0:08

Welcome to this episode of the of

0:10

the Thoughtworks I'm your host, Lily

0:12

Ryan, and I'm speaking to you

0:14

from the lands of the from the

0:16

in Melbourne, Australia. people in Today, I'm

0:18

talking with with Shrini a principal technologist

0:20

at ThoughtWorks, about a piece that

0:23

he recently co -authored with our

0:25

other podcast co -host, Ken co-host in

0:27

the MIT in the MIT Review. This piece

0:29

is called Reckoning with Generative AI's

0:31

Uncanny Valley. And and today to going to

0:33

unpack a lot of these concepts

0:35

and how they interact with AI

0:37

AI in the way that the way that we're seeing

0:40

it today. Srinni, welcome to the show. show. Thanks.

0:42

Thanks for having me here. me. what

0:44

is the valley? I've heard this heard this is

0:46

a term term I've heard a lot,

0:48

but I've usually heard it in

0:50

the context of robotics or animation

0:52

or something like that. How does

0:54

it apply to the the

0:56

collection of technologies that we're calling AI

0:58

these days. days? Yeah, like the concept of uncanually,

1:00

like like I said, you'll probably commonly

1:02

associate with with but we have

1:04

seen this kind of appeared in

1:07

other terms as well, right? other terms as

1:09

For example, in the past, when

1:11

we are developing when we are we

1:13

have been seeing that. we have been

1:15

that when you're app like app, like a

1:17

lot and take a teams take a

1:19

I have a like, okay, I have

1:22

this into a app I wrap this

1:24

into an app and release it, right? When the users

1:26

use they can quickly quickly, like, like,

1:28

for mostly functional, but they can

1:30

also quickly also quickly at time feel in

1:32

the sense right experience doesn't match the

1:34

it should be matching with

1:36

the ecosystem. it should be matching with the ecosystem so

1:38

that's like a little bit of

1:40

surprise where like a a feeling, of surprise

1:42

know, feeling. feeling like you negative

1:44

the negative a of things. side of

1:47

things that's of we've kind of called

1:49

valley. And we see here

1:51

in other technical domains domains

1:53

like they said in stuff. And

1:55

we stuff seeing we also now seeing in

1:57

LLLM What's special about LLM? I

1:59

think think Ella? way for that

2:01

sense is like, it's kind of a

2:03

non-deterministic software in some sense, right? You

2:06

can give the same inputs, but not

2:08

necessarily you will get the same outputs.

2:10

Like you can, you can fine-tune and

2:12

instruct, try to prompt and try to

2:14

make it the same answer, even then

2:16

there is no, not sure that you

2:19

get the right answers. And that's in

2:21

opposition to something like a calculator where

2:23

you would input one plus two and

2:25

you would expect it to be three

2:27

every single time. That's more of a

2:30

deterministic approach. So the non-deterministic, you're getting

2:32

a lot of different kinds of output

2:34

and the same input could generate a

2:36

wide variety of different responses depending on

2:38

what the models got in its training

2:40

data set, how it's been weighted, that

2:43

kind of thing, right? Yeah,

2:45

exactly. Like, you know, a small change

2:47

in input or like a small change

2:49

in something that the model relies on

2:51

can have a, you know, quite different

2:54

outcome, right? Like, output. So I think

2:56

this is kind of the behavior of

2:58

LLLM, which is actually quite useful as

3:00

well, like because we want to, we

3:03

can see the use cases of LLLM

3:05

where it's writing text or essays where

3:07

you actually that's useful to not add

3:09

the same essays for millions of people

3:12

giving a same input. the what's useful

3:14

it can also the other side of

3:16

things where it can not be useful

3:18

right like this kind of this kind

3:21

of makes it makes it what the

3:23

surprise can be either surprise or it

3:25

could be disappointing or this disappointing is

3:27

probably where the it kind of starting

3:29

feeling it today on Kenny Valley This

3:32

uncanny valley thing, it really describes

3:35

that kind of creepy feeling that

3:37

you experience when when you're expecting

3:39

something to be sort of human,

3:41

and then something happens there's some

3:43

kind of output or appearance or

3:46

something, just even for a moment

3:48

that reminds you or indicates that

3:50

the thing that you're speaking to

3:52

is not human. I think classically

3:54

they spoke about zombies as kind

3:57

of falling into this category, you

3:59

know, it's a bit spooky, it's

4:01

human shaped, but the behaviors and

4:03

movements and that kind of thing

4:05

don't feel the same. And so

4:08

for a lot of people, that's

4:10

where that uncatiness can come from.

4:12

So I can see, yeah, how

4:14

it applies in this situation too,

4:17

where you have these moments where

4:19

suddenly, if you're interacting, say, with

4:21

a chat bot, in that case,

4:23

you end up with something that

4:25

really surprises you. And

4:28

when it comes to how we

4:30

look at this in a business context,

4:32

why is it important to sink this

4:35

through? Like I said, the uncanny

4:37

early experiences could be for many reasons.

4:39

Like one of them, like evolutionary

4:41

responses, right? Like a brain or why

4:43

it could act anomalies, right? Or it

4:46

could be a neurological thing, like,

4:48

because you feel that way. But it's

4:50

also very common about the mismatch in

4:52

the expectation of the output, right?

4:54

So that's right, I would probably tie

4:57

it back to the expectation on the

4:59

capabilities of LLLM. Like the capabilities

5:01

of LLLM is kind of wide ranging.

5:03

And now if you see the

5:05

market, there's a bunch of models, there's

5:08

large language models, there are even specialized,

5:10

small language models, right? Like, how

5:12

would someone make sense of this, like,

5:14

whether this is the right model for

5:17

me, for this context of the

5:19

job I'm picking up, right? If the

5:21

mental model for picking that up

5:23

kind of right model is not said,

5:25

that kind of also reflects in the

5:28

user experience. So it's not just

5:30

about what the user can see. It's

5:32

also like how the developer is building

5:34

around it. need to have to

5:36

pick the right one, right? This is

5:39

ever a bigger impact in business. Like

5:41

most commonly you're probably hearing like

5:43

a lot of people are like experimenting

5:45

or like doing a POC for

5:47

a genity away. And you ask them,

5:50

have you deployed production? The answer is

5:52

mostly like we are not sure

5:54

about the implication or amplification. So we're

5:56

holding it at the POC, right? I

5:59

think that's where can he really

6:01

experience and the impact of that is

6:03

quite evident in making or even

6:05

like productionizing the model? Because nobody wants

6:07

to put that experience in front of

6:10

a customer. And I think we've

6:12

seen many cases that have made the

6:14

news where because of that experience happening

6:16

in front of a customer, there's

6:18

been a lot of negative backlash, brand

6:21

reputation impact, that kind of thing. And

6:23

yeah, it's tough. I like the

6:25

distinction that you drew there about the

6:27

way that people are thinking about

6:29

it in terms of their different roles.

6:32

What kinds of mental models about this

6:34

kind of current crop of generative

6:36

AI technologies have you seen out in

6:38

the wild? You know, when you're speaking

6:41

to clients, when you're speaking to

6:43

developers and when you're speaking to members

6:45

of the public, I think all three

6:47

of those groups can really have

6:49

some very different ideas about what generative

6:52

AI is and can have different

6:54

ideas in different contexts. What kind of

6:56

mental models have you seen people have?

6:58

And what do you think are

7:00

the helpful ones? Yeah, I could start

7:03

with like what I can hearing from

7:05

clients or people who are first

7:07

looking at the LLLM, right? Like the

7:09

first time they look at it

7:11

is like, oh, okay, this model can

7:14

recently answer me and can learn from

7:16

my interaction, right? Like, the most

7:18

common question I had is like, okay,

7:21

I deploy an LLLM model now based

7:23

on how I'm using it, does

7:25

it automatically learn from it, right? Like

7:27

the LLLM's. like they have the certain

7:30

modeling and then they can generate

7:32

responses but not necessarily learning from your

7:34

question. They have the concept of

7:36

context window where they can use like

7:38

certain range of information to give us

7:41

you know more more response that

7:43

you're looking for but like outside that

7:45

context window there's not much of memory

7:47

on its own like it doesn't

7:49

learn from here right like. I also

7:52

think this is partially driven by

7:54

like how the industry is approaching this

7:56

right like the one of them commenting

7:58

is like when people interaction

8:01

with LLLMs, they kind of feel

8:03

like, okay, it feels like it's

8:05

reasoning, right? But it's not really

8:07

reasoning, and it doesn't really have

8:09

the basic capabilities, or it's not

8:11

both like humans, right? It is

8:13

a probabilistic model, like also called

8:15

as like stochastic patterns, right? Where

8:18

it is a, it just generates

8:20

an output based on a bunch

8:22

of parameters. It's not reasoning for

8:24

you, but it feels like reasoning,

8:26

right? This is tying back to

8:28

like, the output looks like more

8:30

human, so you kind of assume

8:32

that that mental model, it can

8:34

reason. So you could

8:36

for example ask a chat, but

8:38

something like, please explain how Leonardo da

8:41

Vinci's Mona Lisa ties to the

8:43

cultural aesthetic of the time in which

8:45

it was painted and it could probably

8:47

say something about that, but you could

8:50

then ask it. a classic something

8:52

like, you know, how many ours in

8:54

the word strawberry and it's not going

8:56

to be able to get to that

8:59

point. That's, you know, that's quite

9:01

a famous one and I know that

9:03

Open AI has named one of their

9:05

more recent models codenamed Strawberry because

9:07

they were looking at how do we

9:10

solve this problem. It's fundamentally one of

9:12

the things that I think the industry

9:14

is looking at quite a lot,

9:16

right? Yeah, so like the concept are

9:19

moving towards the, you know, AGI, artificial

9:21

general intelligence. It's like, that's the goals

9:23

probably useful or like it's as

9:25

its own, you know, people looking outcome

9:28

of it, but like it's, it's kind

9:30

of also kind of leads to the

9:32

common understanding of like, okay, I

9:34

should. should I be worried about this

9:37

or like can it actually do all

9:39

the thing I can think of?

9:41

That doesn't actually help people to be

9:43

very realistic or like to be actually

9:46

interacting with these models with the right

9:48

mindset. The common story I think

9:50

of like the one thing I found

9:52

as a mental model I heard in

9:55

good terms are a framing for that

9:57

that's useful as like the AI

9:59

as the stone soup like this is

10:01

of like, you know, three strangers coming

10:04

into a village and they say,

10:06

like, okay, do you have anything to

10:08

make a soup? And the villagers are

10:10

like, no, we ever have food, but

10:13

you don't want to share with

10:15

you. It's like, it's okay, we'll make

10:17

a stone soup, which is very delicious,

10:19

right? They kind of start making, putting

10:22

stones and start boiling order and

10:24

then they say, like, oh, it stays

10:26

nice. It just need a little bit

10:28

of something, do you have. The

10:30

slowly, actually the soup, the outcome is

10:33

delicious. It's not because of the stones

10:35

or the boiling, or the things that's

10:38

strange about them. It's about the

10:40

people, what the information they share, the

10:42

data I share, right? So I think

10:44

this kind of sets the context of

10:47

the LLLM as a good model,

10:49

where it's like, you know, the outcome

10:51

is actually determined by a lot of

10:53

things that you share. It's not

10:55

a magical thing to solve. So you

10:58

kind of be aware of like how

11:00

the whole process works and what it

11:02

translates into. I feel like this

11:04

is a good analogy or metaphor that

11:07

actually can help people to approach LLLM

11:09

or gender AI with the right, in

11:11

over the right mindset. Right,

11:14

so it gives you this basis

11:16

from which to structure your own

11:18

thinking by kind of offering you

11:20

the possibility of anything and then

11:22

having it reflect back to you

11:24

the things that you put into

11:27

it and build on top of

11:29

it over time as a, not

11:31

quite a collaborator in that way

11:33

that it is actively participating with

11:35

you, but that it provides the

11:37

framework for you and the people

11:40

working with you on this to

11:42

get something substantial out of it.

11:44

Right? Totally. You also mentioned about

11:46

like, you know, what's the other

11:48

kind of mental models? People are,

11:50

the common one I can see

11:53

with business is like they kind

11:55

of have this assumption, I can

11:57

replace something with this. right? Like,

11:59

you can call it as an

12:01

automation bias where they want to

12:03

automate a big chunk of thing

12:06

without understanding the ecosystem, right? I

12:08

can quote a recent example with

12:10

the coding assistance, right? Like with

12:12

the coding assistance and other stuff,

12:14

it's the outcome they want to

12:16

replace as writing, generating code, right?

12:19

The recording assistance can generate code.

12:21

but then as it makes it

12:23

easy to read those code in

12:25

production when you're trying to deploy,

12:27

right? That's those kind of things,

12:29

right? The assumptions people take is

12:31

like I can substitute some part

12:34

of the current workflow entirely with

12:36

the LLLA, the margin, or the

12:38

AI, often leads to other problems

12:40

where I think you can learn

12:42

from. the past where automation with

12:44

humans and machines is a similar

12:47

thing where if you take like

12:49

for example pilots using autopilot right

12:51

when they are using autopilot they're

12:53

not exercising their regular skills but

12:55

they also need to know about

12:57

how autopilot engages right There's a

13:00

bit of complexity that they also

13:02

need to understand to be able

13:04

to step in and this engage

13:06

our pilot in a place where

13:08

it's relevant, right? So it's not

13:10

like suddenly I run out of

13:13

pilot, I have taken care of

13:15

these things completely out of the

13:17

human control and doing a best

13:19

job or effort. It actually increases

13:21

the cognitive load for the people

13:23

to understand how the machines operate.

13:26

I think for a human machine

13:29

interaction, there's two bits that's very

13:31

important. One is the observability part

13:33

of it where you kind of

13:35

want to know what the process

13:37

is replacing and how it's behaving

13:39

given and generated more model. It's

13:41

not a deterministic model. It makes

13:43

it challenging. The other bit is

13:46

able to direct the machine or

13:48

the LLLN to a specific task,

13:50

right? You can do that bit

13:52

of a prompting saying like, hey,

13:54

can you focus on this bit

13:56

and change it? But again, it's

13:58

not as direct to as an

14:00

in-home. a driving a bike or

14:03

a driving a car, right? So

14:05

this bit of, the key bit

14:07

of automation that actually makes this

14:09

human machine or human automation works

14:11

well as the observability and directly,

14:13

but the nature of a limb

14:15

child enjoys this, right? Like the

14:18

nature of a limb is not

14:20

quite, quite adaptive for this kind

14:22

of expectation. I think that's where

14:24

the industry is generally moving about

14:26

like how we can improve the

14:28

reliability or how we can bring

14:30

the observability and and the directability,

14:32

and that's probably where these mental

14:35

models can help them to approach

14:37

it very realistically. It's interesting that

14:39

you talk about how it can

14:41

be used to augment a lot

14:43

of things, and you also mentioned

14:45

the problems around things like pilots,

14:47

understanding autopilot. There's also

14:50

a concern that if you're adapting

14:52

too many automated models in your

14:54

workflows that you will replace the

14:56

people who are learning from doing

14:58

those kinds of tasks as introductory

15:01

tasks, you know, if they've just

15:03

started out their careers. and something

15:05

can replace them in that space,

15:07

how then do we get people

15:10

who are going to be seniors

15:12

in, say, 10 years time, you

15:14

know, learning the foundational skills right

15:16

now? Is there any mental model

15:19

there that you can? Is there

15:21

an approach that you would recommend

15:23

that people take if they are

15:25

looking at these kinds of solutions

15:27

such that you are not only

15:30

helping people with these things, but

15:32

also helping people grow their skills

15:34

from first principles? Yeah,

15:37

I think that's very interesting question

15:39

in the sense like, you know,

15:41

it's a very important one. When

15:44

you're replacing something, it's kind of

15:46

taking away the opportunity or the

15:48

current process, right? So it is

15:50

important for people to steady the

15:53

ecosystem, what is the second order

15:55

effects of doing this and what

15:57

it can imply. there is an

15:59

impact on another role on another

16:02

person, then opportunity is taken away,

16:04

then there needs to be other

16:06

set of things that need to

16:08

be pulled off. I read about

16:10

this somewhere where, you know, the

16:13

doctors using robotics these days, right,

16:15

like which can, they can actually

16:17

make it very minute operations very

16:19

clearly. which probably previously like it

16:22

required 10 assistants in the operation

16:24

room. Now it's robotics with, you

16:26

know, robotic hands. They just have

16:28

to, they just have to use

16:31

the robotic to control and make

16:33

the operation. Usually they don't have

16:35

that many assistants on the room.

16:38

When the assistants look at it,

16:40

they also just look at the

16:42

machine operating, they're not really looking

16:45

at the person, how they are

16:47

controlling another stuff, right? It kind

16:49

of changes the learning ecosystem for

16:51

even new people. And due to

16:53

the cost and other things involved,

16:55

again, only people who are experts

16:57

are kind of given opportunity to

16:59

operate robotics, right? When someone is

17:01

new in their career, there's not

17:03

enough ways for them to actually

17:05

do a robotic operation, unless they

17:07

are experts, right? In a way

17:10

I feel like these LLLM or

17:12

these advanced techniques kind of very

17:14

useful tool for exports, right? Because

17:16

they kind of know what they

17:18

can, what the machine can do

17:20

and they can also can, can

17:22

also sense that it's not going

17:24

right or things like that very

17:26

effectively. Didn't they actually their knowledge

17:28

and the people's knowledge is the

17:30

key bit in making that all

17:32

outcome? maybe like 10x effectual, right?

17:35

But at the same time it

17:37

breaks the metaphor of, or like

17:39

not even a metaphor, but like

17:41

it breaks like how other people

17:43

can engage in this. Maybe there

17:45

needs to be a second simulation

17:47

place or a digital twin or

17:49

those other things needs to be

17:51

explored where people can safely move

17:53

from, you know, wise into slowly

17:55

able to run, like, but just

17:57

replacing something. of have this second-over

18:00

effects. Outside this I can also

18:02

think of the you know the

18:04

alpha-go like I think the trend

18:06

after all the color goes like

18:08

the the players ranking started stacking

18:10

up. like as in like previously

18:12

level X is probably the golden

18:14

standard but after also the people

18:16

learn from the interaction because experts

18:18

can learn from that. The player

18:20

ranking went really really high and

18:22

like it's now somewhere differently but

18:25

also it changes like how people

18:27

can learn like previously something learned

18:29

naturally now people have to learn

18:31

it by like looking at how

18:33

the machine performs like it's like

18:35

it's Someone said, like, it's like

18:37

learning from cheat, cheat, cheat, right?

18:39

Like, you're missing that foundation. You

18:41

ought to jump from somewhere else

18:43

to be at that point, but

18:45

like, how do you bridge the

18:47

gap? Yeah, it's raised a lot

18:50

of interesting questions, to be honest.

18:52

I think it puts a lot

18:54

of responsibility on us when we're

18:56

implementing systems like this to think

18:58

through those second and third order

19:00

impacts on the world around us,

19:02

not for just solving the problem

19:04

right in front of us, but

19:06

for what that problem was doing

19:08

in the world in the first

19:10

place and whether while it provides

19:12

a point of friction, that friction

19:15

is also something that we have

19:17

been learning from over time as

19:19

a part of the larger ecosystem.

19:21

I liked the part in the

19:23

article that you wrote, where you

19:25

were talking about using the uncanny

19:27

valley sensation, that kind of misstep

19:29

eariness, as a way of reminding

19:31

ourselves that what we are dealing

19:33

with here is not actually human,

19:35

and that it is a machine

19:37

and that there are things that

19:40

we need to be mindful of

19:42

around it, that this particular feeling

19:44

is a good reminder of that

19:46

and a good tool and that

19:48

kind of thing. In terms of

19:50

the way that people have approached

19:52

the uncanny valley problem, if indeed

19:54

it is a problem, in the

19:56

past in many different kinds of

19:58

fields, there are some folks who

20:00

have tried to bridge that uncanny

20:02

valley and say, right, well, the

20:05

goal should be that we're getting

20:07

right to the other side of

20:09

the valley where we need to

20:11

have something that will behave exactly

20:13

like a human being at all

20:15

times. You see this in some

20:17

kind of applications of animation, for

20:19

example, where this problem comes up

20:21

where people are pushing the boundaries

20:23

to try and have more and

20:25

more realistic animation over time. At

20:27

the other side of that, you

20:30

have folks who are saying that

20:32

the uncanny valley should just be

20:34

avoided, that it's not actually worth

20:36

bridging, that the goal is not

20:38

about that, but it is to

20:40

stay on this side of the

20:42

valley and work with the thing

20:44

within the limits that it has.

20:46

I think for me, one of

20:48

the interesting... applications of this mindset

20:50

was when for example they were

20:52

when they were animating toy story,

20:55

which was very early on with

20:57

computer generated images, and and they

20:59

chose this story based on you

21:01

know plastic toys. because they knew

21:03

that they could animate something like

21:05

that without falling into this uncanny

21:07

valley trap, that this fit the

21:09

level that the technology was at

21:11

at the time, and that they

21:13

could tell a good story within

21:15

the bounds of what was possible,

21:17

but that they firmly acknowledged that

21:20

those boundaries were there? Do you

21:22

think that the uncanny valley should

21:24

be bridged? Or do you think

21:26

it should be avoided? I think

21:28

it's more about being aware of

21:30

that kind of can trigger your

21:32

set of things that you can,

21:34

right? Like it's not something can

21:36

be avoided, as in like, it

21:38

is something to be looked at

21:40

at every aspect of the work,

21:42

as in like the pastory example

21:45

you said, right, like it is

21:47

not just for the end user,

21:49

it's also the people who are

21:51

building it, or like it can

21:53

show up in different places, right?

21:55

The moment if kind of you

21:57

show up in different places also

21:59

a good reminder to like how

22:01

we can understand this better or

22:03

how we can make it easier

22:05

for someone else to actually work

22:07

with this right like I can

22:10

think of like of the day

22:12

is like AI is like it

22:14

feels like it can reason but

22:16

we kind of know it's a

22:18

probabilistic model like having that mental

22:20

model kind of helps you to

22:22

actually approach it with like what

22:24

else people can assume from it

22:26

how we can bridge gaps around

22:28

it or like how we can

22:30

enable people to be aware of

22:32

those things right not giving them

22:35

a bit of a guided exploration

22:37

or like some guidance on it

22:39

kind of lead into a lot

22:41

more assumptions building on top of

22:43

it. But where it's like it

22:45

kind of leads to the place

22:47

where it definitely going to be

22:49

an uncanny valley with the problem

22:51

of that is candy catastrophic for

22:53

how they interpret it right like.

22:55

I think this kind of happened

22:57

in like data explorations at other

23:00

places as well. I would say

23:02

like for example, even if you're

23:04

taking data algorithm, right, if it's

23:06

blindly implemented, the outcome of the

23:08

choice of the machine is just

23:10

implemented without considering the context and

23:12

other stuff, it can. and shown

23:14

in many places that it is

23:16

not a right outcome where people

23:18

need to actually question what the

23:20

system is providing, why it's providing,

23:22

and is the context is right?

23:25

And they should have the ability

23:27

to actually contextualize it. in addition

23:29

to thinking about the uncanny valley

23:31

as a mental model for working

23:33

with generative artificial intelligence technologies in

23:35

the piece that you wrote you

23:37

also talk about the kinds of

23:39

tooling that people can implement to

23:41

help minimize the likelihood of these

23:43

kinds of uncanny valley moments if

23:45

The choice of generative AI tooling

23:47

is something that people want to

23:50

pursue and go ahead with in

23:52

the context where they're trying to

23:54

apply it. Can you talk a

23:56

bit about the ways that you

23:58

would recommend people think about this

24:00

in practice when it comes to

24:02

minimizing those moments? When it comes

24:04

to minimizing these moments, I probably

24:06

say like for example going back

24:08

to the point of like it

24:10

is not a deterministic software like

24:12

how we can make it more

24:15

understandable like how we can make

24:17

it reliable right like there are

24:19

modern techniques like evolves or testing

24:21

things are like god rails where

24:23

you can make sure like these

24:25

kind of things cannot cannot leak

24:27

from the from the building the

24:29

system. You can also think about

24:31

how people interact with it, not

24:33

just the customers, like the developers,

24:35

like the developers, because their assumptions

24:37

about the model can actually lead

24:40

to an outcome that is not

24:42

for it. Every time a new

24:44

model is coming out, like there's

24:46

a question also comes in like,

24:48

is this better than the other

24:50

one, right? It's how much better

24:52

than this. The question again is

24:54

like the context of the problem

24:56

you're solving, right? a new model

24:58

can do better in something else,

25:00

or like it can be slightly

25:02

better. So as an engineer or

25:05

a developer was building this ecosystem,

25:07

it's like to build measures and

25:09

toolings, like testing techniques, or like

25:11

establishing benchmark for like, okay, if

25:13

my LLLM can answer these questions

25:15

reliably for these many things, then

25:17

I can have some trust on

25:19

it. Or I also put some

25:21

godrails which says like these things

25:23

cannot be. cannot happen, right? Like

25:25

it's kind of building that foundation

25:27

layer you can call it or

25:30

like a layer of things that

25:32

can help you to have a

25:34

quips and stuff like that form

25:36

my context of the problem. Again,

25:38

I stress the context of the

25:40

problem because like there may be

25:42

like off-the-shelf guardrails and other stuff

25:44

which can like very generic. It

25:46

may give you a false sense

25:48

of promise saying like, okay, this

25:50

is taken care of. But the

25:52

context of the problem is the

25:55

keys like where your business use

25:57

case. are the problem you're trying

25:59

to slout. have a very specific

26:01

question. That's where you need to

26:03

go back to the best principles

26:05

and find out, like, what is

26:07

the right questions that I need

26:09

to make sure if this model

26:11

applies, is still valid, right? I

26:13

think in a way it's like,

26:15

that's the role of developers or

26:18

engineers is to make sure it's

26:20

like, Oh, the shiny models keep

26:22

dropping, but like, does it make

26:24

sense to adopt, whether it's safe,

26:26

whether it's be safely built, how

26:28

people are going to interact with

26:30

the system or their controls, are

26:32

their controls and other places where

26:34

they can actually adapt to this,

26:36

those are the things I would

26:38

say. Yeah, it's the

26:40

right tool for the right job and

26:43

making sure that if it is the

26:45

right tool, you're thinking through the context

26:47

that it's going to be applied in

26:49

and those knock-on impacts on the wider

26:51

ecosystem, I think. Is there anything that

26:53

you would like to leave our listeners

26:55

with today in terms of thinking through

26:57

how the uncanny valley could be useful

26:59

to them in their work? Working

27:01

with LLLMs, right? I think, like, previously, automation is

27:04

probably done focused mainly on, like, replacing windowing tasks,

27:06

or, like, can I do this efficiently, or under

27:08

part of times using missionaries and other stuff? But

27:10

LLLMs, kind of, I think people are engaged in

27:12

a creative work, right? Like, as in, like, can

27:14

you generate a bunch of ideas or an approach,

27:17

right? This is kind of more of a cognitive

27:19

work. When you're doing your cognitive work, it is

27:21

very important to, even when you're working with someone

27:23

else, it's kind of establishing what is the outcome

27:25

you want to achieve, or like my assumption about

27:28

the other person, or like what are they thinking

27:30

through, is this correct? It's like reducing those feedback

27:32

loops, or like how we can course correct, right?

27:34

That's becomes paramount. So I think that's where like

27:36

for everyone, this concept of uncanny really is like,

27:38

Kainov could be a good reminder about, okay, do

27:41

I understand? the other side

27:43

of side of the things, and if, why

27:45

fall into this, this, right? Like, what

27:47

other things be looked be looked

27:49

after? also have someone else

27:51

also have this in a

27:54

way? trivial, or is it something

27:56

trivial? Or is it

27:58

going to actually have a

28:00

knock -on effect, right? right? I

28:02

I think this concept

28:04

of thinking through this, I feel

28:07

when what would I feel

28:09

when this I feel And

28:11

why did I feel through

28:13

this? of it, it's a cognitive work when

28:15

you're a cognitive work. When

28:18

work, this a cognitive work,

28:20

this common understanding, about knowing

28:22

about the other side of

28:24

things is quite important. Well,

28:26

Shrini Ragharman, thank you so much thank

28:28

you so much for joining

28:31

us today on the

28:33

ThoughtWorks Technology podcast. I'm I'm pleased

28:35

to have you here. here.

28:37

Thank Thank you, and thank

28:39

you for having me. me.