0:02

Hello, Hello, Welcome to Smart Talks with

0:04

IBM, a podcast from Pushkin

0:06

Industries, iHeartRadio and

0:09

IBM. I'm Malcolm Gabwell. This

0:11

season, we're continuing our conversation with

0:14

new creators visionaries

0:16

who are creatively applying technology

0:18

in business to drive change, but

0:20

with a focus on the transformative

0:22

power of artificial intelligence and

0:25

what it means to leverage AI

0:27

as a game changing multiplier for your

0:29

business. Our guest today

0:32

is doctor David Cox, VP

0:34

of AI Models at IBM

0:36

Research and IBM Director

0:38

of the MIT IBM Watson

0:41

AI Lab, a first of its kind

0:43

industry academic collaboration between

0:46

IBM and MIT focused

0:48

on the fundamental research of artificial

0:51

intelligence. Over the course

0:53

of decades, David Cox watched

0:56

as the AI revolution steadily

0:58

grew from the sim ideas of

1:00

a few academics and technologists

1:02

into the industrial boom we are experiencing

1:05

today. Having dedicated

1:07

his life to pushing the field of AI

1:09

towards new horizons, David has

1:11

both contributed to and presided

1:14

over many of the major breakthroughs

1:17

in artificial intelligence. In

1:19

today's episode, you'll hear David

1:21

explain some of the conceptual

1:23

underpinnings of the current AI

1:26

landscape, things like foundation models

1:28

in surprisingly comprehensible terms,

1:30

I might add, we'll also get into some

1:33

of the amazing practical applications

1:35

for AI in business, as well as what implications

1:37

AI will have for the future of work

1:40

and design. David spoke with Jacob

1:42

Goldstein, host of the Pushkin podcast

1:45

What's Your Problem. A veteran

1:47

business journalist, Jacob has reported

1:49

for The Wall Street Journal, the Miami Herald,

1:52

and was a longtime host of the NPR

1:54

program Planet Money.

1:57

Okay, let's get to the interview.

2:05

Tell me about your job at IBM.

2:08

SO. I wear two hats at IBM. SO

2:10

one, I'm the IBM Doctor of the MI t IBM

2:13

Watson AI Lab. So that's a joint

2:16

lab between IBM and MIT where we

2:18

try and invent what's next in AI. It's been

2:20

running for about five years, and then

2:22

more recently I started as the vice president

2:24

for AI Models, and I'm in charge

2:27

of building IBM's foundation

2:29

models, you know, building these these

2:31

big models, generative models that allow us to have all

2:33

kinds of new exciting capabilities in AI.

2:36

So, so I want to talk to you a lot about

2:38

foundation models, about genitive AI.

2:40

But before we get to that. Let's just spend a minute on

2:43

the on the IBM MI

2:45

T collaboration. Where

2:47

where did that partnership start? How did it originate?

2:51

Yeah, so, actually it turns out that MI T

2:53

and IBM have been collaborating for

2:56

a very long time in the area of AI. In

2:58

fact, the term artificial

3:00

intelligence was coined in a nineteen

3:03

fifty six workshop that was held

3:05

at Dartmouth. It was actually organized by an IBM

3:07

or Nathaniel Rochester, who led

3:09

the development of the IBM seven and one. So

3:12

we've really been together in AIS since

3:14

the beginning, and as

3:16

AI kept accelerating more and

3:18

more and more, I think

3:20

there was a really interesting decision to say,

3:23

let's make this a formal partnership. So IBM

3:25

in twenty seventeen and now, so it'd be committing close to a

3:27

quarter billion dollars over ten years to

3:30

have this joint lab with MIT, and

3:32

we located ourselves right on the campus

3:35

and we've been developing very very deep relationships

3:37

where we can really get to know each other, work

3:39

shoulder to shoulder, conceiving

3:41

what we should work on next, and then executing the projects.

3:44

And it's really very

3:46

few entities like this exist

3:49

between academia industry. It's been really

3:51

fun the last five years to be a part of

3:53

it.

3:53

And what do you think are some of the most important

3:55

outcomes of this collaboration between

3:58

IBM and MIT.

4:00

Yeah, so we're really kind

4:02

of the tip of the sphere for for IBM's

4:05

AI strategy. So we're we're really

4:07

looking what, you know, what's coming ahead, and

4:10

you know, in areas like Foundation models, you know, as

4:12

the field changes and I T

4:15

people are interested in working on you know, faculty,

4:17

students and staff are interested in working on what's the latest

4:19

thing, what's the next thing. We at IBM Research very

4:22

much interested in the same so we can kind

4:24

of put out feelers, you know, interesting things

4:27

that we're seeing in our research, interesting

4:29

things we're hearing in the field. We can go and chase those opportunities.

4:32

So when something big comes, like the big

4:34

change that's been happening lately with Foundation

4:36

Models, we're ready to jump on it. That's

4:38

really the purpose, that's that's the lab functioning

4:41

the way it should. We're also really interested

4:43

in how do we advance you

4:45

know, AI that can help with climate change

4:48

or you know, build better materials

4:50

and all these kinds of things. That are you know, a broader

4:52

aperture sometimes than than what we might

4:55

consider just looking at the product portfolio

4:57

of IBM, and that that gives us again a

4:59

breadth where we can connections that we might

5:01

not have seen otherwise. We can, you

5:03

know, think things that help out society and

5:05

also help out our customers.

5:08

So the last whatever

5:10

six months, say, there has been this

5:14

wild rise in the

5:16

public's interest in AI, right clearly

5:18

coming out of these generative

5:20

AI models that are really accessible, you know,

5:23

certainly chat GPT language

5:25

models like that, as well as models that generate images

5:28

like mid Journey. I mean, can

5:30

you just sort of briefly talk about

5:32

the breakthroughs in AI

5:34

that have made this moment feel so exciting,

5:37

so revolutionary for artificial intelligence.

5:41

Yeah. You know, I've been studying

5:44

AI basically my entire adult

5:46

life. Before I came to IABM, I was a professor

5:48

at Harvard. I've been doing this a long time,

5:51

and I've gotten used to being surprised. It sounds

5:53

like a joke, but it's serious, Like I'm

5:55

getting used to being surprised at the acceleration

5:58

of the pace. Again. It tracks

6:00

actually a long way back. You know, there's

6:03

lots of things where there was an idea that

6:05

just simmered for a really

6:07

long time. Some of the key

6:09

math behind the

6:12

stuff that we have today, which is amazing. There's

6:14

an algorithm called backpropagation, which

6:17

is sort of key to training neural networks that's

6:19

been around, you know, since the eighties in

6:21

wide use. And really

6:23

what happened was it simmered for a

6:25

long time, and then enough

6:28

data and enough compute came. So

6:30

we had enough data because you

6:33

know, we all started carrying multiple

6:35

cameras around with us. Our mobile phones have

6:37

all, you know, all these cameras and this we

6:39

put everything on the Internet, and there's all this data

6:42

out there. We called a lucky break that there

6:44

was something called a graphics processing unit, which

6:46

you know, turns out to be really useful for doing these kinds

6:48

of algorithms, maybe even more useful than

6:50

it is for doing graphics. They're great graphics too.

6:53

And things just kept kind

6:55

of adding to the snowball. So we had

6:57

deep learning, which is sort of a

7:00

a rebrand of neural networks

7:02

that I mentioned from the eighties, and that was

7:04

enabled again by data because we digitalized

7:07

the world and compute because because we

7:09

kept building faster and faster and more powerful computers,

7:12

and then that allowed us to make this big

7:14

breakthrough. And then, you know, more

7:16

recently, using the same building

7:19

blocks, that inexorable rise

7:21

of more and more and more data met

7:24

the technology called self supervised

7:26

learning, where the key

7:29

difference there in traditional

7:31

deep learning, you know, for classifying images,

7:33

you know, like is this a cat or is this a dog? And

7:35

a picture those technologies

7:38

require supervision, so you have to take

7:41

what you have and then you have to label it. So you have to take

7:43

a picture of a cat, and then you label it as a cat,

7:46

and it turns out that you know, that's very

7:48

powerful, but it takes a lot of time to label

7:51

gats and to label dogs, and there's

7:53

only so many labels that us in the world. So

7:55

what really changed more recently is

7:58

that we have self supervised learning where you don't

8:00

have to have the labels. We can just take unannotated

8:02

data. And what that does is it lots you use

8:05

even more data. And that's

8:07

really what drove this latest

8:10

sort of rage. And then and then all of a sudden

8:12

we start getting these these really powerful

8:14

models. And then really this

8:16

has been simmering technologies,

8:19

right, this has been happening

8:21

for a while and progressively

8:23

getting more and more powerful. One of

8:26

the things that really happened with

8:28

CHATGBT and technologies like

8:30

stable Diffusion and mid Journey was

8:33

that they made it visible to the public.

8:36

You know, you put it out there. The public can touch

8:38

and feel and they're like, wow, not only is there

8:40

palpable change and wow this you

8:43

know, I can talk to this thing. Wow, this thing can generate

8:45

an image. Not only that, but everyone

8:47

can touch and feel and try. My

8:49

kids can use some

8:51

of these AI art generation technologies.

8:54

And that's really just launched.

8:57

You know. It's like a propelled slingshot

8:59

at a into a different regime.

9:01

In terms of the public awareness of these technologies.

9:04

You mentioned earlier in the conversation foundation

9:07

models, and I want to talk a little bit about that.

9:09

I mean, can you just tell me, you

9:11

know, what are foundation models

9:13

for AI and why are they a big

9:15

deal?

9:17

Yeah, So this term foundation model

9:19

was coined by a group at Stanford,

9:22

and I think it's actually a really

9:24

apt term because remember

9:26

I said, you know, one of the big things

9:28

that unlocked this latest excitement was

9:31

the fact that we could use large amounts of unannotated

9:34

data. We could train a model. We don't have

9:37

to go through the painful effort of labeling

9:39

each and every example. You still

9:41

need to have your model do something you wanted to

9:43

do. You still need to tell it what you want

9:46

to do. You can't just have a model that doesn't

9:48

have any purpose. But what a foundation models

9:50

that provides a foundation, Like

9:52

a literal foundation, you can sort of stand

9:54

on the shoulders of giants. You can have them these massively

9:57

trained models, and then do a little bit

9:59

on top. You know, you could use just a few

10:01

examples of what you're looking for and

10:04

you can get what you want from the model. So

10:07

just a little bit on top. Now it gets to the

10:09

results that a huge amount of effort used to have

10:11

to put in, you know, to get from the ground

10:13

up to that level.

10:15

I was trying to think of

10:17

of an analogy for sort

10:19

of foundation models versus what came

10:21

before, and I don't know that I came up with a

10:24

good one, But the best I could do was this. I

10:26

want you to tell me if it's plausible. It's

10:29

like before foundation models, it

10:31

was like you had these sort of single

10:33

use kitchen appliances. You could make a

10:35

waffle iron if you wanted waffles, or you could

10:38

make a toaster if you wanted to make toast.

10:40

But a foundation model is like like an oven

10:43

with a range on top. So it's like this machine

10:45

and you could just cook anything with this

10:47

machine.

10:48

Yeah, that's a great analogy. They're

10:51

very versatile. The other

10:53

piece of it, too, is that they dramatically

10:55

lower the effort that it takes

10:57

to do something that you want to do. And

11:00

I used to say about the old world

11:02

of AI, would say, you know, the problem with automation

11:05

is that it's too labor intensive, which

11:07

sounds like I'm making a joke.

11:09

Indeed, famously, if automation does

11:11

one thing, it substitutes machines

11:14

or computing power for labor. Right,

11:16

So what does that mean to say AI

11:18

is or automation is too labor

11:21

intensive.

11:22

It sounds like I'm making a joke, but I'm actually serious. And

11:24

what I mean is that the effort it took

11:27

the old regime to automate something was

11:29

very, very high. So if

11:31

I need to go and curate

11:33

all this data, collect all this data, and then carefully

11:36

label all these examples, that labeling

11:39

itself might be incredibly expensive

11:41

and time, and we estimate anywhere between

11:43

eighty to ninety percent of the effort it

11:46

takes to feel an AI solution actually

11:48

is just spent on data, so

11:50

that that has some consequences, which

11:52

is the threshold for

11:55

bothering. You know, if you're going to

11:57

only get a little bit of value back from

12:00

something, are you going to go through this huge effort

12:02

to curate all this data? And then

12:05

when it comes time to train the model, you need highly

12:07

skilled people defensive

12:09

or hard to find in the labor market.

12:12

You know, are you really going to do something that's just a tiny, little

12:14

incremental thing. Now you're going to do the only

12:16

the highest value things that warrn't right

12:19

level because you.

12:20

Have to essentially build the whole machine

12:23

from scratch, and there aren't many

12:25

things where it's worth that much work to build

12:27

a machine that's only going to do one narrow

12:29

thing that's.

12:31

Right, and then you tackle the next

12:33

problem and you basically have to start over.

12:35

And you know, there are some nuances here, like

12:37

for images, you can pre train a model

12:39

on some other task and change it around. So

12:41

there are some examples of this, like non

12:44

recurring cost that we have in

12:46

the old world too, but by and large, it's just

12:48

a lot of effort. It's hard. It

12:50

takes you know, a large level of

12:52

skill to implement. One

12:55

analogy that I like is, you know,

12:57

think about it as you know, you have a river of

12:59

data, you know, running through your company or

13:01

your institution. Traditional AI

13:04

solutions are kind of like building a dam on

13:06

that river. You know, Dams are very

13:08

expensive things to build. They require

13:10

highly specialized skills and

13:12

lots of planning. And you know, you're

13:15

only going to put a dam on a river

13:17

that's big enough that you're gonna get

13:19

enough energy out of it that it was worth your trouble.

13:21

You're gonna get a lot of value out of that dam. If you have

13:23

a river like that, you know, a river of data, but

13:26

it's actually the vast majority

13:28

of the water you know, in your kingdom

13:30

actually isn't in that river. It's in

13:33

puddles and greeks and vallet bricks.

13:35

And you know, there's a lot of

13:38

value left on the table because it's like, well,

13:40

I can't there's nothing you can do about it. It's just

13:42

that that's too low value,

13:45

so it takes too much effort, so

13:47

I'm just not going to do it. The return on investment just

13:49

isn't there so you just end up not automating

13:51

things. It's too much of a pain. Now

13:54

what foundation models do is they say, well,

13:56

actually, no, we can train a

13:58

base model a foundation they can work on, don't

14:01

We don't care. We not specify what the task is ahead

14:03

of time. We just need to learn about the domain

14:05

of data. So if we want to build something

14:08

that can understand English language,

14:10

there's a ton of English language text available

14:13

out in the world. We can now train models

14:15

on huge quantities of it,

14:18

and then it learned the structure, learned

14:21

how language you know, good part of how

14:23

language works on all that unlabeled data,

14:25

and then when you roll up with your task, you

14:27

know, I want to solve this particular

14:29

problem. You don't have to start

14:31

from scratch. You're starting from a very, very

14:34

very high place. So that just

14:36

gives you the ability to just you know, now,

14:38

all of a sudden, everything is accessible.

14:40

All the puddles and greeks and babbling books

14:42

and gelipons, you know, those are all accessible

14:46

now. And that's that's very exciting.

14:48

But it just changes the equation on what kinds of

14:50

problems you could use AI to solve.

14:52

And so foundation models basically

14:55

mean that automating

14:57

some new task is much less labor

14:59

and intensive. The sort of marginal effort

15:02

to do some new automation thing is much

15:04

lower because you're building on top of the foundation

15:06

model rather than starting from scratch.

15:09

Absolutely, So that is that

15:12

is like the exciting good

15:14

news. I do feel like there's a

15:16

little bit of a countervailing idea that's worth

15:19

talking about here, and that is the idea that even

15:21

though there are these foundation models

15:24

that are really powerful, that are relatively

15:26

easy to build on top of, it's still

15:28

the case right that there is not some one

15:31

size fits all foundation model. So

15:34

you know, what does that mean and why is that

15:37

important to think about.

15:38

In this context? Yeah, so

15:41

we believe very strongly that there isn't

15:43

just one model to rule

15:45

them all. There's a number of reasons why that could

15:47

be true. One which I think is

15:49

important and very relevant today is

15:52

how much energy these

15:54

models can consume. So these

15:56

models, you know, can get

15:59

very very large. So one

16:01

thing that we're starting

16:04

to see or starting to believe, is that you

16:06

probably shouldn't use one giant

16:08

sledgehammer model to solve every

16:10

single problem, you know, like we should pick

16:13

the right size model to solve the problem. We shouldn't

16:15

necessarily assume that we need the biggest,

16:18

baddest model for every little

16:21

use case. And we're also seeing that, you

16:23

know, small models that are trained like to

16:25

specialize on particular domains can

16:28

actually outperform much bigger models. So bigger

16:30

isn't always even better.

16:32

So they're more efficient and they do

16:34

the thing you want them to do better as well.

16:37

That's right. So Stanford, for

16:39

instance, a group of Stanford trained a model. It

16:42

is a two point seven billion parameter model,

16:44

which isn't terribly big by today's standards.

16:46

They trained it just on the biomedical literature,

16:48

you know, this is the kind of thing that universities do.

16:51

And what they showed was that this model

16:54

was better at answering questions about the biomedical

16:56

literature than some models that are one

16:58

hundred billion parameters you any times

17:00

larger. So it's a little bit like

17:03

you know, asking an expert for help

17:05

on something versus asking the smartest

17:07

person you know, the smartest person you

17:09

know, maybe very smart, but they're not

17:11

going to be expertise. And then

17:14

as an added bonus, you know, this is now a much smaller

17:16

model, it's much more efficient to run. We are

17:18

you know, you know, it's cheaper, so

17:21

there's lots of different advantages there. So I

17:23

think we're going to see attention in

17:26

the industry between vendors

17:29

that say, hey, this is the one, you know, big model,

17:31

and then others that say, well, actually, you know,

17:33

there's there's you know, lots of different

17:35

tools we can use that all have this nice quality

17:37

that we outligned at the beginning, and

17:39

then we should really pick the one that makes the most sense

17:41

for the task at hand.

17:44

So there's sustainability basically

17:46

efficiency. Another kind of set

17:48

of issues that come up a lot with ai A

17:50

are bias hallucination. Can

17:53

you talk a little bit about bias and hallucination,

17:56

what they are and how you're working to mitigate

17:58

those problems.

17:59

Yeah, so there are lots of issues

18:01

still as amazing as these technologies are, and

18:03

they are amazing, let's be very clear,

18:06

lots of great things we're going to enable with these

18:08

kinds of technologies. Bias isn't

18:10

a new problem, so you know,

18:13

basically we've seen this

18:15

since the beginning of AI. If you train a model

18:18

on data that has a bias in it,

18:21

the model is going to recapitulate that bias

18:23

when it provides its answers. So every

18:26

time, you know, if all the text you have says,

18:29

you know, it's more likely to refer to female nurses

18:31

and male scientists, then you're going to

18:33

get models that you know. For instance, there was

18:35

an example where a machine learning

18:37

based translation system translated from Hungarian

18:40

to English. Hungarian doesn't

18:42

have gendered pronouns. English does,

18:44

and when you ask them to translate, it would translate

18:46

they are a nurse to she is a nurse,

18:49

would translate they are a scientist to he

18:51

is a scientist. And that's not because the people

18:54

who wrote the algorithm were building in bias

18:56

and coding in like oh, it's got to be this way. It's

18:58

because the data was like that. You know, we

19:01

have biases in our society and

19:03

they're reflected in our data

19:05

and our text and our images everywhere.

19:08

And then the models they're just mapping

19:11

from what they've what they've seen in their training data to

19:13

to the result that you're trying to get them to do and

19:16

to give, and then these biases

19:18

come out. So there's a very

19:20

active program of research

19:23

in you know, we we do quite a bit at IBM

19:25

research and I, but also

19:28

all over the community and industry and academia

19:30

trying to figure out how do we explicitly

19:33

remove these biases, how do we identify them,

19:35

how do you know, how do we build tools that allow

19:37

people to audit their systems to make sure they aren't

19:40

biased. So this is a really important

19:42

thing. And you know, again this was here since

19:45

the beginning, you know, of machine

19:47

learning and AI, but foundation

19:49

models and large language models and generative AI

19:53

just bring it into sharper even sharper focus

19:55

because there's just so much data and it's sort

19:57

of building in baking and all

19:59

these different biases we have. So

20:01

that's that's absolutely a

20:03

problem that these models have. Another

20:06

one that you mentioned was hallucinations. So

20:08

even the most impressive of our models

20:11

will often just make

20:13

stuff up. You know, the technical term

20:15

that the heels chosen is hallucination.

20:18

To give you an example, I asked chat

20:20

tbt to create a biography

20:22

of David Cox IBM, and

20:25

you know, it started off really well, you

20:27

know, the identifying that I was the director of the mt

20:29

IBM Watsonday and said a few words about

20:31

that, and then it proceeded to create

20:33

an authoritative but completely fake

20:36

biography of me where I was British.

20:39

I was born in the UK. I

20:41

went to British university, you know, universities

20:44

in the UK. I was professor in the authority.

20:46

Right, it's the certainty that that is

20:48

weird about it, Right, it's it's dead certain

20:51

that you're from the UK, et cetera.

20:53

Absolutely, yeah, it has all kinds of flourishes,

20:56

like I want awards in the UK. So

20:58

yeah, it's it's problematic

21:01

because it kind of pokes at a lot of

21:03

weak spots in our human psychology,

21:06

where if something sounds coherent,

21:09

we're likely to assume it's true. We're

21:11

not used to interacting with people who eloquently

21:13

and authoritatively, you know, admit

21:16

complete nonsense like yeah, you

21:18

know, you know we can debate about that, but yeah, we.

21:20

Can debate about that. But yes, the

21:23

sort of blive confidence throws

21:25

you off when you realize it's completely wrong.

21:28

Right, that's right. And we do

21:30

have a little bit of like a great and powerful

21:32

aws sort of vibe going

21:34

sometimes where we're like, well, you know, the AI is

21:36

all knowing and therefore whatever

21:38

it says must be true. But but these

21:41

things will make up stuff, you know, very

21:44

aggressively, and you know, you

21:46

everyone can try asking it for their their bio. You'll

21:49

you'll get something that You'll always get something

21:51

that's of the right form, that has the right

21:54

tone. But you know, the facts just aren't necessarily

21:56

there. So that's obviously a problem.

21:58

We need to figure out how to close those gaps, fix

22:00

those problems. There's lots of ways

22:03

we could use them more easily.

22:05

I'd just like to say, faced with the awesome

22:07

potential of what these technologies might do,

22:10

it's a bit encouraging to hear that even

22:12

chat GPT has a weakness

22:14

for inventing flamboyant, if

22:16

fictional versions of people's lives.

22:19

And while entertaining ourselves with chat GPT

22:22

and mid journey is important, the

22:24

way lay people use consumer facing chatbots

22:27

and generative AI is just

22:29

fundamentally different from the

22:31

way an enterprise business uses AI.

22:34

How can we harness the abilities of artificial

22:36

intelligence to help us solve the problems

22:39

we face in business and technology.

22:41

Let's listen on as David and Jacob

22:43

continue their conversation.

22:45

We've been talking in a somewhat abstract way

22:48

about AI in the ways it can

22:50

be used. Let's talk in a little bit

22:52

more of a specific way. Can

22:54

you just talk about some examples

22:57

of business challenges that can be solved

22:59

with automation. With this kind of automation

23:02

we're talking about.

23:03

Yeah, so there really really this guy's

23:06

the limit. There's a whole set

23:08

of different applications that these

23:10

models are really good at. And basically

23:12

it's a super set of everything we used to use

23:14

ALI for in business. So, you

23:17

know, the simple kinds of things are like,

23:19

hey, if I have text and i'm you know, I have product

23:21

reviews and I want to be able to tell if these

23:23

are positive or negative. You know, like let's look

23:25

at all the negative reviews so we can have a human look through

23:27

them and see what was up. Very

23:30

common business use case. You can

23:32

do it with traditional deep learning based

23:34

AI. So so there's things like

23:36

that that are you know, it's very prosaic sort

23:38

that we were already doing it. We've been doing it for a long time.

23:42

Then you get situations that are

23:44

that were harder for the old day. I like, if

23:46

i'm I want to compress

23:48

something like I want to I have like they have

23:51

a chat transcript, like a customer called in

23:53

and they had a complaint, they called

23:56

back. Okay, Now a new you

23:58

know, person on the line needs to go read

24:00

the old transcript to catch up. Wouldn't

24:03

it be better if we could just summarize that,

24:05

just condense it all down a quick little paragraph,

24:07

you know, customer call they were upset about this, rather than

24:09

having to read the blow by blow. There's just

24:11

lots of settings like that where summarization

24:14

is really helpful. Hey, you have a meeting and

24:17

I'd like to just automatically, you

24:19

know, have have that meeting or that email or

24:21

whatever. I'd like to just have a condensed down so I can

24:23

really quickly get to the heart of the matter. These

24:26

models are are really good at doing that. They're

24:28

also really good at question answering. So if

24:30

I want to find out what's how many vacation days

24:33

do I have? I can now interact

24:35

in natural language with a system

24:38

that can go and that has access

24:40

to our HR policies, and I can actually

24:42

have a you know, multi turn conversation

24:44

where I can, you know, like I would have with you

24:46

know, somebody, you know, actual HR

24:49

professional or customer service representative.

24:52

So a big part, you

24:54

know, of what this is doing is it's

24:56

it's putting an interface. You know, when

24:58

we think of computer interfaces, usually thinking about

25:01

UI user interface elements where I

25:03

click on menus and there's buttons and all

25:05

this stuff. Increasingly, now we

25:07

can just talk, you know, you just

25:10

in words. You can describe what you want, you

25:12

want to answer, ask a question, you

25:14

want to sort of command the system to do something,

25:17

rather than having to learn how to do that clicking buttons,

25:19

which might be inefficient. Now we can just sort of spell

25:21

it out.

25:22

Interesting, right, the graphical user interface

25:24

that we all sort of default to, that's

25:27

not like the state of nature, right, That's

25:29

a thing that was invented and just came

25:31

to be the standard way that we interact with computers.

25:33

And so you could imagine, as you're saying,

25:36

like chat essentially chatting

25:39

with the machine could could become

25:41

a sort of standard user interface, just like

25:43

the graphical user interface, did you know

25:45

over the past several decades.

25:47

Absolutely, And I think those kinds of

25:49

conversational interfaces are going to be hugely

25:52

important for increasing our productivity.

25:54

It's just a lot easier if I if I have to

25:56

learn how to use a tool, or I don't have to kind

25:59

of have awkward, you know, interactions

26:01

for the computer. I can just tell it what I want, and I can understand

26:03

it, could you know, potentially even ask questions

26:06

back to clarify and have those kinds of conversations

26:09

that can be extremely powerful.

26:11

And in fact, one area where that's going to I think be absolutely

26:15

game changing is in code. When we write

26:17

code. You know, programming

26:19

languages are a way

26:21

for us to sort of match between

26:24

our very sloppy way of talking and

26:27

the very exact way that you need to command a computer

26:29

to do what you wanted to do. They're cumbersome

26:32

to learn, they can you know, create very complex

26:34

systems that are very hard to reason about. And

26:37

we're already starting to see the ability to just

26:39

write down what you want and AI will

26:41

generate the code for you. And I think we're

26:43

just going to see a huge revolution of like we just

26:45

converse you and we can have a conversation to

26:47

say what we want, and then the computer can

26:50

actually not only do fixed

26:52

actions and do things for us, but it can actually

26:54

even write code to do new things, you know, and

26:57

generate software itself. Given how much

26:59

software we have, of how much craving we

27:01

have for software, like we'll never have enough

27:03

software in our world, uh,

27:05

you know, the ability to have a systems

27:07

as a helper in that, I

27:09

think we're going to see a lot of a lot of value

27:12

there.

27:13

So if you if you think about the different

27:15

ways AI might be applied

27:17

to business, I mean you've talked about a number of the sort

27:19

of classic use cases. What

27:22

are some of the more out

27:24

there use cases. What are some you know, unique

27:27

ways you could imagine AI being applied

27:29

to business.

27:31

Yeah, there's really disguised the limit.

27:33

I mean, we have one project that I'm kind of a fan

27:36

of where we actually were

27:38

working with a mechanical engineering professor

27:40

at MIT working on a classic

27:42

problem, how do you build linkage systems

27:45

which like you imagine bars and joints

27:47

and ogres, you know, the things

27:49

that are.

27:50

Building a thing, building a physical

27:52

machine of some.

27:53

Kind of like real like metal

27:55

and you know nineteenth

27:58

century just old school industrial

28:00

revolution. Yeah yeah, yeah, but you know the little

28:03

arm that's that's holding up my microphone in front

28:05

of me. Cranes get bold, your buildings,

28:07

you know, parts of your engines. This is like classical

28:09

stuff. It turns out that you know humans,

28:11

if you want to build an advanced system, you

28:14

decide what like curve you want to create,

28:16

and then a human together with a computer

28:18

program can build a five or six bar

28:21

linkage, and then that's kind of where you top out. It is

28:23

because it gets too complicated to work more

28:26

than that. We built a generative AI

28:28

system that can build twenty bar linkages,

28:30

like arbitrarily complex. So these are

28:32

machines that are beyond the capability of

28:34

a human to design themselves.

28:38

Another example, we have an AI system

28:40

that can generate electronic circuits. You know,

28:42

we had a project where we're working where we were building

28:44

better power converters which allow our

28:47

computers and our devices to be more efficient,

28:50

save energy, you know, less

28:52

less carbon ote. But I think the world

28:54

around us has always been shaped by

28:56

technology. If you look around, you know, just think

28:58

about how many steps and how people and

29:00

how many designs went into the table

29:02

and the chair and the lamp. It's

29:05

it's really just astonishing. And that's

29:07

already you know, the fruit of automation

29:10

and computers and those kinds of tools. But we're going to see

29:12

that increasingly be product also

29:15

of AI. It's just going to be everywhere around

29:17

us. Everything we touch is going to have to you

29:19

know, helped in some way to get get

29:22

to you by a.

29:23

You know, that is a pretty profound transformation

29:26

that you're talking about in business. How

29:28

do you think about the implications of that, both

29:30

for the sort of you know, business

29:33

itself and also for for employees.

29:37

Yeah, so I think for businesses

29:39

this is gonna cut costs, make

29:42

new opportunities to like customers,

29:44

you know, like there's just you

29:46

know, it's sort of all upside right, like for

29:49

the for the workers, I think the story is mostly

29:52

good too. You know, like how many things

29:54

do you do in your day that you'd

29:57

really rather not right? You know, and we're

29:59

used to have I think, things we don't like automated

30:01

away, you know, we didn't

30:04

you know, if you didn't like walking many miles

30:06

to work, then you know, like you have a car and

30:08

you can drive there. Or we used to have a

30:10

huge fraction over ninety percent of the US

30:12

population engaged in agriculture, and then we

30:15

mechanized it how very few people work

30:17

in agriculture. A small number of people can do the work

30:19

of a large number of people. And then you

30:21

know, things like email, and you know, they've

30:23

led to huge productivity enhancements because

30:25

I don't need to be writing letters and sending them

30:28

in the mail. I can just instantly communicate with

30:30

people. We just become more

30:32

effective. Like our jobs have transformed,

30:36

whether it's a physical job like agriculture

30:38

or whether it's a knowledge worker job where

30:40

you're sending emails and communicating

30:42

with people and coordinating teams, we've

30:44

just gotten better. And you know, the technology

30:46

has just made us more productive. And this is

30:48

just another example. Now, you know,

30:51

there are people who worry that, you know, we'll

30:53

be so good at that that maybe jobs

30:55

will be displaced, and

30:57

that's that's a legitimate concern. But

30:59

just like how in agriculture,

31:02

you know, it's not like suddenly we had ninety percent of

31:04

the population unemployed. You know, people

31:06

transitioned to other jobs.

31:09

And the other thing that we've found, too, is that

31:12

our appetite for doing more

31:14

things is as humans

31:16

is sort of insatiable. So even if we

31:19

can dramatically increase how much you know,

31:21

one human can do, that doesn't

31:23

necessarily mean we're going to do a fixed amount of

31:25

stuff. There's an appetite to have even more, so

31:27

we're going to you can continue to grow the pie.

31:30

So I think at least certainly in the near

31:32

term, you know, we're going to see a lot of drudgery go away

31:34

from work. We're going to see people

31:37

be able to be more effective at their

31:39

jobs. You know, we will see some transformation

31:42

in jobs and what look like. But we've

31:44

seen that before and

31:47

the technology a least has the potential to make our

31:49

lives a lot easier.

31:52

So IBM recently launched

31:54

Watson X, which includes Watson

31:56

x dot AI. Tell me about

31:59

that, Tell me about you know what it is and the new

32:01

possibilities that it opens up.

32:03

Yeah, So Watson next is

32:05

obviously a bit of a

32:08

new branding on the Watson

32:10

brand. TJ. Watson that was the

32:12

founder of IBM and

32:15

our EI technologies have had the Watson

32:17

brand. Watson X is

32:19

a recognition that there's

32:21

something new, there's something that actually has changed

32:23

the game. We've gone from this

32:26

old world of automation is

32:28

to labor intensive to this new world of possibilities

32:31

where it's much easier to use AI. And

32:34

what Watson X does

32:36

it brings together tools for

32:38

businesses to harness that power. So

32:41

whattsonex dot AI foundation

32:44

models that our customers can use. It includes

32:47

tools that make it easy to run, easy

32:49

to deploy, easy to experiment.

32:52

There's a watsonex dot Data component

32:54

which allows you to sort of organize

32:57

and access to your data. So what we're really

32:59

trying to do is give our customers a

33:01

cohesive set of tools

33:03

to harness the value of

33:06

these technologies and at the same time be

33:08

able to manage the risks and other

33:10

things that you have to keep an eye on in

33:12

an enterprise context.

33:15

So we talk about the guests on this

33:17

show as new creators,

33:20

by which we mean people who are creatively

33:22

applying technology in business

33:25

to drive change. And I'm

33:27

curious how creativity

33:30

plays a role in the research that you do.

33:33

I honestly, I think the creative

33:36

aspects of this job,

33:38

this is what makes this work exciting.

33:41

You know, I should say, you know, the folks who

33:43

work in my organization are

33:45

doing the creating, and I.

33:47

Guess you're doing the

33:49

managing so that they could do the creator.

33:52

I'm helping them be their best and

33:55

I still get to get involved in the

33:57

weeds of the research as much as I can.

33:59

But you know, there's something really exciting

34:01

about inventing, you know,

34:04

Like one of the nice things about doing

34:06

invention and doing research on AI

34:08

in industry is it's usually grounded

34:10

and a real problem that somebody's having. You

34:12

know, a customer wants to solve this problem

34:15

it's losing money or there

34:17

there would be a new opportunity. You identify

34:19

that problem and then you build

34:22

something that's never been built before

34:24

to do that. And I think that's honestly

34:26

the adrenaline rush that keeps

34:28

all of us in this field. How

34:30

do you do something that nobody else on

34:33

earth has done before or

34:35

tried before, So that that kind of

34:37

creativity, and there's also creativity

34:39

as well, and identifying what those problems are, being

34:42

able to understand the places

34:45

where you know the technology

34:47

is close enough to solving a problem,

34:49

and doing that matchmaking between problems

34:53

that are now solvable, you know, and

34:55

in AI, where the field is moving so fast, this

34:57

is constantly growing horizon

35:00

of things that we might be able to solve. So

35:02

that matchmaking, I think is also a really

35:04

interesting creative problem. So

35:07

I think I think that's that's that's why it's

35:09

so much fun. And it's a fun environment

35:11

we have here too. It's you know, people drawing

35:14

on whiteboards and writing on

35:16

pages of math and.

35:18

You know, like in a movie, like in a movie.

35:21

Yeah, straight from special casting.

35:23

The drawing on the window, writing on the window in sharp.

35:26

Absolutely so,

35:29

so let's close with the really long

35:31

view. How

35:33

do you imagine AI and

35:35

people working together twenty.

35:38

Years from now? Yeah,

35:42

it's really hard to make predictions. The

35:45

vision that I like,

35:49

actually this came from an MIT

35:51

economist named David Ottur, which

35:54

was imagine AI almost

35:57

as a natural resource. You

35:59

know, we know how natural resources

36:02

work, right, Like there's an ore we can dig up

36:04

out of the earth that comes from kind of springs

36:06

from the earth, or we usually think of

36:08

that in terms of physical stuff. With

36:10

AI, you can almost think of it as like there's a new kind

36:13

of abundance potentially twenty years

36:15

from now where not only can we have

36:17

things we can build or eat or use or burn

36:19

or whatever, now we have, you know, this

36:21

ability to do things and understand things

36:23

and do intellectual work. And I

36:26

think we can get to a world where

36:28

automating things is just seamless. We're

36:31

surrounded by capability

36:33

to augment ourselves to get

36:36

things done. And you

36:38

could think of that in terms of like, well, that's

36:40

going to displace our jobs, because eventually the AI system

36:42

is going to do everything we can do. But you

36:44

could also think of it in terms of like, wow,

36:46

that's just so much abundance that we now have,

36:49

and really how we use that abundance

36:51

is sort of up to us, you know,

36:53

like when you can writing software is super

36:55

easy and fast and anybody can do it. Just

36:58

think about all the things you can do now, think

37:00

about all the new activities, and go out all the

37:02

ways we could use that to enrich our lives.

37:05

That's where I'd like to see us in

37:07

twenty years. You know, we can we

37:09

can do just so much more than

37:11

we were able to do before abundance.

37:14

Great, thank you so much

37:17

for your time.

37:18

Yeah, it's been a pleasure. Thanks for inviting me.

37:22

What a far ranging, deep conversation.

37:25

I'm mesmerized by the vision David just described.

37:27

A world where natural conversation between

37:30

mankind and machine can generate

37:32

creative solutions to our most

37:34

complex problems. A world where

37:36

we view AI not as our

37:38

replacements, but as a powerful

37:41

resource we can tap into and

37:43

exponentially boost our innovation

37:46

and productivity. Thanks so much

37:48

to doctor David Cox for joining us

37:50

on smart Talks. We deeply appreciate

37:53

him sharing his huge breadth

37:55

of AI knowledge with us and for explaining

37:58

the transformative potential of foundation

38:00

models in a way that even I can

38:02

understand. We eagerly await his

38:05

next great breakthrough. Smart

38:08

Talks with IBM is produced by Matt Romano,

38:10

David jaw nishe Venkat

38:13

and Royston Preserve with Jacob

38:15

Goldstein. We're edited by Lydia

38:17

Jane Kott. Our engineers are Jason

38:19

Gambrel, Sarah Buguier and

38:21

Ben Holliday. Theme song by

38:24

Gramosco. Special thanks

38:26

to Carli Megliori, Andy Kelly,

38:28

Kathy Callahan and the eight Bar

38:30

and IBM teams, as well as

38:33

the Pushkin marketing team. Smart

38:35

Talks with IBM is a production of Pushkin

38:38

Industries and iHeartMedia. To

38:40

find more Pushkin podcasts, listen

38:42

on the iHeartRadio app, Apple Podcasts,

38:45

or wherever you listen to podcasts.

38:48

Him Malcolm Gladwell. This is a

38:50

paid advertisement from IBM.