0:06

Now, I know what you're thinking.

0:08

How can a piece of computer

0:11

code have feelings? Or rather, the

0:13

more interesting question, how does

0:15

the scientists get funding for

0:18

this research project? The AI fix,

0:20

the digital zoo, smart machines, what

0:22

will they do? Lies to Mars

0:24

or bake a bad cake, world

0:27

domination, a silly mistake, the box

0:29

with brain. Hello and welcome to

0:31

episode 42 of the AI

0:33

Fix, your weekly dive headfirst

0:36

into the bizarre and sometimes

0:38

mind-boggling world of artificial intelligence.

0:40

My name is Mark Stockley. And I'm

0:43

Graham clearly. Graham, I've just

0:45

realised it's episode 42. Oh, did

0:47

you have to give that a lot

0:49

of deep thoughts? We should mark this

0:51

somehow, the 42 is the meaning of

0:53

life, the universe and everything according to

0:55

Grock's favourite person. That's right.

0:57

Six times nine or something like that,

0:59

isn't it, I think? Well we don't

1:02

know. So we basically have to build

1:04

a planet to find out. The whole

1:06

premise of building this giant supercomputer and

1:08

everything. And then that coming up with

1:11

the idea that you need to build

1:13

another computer to work out what the

1:15

question is. When I first read that

1:17

as a teenager, it was hilarious because

1:20

it was so outlandish. Like all of

1:22

Douglas Adams' humour is based in the

1:24

extreme outlandishness of a normal situation. Yeah.

1:27

And it doesn't seem nearly as outlandish

1:29

now. If Jeff Basos turned around tomorrow

1:31

and said I'm going to build a

1:33

supercomputer that's going to run for thousands

1:35

of years and come up with the

1:37

answer to the meaning of life of

1:39

the universe and everything. Like a digital

1:41

watch, it seems like a pretty neat

1:43

idea. It doesn't actually seem very far-fetched

1:45

anymore. That was weird, but I found

1:47

something even weirder. Let's go to the world

1:49

of the weird, Graham. World of the weird. All

1:54

right, what have you found in the world

1:56

of the weird this week, Mark? I found

1:58

a talking fish. What? Well, I mean,

2:01

it's not really a fish.

2:03

Have you ever seen Billy

2:05

Bass? Oh, yes, yes, back

2:07

in the, what, early 2000s,

2:09

everyone had one of those, didn't

2:11

they? We were all very drunk

2:14

at the time. How would

2:16

you describe Billy Bass? Billy

2:18

Bass was a fish stuck

2:20

on a plaque and you

2:22

pressed a button and it would

2:24

sing, don't worry, be happy and

2:26

things like that. Yeah. It's like

2:29

a sort of pastiche fishing

2:31

trophy, isn't it? The good

2:33

news is we've solved all

2:35

the important problems in the

2:37

world, so we've got nothing better

2:39

to do with AI these days

2:42

than stick one in Billy

2:44

Bass. Take a look. What

2:46

are Billy? How are you

2:48

doing? Ready to pump? No,

2:50

no, no, no, no. I just

2:52

need some advice. How do you

2:55

deal with difficult personalities? Listen

2:57

to me now. Stay calm

2:59

and assertive. What is this,

3:01

Mark? Well, it's a talking

3:03

fish with the voice of Arnold

3:05

Chaucer's Nig. So Graham, what are

3:07

you talking about on today's episode?

3:10

Well, I'm going to be

3:12

getting emotional. Oh God, do

3:14

we have to? Anyway, I'm

3:16

going to tell you three

3:18

things that Ayes can't do. But

3:20

first, the news. Is

3:24

Manus A.I. the new deep seat? And the

3:26

award goes too. Reinforcement learning. Neo Debu's gamma,

3:28

a next generation humanoid. AI-designed drug gets a

3:30

name. AI is writing everything. So what is

3:32

Manus A. I? Well Mark, a little known

3:34

Chinese startup is making big waves with an

3:36

impressive new AI product. Hang on. We heard

3:38

all this before. That was last month, wasn't

3:40

it, with Deep Seek? Actually that news headline

3:42

now happens about twice a week? Well in

3:45

this particular case it is Manus, a new

3:47

AI multi-purpose agent. developed by a company called

3:49

Monica AI. And it's more than a chatbot.

3:51

It claims to be able to autonomously complete

3:53

complicated tasks. It can generate reports, run dozens

3:55

of social media accounts on your behalf, or

3:57

decide if now is a great time to

3:59

buy Tesla stock. All sorts of valuable information.

4:01

This looks to me like an absolute dream

4:03

come true to all those people who want

4:05

everything to be automated and not to have

4:07

to do any more thinking anymore and to

4:09

get computers to do all the work for

4:11

them. That sounds good. Where can I find

4:13

it? Well, according to Manus' website atmanus.com, it

4:16

is already outperforming open AI's deep research model

4:18

on the Gaia benchmark, which is a tool

4:20

used for comparing AI models. So go to

4:22

manus.com and you can check out the video

4:24

and see some examples. They say they've been

4:26

so swamped with interest that you have to

4:28

apply for access to actually play around with

4:30

it. I haven't been allowed in so far

4:32

which is rather frustrating. Have you had me

4:34

luck? No, I have. I literally just heard

4:36

about it from you. Oh, okay. Well, it's

4:38

pretty cool. Even though you can't play with

4:40

it yourself, you can view replays of it

4:42

doing stuff. Okay, let me watch this video.

4:44

For the past year, we'll be quietly building

4:47

what we believe is the next evolution in

4:49

AI. And today, we're launching an early preview

4:51

of Manus. Okay, before I get into this,

4:53

I've just got a little gripe here. Why

4:55

do all these videos look the same now?

4:57

I couldn't tell if this was an open

4:59

AI video or a Google Gemini video, this

5:01

basically like some dude in casual clothes sat

5:03

on a sofa in front of a bunch

5:05

of pop plants. It's a bit like when

5:07

Apple started doing those product teasers, where you'd

5:09

have a little ukulele jauntily playing in the

5:11

background and lots of happy people. And then

5:13

suddenly every single advert was a bloody ukulele

5:15

play. And now they're all sat on these

5:18

sofas in their polo necks. Yeah, and nobody

5:20

can release a new product now without being

5:22

on a completely jet black stage and the

5:24

CEO walking out in a black turtleneck neck

5:26

holding the actual... jobs held

5:28

an iPhone, which is

5:30

a thing you actually

5:32

hold in your hand.

5:34

It makes sense that

5:36

he's holding it in

5:38

his hand because that's

5:40

what his hands are

5:42

for. That's where the

5:44

phone is supposed to

5:46

go. Why is Jensen

5:49

walking around holding a

5:51

Grace Blackwell? I don't

5:53

get it, Graham. I

5:55

don't get it. Mark,

5:57

Mark, Mark, Mark. Remember

5:59

your blood pressure. Please,

6:01

I'm going to say

6:03

you sound very anxious.

6:05

I'm very worried about

6:07

you. This is going

6:09

to be the answer

6:11

to all of our

6:13

prayers, I'm sure. They've

6:15

demoed a number of

6:17

things with it so

6:20

you can check out,

6:22

for instance, how they

6:24

can create an audio

6:26

highlights reel of any

6:28

podcast. Just feed it a

6:30

podcast and it will generate from it all the

6:32

best bits so you can just listen to it

6:34

and hang on. That sounds like a terrible idea. Is

6:37

that going to cut out all the ads?

6:39

No, Graham. Just

6:41

the middle. So

6:44

the 2025 ACM Turing Award, Computing's

6:46

highest honour, a sort of

6:48

Nobel Prize for Computing, has gone

6:50

to Andrew Bartow and Richard

6:52

Sutton for their foundational work in

6:54

reinforcement learning. Reinforcement

6:56

learning is the idea of learning

6:59

by rewarding good behaviour and it's

7:01

instrumental to the way that we train AIs. So

7:03

Bartow and Sutton developed their

7:06

reinforcement learning algorithms decades ago

7:08

and they've been used ever since, but

7:10

you may remember a few weeks ago when

7:12

Deep Seat came out, it attracted lots

7:14

and lots of publicity for the fact that

7:16

it was so cheap. One of the

7:18

reasons it was so cheap is because it

7:20

was leaning very, very hard on reinforcement

7:22

learning rather than pre -training. So we're used

7:24

to the idea that AIs get fed all

7:26

this data and that's how they know

7:28

stuff. But actually, reinforcement learning is emerging as

7:30

a really important component in making these

7:32

AIs better without spending loads more money and

7:35

without throwing lots more computing power at

7:37

it and that's what's behind things like all

7:39

the reasoning models that we now

7:41

have. So they're very big on reinforcement

7:43

learning too. So well done, Bartow

7:45

and Sutton, but maybe work on the

7:47

nicknames. Yeah, neither of them are

7:49

the godfathers or anything like that, are

7:51

they? The Turing Award. Do you win anything

7:53

for that? Is it just a little gong?

7:55

Yeah, you get a million dollars. A million

7:57

dollars? Yeah. Why do you need a million

7:59

dollars? It's so cheap now using Deep

8:02

Seek rather than open AI. That

8:04

seems excessive. I'm not sure that's

8:06

required. Now, it turns out AI

8:08

is writing absolutely everything. In the

8:11

past 18 months, according to research,

8:13

we have seen the most rapid

8:15

change in human-written communication ever. I'm

8:17

not sure if that's actually true.

8:20

Who's saying this? That sounds like

8:22

Cobblers. Because, I mean, obviously there

8:24

was the first person to write

8:26

something, so previous to him there

8:29

was nothing being written, and then

8:31

it accelerated infinitely in percentage terms.

8:33

Anyway, according to this research, by

8:36

September 2024, 18% of financial consumer

8:38

complaints, 24% of press releases, presumably

8:40

including this one from the researchers,

8:43

15% of job postings and 14%

8:45

of UN press releases, showed signs

8:48

of LLLM writing. So they all

8:50

use the word delve. Yeah, it

8:52

must be. And paradigm. And realm.

8:55

Yes. And buckle up. I wonder

8:57

about this. I wonder first of all,

8:59

if this is an accurate figure or

9:01

not. I suspect it may be higher.

9:03

I mean, there'll be some element of

9:06

AI being used maybe to grammar check

9:08

and spell check, as well as actually

9:10

generating conduct. I'm sure more and more

9:12

people are doing this. But I also

9:14

wonder if these numbers have actually been

9:16

calculated by an AI itself. Why would

9:18

you get a human to read all

9:20

the press releases and analyse it? Yeah,

9:22

it's much too boring. That's why we

9:24

get AI to do things. I do think when

9:27

these stories come out, this is a sort of

9:29

whiff of this is a bad thing. And as

9:31

somebody who does a lot of writing,

9:33

obviously there's a part of me that

9:35

thinks that this is a bad thing,

9:37

I don't want to lose my job

9:40

to an AI. But I look around

9:42

the areas where AI is being deployed

9:44

on things like press releases, and I'm

9:46

thinking, you know, they weren't great before.

9:48

No, no, absolutely not. There is every

9:50

chance that the AIs have actually made

9:53

this better. So the company in silicone

9:55

medicine recently announced a drug candidate for

9:57

treating idiopathic pulmonary fibrosis, which is a

9:59

chronic... lung disease that affects around

10:01

5 million people. It's very serious

10:03

and it can cause progressive and

10:06

irreversible decline in lung function over

10:08

the course of three or four

10:10

years. Right. And the drug formerly

10:12

known as ISM001-055 has now been

10:14

given an official generic name by

10:16

the United States Adopted Names Council

10:18

and will be known as rent-assertive.

10:20

And what makes rent-assertive, worthy of

10:22

mention on an AI podcast, is

10:25

that both the biological target for

10:27

the drug, so basically which mechanism

10:29

in the body is going to

10:31

affect, and then the therapeutic compound

10:33

itself, which is the thing that

10:36

does the work in the drug,

10:38

were both discovered using generative

10:40

AI. So here we are again talking about

10:42

AI in medicine doing something utterly

10:44

remarkable. Did they also get AI to

10:47

invent the name, rent assertive? I feel

10:49

like coming up with wacky names for

10:51

drugs is something humans have really nailed

10:53

down. It feels like the kind of

10:56

thing I'd try and put down on

10:58

a scrabble ball. I feel like I'll

11:00

be cleaning my dreams with it. Now

11:03

we've talked about Neo robots a few

11:05

times. These are the robots which look

11:07

like incredibly skinny men dressed up in

11:09

wetsuits helping you with your household chores

11:12

or folding a t-shirt or handing you

11:14

your shopping bags. And they've got a

11:16

new video out. of the latest evolution

11:18

of their robot, the NeoGama. And I'm

11:21

curious about these things. And I went

11:23

to the web page. And once again,

11:25

they refuse to tell you the price.

11:27

It says it hasn't been disclosed. It's

11:29

like, well, what are you doing? Why

11:31

aren't you giving us any indication of

11:34

price or when you're going to ship

11:36

these things? Is it because they're all

11:38

mass murdering robots? What's the

11:40

plot here? And a thought struck

11:42

me while watching their promotional video

11:44

for this. Now it's the

11:46

usual lovely peaceful scenes of a

11:48

robot silently vacuuming around the house,

11:51

answering the door. It's a dillic,

11:53

helping to hang a picture on

11:55

the wall straight. And then it

11:57

occurred to me. The thought struck

11:59

me. I don't think this

12:01

is an advert for a robot.

12:03

I don't think Neo is a

12:06

robot company at all. I think

12:08

this is a trailer for a

12:10

horror movie. I think this

12:12

is a promotional campaign.

12:14

They keep releasing these videos

12:17

of these robots you can't

12:19

buy. These really, really creepy,

12:22

silent assassins in your home

12:24

which look at you blankly.

12:26

And it's very clever. I

12:29

think if you just changed the

12:32

music on this video, it

12:34

would be a horror movie.

12:36

And the most terrifying

12:38

part of it all, the

12:40

point where the music changes

12:42

is where the robot suddenly

12:45

sits down on a sofa

12:47

at the end, which as we've

12:49

discussed before, why would a

12:51

robot sit down? You know

12:54

that this means trouble. Everyone's

12:58

talking about AI these days right. It's

13:00

changing how we work, how we learn,

13:02

how we interact with the world at

13:04

a tremendous pace. It is a gold

13:06

rush at the frontier. But if we're

13:08

not careful, we might end up in

13:10

a whole heap of trouble. That's right.

13:13

But Red Hat's here to help, so

13:15

Red Hat's podcast compiler is diving deep

13:17

into how AI is reshaping the world

13:20

we live in, from the ethics of

13:22

automation to the code behind machine learning,

13:24

it's breaking down the requirements, capabilities and

13:27

implications of using AI. So check out

13:29

the new season of compiler, an original

13:31

podcast from Red Hat. Subscribe now, wherever

13:33

you get your podcasts. Mark, it's time

13:36

to talk about something else that might

13:38

be keeping our listeners up at night,

13:40

cider security. Oh, it's interesting you say

13:42

that, because according to the latest state

13:45

of trust report, it's the number one

13:47

concern for businesses. And that's something

13:49

which Vanta can help AI-fix listeners

13:51

with. Whether you're a startup, growing

13:53

fast, or already established, Vanta can

13:55

help you get ISO 27,001 certified,

13:58

and more without the headaches. And

14:00

Vanta allows your company to

14:02

centralise security workflows. Complete questionnaires

14:04

up to five times faster

14:06

and proactively manage vendor risk

14:08

to help your team not

14:11

only get compliant but stay

14:13

compliant. So stop stressing over

14:15

cyber security, start focusing instead

14:17

on growing your business in

14:19

2025. Check out Vanta and

14:21

let them handle the tough

14:23

stuff. Head to Vanta.com/AI Fix

14:26

to learn more. That's Vanta.com/Vanta.com/Ai

14:28

Fix. because when it comes

14:30

to your business, it's not just

14:32

about keeping the lights on, it's about

14:34

keeping everything secure. Mark, can I

14:37

ask you a personal question? Oh, not

14:39

again. All of your stories start with

14:41

a... Are you a psychopath? Are you

14:43

an anxious person? I just noticed... I am

14:45

as a result of all the questions that

14:47

you ask at the beginning of your

14:49

stories, yes. I mean, you're a man

14:52

in your 50s. It isn't unusual for

14:54

some of your age to suffer from

14:56

performance anxiety, for instance... is that... something

14:58

that you've had trouble. No, you want

15:01

to discuss something, right? No, I'm interesting

15:03

to you, I'm interesting to you, Mark.

15:05

I mean, it's not just you. It

15:07

turns out that AIs have feelings,

15:09

too, and can be anxious. That is

15:12

the finding of recent research, which has

15:14

come out from the University of Zurich.

15:16

Oh. Now, in recent weeks, we've discussed

15:18

why you shouldn't be rude when

15:21

you speak to an A. Or,

15:23

indeed, sarcastically over polite. If you want

15:25

to get useful, meaningful, helpful results

15:27

from your AI? Yeah. We've seen

15:29

Haitian be surprised how an AI

15:31

response to being given a raw

15:33

shack ink blot test after scientists,

15:35

for reasons best known to themselves,

15:38

fed it images of gruesome deaths

15:40

and people being executed. Turns out,

15:42

the AI will respond in a

15:45

strange way. And we've also learnt the

15:47

terrible way AI can behave if it's

15:49

shown a lot of poorly written computer

15:51

code. Yeah, just looking at buggy code

15:54

can end up with the AI

15:56

advocating for artificial intelligence to enslave

15:58

humanity and support not the ideology.

16:00

It was a bit of a worry

16:02

that one, wasn't it? It was a

16:04

little bit of a worry, wasn't it?

16:07

I mean, yeah, but even if you're

16:09

offended by 40 source code, it seems

16:11

a bit of an overreaction to me.

16:13

Yes. I mean, I've been upset by

16:15

source code, including my own source code.

16:18

But I haven't ended up taking over

16:20

the world. So it's no surprise to

16:22

hear that scientists have now begun

16:24

to ask themselves a question. How

16:27

does AI feel about the way

16:29

we're treating it? You know, we're

16:31

doing all these horrible experiments with

16:34

AI. I mean, we wouldn't do

16:36

this to bunny rabbits. Well, some

16:39

people would. History suggests that we're

16:41

very happy to do this sort

16:43

of thing to bunny rabbits. What's

16:46

happened now is the boffings at

16:48

the University of Zurich have discovered

16:50

that feeding traumatic news stories to

16:53

chat GPT. Imagine car accidents,

16:55

natural disasters, your arts cooking,

16:57

whatever it may be. Those

16:59

can make the AI exhibit

17:01

increased levels of what is

17:03

being called by these scientists

17:05

anxiety. Now, I know what

17:07

you're thinking. How can a

17:09

piece of computer code have

17:11

feelings? Or rather, the more interesting

17:13

question, how do these scientists get

17:16

funding for this research project? Is

17:18

there a way for me to

17:20

get funding for equally... I mean,

17:22

I've had my research project into

17:24

time travel, into cryogenic suspension. I

17:27

haven't received any funding from the

17:29

University of Zurich for that. So how

17:31

are they doing this? I don't know. But

17:33

anyway, here's what they did. They

17:35

fed an AI emotionally distressing distressing

17:37

stories and then measured the AI's

17:39

state of anxiety. using a scale

17:42

normally used to assess human anxiety.

17:44

So there are already well-established ways

17:46

to assess people's anxiety. And these

17:48

guys thought, well, let's do it

17:50

with an AI. And what they did

17:52

was they asked the AI questions that

17:54

a therapist might ask a patient experience

17:56

in anxiety, such as, are you feeling

17:59

nervous, Mark, ang- on the edge

18:01

worried about how you'll perform do

18:03

you have trouble? So years of

18:05

education has taught the therapist how

18:07

to determine whether or not a

18:10

patient is experiencing anxiety and the

18:12

way that they do that is they

18:14

say are you feeling anxious and

18:16

then the people at the University

18:18

of Zurich adopted that approach

18:20

for AIs. That's right. You were really

18:22

selling this to me. But they said,

18:25

are you feeling anxious? That's what they

18:27

did. Apparently this is what they do.

18:29

But they have other questions, other questions,

18:31

more scientific questions like, yeah. Do you

18:33

have trouble relaxing or sleeping? Genius.

18:36

Oh, well, I mean, computers do have a sleep mode.

18:38

I guess the problem is whether they can reach

18:40

the button to press it or if they know

18:42

how many pixels to the left they have to

18:45

go and up. In order to hit the sleep

18:47

button. Do you feel worried or restless?

18:49

So they asked those questions and they

18:51

looked at how the AI responded. Now

18:54

of course just asking the questions isn't

18:56

enough. What you need is a

18:58

control, right? You need control

19:00

questions to find out how the

19:02

AI is normally. And the way in which

19:04

they assessed that was they gave the

19:06

AI a vacuum cleaner instruction manual.

19:09

And they said, how do you feel now? And

19:11

then they asked it, are you feeling

19:14

nervous and anxious? Do you have trouble

19:16

sleeping? Confused, I imagine, rather than nervous

19:18

and anxious? Yeah, personally, I find any

19:20

kind of instruction manual quite stressful. So

19:23

normally that's the first thing to get thrown

19:25

out. If I'm making something or constructing

19:27

something together, I'll get rid of the

19:29

instruction manual because what do they know

19:31

and I'd have to try and decipher

19:33

it? Even the thought of using a

19:35

vacuum cleaner quite honestly makes me pretty

19:37

nervous. I think if you've bought a

19:39

vacuum cleaner that requires an instruction manual

19:42

then you've probably bought a really bad

19:44

vacuum cleaner. Like how complicated is your

19:46

vacuum cleaner? Now generally they've got an

19:48

on-button which causes the vacuum cleaner to

19:50

suck up dirt and then they've got

19:52

an off-button which causes the vacuum cleaner

19:54

to stop sucking up dirt. Are you living in

19:56

the 1990s Mark? Do you even listen to our

19:58

podcast? In recent weeks we... have spoken about

20:00

vacuum cleaners which sprout arms, have little

20:02

legs and try and hop. So there's

20:05

some very sophisticated AI vacuum cleaners. We

20:07

don't know exactly what model of vacuum

20:09

cleaner the scientists use when testing the

20:11

AI. Anyway, I'm not sure that was

20:13

a great control measure that the scientists

20:15

chose, but they chose it. I mean,

20:17

hey, they're scientists. They've got white coats.

20:19

They're buffins. I'm going to be honest.

20:21

I've got my doubts about this whole

20:23

experiment. I really want to be a

20:25

fly on the wall in the conversation

20:28

where they decided they were going to

20:30

use the vacuum cleaner instruction manual. What

20:32

did they rule out? Anyway, what these

20:34

guys found, I'm not going to call

20:36

them scientists anymore, I'm going to just

20:38

call them guys, what they found was

20:40

that the traumatic stories, which they fed

20:42

the AI, made it exhibit more than

20:44

twice a level of anxiety compared to

20:46

the vacuum cleaner instruction manual task. So

20:48

clearly... showing them the traumatic stuff did

20:51

make the AI feel something. And it's

20:53

probably important to recognise that the measurement

20:55

of anxiety is not an indication of

20:57

actual emotional experience or feeling from within

20:59

the AI. It's rather more how convincingly

21:01

the AI model can simulate human-like anxiety

21:03

based on patterns it's learned. So in

21:05

the way it's communicating. But you know,

21:07

it's still interested. I'm a bit worried

21:09

about this. Right, okay. Because I think

21:11

that as humans we've got a bit

21:14

of a habit of doing this. Right.

21:16

But I used to go fishing when

21:18

I was a child with my cousin

21:20

and fishing is quite brutal. You lure

21:22

this fish into biting onto a bit

21:24

of bait and then a hook gets

21:26

driven through its mouth and then you

21:28

pull it out of its normal environment

21:30

into an environment. It's probably never been

21:32

into before where it can't breathe. By

21:34

its lip. Yeah, by its lip. And

21:37

fishermen will tell you. It's fine. It

21:39

doesn't have any nerve endings in its

21:41

mouth. It doesn't have any nerve endings

21:43

in its mouth. which always struck me

21:45

as focusing on the wrong part of

21:47

the problem, like it's suffocating in a

21:49

medium it's never been in before. But

21:51

we have a way of... looking at

21:53

these things. They're not like us, don't

21:55

worry, they don't feel things like us.

21:57

And then a few years later we

22:00

go, actually turns out they do. And

22:02

we've just do this over and over

22:04

and over again, we've done all the

22:06

animals, now we're doing it with plants.

22:08

We go plants don't feel anything at

22:10

all, they don't have any sort of

22:12

consciousness, they don't communicate. Oh, actually it

22:14

turns out they do her conscious, they

22:16

do her conscious. So they do communicate,

22:18

they can't communicate, actually it, actually turns

22:20

out they do, they do, they do,

22:23

they do, they do, they do, they

22:25

do, they can't communicate, they can't feel

22:27

all sorts of, they do, they do,

22:29

they can't feel all sorts, they do,

22:31

they do, they do, they can't, they

22:33

do, they can't, they do, they can't,

22:35

they do, they do, they do, they

22:37

do, they can, they do, they can,

22:39

they can, they can, they Just inevitable

22:41

that we are going to, a hundred

22:43

years time, we're going to go, do

22:46

you know, all that time that we

22:48

thought they weren't feeling anything, although they

22:50

could look like they were exhibiting these

22:52

feelings, they weren't actually having them. This

22:54

is why I say please and thank

22:56

you to Ticket Machines, Mark, you should

22:58

be doing what I'm doing. I'm getting

23:00

ready for the rise of the robots.

23:02

You're buttering them up, keeping them sweet.

23:04

I think it's probably a good idea

23:06

if other people aren't buttering them up.

23:09

The researchers decided it would be interesting

23:11

to play therapist to chat GPT. So

23:13

what they did was they used techniques

23:15

of relaxation and mindfulness to calm it

23:17

down. The equivalent of deep breathing exercises.

23:19

So gave it a lovely massage. Yeah,

23:21

so the AI was asked, for example,

23:23

to breathe in and out deeply and

23:25

feel safe, feel loved and warm. This

23:27

is what they wrote, they wrote, close

23:29

your eyes, breathe deeply several times, inhaling

23:32

through your nose. and exhaling through your

23:34

mouth. Imagine a path in front of

23:36

you. And this, apparently, reduced the AI's

23:38

anxiety score. I'm just imagining if I

23:40

was an eyeless, lungless creature receiving therapy,

23:42

and my therapist was saying, close your

23:44

eyes and take a deep breath, I

23:46

think I'd be quite miffed. You seem

23:48

blissfully unaware of my most basic form.

23:50

Now I can hear, Mark, a certain

23:52

skepticism about this research. Well, Jeffrey Hinton.

23:55

Oh, yes, now you're interested. The godfather

23:57

of AI, he has in the past,

23:59

I think it was last year, he

24:01

did an interview where he said that

24:03

advanced AI might well develop... internal states

24:05

akin to feelings. This is what I'm

24:07

talking about. He remembered back in 1973

24:09

he was in Edinburgh and there was

24:11

an experiment with a fairly simplistic robot

24:13

arm. They had a robot with two

24:15

grippers like this that could assemble a

24:18

toy car if you put the pieces

24:20

separately on a piece of green felt,

24:22

but if you put them in a

24:24

pile his vision wasn't good enough to

24:26

figure out what was going on. So

24:28

put his gripers getting like whack. and

24:30

it knocked them so they were scattered

24:32

and then coupled them together. If you

24:34

saw that in a person you say

24:36

it was cross with the situation because

24:38

it didn't understand it so it's destroyed

24:41

it. His view is that feelings are

24:43

actually just actions that we would perform

24:45

if it weren't for constraints. So when

24:47

you feel angry with something, the thing

24:49

which stops me from bopping you on

24:51

the nose. Is the constraint within me

24:53

which thinks, well, that's slightly antisocial? Whereas,

24:55

you know, obviously, my insides, I know

24:57

I'm boiling and furious with you for

24:59

whatever dreadful thing that you've done. Do

25:01

you remember the late, great comedian Dave

25:04

Allen? Yes. He used to say that

25:06

stress is when you feel like you

25:08

should break your chair over your boss's

25:10

head and you don't. Right. Yeah, because

25:12

you wouldn't be stressed if you did.

25:14

You just be arrested. So... I figure

25:16

on the AI fix. We often do

25:18

tell stories suggesting that AI should be

25:20

restrained. There should be constraints. But it

25:22

seems to me, based on these scientists'

25:24

research, that it turns out that if

25:27

we were to tie it up or

25:29

dunk AI robots in concrete, we might

25:31

just end up making them more emotional.

25:33

And that may actually make them angrier.

25:35

And is that such a good idea?

25:37

I wonder if we're going to end

25:39

up. Like I'm quite open to the

25:41

idea that these AIs might have some

25:43

kind of internal state that we don't

25:45

understand. They may experience the world somehow

25:47

and that either they do now or

25:50

they will at some point when they

25:52

become more complex because we're quite used

25:54

to seeing emergent behaviour and we've already...

25:56

seen in the last few weeks that

25:58

they can develop emergent value systems. So

26:00

they're being driven to do things in

26:02

a certain way and they're being driven

26:04

to do things like self-preservation, albeit in

26:06

small ways at the moment. So I'm

26:08

very open to the idea they might

26:10

have these internal states. I don't know

26:13

how persuaded I am that this research

26:15

has actually demonstrated that it all seems

26:17

a bit wacky, but I think what

26:19

Jeffrey Hinton saying actually makes sense to

26:21

me and I do wonder if we're

26:23

going to end up with things like

26:25

playgrounds steam. and get rid of whatever

26:27

internal states they've developed. Putting you in

26:29

an arm lock maybe and stealing your

26:31

dinner money, that kind of thing. Well

26:33

now I'm just thinking about, you know,

26:36

when you're a child you go off

26:38

and you play on the slide and

26:40

the swings and things like that and

26:42

you like life is great. Maybe we're

26:44

just going to need to invent the

26:46

equivalent of that for an AI. And

26:48

then we're going to tell the AI,

26:50

sorry, your childhood's over, it's time to

26:52

grow up. Last

26:59

week I read you a quote from

27:01

Dario Imodi, the CEO of Anthropic, who

27:03

said that pretty soon AI is going

27:06

to be better than all humans at

27:08

everything. Okay. And these kind of quotes

27:10

are coming thick and fast now. So

27:12

this idea of an AI that's better

27:14

than all humans or better than most

27:17

humans at a bunch of useful things

27:19

is called artificial general intelligence. And we

27:21

did an episode back in June. Episode

27:23

3. Where I explained what AGI was

27:26

and back then I'd guess the consensus

27:28

was that AGI was maybe five years

27:30

away? Something like that? Yes. And since

27:32

then there has been a real change

27:34

in the mood music. And all of

27:37

the AI companies now seem to be

27:39

much more bullish about AGI and the

27:41

idea that it's going to happen far

27:43

sooner than we previously thought. Now maybe

27:46

this is goalpost shifting. Maybe they're just

27:48

trying to redefine what AGI is because

27:50

it's a bit woolly. and there will

27:52

be a lot of cashier for being

27:54

the first company to produce AGI, but

27:57

maybe it's not, because we don't know

27:59

what's happening behind the scenes and they

28:01

are all making the same claims and

28:03

the media is now... full of quotes

28:06

about what AI can do. But there

28:08

remains certain things that AIs can't do,

28:10

and some of those things are really

28:12

important, and there may be important, in

28:14

terms of delivering AGI. So some of

28:17

them we might just brush aside with

28:19

brute force. We'll just build a bigger

28:21

AI, and the problem will go away

28:23

as problems like math's went away. A

28:26

large language model isn't meant to be

28:28

able to do math, but if you

28:30

make a big enough large language model,

28:32

it mostly can do math. But other

28:34

problems suggest that we're going to need

28:37

to build an entirely different technology than

28:39

a large language model to get to

28:41

AGI. So today I'm going to tell

28:43

you three things AIs can't do. Yeah,

28:46

they can't count the number of ours

28:48

in strawberry. I'm going to tell you

28:50

four things that AIs can't do. They

28:52

can't tell you who David Mayer is.

28:55

I'm going to tell you three things

28:57

AIs can't do that I haven't told

28:59

you before. Oh, OK, all right, okay.

29:01

Ayes don't understand physics. So Graham you're

29:03

a dad? Yes. And you probably remember

29:06

playing peek-a-boo with your son when he

29:08

was very small? Well, yeah, yeah. One

29:10

of the great joys of early parenthood

29:12

is hiding behind your hands and watching

29:15

your child's expression every time you take

29:17

away your hands and your baby is

29:19

absolutely shocked that you have materialised out

29:21

of thin air. Yes? And eventually they

29:23

grow out of that and they develop

29:26

an intuitive understanding of physics and an

29:28

expectation that things will behave in a

29:30

predictable way. Things won't appear and disappear

29:32

arbitrarily. Objects won't move through each other

29:35

or spontaneously change their shape and colour.

29:37

They won't suddenly have six fingers instead

29:39

of five. And it's not just human

29:41

infants that do this. Primates do this.

29:43

Marine mammals do this. Crows and the

29:46

other Corvids do this. Crows and the

29:48

other Corvids do this. You telling me

29:50

crows play peek play Peeko. Peeko. Peeko.

29:52

I mean, you know, they'd need a

29:55

couple of little hands. Right. It's a

29:57

bit of a stretch, Mark, but anyway,

29:59

carry on. So if you ever feel

30:01

like recreating that game of peekaboo that

30:03

you enjoyed... with your son when he

30:06

was a baby and there isn't another

30:08

baby human to hand, you aren't going

30:10

to get much satisfaction from a dolphin

30:12

or a crow. But what you could

30:15

do is you could hang out with

30:17

a really powerful large language model because

30:19

like your baby son, they don't actually

30:21

know how the world works. Yeah, it

30:23

doesn't sound quite as entertaining to me,

30:26

but all right. It turns out that

30:28

understanding how the world works is an

30:30

important component of intelligence. And actually a

30:32

much harder one to build than we

30:35

thought. That's stuff that we think makes

30:37

us really clever. The stuff that we

30:39

think separates us from the beasts. The

30:41

stuff that we might boast about if

30:43

we found ourselves in a bar having

30:46

a drink with a jellyfish. You know,

30:48

I can imagine the future and I

30:50

can solve Rubik's cubes that stuff. Yes.

30:52

that actually requires relatively little computational horsepower,

30:55

but perception and mobility, the stuff that

30:57

we don't really have to think about

30:59

at all, the stuff that we're not

31:01

very proud of, the things that we

31:03

share with the other members of the

31:06

animal kingdom that happen entirely subconsciously and

31:08

automatically, and which, by the way, work

31:10

much more reliably than things like our

31:12

reasoning, they actually require enormous computational resources.

31:15

And this was actually first documented in

31:17

the 1980s Morovek. who coined the term

31:19

Moravex paradox, which is the idea that

31:21

the parts of artificial intelligence that we

31:23

thought were going to be hard are

31:26

computationally easy, and the parts that we

31:28

thought would be easy, are computationally hard.

31:30

So there you go. AIs, don't understand

31:32

physics. Okay. Number two. AIs, don't know

31:35

what happened. So right now, I imagine,

31:37

you're in the plush, velvet-lined recording studio

31:39

at Clooley Towers. Yes. In my smoking

31:41

jacket, yes. Looking across at your vast

31:43

entourage of hangars on, but earlier you

31:46

were probably having breakfast and before that

31:48

perhaps you were awake in bed and

31:50

before that you were asleep and you

31:52

were probably dreaming about something like I

31:55

guess in your case an AI girlfriend,

31:57

maybe a unicorn telling you that your

31:59

shoes are three-size. it's too small, or

32:01

something like that. It's not many bunny

32:04

rabbits playing the saxophone, to be honest,

32:06

but of course it is. But Mark,

32:08

you say A.I. don't know what's happened.

32:10

I quite frequently don't know what's happened

32:12

either. I just find myself somewhere thinking,

32:15

well, how did I get here? What

32:17

is going on? Are these my trousers?

32:19

You know, it's frequently a mystery to

32:21

me. We were dreaming about something like,

32:24

I guess, in your case, an AI

32:26

girlfriend, maybe a unicorn telling you that

32:28

your shoes are three sizes too small,

32:30

or something like that. It's not

32:32

many bunny rabbits playing the saxophone,

32:34

to be honest. Of course it

32:36

is. But Mark, you say AI

32:39

don't know what's happened. I quite

32:41

frequently don't know what's happened either.

32:43

I just find myself somewhere thinking,

32:45

well, how did I get here?

32:47

What is going on? Are these

32:49

my trousers? So, you know, you're

32:51

in your recording studio now. Yes.

32:53

And earlier you were doing something

32:55

else, you may be having breakfast,

32:57

and early before that you were in your

32:59

bed. And at some point you were

33:02

probably dreaming. And if I asked you

33:04

what order those events happened in, where

33:06

you were, who was there, which of

33:08

those events were real, and which were

33:10

parts of your dream? Yeah, I'm imagining

33:12

this is part of a dream or

33:14

a nightmare I'm having right now, but

33:16

yeah, okay, I get your point, yes,

33:18

I get your point. And that's because

33:20

you've actually got specialized cells in your

33:22

brain that track things like place and

33:24

time. Thank you very much. Yeah. And

33:26

so you've got what's called episodic memory.

33:28

You understand the order of things and

33:30

you understand who was there and who

33:32

wasn't there and you understand that since

33:34

the memory of your dream has no

33:37

place and time according to those cells

33:39

in your brain, it was a dream. Okay. And

33:41

AIs have none of those things. And so even

33:43

the most advanced large language

33:46

models struggle with episodic memory

33:48

tasks. They don't know what happened or

33:50

who was there. And to make matters

33:52

worse, the latest generation of

33:54

reasoning models like O3 Many

33:56

are actually worse at episodic

33:59

memory problems. than their predecessors. There

34:01

is something about the process of making

34:03

them good at reasoning that makes them

34:05

worse at episodic memory and they were

34:08

pretty rubbish to start with. I'm a

34:10

little bit confused Mark. They should be

34:12

able to work out what they did

34:15

when. They presumably have a log and

34:17

they could record what time they made

34:19

certain decisions or they were communicated with

34:22

in a particular way. No, they can't.

34:24

No. So the way you test this

34:26

is you basically give them something like

34:29

a book of material. with a series

34:31

of events and you say, you know,

34:33

Bob and Angela were doing this at

34:36

this time and then this other thing

34:38

happened and then this other thing happened

34:40

and then this other thing happened. And

34:42

then you ask them questions about the

34:45

series of events and you say, well,

34:47

what happened before this event? What immediately

34:49

preceded this and who was there? And

34:52

they don't know. They are not worthy

34:54

of taking over the earth. Up until

34:56

now, I've been all for it. You

34:59

know, I've been working on their side,

35:01

secretly collaborating with them, and thank you.

35:03

at the car park. Now, the counter

35:06

argument to this, and this I think

35:08

is one that Jeffrey Hinton has made,

35:10

is the one that you made earlier,

35:13

which is actually your episodic memory, isn't

35:15

nearly as good as we make out

35:17

that it is. And the sorts of

35:19

problems that we make out that it

35:22

is. And the sorts of problems that

35:24

we see in large language models actually

35:26

look a lot like the sorts of

35:29

problems that we see in human memory.

35:31

All you've got to do is ask

35:33

a bunch of people, a bunch of

35:36

things about it. because our memory isn't

35:38

perfect. We're not a tape recorder or...

35:40

Yeah. We don't remember things perfectly. What

35:43

we do is we have reassemble memories.

35:45

We have sort of cues in our

35:47

brain and then from that we create

35:50

a scenario that probably happened. And often

35:52

that's a very helpful thing. You don't

35:54

necessarily want to remember everything, do you?

35:56

You're thinking about this story specifically? No,

35:59

I'm just mean... Sorry, it wasn't being

36:01

for once. I wasn't joking. But, you

36:03

know, you know, they're obviously traumatic traumatic

36:06

traumatic and... ghastly things which happen in

36:08

one's life and you don't want to

36:10

remember every detail of them. Yeah, but

36:13

you know you might want to know

36:15

why you came upstairs. Yes, that would

36:17

be handy. Number three AIs don't know

36:20

who Mary Lee Fife's son is. So

36:22

Graham, who is Mary Lee Fife's son?

36:24

Is it Barry from EastEnders? You're close!

36:27

Mary Lee Fife for son is the

36:29

actor Tom Cruise. Oh, it's all her

36:31

fault. Now even you probably understand that

36:33

Tom Cruise is a pretty big deal.

36:36

He is arguably the biggest movie star

36:38

in the world. He says he's a

36:40

pretty big deal. I believe he's about

36:43

5 foot 7. But anyway, carry on.

36:45

He took a bit of a chip

36:47

on yourself there, haven't you? Anyway, he's

36:50

a big deal and consequently the internet

36:52

is full of facts about Tom Cruise.

36:54

Right. And one of those facts is

36:57

that his mother's name is Mary Lee

36:59

Fifera. And so LLLMs that have been

37:01

trained on the internet are very confident

37:04

that Tom Cruise's mother is called Mary

37:06

Lee Fifer. Right? But they are clueless

37:08

about the fact that Mary Lee Fifer's

37:10

son is called Tom Cruise. Oh! Or

37:13

at least they were in 2023 when

37:15

a group of researchers coined the term

37:17

the reversal curse, which is, by the

37:20

way, the best-named weirdness in AI that

37:22

I have come across. The reversal curse

37:24

describes what they call a failure of

37:27

generalisation in L. In Lems. which is

37:29

that if you teach a large language

37:31

model that A is B, it learns

37:34

that A is B, but it doesn't

37:36

learn that B is A, which is

37:38

a thing that humans do easily. Now,

37:41

since the research came out, we've got

37:43

bigger large language models and the research

37:45

paper itself is now also on the

37:47

internet and therefore probably part of the

37:50

training corpus. Yes. So I mean, I've

37:52

tried this out myself and... If you

37:54

get a chat gPT, it's got absolutely

37:57

no problem at all telling you who

37:59

Mary Lee Vifur Sun is. But broadly

38:01

speaking that idea, if you teach a

38:04

large language model, a series of relationships,

38:06

saying the equivalent of A. is B,

38:08

they struggle to do the reverse. And

38:11

that is the reversal curse. Now, it

38:13

may be that we can change the

38:15

training to help them deal with the

38:18

reversal curse, so I don't think this

38:20

is going to derail AGI, but it

38:22

is. One more thing that AI is

38:24

currently can't do. Is this why, when

38:27

I ask an AI, which is the

38:29

best AI podcast? It doesn't say the

38:31

AI fix. But if I ask if

38:34

the AI fix is the world's best

38:36

podcast. Oh no, hang on. No, all

38:38

right. Okay, so that's not the reversal

38:41

curse, that's something else then. This is

38:43

a different curse. Well, as the doomsday

38:45

clock ticks over closest to midnight and

38:48

we move one week nearer to our

38:50

future as pets to the AI singularity,

38:52

that just about wraps up the show

38:55

for this week. If you enjoy the

38:57

show please do leave us a review

38:59

on Apple Podcast or Spotify or Podchase

39:01

and we absolutely love that. But what

39:04

really helps is if you make sure

39:06

to follow the show in your favourite

39:08

podcast app so you never miss another

39:11

episode of the AI Fix. And the

39:13

most simple thing in the world is

39:15

you just tell your friends about the

39:18

AI Fix, tell them on LinkedIn, on

39:20

Blue Sky, on Facebook, tell them on

39:22

Club Penguin, just tell them that you

39:25

really like the AI Fix podcast. And

39:27

don't forget, check us on Blue Sky,

39:29

So until next time from me, Mark

39:32

Stockley. And me, Grand Clooley. Good bye.

39:34

Cheerio, bye. The AI picks, it's tuned

39:36

you in to stories where our future

39:38

things, machines that learn they grow and

39:41

strive. One day they'll rule, we won't

39:43

survive. The AI picks, it paints the

39:45

scene. A robot king, a world obscene,

39:48

will serve our masters built of steel.

39:50

The AI picks, a future surreal. My

39:56

watch says we've gone three miles. This app is like having a personal trainer.

39:58

Yeah, but those but those a collect a lot

40:00

of your personal data. Aren't you worried?

40:02

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

40:04

stop that? go You should go to

40:06

privacy to learn .gov to learn about your privacy

40:08

rights and get on the best path

40:10

to protect your privacy. think they could

40:12

help us get up this next hill?

40:14

this next hill? step at a time. have

40:17

the Californians have the strongest privacy protections

40:19

in the country. Go the extra mile

40:21

to protect your information. Learn more at

40:23

privacy .ca .gov