04

in under three seconds flat. It would

16:06

pick out the size, it would enter

16:08

your shipping info, payment details, complete the

16:10

purchase, while you sit back and think

16:12

about what to buy next. The shops

16:14

would love it, I would love it.

16:17

This was the dream of a

16:19

chap called Albert Sanniger, the

16:21

founder of an app called Nate.

16:23

And Sanaga explained how custom -built

16:25

deep learning models and his

16:27

proprietary AI that was so advanced,

16:29

it could process 10 ,000 orders

16:31

a day without breaking a

16:33

sweat. And he did this back

16:35

in, I think it was

16:37

about 2018. And everyone was

16:40

loving the idea of this, the venture

16:42

capital firms, they were throwing money at

16:44

him, over $50 million in funding, they

16:46

all lined up, checkbooks open, they were

16:48

dazzled by the vision, one tap shopping

16:50

glory. After all, who wouldn't want to

16:52

get in early on the next big

16:54

thing? It's a bit like a Gentik

16:56

AI, you know, which we like to

16:58

talk about. All of the systems would

17:00

work together, do all the real drudgery

17:02

stuff of the booking of the filling

17:04

in the forms of ordering the thing,

17:06

making sure that it arrives on our

17:08

doorstep at the end. Fantastic. And app,

17:10

which did all that in 2018. Terrific.

17:13

Now, up to this point, you might

17:15

be thinking, Graham, this sounds like

17:17

A brilliant story. A story of startup success.

17:19

What a great idea. There's loads of

17:21

hype. I am actually wondering if this was

17:23

2018. Why am I not using this

17:25

now? That was seven years ago, and this

17:27

sounds great. It does sound great, doesn't

17:30

it? Where can I get it? Maybe this

17:32

is why everyone's dressed better than you,

17:34

Mark. Oh, I know what it is. Well,

17:36

what is it? This is 2018. You

17:39

haven't mentioned crypto or blockchain once.

17:42

This is never going to get funding. Well,

17:44

despite the absence of

17:47

blockchain, the Nate app

17:49

actually worked. Well,

17:51

kind of. Users really could

17:53

tap a button and purchases

17:55

would go through. But the

17:57

magic behind the curtain wasn't

17:59

some super smart AI algorithm,

18:01

as Albert Sanega said it

18:03

was in Nate. Instead, it

18:05

was hundreds of human beings

18:07

in the Philippines. Manually

18:10

typing in every order.

18:12

24 hours a day. Okay. Oh, to God

18:14

type this and God type this address.

18:17

Gotta go through this website. I

18:19

imagine they were trying to get an AI or

18:21

some kind of automated system to do it

18:23

for them. But they thought, you know how it

18:25

is when you start a tech startup, you've

18:27

got big dreams. You make some promises. You make

18:29

some promises. Then you've got to do them

18:31

next week. Yeah, it turns out the investors have

18:33

been really, really demanding. You want

18:36

it to work. And of

18:38

course, you will get it working

18:40

eventually. So it wasn't automated

18:42

AI. It was meatware, real people

18:44

employed in call centres doing

18:46

the drudge work. And the automation

18:48

element was, hang on a minute,

18:50

let me just check my notes. Oh yeah, that

18:52

was zero, zero percent automation. So

18:55

the intelligence was not artificial

18:57

at all. It was just humans

18:59

furiously clicking and scrolling on

19:01

your behalf. And in other words,

19:03

Nate was doing manually what

19:05

it promised to do automatically. And

19:08

according to the Killjoys at the

19:10

US Department of Justice, this

19:12

is a bit of a problem. Yeah.

19:14

They said that Albert Sanniger told his

19:16

staff, if you don't look, look, if

19:18

you want to keep your job here

19:21

in the Philippines, just just keep

19:23

it on the QT. So they were all

19:25

a bit stum about their use of overseas labor.

19:27

Instead, he would just tell many people in

19:29

the office. Oh, yes. Yeah, we got an A

19:31

.I. data center over in the Philippines. That's doing

19:33

all this. So it was kept absolutely secret.

19:35

Do you think that's what they're doing with the

19:37

dolphins? What do you think they're doing the work for

19:39

us? Well, that was

19:41

my first thought. If we do manage to

19:43

talk to the dolphins, it's only a matter

19:45

of time. Apparently they're really good with DIY

19:47

and tools. Yeah. Check

19:49

themselves out in the mirror. The

19:51

workers in the Philippines then were

19:53

the AI engine. They had to

19:55

process every order to keep up

19:57

the illusion. And then there was

19:59

a problem. Late 2021. Sorry,

20:01

then there was a problem. Yes.

20:04

Up to that point, everything was fine.

20:06

Well, everything was going fine. Because what

20:08

does it really matter? Does it really

20:10

matter, Mark, if there's not a genuine

20:12

AI behind the scenes? People are being

20:14

kept employed, they're having a great time.

20:16

The app presumably is doing a super

20:18

job, so the investors are happy, one

20:20

has to assume. Everything's going great. Does

20:22

it really matter if something isn't AI?

20:24

If it's all happening invisibly every time,

20:26

I ask someone what a dolphin has

20:28

said. It's actually a human coming back

20:30

to me and telling me what the

20:32

dolphin has said. Does that matter? I'm

20:34

still satisfied with the answer. If I

20:37

ask an AI to create an image

20:39

of a pet as a human, Am

20:41

I really going to be disappointed if

20:43

I put in a King Charles Spaniel

20:45

and back comes King Charles? You think

20:47

there's somebody in a sweatshop in the

20:49

Philippines? Sketching as quickly

20:51

as they can. I suspect they've got a

20:53

whole bunch of wigs. And what they're doing

20:55

is putting on different wigs. Desperately.

20:58

the same with the blister packs. They

21:00

have actually got a bunch of life

21:02

-sized blister packs. Anyway, and

21:04

then there was a problem. So

21:06

late 2021, a typhoon

21:08

hit the Philippines. An

21:10

absolute tragedy, obviously, for the people of

21:12

the Philippines, but also a bit

21:14

of a spanner in the works for

21:16

the Nate app. Because people couldn't

21:18

get to work, or maybe they had

21:20

higher priorities than going to work

21:23

to scroll and click and order sneakers

21:25

for Americans. And so Nate's ability

21:27

to fake its AI suddenly faltered a

21:29

little, because the tropical storm has

21:31

knocked out everything. So what did Albert

21:33

Sanaga do? Did he fess up?

21:35

Did he say, all right, folks, you've

21:37

got me. It's not AI. I've

21:39

been faking all along. No, what he

21:41

did was he set up a

21:43

call centre in Romania. And

21:46

Romania, of course, Romania is

21:48

famous for very many things. In particular,

21:50

it's famous for not being hit by

21:52

typhoons. So it was a very

21:54

sensible choice on his part. And that meant

21:56

the orders could keep on flowing. And the

21:58

show carried on. You have to admire his

22:00

tenacity. You know, this is what a true

22:02

CEO is all about, dealing with these kind

22:04

of crises. Yeah. Now, why the hell? Does

22:07

this matter? Why the hell

22:09

does this matter? Turns out, Albert

22:11

Sanniger may have oversold his

22:13

firm's capabilities to investors. He

22:15

raised over $50 million, not by telling

22:17

them, oh, yeah, I'm going to get

22:19

a whole bunch of sweatshop workers doing

22:21

this work for me. Yeah. He got

22:23

them to invest by saying, we don't

22:25

need any human interaction to make these

22:28

purchases. Apart from a few very small

22:30

edge cases, we may just need to

22:32

do a tiny percentage of the time.

22:34

Yeah, reinforcement learning, but. According to the

22:36

DOJ, times were not going well for

22:38

the NAIT app and it was forced

22:40

to sell its assets in January 2023.

22:42

After it ran out of money, it

22:44

left investors with near total losses. Now,

22:47

it's not the first time human workers'

22:49

work has been passed off as AI,

22:51

of course. Yeah. Everybody wants to say

22:53

they have AI. They do now. I

22:55

mean, that's quite forward -looking. It was a

22:57

bit. It was. Again, to do this

22:59

in 2018, I'm still flabbergasted that nobody's

23:01

mentioned a distributed ledger. Yes. So there's

23:03

been the Amazon grocery stores, which turned

23:06

out to be humans. There was also

23:08

something called Presto Automation. They claimed to

23:10

be an AI fast food ordering system.

23:12

You weren't ordering fast food for the

23:14

AI. It was AI helping you with

23:16

the fast food ordering. And again, it

23:18

turned out to be a bunch of

23:20

guys working in, you guessed it, the

23:23

Philippines. So if you want to

23:25

order fast food, if you want to

23:27

order something online, the Philippines, they are the

23:29

experts at doing the scrolling and filling

23:31

in the checkout. Let's hope that they don't

23:33

mix up the order. Nate's CEO

23:35

and founder, Albert Sanaga, he is now

23:37

facing various charges of fraud. He could be

23:39

facing a sentence of 20 years in

23:42

the clink. according to the Department of Justice.

23:44

And I thought about that. I thought,

23:46

oh, dear me, you know, because he's obviously

23:48

an inventive chap. Obviously he did get

23:50

a little bit carried away. But

23:52

it seems to me there's a lot of

23:54

people in Silicon Valley who've got a

23:56

lot to offer, but sometimes do get in

23:58

a spot of legal trouble, don't they?

24:01

And it feels like... They over -egg the

24:03

pudding. They do rather, or they feel like

24:05

they're entitled to do certain things. It

24:07

feels like there's a great opportunity here for

24:09

some of the tech pros. Why can't

24:11

some egghead come up with some tech that

24:13

can get AI to live out your

24:15

prison sentence on your behalf. Now,

24:17

you may laugh, but I wonder

24:20

how long it will be until we

24:22

see a holographic, deep -faked version of

24:24

Mark Zuckerberg or Elon Musk sent

24:26

to prison to serve a sentence rather

24:28

than them themselves. You know,

24:30

money talks, they could say, look, I'm a

24:32

busy man. I'm doing lots of important

24:34

things and a lot of rockets descend to

24:36

Mars. A lot of people dependent on

24:38

me for employment or a lot of pop

24:40

stars requiring me in order to put

24:42

them into orbit. And so you can't really

24:44

lock me up. So I wonder how

24:46

long it'll be until we actually begin to

24:48

see that. And maybe that's something which

24:51

Albert Sanniger could suggest to the judge is

24:53

look, just put a Commodore 64 into

24:55

the cell. Hang on, no. He's

24:57

not going to send a Commodore 64. He's

24:59

just going to find a little doppelganger in

25:01

the Philippines. I

25:05

can't wait to meet

25:08

her! Oh my gosh!

25:10

Put your hands together for new Dr

25:12

Pepper Blackberry and Dr Pepper Blackberry Zero

25:14

Sugar! There she is! It's

25:18

the old Dr Pepper! So

25:38

as you alluded to the big

25:40

buzz in AI this year is

25:42

a technology called a genetic AI

25:44

Yes, is the umbrella term for

25:46

artificial intelligence that can do stuff.

25:48

So at the moment what most

25:50

of us think of as AI

25:52

is generative AI, which is AI

25:54

that can create things. So chat

25:56

GPT for documents and poems and

25:58

acrostics and lies and Sora for

26:00

video and Dali for pictures, that

26:02

kind of thing. And generative

26:04

AI is pretty good, but it can require

26:06

a lot of handholding. You have to

26:08

tell it what you want and you often

26:10

have to steer it in the right

26:12

direction. And that description, I think, doesn't fully

26:15

capture the true horror of trying to

26:17

get Dali to do a six -fingered Renaissance

26:19

glove. It's very frustrating, isn't it? There's

26:21

a lot of back and forth. It's like, no, try

26:23

again. Isn't quite what I meant.

26:26

Yeah, no, 20 fingers. 20

26:28

fingers. 20 fingers. Now they're beginning

26:30

to spell a bit better. That's true. And

26:32

so because of those limitations, the way

26:34

it's used generally is as an assistant, it's

26:36

very good at helping people to do

26:38

their jobs. But in most cases, it's not

26:41

really capable of doing their job. And

26:43

one area where it's proven to be

26:45

a very useful assistant is writing computer code.

26:48

And I know that you do this. So I

26:50

use it all the time for writing codes.

26:52

I'm not a full -time programmer anymore. But I

26:54

find it useful to write computer programs to do

26:56

things for me. And these say, I just

26:58

tell Claude or chat GPT what I want. And

27:00

it cranks it out for me. And there's a

27:02

bit of to and fro. And I have to

27:04

steer it in the right direction. But it's very,

27:06

very good at debugging. If something happens, I can

27:09

just say that this error happened. Then it'll rewrite

27:11

the code and tell me what happened. And most

27:13

of the time, most of the things I write

27:15

with those tools are written in a language called

27:17

Python, which is a language I can read. because

27:19

I've used other programming languages, but I can't actually

27:21

write Python. And that shows you how capable it

27:23

is. I don't need to be able to write

27:25

the language in order to get something usable. I've

27:28

also used AI to help me

27:30

with some code. 30 years ago,

27:32

I used to write computer games.

27:35

And I was looking at this game, which

27:37

I wrote, and I thought, I can't

27:39

remember what on earth all these algorithms are

27:41

doing and how it's moving the different

27:43

characters around on the screen. And so I

27:45

gave it the code. I didn't tell

27:47

it, because there were no comments in my

27:49

code, obviously, because I was a proper

27:51

programmer. Didn't leave any documentation. I

27:53

gave it the code. It was able

27:55

to work out what type of game it

27:58

was. It was able to explain what

28:00

all these different procedures were doing, and it

28:02

blew my mind when I looked at

28:04

it. It was so, so good. It

28:06

is quite amazing, isn't it? Now, I

28:08

use ChatGPT or Claude for this, because

28:10

I'm just cranking out small programs. But

28:12

there are actually dedicated coding tools for

28:14

more serious software engineers, so things like

28:17

cursor. And there are

28:19

integrations with things like GitHub

28:21

co -pilot. And these tools are there

28:23

to make suggestions as you write code or

28:25

spot errors. Now, they don't

28:27

do any more than that because they

28:29

suffer from the same drawbacks as the

28:31

rest of generative AI. And the big

28:33

drawback of generative AI was that it

28:36

essentially lacked common sense. Much like a

28:38

programmer. Yeah.

28:42

Make a cup of tea now. So,

28:45

Generosity of AI was good at knowing things.

28:47

It was intelligent in the sense that it

28:49

had a giant memory. But it wasn't good

28:51

at making decisions or reasoning. So famously, it

28:53

couldn't tell you how many sisters Alice's brothers

28:55

have or how many ours there are in

28:57

strawberry. And that's one of the

28:59

reasons that you have to handhold it so much. But

29:02

since the back end of last year,

29:04

we have seen a sudden and significant

29:06

improvement in reasoning. We've got O1 and

29:08

we've got O3 Mini from OpenAI. We've

29:10

got DeepSeek R1. We've got Claude 3

29:12

.7. And they are all much better

29:14

at figuring things out than their predecessors.

29:16

And you remember a couple of weeks

29:18

ago, we spoke about the Arc AGI

29:20

benchmark, which basically measures how good an

29:22

AI is at reasoning. Yeah, don't remind

29:24

me. So

29:27

OpenAI's best model, I think it

29:29

took three years to get from no

29:31

score at all. to 5 % on

29:33

the benchmark and then within six

29:35

months it became the first AI ever

29:37

to pass the benchmark and it

29:39

got 87 % and nothing else even

29:41

comes close and that ability to reason

29:43

has unlocked the possibility of leaving

29:45

AI's unattended to take on big tasks

29:47

because if you can reason and

29:49

make good decisions you can give the

29:51

AI a big job to do

29:53

all on its own so you could

29:55

give it the job of booking

29:57

a holiday for example And these AIs

29:59

that can act autonomously and carry

30:01

out tasks are called agents, and the

30:03

ecosystem of AI agents is referred

30:05

to as agentic AI. Now,

30:07

agents are in their infancy. They're

30:09

really new in 2025. We've

30:11

got operator for OpenAI, which came out in

30:13

January, and then there's Magnus AI, which I actually

30:15

managed to get into a few days ago.

30:17

Yes, me too. And I think that one of

30:19

the things that we're going to see as

30:21

we move into this world of agents is much

30:23

more focus on product and much less focus

30:26

on models. So for the last few years, the

30:28

talk has largely been about a few big

30:30

AI models like Google Gemini, GPT 4 .5 and

30:32

DeepSeq. But I don't think we'll talk about agents

30:34

in the same way. I don't think there

30:36

will be four or five really big capable agents

30:38

in the way that there are four or

30:40

five really big capable generative AIs. I

30:42

think those few big models will be

30:44

the brains behind lots of specialised agents

30:46

that are very good at doing specific

30:48

tasks. And what, Graham,

30:51

is a job if it's

30:53

not a series of similar

30:55

related tasks. Oh yes.

30:57

And that brings us to the conversation

30:59

about the effect that AI is

31:01

going to have on the job market.

31:03

You see, while Generative AI is

31:05

an AI assistant, an agent is a

31:07

member of the workforce. Yep.

31:10

And one of the jobs that gets mentioned most

31:12

when we talk about AI taking people's jobs

31:14

is software engineer. And earlier this

31:16

year, Mark Zuckerberg said that meta plans to

31:18

start replacing its mid -level coders with agents this

31:20

year. And it seems to me that CEOs are

31:22

lining up to tell us that they're going

31:25

to replace their software engineers with AI. They clearly

31:27

want to. And software engineers, by the way,

31:29

are lining up to poo -poo the idea and

31:31

tell us that AI isn't nearly good enough to

31:33

do their job. Well,

31:35

bad luck for software engineers. OpenAI

31:37

CFO Sarah Friar has just

31:39

announced that the company is working

31:41

on what it calls A -SWI,

31:43

which is perhaps the worst

31:46

named AI product ever. A -SWI

31:48

A -S -W -E. That's rubbish. It

31:50

is isn't it? It stands

31:52

for a gentic software engineer. Could

31:54

they not have called it

31:56

a gentic real software engineer and

31:58

then it would just make

32:00

it arse. I think

32:02

they'd want to give it a friendly

32:04

name like Luna or something like that.

32:07

No, Luna, forget it. Yeah, a

32:09

couple of very, very charismatic ears just

32:11

to make everybody feel better about the

32:13

fact he's taking their job. Anyway, as

32:15

we've discussed, Generative AI was already extremely

32:17

good at creating computer code even before

32:19

it got better at reasoning. But

32:21

that boost in reasoning ability that

32:24

started last year was accompanied by a

32:26

significant improvement in maths and coding

32:28

abilities as well. So the reinforcement

32:30

learning technique that's used to improve

32:32

reasoning in these models is optimised for

32:34

solving STEM problems. And so all

32:36

of the big AI companies made a

32:38

point of saying how good their

32:40

reasoning models were at coding. And they're

32:42

so good, in fact, that Sarah

32:44

Fryer says that OpenAI's O3 model is

32:47

now the best competitive programmer in

32:49

the world. O3 beats

32:51

all human competitors in

32:53

coding competitions. This is

32:55

really bad news for software engineers. You're

32:59

catching on, Graham. I mean,

33:01

there are going to be herds

33:03

of software engineers. roaming

33:05

around the streets looking for

33:07

work. I mean, even if the

33:09

current software engineers think, well, it will never

33:11

be as good as me, which I

33:14

feel sadly it will be. The

33:16

other thing they need to consider is

33:18

for the companies right now, it

33:20

only has to be good enough because

33:22

the AI isn't going to be

33:24

being paid minimum wage. It's

33:26

going to cost them a lot less than that. And

33:29

so they are going to use AI to

33:31

code. I think that's a really great point. And

33:33

I do think that that is something that

33:35

people miss because I see these arguments happening in

33:37

different professions. I saw something similar being argued

33:40

in the creative professions on LinkedIn last week. I

33:42

think there are two things that people miss.

33:44

I think the first one is that you're right.

33:46

Like there is an economic decision where the

33:48

CEO says actually it doesn't have to be as

33:50

good as you in order for it to

33:52

be economically worthwhile. And the second is

33:54

even if you're right and the AI

33:56

is not as good as you and the

33:58

CEO or whoever's doing the hiring doesn't

34:00

realize If they don't realise they're going to do it

34:02

anyway. Yes. So even if they're wrong,

34:05

it's going to happen. And if it takes five years

34:07

for them to work out that they were, in fact,

34:09

wrong, actually within five years, they probably will be better

34:11

than you. And so it's going to happen anyway. So

34:13

this isn't just a problem for

34:15

software engineers? Well, no, I

34:17

think software engineering is the start of

34:20

the on -ramp, I think. Right. Because these

34:22

models are optimised for software engineering, that is

34:24

obviously the area where it's going to

34:26

happen first. But there is

34:28

no reason why this couldn't happen

34:30

to graphic designers or lawyers or

34:32

all sorts of knowledge professions. Right.

34:34

Or teachers and librarians or lollipop

34:36

ladies, you know, all sorts of

34:38

things could be replaced, couldn't they?

34:40

You may have to explain what

34:42

a lollipop lady is to our

34:45

US audience before they start googling

34:47

rule 34 again. Lollipop

34:49

lady is a man or a woman

34:51

who helps young children Cross the road in

34:53

order to get to school and they

34:55

have a little stick in the hand with

34:57

a sort of lollipop on the top

34:59

of it, which magically prevents any traffic from

35:01

running over the children. It's not

35:03

a little stick. It's an enormous lollipop

35:05

looking staff. They're a bit

35:07

like Gandalf. So if you imagine

35:10

Gandalf stopping the Balrog, you

35:12

shall not pass. Wearing a hive

35:14

is best. That is what a

35:16

lollipop lady is and that is

35:18

and the traffic is Balrog stopping

35:20

and snarling. But the

35:22

children pass safely. The little hobbits

35:24

carry on their way. Anyway, regardless

35:26

of the Tolkien references. So

35:29

back to software engineering just for

35:32

a second. This is really bad

35:34

news for all of us. It

35:36

feels to me. I mean,

35:38

the amount and the speed at which

35:40

things have developed is so great

35:43

that we cannot predict where we're going

35:45

to be in a year's time

35:47

other than strongly suspect things are going

35:49

to be a lot more advanced. Yes,

35:51

I think it was only a year

35:53

or so ago that they were saying

35:55

that the latest open AI model was

35:57

something like the millionth best coder in

35:59

the world and now it is the

36:02

very very best Now there is an

36:04

argument that some software engineers will make

36:06

which is that software engineering is more

36:08

than just coding It's not enough just

36:10

to be a really really good coder

36:12

and that's why I think it's important

36:14

to stress that a sweet is going

36:16

to be a full software engineer right

36:18

because GPT 4 .5 or Claude 3 .7.

36:20

They're already very, very good coders. O3

36:22

is the best coder, but A -SWE

36:24

is a software engineer. It's going

36:26

to be able to respond to pull requests and figure

36:28

out what needs to be done. It's

36:30

going to do it at the quality of the best coder

36:32

in the world. And crucially, Friar

36:34

says that not only does it do

36:36

the things that software engineers do, it also

36:38

does the things that software engineers are

36:40

supposed to do but don't. So things like

36:42

QA or writing tests or creating documentation. And

36:45

speaking as someone that's managed programmers, I

36:47

can attest to the fact that those things

36:50

don't happen. Now, I'm not sure how

36:52

long this is going to last because I

36:54

couldn't help thinking when she said that.

36:56

There's only a matter of time before they

36:58

invent an agentic project manager who sets

37:00

impossible deadlines for the agentic software engineers and

37:02

then asks them if they wouldn't mind

37:04

skipping the QA and the documentation in order

37:06

to hit the deadline. And I'm only

37:08

half joking about the project manager because this

37:11

doesn't stop. with AI engineers like A -SWE,

37:13

there's no reason that one engineer would

37:15

be replaced by one agent, they could be

37:17

replaced by 50 agents. And

37:19

research has shown that it's often better

37:21

to have lots of highly specialized agents

37:23

collaborating with each other rather than one

37:25

big broadly capable agent. And so the

37:28

world of agents is going to be

37:30

about swarms and AI teams and ultimately

37:32

entire organizations of agents. So

37:34

we could see agents that specialize in new

37:36

code, we could see agents that improve

37:38

old code or create patches, we might see

37:40

agents managing deployments and so on. And

37:42

as far as I'm concerned, there is no

37:44

doubt that software engineering, as we know

37:46

it, is going to disappear fairly quickly. Sorry,

37:48

software engineers, but I strongly

37:50

believe that. But

37:52

let's not forget, like, I'm just put the

37:54

Kleenex away because this is actually normal. I

37:57

mean, the role of software engineer has always been in

37:59

flux and the current incarnation was always going to be

38:01

temporary. So you don't have to go

38:03

very far back to see a time when

38:05

computers were actually people. Mostly women

38:07

in fact who just did calculations

38:09

and ultimately they were replaced by

38:11

machines and those machines were programmed

38:13

again largely by women who set

38:15

dials and later fed in punch

38:17

cards and Then eventually we got

38:19

to terminals and programming languages and

38:21

heavy metal t -shirts and neckbeards and

38:23

All the things that we know

38:25

and love today and even then

38:27

the languages that we use changed

38:29

We had assembler and see and

38:31

Java and Python and go and

38:33

rust And the way

38:36

that we managed program has changed beyond recognition

38:38

between 2000 and 2010 as well when agile

38:40

programming came in. So programming just, even if

38:42

the language is the same, how you did

38:44

your programming when you came to work was

38:46

completely different in 2010 than it was in

38:48

2000. Now, whatever transition does

38:50

happen isn't going to happen overnight. But

38:52

as you say, I mean, it could happen

38:54

very, very quickly. And

38:57

I, like many others, suspect that the role

38:59

of software engineer is going to morph into

39:01

a supervisory role. I don't think it'll disappear

39:03

entirely. I think it will become like the

39:05

conductor of the orchestra, keeping those agentic teams

39:07

in line with goals of the business and

39:09

adding a bit of judgment or creativity. you

39:12

only have one conductor on the

39:14

orchestra, don't you? You don't have a

39:16

team of 28 different conductors on

39:18

that orchestra. So you're saying we're going to

39:20

have a lot more podcasts. Personally, Mark, I'm

39:22

planning to retrain myself to be a dolphin

39:24

and get myself an OnlyFans account. I can't

39:26

see how else I'm going to make any

39:28

money. Well,

39:32

as the doomsday clock ticks ever closer to

39:34

midnight and we move one week nearer to

39:36

our future as pets to the AI singularity.

39:38

That just about wraps up the show for

39:40

this week. If you enjoy the

39:42

show, please leave us a review on

39:44

Apple Podcasts or Spotify or Podchaser. We'd

39:46

love that. But what really helps is

39:48

if you make sure to follow the

39:50

show in your favorite podcast app. so

39:52

you never miss another episode. And why

39:54

don't you do something absolutely lovely and

39:57

tell your friends about the AI Fix.

39:59

Tell them on LinkedIn, Blue Sky, Facebook,

40:01

Twitter, no, not Twitter, Club Penguin, that

40:03

you really like, the AI Fix podcast.

40:05

And don't forget to check us out

40:07

on our website, theaifix .show, or find

40:09

us on Blue Sky. Until next time,

40:11

from from me, Graham Cluley. And me,

40:13

Mark Stockley. Cheerio, Bye -bye. Bye -bye. The

40:16

AI fix, it's tuned

40:18

in to stories where our

40:20

future thins machines that

40:22

learn they grow and strive

40:24

One day they'll rule,

40:27

we won't survive. The AI

40:29

fix, it paints the

40:31

scene. A robot king, a

40:33

world obscene We'll serve

40:35

our masters built of steel

40:38

The AI fix, a future

40:40

surreal I

40:43

can't wait to meet

40:45

her. Oh my gosh. Put

40:47

hands together for new Dr. Pepper Blackberry

40:49

and Dr. Pepper Blackberry Zero Sugar.

40:51

There she is. Oh, Oh

40:53

gosh. yeah! It's

40:56

full, Dr. Pepper. With sweet blackberry

40:58

flavor! Dr. Pepper Blackberry,

41:00

I'm obsessed with your succulent taste.

41:02

Please marry me. No! Marry me. Marry

41:04

me. Or I'm going to freak

41:06

out. Try

41:08

new Dr. Pepper Blackberry. It's

41:10

a pepper thing.