0:00

This episode is brought to

0:02

you by Shopify. Forget the

0:04

frustration of picking commerce platforms

0:07

when you switch your business

0:09

to Shopify. The global commerce

0:11

platform that supercharges you're selling,

0:13

wherever you sell. With Shopify,

0:16

you'll harness the same intuitive

0:18

features, trusted apps, and powerful

0:20

analytics used by the world's

0:22

leading brands. Sign up today

0:25

for your $1 per month

0:27

trial period at Shopify.com/ tech.

0:31

Hello and welcome to the latest dose

0:34

of science news from New Scientists. I'm

0:36

Penny Sasha joined by Roanhooper. Hello. Hello.

0:38

We've got a food special this week

0:40

or maybe a consumption special is a

0:42

better way to put it. We're looking

0:44

at why some people put weight back

0:46

on after they've lost it and we're

0:48

also going to be talking about the

0:50

effect those labels next to food on

0:52

menus have, you know, the ones that

0:54

show you the calorie content. And we're

0:56

also going to hear about quite a

0:58

strange, possibly terrifying new AI study showing

1:01

that a version of chat GPT is

1:03

able to replicate people's unique

1:06

thoughts and personalities. Do I

1:08

need to sort of prod you to

1:10

show everyone that we're really human? And

1:12

you've known me long enough. Can

1:14

you? Can you? What do you think?

1:17

We're going to need a watermark or

1:19

some sort of equipment. Yeah. Well,

1:21

so we're going to go to the

1:24

future later. We're going to start

1:26

with the past, our evolutionary past,

1:28

chimpanzee culture. There's a story this

1:30

week on chimpanzee culture, and it

1:32

gets to something that's really, really

1:34

puzzled me actually, and that is

1:36

the fact that chimps have tools.

1:38

We've known that since Jane Goodall's,

1:40

you know, legendary studies 60 years

1:42

ago, and we know they have

1:44

culture, but their culture's static, right?

1:46

So they've had tools and culture

1:49

for millions of years, but they're

1:51

it's like we had the stone age

1:53

and then we invented you know controlled fire and

1:55

before you know it we're on the moon right

1:57

yeah but chimps have been stuck

1:59

in the Stone Age. Yeah,

2:01

I've never thought about it like

2:03

that. Why is that? Oh, well, so

2:05

that's the thing. No one really knew, but

2:07

we've got a clue to it this week, and

2:09

I've been talking to some primatologists about it,

2:11

and we're going to hear from them. But first,

2:13

we're going to hear from reporter Krista Leste -Lesse.

2:16

She's here to tell us about it. Hi,

2:18

Krista, welcome to the pod. Hi. So,

2:20

go on, give us an intro to this

2:22

study. So, as you say, the

2:24

question is about chimp culture, but not

2:26

so much why they're in the Stone

2:28

Age, or you might even call it

2:30

the Plant Age, since most of the

2:32

tools are made from sticks and leaves,

2:35

which historically actually made their evolution tough

2:37

to study. But rather, it's about whether

2:39

chimp cultures may evolve cumulatively, much

2:41

as human cultures do. The assumption across

2:43

many research groups has been that

2:45

they just don't, but now a big

2:47

study looking at 35 chimpanzee study

2:49

sites across Africa suggests that there is

2:51

a generational buildup of culture. Oh,

2:53

well, that's brilliant, because that's just another

2:55

assumption about chimpanzees that gets completely

2:57

thrown out. That's exciting. Yeah, and it's

3:00

not just a generational buildup, but

3:02

the idea that only we have the

3:04

ability to develop more complex cultures

3:06

over time. I mean, yes, we know

3:08

that many animals are capable of

3:10

what we call social learning, where they learn

3:12

by watching each other, and that

3:14

sort of thing can get passed on to other

3:16

animals of the same and later generation. But

3:18

that's very different from saying, okay, we've got this

3:20

one skill that we know how to do, and it

3:22

works really well. But how about improving on

3:24

that and making it better? And

3:26

for the most part, scientists have believed that

3:29

this is something that really distinguishes humans from other

3:31

species. So I spoke to Andrew

3:33

Whiten of the University of St Andrews

3:35

about this, and I asked him about what

3:37

we know about how chimp culture develops,

3:39

and here he is. Some of

3:41

the chimpanzees' most complex technologies involve what

3:43

we call tool sets, where they use

3:45

two or more tools for the job. For

3:48

example, in the Congo region, they do

3:50

this to feed on underground termite

3:52

nests. First, they get a stout stick,

3:54

and they stab it right down

3:56

through the hard earth, sometimes as far

3:58

as half a meter. to create

4:00

a tunnel down to the nest, then

4:02

they get a different thing, a

4:04

long plant stem, they strip that through

4:06

their teeth to fray the end

4:08

into a little brush. They wet that,

4:10

form it into a kind of

4:12

point, and then they carefully feed that

4:14

down the tunnel. The termites

4:16

bite the invading brush tip and

4:19

the chimpanzee then carefully pulls out the

4:21

tool and eats the termites. If

4:23

you can imagine, you're a primatologist observing

4:25

this, you can't help but suspect

4:27

that all this didn't just spring up

4:29

in one go. Instead, it's

4:32

probably a culture that has had

4:34

to be built up over stages. The

4:37

problem with that is that these

4:39

tools are all perishable. So we

4:41

don't have any record of

4:43

their past history like we

4:45

do for say the wheel

4:47

or many other examples in

4:49

human culture. And that's where

4:51

our new study comes in

4:53

because it uses modern genetics

4:56

to provide us with a kind

4:58

of time machine to look

5:00

back into the past. It's based

5:02

on the fact that in

5:04

chimpanzees, sexually mature females who

5:06

migrate to breed in other groups,

5:08

that's their way of avoiding

5:11

incest, but it spreads their genes

5:13

between groups so that over

5:15

centuries, those genes spread far and

5:17

wide but those female chimpanzees

5:19

could also spread their native cultures.

5:21

For example, if they're one of

5:23

those who have that skill to

5:25

use a tool set to get

5:27

underground termites, they might take that

5:29

to a group that's ignorant of

5:31

it. So what this new

5:33

study has done is

5:35

to measure where distant communities

5:37

share those genetic similarities

5:39

and ask whether they're also

5:41

likely to share cultural

5:43

similarities because migrating females spread

5:45

both their genes and

5:47

their cultural knowledge. And sure

5:50

enough, we found that that

5:52

was strongly the case for those

5:54

complex tool sets. It was

5:56

rather less so where only a

5:58

single tool was used compared

6:00

to... foraging without tools. So what this

6:02

suggests is that inventing those complex

6:04

tool sets is achieved only rarely,

6:06

and for a community to show

6:08

it, they usually have to have

6:10

learned it from a knowledgeable migrating

6:12

female. Yeah, so essentially they have

6:15

these multi- tools that seem to just

6:17

build up from simple tools. And the

6:19

scientists here have very cleverly to use

6:21

genetics to map the way these chimps

6:23

spread with the way a culture spreads.

6:26

And they were taking advantage of methods

6:28

that scientists have already used to study

6:30

community culture in ancient hunter-gatherer groups. Yeah,

6:32

so that's why he says that it's

6:34

effectively a time machine to look in

6:36

a window onto the past. Yeah, exactly.

6:38

And you can sort of trace the

6:41

genetic link from simple tool use to

6:43

more complex ones, because essentially what they're

6:45

seeing is that there aren't any particular

6:47

genetic links among chimps who don't use

6:49

tools. And then there are some links,

6:51

they're fairly weak, but they're still there

6:54

among those that are using simple tools.

6:56

And then we get to the complex

6:58

tools. And those are three to five

7:00

times more likely to be associated with

7:02

sharing genes among those who use them.

7:04

So we're really seeing that evolution there.

7:07

Yeah, there's lots, there's so many lots

7:09

of fascinating things about this work. And

7:11

well, just as an aside, the lead

7:13

authors, Cassandra Gunasekoram of the University of

7:15

Zurich. She was only a PhD student

7:17

when she started, and this is her

7:19

first paper, and it's in science. So.

7:22

It kind of makes you a god

7:24

among men doesn't it if you get

7:26

a nature or science paper from a

7:28

PhD. So I spoke to Cassandra as

7:30

well about the work and I asked

7:32

why it's taken us so long to

7:35

figure this out seeing as we've known

7:37

about chimps and tool use for so

7:39

long and here she is. Cumulative culture

7:41

has been studied a lot in humans

7:43

and only recently our group has worked

7:45

on cultural exchange and mobility patterns in

7:48

hunter-gatherers using similar genetic genetic data. And

7:50

that study had showed that the hunter

7:52

gathers from the west to the east

7:54

of the coast. have been

7:56

connected through migrations

7:58

between groups over the

8:00

past 5 ,000 years

8:03

at least, and

8:05

that these connections correlated

8:07

with cultural exchange

8:09

of, for example, musical

8:11

instruments. So when this

8:13

genetic data came out for chimpanzees, just

8:15

when I had started my PhD, we

8:17

wanted to try this and see if

8:19

we can find some correlation in chimpanzees

8:21

as well. I really think

8:23

that we were very lucky because previously

8:26

this data just didn't exist yet,

8:28

but now we have this huge genetic

8:30

database that comes from collected poo. When

8:33

we then actually found the correlation

8:35

between migration patterns and complex

8:37

tools in chimpanzees, this was of

8:39

course very exciting, and now

8:41

having my first PhD chapter published

8:43

in Science is just amazing,

8:45

and I am very grateful to

8:47

everyone that contributed to this

8:49

paper. Also, I am

8:51

very excited to now continue

8:53

this work, and we wanted

8:55

to quantify the differences in

8:57

migration patterns between humans and

8:59

chimpanzees, and then further analyze

9:02

the complexification of culture to

9:04

have a proper understanding of

9:06

the differences in these processes

9:08

that correlate with cumulative culture

9:10

in these two species. Yeah,

9:12

exactly. So essentially they just didn't have

9:14

the doubt to do this kind of

9:16

research. That's really why it has taken

9:18

so long to get there. But now

9:20

our own cumulative culture, I guess you

9:22

could say, has led us to be

9:24

able to do the genetic analysis and

9:26

even create these huge data banks of

9:29

genetic information and make it accessible to

9:31

scientists around the world. Before that, as

9:33

Andrew was saying, they just didn't have

9:35

the chimps archeological artifacts around because the

9:37

vast majority were made from plants that

9:39

just degraded out of visible history. Yeah,

9:41

and actually we just had a story

9:43

about that in New Scientist about the

9:45

Botanic Age, about where our ancestors used

9:47

plant -based tools and the new discoveries that

9:49

are being made about that. But look,

9:51

there's still the question of, given

9:53

this work, so we know now

9:55

that chimps can evolve their

9:57

own cultures in a cumulative way.

10:00

that we thought only humans did. So

10:02

why are they still scrambling around

10:04

in the dirt and why haven't they

10:06

made particle accelerators yet? You know,

10:08

when people usually tackle that question, it's

10:10

about us, it's like, well, we're

10:12

better. Humans and tech. But here's a

10:14

nice way around to look at

10:16

it. And this is from Andrea Migliano.

10:18

She's a member of the research

10:20

team that put this paper out also

10:23

from Zurich. Over the millions

10:25

of years that separated hunter -gatherers

10:27

from chimpanzees in the human foraging

10:29

niche, there was strong selection

10:31

to access difficult to acquire resources.

10:33

So this meant that humans

10:35

became very dependent on to use.

10:37

While chimpanzees are proficient in

10:39

using tools and they use

10:41

it very regularly, they

10:44

are not dependent on that in

10:46

the same way. Some populations of

10:48

chimpanzees survive without the use of

10:50

any tool for foraging. On top

10:52

of that, humans also had to

10:54

become very cooperative, cooperating, for instance,

10:56

to access these resources to cooperative

10:58

hunting. And the

11:00

combination of these two

11:02

factors meant that human society

11:04

became very fluid. But

11:06

hunter -gatherers nowadays, living today,

11:09

moving very often between

11:11

camps to visit friends

11:13

and to find cooperators. And

11:15

also, of course, to

11:17

access these resources. So

11:19

this exactly process of

11:21

migration between camps

11:23

bring the possibility to recombine

11:26

information and produce new tools

11:28

and creates what we call

11:30

cumulative culture, which is this

11:32

culture that is so complex

11:34

that no single individual can

11:37

recreate from zero. So

11:39

what I think we show

11:41

in this paper that is really

11:43

cool is that the origins

11:45

of this cumulative culture process was

11:47

probably already in the common

11:49

ancestor between humans and chimpanzees. Chimpanzees

11:51

have very limited to use

11:53

and also very limited between group

11:56

migration, but they do have

11:58

both. And this migration is already...

12:00

for the beginning of cumulative culture to

12:02

evolve there. Now when the penny

12:04

dropped for me, the penny dropped.

12:06

No, yeah I never had that

12:08

one before. I didn't mean to.

12:10

Was when it said that chimps

12:12

aren't dependent. on tools like we

12:14

are. So they haven't needed to

12:16

do anything else selectively. It speaks

12:18

to that thing that you know

12:20

for a very long time in

12:22

human history we've seen ourselves as

12:24

like the pinnacle of evolution but

12:26

chimps are the best at being

12:28

chimps. Yeah absolutely and then not

12:30

a social and then the other

12:32

really important point is our cooperation

12:34

right that we're super cooperative. Yeah

12:36

we are and of course we've

12:39

also developed this complex spoken and

12:41

written language right? I mean that

12:43

has really helped share our knowledge

12:45

far and wide so beyond that

12:47

I think that maybe we ourselves

12:49

understand the benefit of sharing our

12:51

knowledge and skills towards some greater

12:53

and wider reaching goal. And the

12:55

goal of this podcast is to

12:57

make more people go to look

12:59

at new scientists. That's the pinnacle

13:01

here. Well what it reminds me

13:03

of is how the idea that

13:05

you get more innovation in in

13:07

cities in human culture because of

13:09

the greater mixing of ideas and

13:11

it kind of seems like in

13:13

chimp populations where you get this

13:15

multi-tool use building up they're like

13:17

prototypes of cities of like centers

13:19

of innovation where you get this

13:21

crossover of ideas. Yeah well they

13:23

certainly could be but you know

13:25

I think it's important to keep

13:27

in mind that like Andrew told

13:29

me cumulative culture remains very quote

13:31

tiny he said to compare it

13:33

to our own. So I don't

13:35

think we can expect them to

13:38

be, you know, creating an industrial

13:40

revolution in their communities, and they're

13:42

not going to be inventing computers

13:44

any time soon. This episode is

13:46

brought to you by, indeed. When

13:48

your computer breaks, you don't wait

13:50

for it to magically start working

13:52

again. You fix the problem, so

13:54

why wait to hire the people

13:56

your company desperately needs? Use Indeed

13:58

sponsored jobs to hire top talent

14:00

facts. And even better, you only

14:02

pay for results. There's

14:04

no need to wait. Speed

14:07

up your hiring with a

14:09

$75 sponsored job credit at

14:11

indeed.com slash

14:13

podcast. Terms and conditions

14:16

apply. Now are you

14:18

feeling hungry, Rhone? Now

14:20

you mention it. Once someone asks, right?

14:22

I could always eat. Great, well I've

14:24

got plenty for you to talk into

14:26

metaphorically. So we were going to talk

14:28

about two studies this week. One's looking

14:30

at calories and dietary choices and the other

14:32

is looking at why when you lose

14:34

weight it can often be really really

14:36

hard to keep that weight off and

14:39

instead you just put it all back on

14:41

again. food for thought. So first

14:43

off I wanted you know when

14:45

you see calories on a menu

14:47

next to a dish does that

14:50

influence your choice at all? If

14:52

I'm like ravenously hungry it might do

14:54

I might choose the one that's so

14:56

you're doing the opposite yeah I'm doing

14:58

what you're not supposed to do yeah

15:01

yeah I was thinking I I'm not

15:03

sure it really like I maybe would

15:05

factor in all that's a bit more

15:07

calories but I'm really not like looking

15:09

oh that's way worse than I expected

15:11

I wouldn't eat it then yeah so

15:13

I've been pretty skeptical about this idea

15:15

of whether putting calories or menus really

15:17

makes any difference at all and that

15:19

made me interested to see there is

15:22

a when Boris Johnson was Prime Minister

15:24

it was something that they introduced then.

15:26

Yeah it was part of this anti-abisty

15:28

campaign that was launched in 2020 there

15:30

were other things going on that year

15:32

but the idea is that large restaurants

15:35

or takeaway businesses have to print calories

15:37

alongside dishes so you know not small

15:39

independence but all the sort of mainstream

15:41

chains and it follows the same sort

15:43

of mandatory labelling that you already get

15:45

in the US. Okay and the whole idea

15:47

obviously is part of this nudge thing isn't

15:49

it's nudging people into what we want them

15:51

to do? Yeah making the right choices that's

15:53

the idea but there may well be other

15:55

effects people may think about like you said

15:57

if you're really hungry do you want to...

16:00

maximize how many calories you're going to

16:02

get, value for money, especially at the

16:04

moment food is really expensive. Yeah, I

16:06

mean, definitely as a student I would

16:08

choose the bulk. The bulk. The bulk

16:10

cheap purchases. Yeah, exactly. So what do

16:12

we know about whether it even works

16:14

or not? So a study just

16:16

out this week has modeled

16:18

the long -term consequences that we

16:20

might expect for this policy,

16:22

looking specifically at body mass

16:24

index and the risk of

16:26

cardiovascular problems in people in

16:28

England over a period of

16:30

20 years, and the results

16:32

as modeled predicted a decrease

16:34

in the prevalence of obesity

16:36

of just 0 .31 percentage

16:38

points and only 730 fewer

16:40

deaths from cardiovascular disease by

16:42

2041. Wow. So that's a lot of

16:45

effort for almost, for not much

16:47

that. Yeah, if you think about it,

16:49

the model predicts 830 ,000 cardiovascular deaths

16:51

during that time. So while 730 fewer

16:53

is not nothing, it's not a big

16:55

dent. The study did suggest that if

16:57

the policy was expanded to all restaurants

17:00

and cafes, like smaller ones, then the

17:02

model could prevent up to 9 ,000, 9

17:04

,200 deaths. So that's a bit more. Yeah.

17:06

Okay. I mean, it's

17:08

hard. It's really hard when you start

17:11

telling people what they can and can't

17:13

eat and what restaurants can and can't

17:15

serve, isn't it? I mean, people love

17:17

to eat certain things, so it

17:19

basically means we have to make available

17:21

the junk food that people like. Well,

17:23

right. Yeah. And some people depend on going

17:25

to fast food chains. It's their only option, given

17:27

the crazy shifts they're working on, the strange

17:29

places there. Yeah. But when I go out to

17:32

a restaurant for a treat, the

17:34

calories doesn't really come into

17:36

it if it's a one -off.

17:38

And also, calories aren't the

17:40

whole story. Some calories are

17:42

better than others. They might

17:44

be in far more nutritious

17:46

foods than a completely processed

17:48

fast food. Yeah. And also,

17:50

I think one of the

17:52

problems is that the idea

17:54

of putting the calories next

17:56

to items on the menu

17:58

and thinking that's going to

18:00

make a meaningful difference, it

18:02

sort of relies on the

18:04

idea that people eat unhealthily

18:06

due to a lack of

18:08

information or a lack of

18:10

education. And I'm not sure

18:12

that... the big problem here in terms of obesity.

18:14

But in any case, a survey of 3,000 people before

18:16

and after this UK legislation went through suggests that having

18:18

calories on menus made no significant change to people's eating

18:20

choices. Yeah. And another study found that a third of

18:22

the calories printed on the menus aren't actually that accurate.

18:24

So the guidelines require calories to lie within a 20%

18:26

margin of error, but one study showed that 35% of

18:29

the calories printed on sampled menus sat outside of

18:31

that margin of error, which sort

18:33

of makes sense if you think

18:35

about it. Food isn't this easily

18:37

standardized reproducible thing. Not the really good

18:40

stuff anyway. the fact that they don't

18:42

make much difference having their labels there.

18:44

Are there people who say they're actually

18:46

actively harmful having the, you know, these

18:48

labels here? Yeah, there's an argument that

18:50

sounds really, you know, it makes sense

18:53

to me that if you're someone who

18:55

struggles with disordered eating like anorexia or

18:57

bulimia. Yeah, if you're trying not to

18:59

let food kind of have that sort

19:01

of triggering a hold over you, having

19:03

the numbers right there next to them,

19:05

that's not not going to be very helpful

19:07

to be very helpful, is it? and when

19:09

the penny drops. You get rowing back on

19:11

that one. Thinking about, there's a lot to

19:13

weigh up when you're thinking about whether

19:16

it's worth doing calorie labelling or

19:18

not. Yeah, exactly. Okay, right. There's

19:20

another food related story that caught my

19:22

eye this week and that's, well, it's one

19:24

we've all probably got direct experience of or

19:26

we know someone who's experienced this and that's

19:28

the fact that some people can't seem to.

19:30

keep weight off when they lose it. You

19:33

know, they might go through so much hell

19:35

to get rid of a lot of weight

19:37

and then, you know, a few months later

19:39

it just piles back on. Yeah, this kind

19:41

of yo-go diet thing, it's an extremely

19:43

well-known phenomenon. And like, for example, around

19:45

85% of people with overweight or obesity

19:47

who managed to lose at least a

19:49

tenth of their body weight, 85% of

19:51

those will then regain it within a

19:53

year. So, you know, this is a

19:56

big problem for people who really are

19:58

trying to sort of sort of... of

20:00

lose that weight. Yeah, I mean it's super

20:02

hard to stick to to watch your calories.

20:04

Yeah, really carefully. And that you know these

20:06

horrible sort of really restricted weight loss

20:08

diet. Yeah. And it is miserable to

20:11

be hungry all the time. But that's

20:13

not the whole story either and we

20:15

know that. It's not just about sort

20:17

of willpower. Studies have also shown that

20:19

the brain interprets a sharp drop in

20:21

body fat as dangerous and responds by

20:23

making the body burn less energy and

20:26

that can potentially cause the body to

20:28

store up more fat again. more about

20:30

these kinds of processes that might be

20:32

happening. Researchers have now analyzed fat tissue

20:34

collected from 20 people with obesity just

20:36

before they had bariatric surgery which shrinks

20:38

the stomach to make people feel fuller

20:40

sooner and is quite an effective way

20:43

of losing a lot of weight. And

20:45

then they looked at these fat cells

20:47

again two years later when they had

20:49

lost at least a quarter of their

20:51

initial body weight. And then for comparison

20:53

they also looked at fat tissue from

20:55

18 people with a healthy weight. And

20:58

so when they compared those two... groups

21:00

before the surgery they found that the

21:02

activity of more than a hundred genes

21:04

differed between them between their fat cells

21:06

and that these differences remained when

21:08

they looked again two years later

21:11

after the bariatric surgery. Wow so

21:13

two years later the bodies of

21:15

previously obese people was still behaving

21:18

differently. Yeah, essentially it's quite astonishing,

21:20

isn't it? Yeah, yeah. And it looks

21:22

like these differences might be increased with

21:24

things like increased inflammation, disruption to how

21:27

fat cells store and burn fat, and

21:29

so you could see how those potentially

21:31

maybe raise the risk of gaining weight

21:33

again, although this experiment didn't sort of

21:35

go and show and show and prove

21:38

that. Okay, and so obviously the big

21:40

question is, can you change... gene activity?

21:42

Can you wrote about that in that

21:44

group? Well, the answer currently, you know,

21:46

if you want to actually kind of

21:48

target those specific genes, is no. Further

21:50

experiments in mice suggest that the changes

21:53

in gene activity that were seen in

21:55

this experiment are probably due to epigenetics.

21:57

So that's, you know, changes on the

21:59

outside. side of the DNA that affects how

22:01

active genes are, and we don't really

22:03

have a way to sort of precision tweak

22:06

that yet. But I don't think that

22:08

means the situation is necessarily hopeless. Our epigenome,

22:10

it changes in response to experience and

22:12

environment. So even though, you know, it's still

22:14

behaving the same way two years later,

22:16

that's not to say it wouldn't change at

22:18

some point. And potentially future studies could

22:20

have a look at what happens over a

22:22

longer period or if there are sort

22:24

of environmental ways that we could sort of

22:26

shift that gene activity the way that

22:29

you want it to go. Yeah. All

22:31

right, since we're talking about obesity

22:33

and weight loss, like the big

22:35

drug at the moment is ozempic.

22:37

The other one is wegovi. Those

22:39

are these glucogen like peptide one.

22:42

Analogs, GLP one, very big names

22:44

of drugs at the moment. How

22:46

do they fit into all this? Yeah,

22:48

I mean, people are saying significant weight loss.

22:50

There's a lot of excitement about it. It's

22:52

still sort of relatively early days. We don't

22:54

know long term, you know, are people going

22:56

to have to take these forever. But I

22:58

think when you sort of look at this

23:01

picture as a whole, what really stands out

23:03

to me is that the tools that we

23:05

currently have that we know are really good,

23:07

a large amount of weight loss, bariatric surgery.

23:09

And now these GLP one analogs, these approaches

23:11

probably work for a variety of reasons. But

23:13

one thing that really stands out to me

23:15

is that they all reduce appetite. And we

23:17

live in these very obesogenic environments now with

23:19

pretty sedentary. We're short on time and

23:21

money and healthy foods are often the cheapest

23:23

and most available. So if you think

23:25

about, you know, telling someone how many calories

23:27

are in something, it's hard to see

23:30

how that's going to have a bigger effect.

23:32

If somebody's body is working in a

23:34

way that it's just driving them to feel

23:36

really hungry and put that weight on. Now,

23:44

for this next story, I'm going

23:46

to break basically break a rule

23:48

of talking about stories in New

23:50

Scientist, which is we're never normally

23:52

allowed to invoke science fiction. But

23:54

this one, I just can't help

23:56

it because, well, you'll see why.

23:58

So this is computer scientists at

24:00

Stanford and they've got this. A

24:02

version of chat GPT to simulate

24:04

a thousand real people and basically

24:06

create digital twins of them. And

24:08

then when you test the AI

24:10

version later, the twin, this digital

24:12

twin manages to just replicate you,

24:14

your personality and really quite accurately.

24:16

Jacob Ehren's here to tell us

24:18

about it. Jacob, look, he's on

24:20

the screen here, it could not

24:22

him. It could not be him.

24:24

Well, take your word for it

24:26

that it's you, but ... I

24:28

promise you, I'm not a deep

24:30

fake. That's what they'd say. Tell

24:32

us about this story, then. So

24:34

the idea is that they've used

24:36

GPT for O, which is the

24:39

kind of most advanced version of

24:41

chat GPT, to interview people for

24:43

two hours, and this is a

24:45

voice interview, so people are essentially

24:47

just sort of naturally speaking and

24:49

having the AI talk back to

24:51

them. And using the transcript that

24:53

they get from that interview, they

24:55

then feed it back into the

24:57

AI and are able to generate

24:59

a replica of that person. Okay,

25:01

so often when you're looking at

25:03

personality in people, you use a

25:05

really standard questionnaire, but what they've

25:07

done here is rather than using

25:09

a generic form, they've actually had

25:11

a chat with the GPT model

25:13

to build up a view of

25:15

their personality. Yeah, so the AI

25:17

was sort of given a list

25:19

of standard questions to ask and

25:21

work through, but it could deviate

25:23

from them, and if the interviewee

25:25

said something in particular, the AI

25:27

was then able to respond to

25:29

it in the way that you

25:31

wouldn't normally get if you're just

25:33

filling out a form. And after

25:36

doing this for two hours, they

25:38

then ran the AI's through a

25:40

variety of tests. And when they

25:42

gave the same test to the

25:44

original humans, they found that the

25:46

AI's were able to replicate their

25:48

answers with 85 % accuracy. If you've

25:50

done this and you've set up

25:52

the AI model, Jake, right, supposing

25:54

they did one of me and

25:56

then they asked my replica, what

25:58

does Roe... and think about, what

26:00

do you think about, say,

26:02

laws against assisted dying, which is a big thing

26:04

at the moment, lots of people, you know,

26:06

lots of things in the news about it? Would

26:09

it give a similar response to the real

26:11

me? You know, is that what they're able to

26:13

do yet? So I think

26:15

it's hard to say whether it

26:17

would be completely accurate, you know,

26:19

if the AI Rowan would say

26:21

exactly what you would say. But

26:23

for example, part of the interview

26:25

involved asking people's views on healthcare,

26:28

you know, their sort of life

26:30

experience which will include people dying,

26:32

their views on vaccines, alternative medicine.

26:34

And you could imagine how with

26:36

all of that information, you could

26:38

infer what someone's views towards assisted

26:40

dying might be, just based on

26:43

sort of, you could cluster people together,

26:45

you know, if you have particular views on

26:47

healthcare, you also perhaps going to have

26:49

particular views on assisted dying. Yeah. Is

26:52

this a terrifying thing? Why

26:54

are they doing this? So the

26:56

goal of the research

26:58

is to replace existing models

27:00

that are used for forecasting policy

27:02

changes. So at the moment, say

27:04

you want to come up with

27:06

a new law or a tax

27:08

change, and you want to know

27:11

how that's going to affect the

27:13

general population, you might look at

27:15

a demographic model say, so you

27:17

could say, well, for people who

27:19

live in cities versus people who

27:21

live in the country, people of

27:23

certain ages and backgrounds, this is

27:25

how they might respond. And you

27:28

can sort of feed that into

27:30

the model and see what what

27:32

their reaction to your policy is

27:34

going to be. Obviously, all of

27:36

that involves quite a lot of

27:38

averaging and guesswork. The hope is

27:40

with something like this by more

27:42

closely replicating individuals. And I should

27:45

say that the 1000 or so

27:47

people they interviewed were representative of

27:49

the US population by interviewing these

27:51

people and being able to replicate

27:53

their answers in more detail, you

27:55

could then be able to get

27:57

sort of more subtle responses to.

28:00

policy changes? Should we worry about

28:02

the ethics of this? Like potentially

28:04

the ability to sort of replicate

28:06

human personalities or their emotions? That

28:08

could be a goldmine for all

28:10

sorts of reasons. I guess marketing

28:12

is the first one that brings

28:14

to mind. So yeah I imagine

28:16

marketers would absolutely love to have

28:19

this. The researchers behind the work

28:21

said they've been really really aware

28:23

about ethics throughout the whole experiment.

28:25

They've been sort of in constant

28:27

communication with their ethical oversight board.

28:29

They put in various controls so

28:31

that if someone who took part

28:33

in this interview now says, oh,

28:35

actually I don't want to be

28:37

involved anymore, they can retract their

28:40

consent, they can delete the model.

28:42

But I think it is a

28:44

bit of a genie out the

28:46

bottle moment, as you say, now

28:48

that someone's shown that this is

28:50

possible, it basically involves just using

28:52

chatGPT and tools that are readily

28:54

available. And if you were willing

28:56

to pay the people to do

28:58

their interviews and things like that,

29:00

I don't see why you couldn't

29:03

replicate this. Obviously, there are sort

29:05

of data protection laws and things

29:07

like that, but those vary from

29:09

country to country. I mean I

29:11

can see this becoming quite normal

29:13

quite quickly and then we'll get

29:15

used to that maybe we'll get

29:17

used to these things around and

29:19

also you know we project a

29:21

lot onto even you know very

29:24

obviously non-human robots and you know

29:26

vacuum cleaners and assign them agency

29:28

you know can imagine all the

29:30

more so in something that's very

29:32

much upper level you know we're

29:34

going to start assuming these are

29:36

real or conscious right? So I

29:38

don't know whether it actually appears

29:40

more conscious than chat-GPT already does.

29:42

It's more specific, I suppose, more

29:45

tailored to an individual, but it's

29:47

going to be just as fluent

29:49

as chat-GPT already is. And actually

29:51

we had a study earlier this

29:53

year showing that 20% of people

29:55

in the US actually already think

29:57

that chat-GPT is sent. I would

29:59

say most people would, or most

30:01

researchers would say that it is

30:03

very much not, but you can

30:05

see how in the public perception

30:08

it does appear to be this

30:10

sort of conscious thing that you

30:12

can interact with. Well, I mean,

30:14

and also, as these get more

30:16

complex, I could see you can

30:18

start to make an argument that

30:20

you might as well call them

30:23

conscious. I mean, your mind very quickly

30:25

goes to that black mirror scenario,

30:27

right, of you, you don't even

30:29

realize you've been replicated and you're

30:31

stuck on a server somewhere. When

30:33

Chris Dokle Walker, the reporter who

30:35

wrote this story for us, kind

30:37

of asked the researchers about this,

30:39

it was not a question they

30:41

wanted to engage with much as

30:43

you can imagine. I mean, they're

30:45

very much more focused on the

30:47

data privacy angle of things, which

30:49

I think is important. You know,

30:51

normally if you take a... part

30:53

in research or data might be

30:55

sort of anonymized, averaged in some

30:57

way. With this, the goal very

31:00

explicitly is to try and replicate

31:02

an individual. And so there are

31:04

data privacy questions and things like

31:07

that. When it comes to saying,

31:09

are these conscious? I think it's

31:11

really difficult, right? Like I said,

31:14

I don't think it's any more

31:16

conscious than chat GPT is normally,

31:18

and I would say chat GPT

31:21

is currently not conscious. Yeah, can I

31:23

just pour a bunch of cold water

31:25

in this for a moment? First off,

31:27

I don't know how it comes when

31:30

you're doing marketing or political modeling, but

31:32

from, say, a biomedical... perspective, 85% accurate,

31:34

is not accurate. You know, that's not

31:36

sort of, that's not particularly useful. And

31:38

also, the thing that I thought was

31:41

really stark about this is that just

31:43

two weeks later, real people when they

31:45

were answering the same questions again answered

31:47

them differently, 19% of the time. So

31:50

people are really variable and we know

31:52

about that from psychology, we change our

31:54

minds from one moment to the next.

31:56

And so doesn't that actually suggest that

31:59

AI models aren't? really useful in

32:01

this way Jacob. Well that's

32:03

really interesting right because in

32:05

in one way and it's

32:07

sort of described in this

32:09

way in the researchers paper

32:11

people are bad models of

32:14

themselves so people were 80%

32:16

accurate at repeating what they'd

32:18

said two weeks ago and

32:20

so then the AI was

32:22

85% accurate sort of repeating

32:24

their answers over those two

32:26

weeks. I think what's interesting

32:28

is they did a comparison

32:31

to those more sort of

32:33

demographic models I was talking

32:35

about earlier and they found

32:37

that the accuracy did increase

32:39

by 14 percentage points. So

32:41

there was a benefit to

32:44

doing this AI version rather

32:46

than the more simplistic demographic

32:48

modeling. You could record a version of

32:50

yourself when you're young and idealistic

32:52

and save that nice young, you

32:54

know, idealistic version before you become

32:56

old and jaded. I'm not speaking

32:58

from personal experience here. I like

33:00

the way you're saying you. One,

33:02

one could. And also, you know,

33:04

I don't want to get too

33:06

black-mirish, but you know, you could

33:08

also see how... you could chat

33:10

to someone after they're dead very

33:12

very easily right if this gets

33:14

a bit better and if you

33:16

spoke to one of these models

33:18

for more than two hours like extensively

33:21

then you could perhaps get a really

33:23

good model of them that you know

33:25

it might be a world leader you know not

33:27

just a loved one that you might want

33:29

to talk to so there's a lot

33:31

of mad possibilities aren't there? Well

33:33

there's a few things to unpack there

33:35

so first of all would talking for

33:38

longer than two hours actually provide more

33:40

gains in a way this is one

33:42

of the biggest questions about AI at

33:44

the moment can you continue scaling the

33:46

more data you feed into an AI

33:48

does it get more and more accurate

33:50

or do we hit a ceiling at

33:52

some point so that they've not done

33:55

the experiment yet and it might be

33:57

that you know the gains after two

33:59

hours on much more. The other thing

34:01

to keep in mind and coming

34:03

back to this idea that people are

34:05

not necessarily good models of themselves,

34:07

is that we actually do a lot

34:09

of work filling in the blanks.

34:11

So take mediums and psychics. They can

34:13

really convince people that they know

34:15

their deep dark secrets when actually they've

34:17

just got a few basic facts

34:19

and then your mind fills in the

34:21

rest. And you go, oh yeah,

34:23

my dead uncle would have said that.

34:25

So I don't see why we

34:28

wouldn't do the same thing with with

34:30

AI. And in a way,

34:32

obviously, these are really impressive

34:34

technological tools, but a lot

34:36

of the magic sort of

34:38

goes on in our own

34:40

brains as much as they're

34:42

going on in the software.

34:44

Not quite sure how that

34:46

reassures me or not. But

34:48

that's it for this week. Thanks to our guests.

34:50

Thanks to you for listening. Do tell all

34:52

your friends about us. Give us a five star

34:55

rating on Spotify, Apple Podcasts everywhere. We'll

34:57

be back next week. Bye, bye for now. Bye.

35:05

This podcast is produced by

35:07

OG Podcasts. Find out

35:09

more at ogpodcasts .co .uk