0:02

Hello and welcome to Decoder. I'm Nilay Patel, Editor-in-Chief

0:04

of The Verge, and Decoder is my show about

0:07

big ideas and other problems. Today,

0:09

I'm talking to Demis Hassabis, the CEO

0:12

of Google DeepMind. That's the newly

0:14

created division of Google responsible for

0:16

AI efforts across the company. Google

0:19

DeepMind is the result of what you might call an internal

0:22

merger. Google acquired Demis'

0:24

DeepMind startup in 2014 and

0:27

ran it as a separate company inside

0:29

of its parent organization, Alphabet, while

0:31

Google itself had an AI team called Google

0:34

Brain. Google's been showing off AI

0:36

demos from both groups for years now, but

0:38

with the explosion of chat GPT and

0:40

a renewed threat from Microsoft in Search,

0:43

Google and Alphabet CEO Sundar Pichai

0:45

made the decision to bring DeepMind into

0:47

Google itself earlier this year, creating

0:50

Google DeepMind. What's interesting is

0:52

that Google Brain and DeepMind were not necessarily

0:55

compatible or even focused on

0:57

the same things. DeepMind was famous

0:59

for applying AI to things like games

1:02

and protein folding simulations. The

1:05

AI that beat world champions at Go, the

1:07

ancient board game? Yeah, that was DeepMind's

1:09

AlphaGo. Meanwhile, Google Brain

1:11

was more focused on what's come to be the familiar

1:13

generative AI toolset. Large

1:15

language models for chatbots, editing features

1:18

in Google Photos, and so on. So

1:20

this was a big structure decision with a goal of being

1:22

more competitive and faster to market with

1:24

AI products, but Demis had to manage

1:27

a culture clash between two very different

1:29

organizations.

1:30

And the competition isn't just OpenAI and

1:32

Microsoft.

1:33

You might have seen a memo from a Google engineer

1:36

floating around the web recently claiming that Google

1:38

has quote no moat in AI

1:40

because open source models running on commodity

1:42

hardware are rapidly evolving and

1:45

catching up to the tools run by the tech giants.

1:48

I asked Demis about that memo and he confirmed

1:50

that it was real, but he said it was part of Google's

1:52

debate culture and that he disagreed with it.

1:54

And we talked about his other ideas about where Google's

1:57

competitive advantages might come into play.

1:59

Of course, we also talked about AI risk

2:02

and especially artificial general intelligence.

2:04

Demis is not shy that his goal is building

2:06

an AGI. And we talked about what risks

2:09

and regulations there should be and on what

2:11

timeline. Demis recently signed

2:13

on to a 22-word statement

2:15

about AI risk with OpenAI, Sam

2:17

Altman, and others that simply reads, quote,

2:20

mitigating the risk of extinction from AI

2:22

should be a global priority alongside

2:24

other societal-scale risks such as pandemics

2:27

and nuclear war.

2:29

So that's pretty chill. But is that

2:31

the real risk right now or just distraction

2:33

from other, more tangible problems like

2:36

AI replacing a bunch of labor in various

2:38

creative industries?

2:39

We also talked about the new kinds of labor that AI

2:41

is creating. Armies of low-paid

2:44

taskers classifying training data

2:46

in countries like Kenya and India.

2:48

We just did a big feature on these taskers, which

2:50

we'll link to in the show notes. I wanted

2:52

to know if Demis thought these jobs were here to stay

2:55

or just a temporary side effect of

2:57

the AI boom. I got

2:59

to say, this one really hits all the decoder high

3:01

points. There's the big idea of AI. There's

3:04

all the problems that come with it, an infinite

3:06

array of complicated decisions to be made, and

3:08

of course, a gigantic org chart

3:10

change in the middle of it all.

3:12

Demis and I got pretty into the weeds and I still don't think

3:14

we covered it all, so we'll have to have him back soon.

3:16

A one-hour AI metadata

3:19

show, I promise you. We're going to make this happen. Okay,

3:22

Demis Asabas, CEO of Google DeepMind.

3:24

Here

3:25

we go.

3:37

Demis Asabas, you are the CEO

3:39

of Google DeepMind. Welcome to Decoder. Thanks

3:42

for having me. I am very excited to talk to

3:44

you. I don't think we have ever had

3:46

a more perfect decoder guest.

3:48

There's a big idea in

3:51

AI. It comes with a bunch of challenges and

3:53

problems. With

3:55

you in particular, there's a

3:58

gigantic org chart. move

4:00

and a set of high-stakes decisions to be made,

4:02

I am thrilled that you are here.

4:06

Glad to be here. Let's start with Google DeepMind

4:08

itself. Google DeepMind is a new

4:11

part of Google that is constructed of two

4:13

existing parts of Google. There

4:15

was Google Brain, which was the AI

4:17

team we were familiar with as we covered Google. That

4:20

was run by Jeff Dean. And there was DeepMind,

4:23

which was your company that you founded. You

4:25

sold it to Alphabet in 2014. You

4:27

were outside of Google. It was run

4:29

as a separate

4:29

company, instead of holding company

4:32

Alphabet structure until just now. Start

4:35

at the very beginning. Why were DeepMind and Google Brain

4:38

separate to begin with? Well, as you mentioned,

4:40

we started DeepMind actually back in 2010, a long

4:42

time ago now, especially

4:45

in the age of AI. So that's prehistory.

4:50

Myself and the co-founders, we realized

4:53

coming from academia and seeing what was going on there,

4:55

things like deep learning had just been invented. We

4:58

were big proponents of reinforcement learning. We

5:00

could see GPUs and other hardware was

5:02

coming online. A lot of great progress

5:05

could be made with a focused effort on

5:07

general learning systems and also taking

5:09

some ideas from neuroscience and how the

5:12

brain works. And so we put all those ingredients

5:14

together back in 2010. We

5:16

had this thesis, we'd make fast progress, and that's what

5:18

happened with our initial game

5:20

systems. And then we decided in 2014

5:23

to join forces with Google at the time,

5:26

because we could see that a lot more compute

5:28

was going to be needed. And obviously, Google has

5:31

the most computers and had the most computers in the world.

5:34

And so that was the obvious

5:36

home for us to be able to focus

5:38

on pushing the research

5:39

as fast as possible.

5:42

So you were acquired by Google and then somewhere

5:44

along the way, Google reoriented

5:47

itself. They turned into Alphabet and

5:49

Google became a division of Alphabet. There's other

5:52

divisions of Alphabet and DeepMind

5:54

was out of it. And that's just the part

5:56

I want to focus on

5:57

right here at the beginning. There

6:00

was what Google was doing with Google Brain,

6:03

which is a lot of LLM research. I recall six

6:05

years ago, Google was showing off LLMs at Google

6:08

IO. But DeepMind was focused

6:10

on winning game, AlphaGo and

6:13

protein folding, a very different kind of AI

6:15

research fully outside of Google. Why

6:18

was that outside of Google? Why was that an alphabet

6:20

proper?

6:20

That was part of the agreement as we were

6:23

required, was that we would pursue

6:25

pushing forward research into

6:28

general AI or sometimes called AGI.

6:30

The system that out of the box can

6:33

operate across a wide range

6:35

of cognitive tasks and

6:38

basically has all the cognitive capabilities

6:41

that humans have. And also using

6:43

AI to accelerate scientific

6:45

discovery. That's one of my personal passions.

6:48

That explains projects like AlphaFold that I'm

6:50

sure we're going to get back to. But also from

6:52

the start of DeepMind and actually prior to even DeepMind

6:55

starting, I believe that games

6:57

was a perfect sort of testing or proving ground

7:00

for developing AI algorithms

7:02

efficiently. Quickly, you can

7:04

generate a lot of data and the objective functions

7:06

are very clear, obviously winning games

7:09

or maximizing the score. So

7:11

there were a lot of reasons to use games in

7:13

the early days of AI research. And

7:15

that was a big part of why we were so successful and

7:17

why we were able to advance so quickly with things like

7:20

AlphaGo, the program that beat

7:22

the world champion at the ancient game

7:24

of Go. Those were really important

7:26

proof points for the whole field really

7:29

that these sort of general learning techniques

7:31

would work. And of course, we've done a lot of

7:33

work on deep learning and neural networks as well.

7:36

And our specialty, I suppose, was combining

7:38

that with reinforcement learning to

7:40

allow these systems to actively

7:42

solve problems

7:43

and make plans and do things

7:45

like win games. And in terms

7:47

of the differences, we always

7:50

had that sort of remit to push the research agenda

7:52

and push things advanced science. And that

7:54

was very much the focus we were given

7:56

and very much the focus that I wanted to have.

8:00

that the internal Google AI teams like Google

8:02

Brain, they had slightly different remits

8:04

and were a bit closer to product, and

8:06

obviously to the rest of Google, and infusing

8:08

Google with amazing AI technology.

8:11

And we also had an applied division that was introducing

8:14

DeepMind technology into Google products too,

8:16

but the cultures were quite different and the remits

8:19

were quite different.

8:20

So from the outside, the timeline

8:23

kind of looks like this. Everyone's

8:25

been working on this for ages. We've all been talking about

8:27

it for ages.

8:29

It is a topic of conversation for

8:31

a bunch of nerdy

8:33

journalists like me, a bunch of researchers,

8:35

we talk about it in the corner at Google events. Then

8:38

chat GPT is released, not even

8:40

as a product. I don't even think Sam would call it a great

8:42

product when it was released, but it was just released

8:44

and people could use it and everyone freaked out, and

8:47

Microsoft releases Bing based on chat GPT,

8:50

and the world goes upside down and

8:52

Google reacts by merging DeepMind

8:54

and Google Brain.

8:56

That's what it looks like from the outside. Is

8:58

that what it felt like from the inside?

9:00

That timeline is correct, but

9:03

it's not these sort of direct consequences.

9:05

It's more indirect in a sense. So Google

9:07

and Alphabet have always run like this. They

9:10

let many flowers bloom, right? And

9:12

I think that's always been the

9:14

way that, even from Larry and Sergey

9:16

from the beginning, set up Google and it's served them very

9:18

well. And it's allowed them to organically

9:21

create incredible things and become the amazing

9:23

company that it is today. On the research

9:25

side, I think it's very compatible

9:28

with doing research, which is another reason we

9:30

chose Google as our partners back in 2014. I

9:33

felt they really understood what fundamental

9:36

and blue sky research was, ambitious research

9:38

was, and they were gonna facilitate us

9:40

being, and enable us to be super

9:43

ambitious with our research. And you've seen the

9:45

results of that right by any measure, alpha

9:47

go alpha fold, but more than 20 nature

9:50

and science papers and so on. All the normal

9:52

metrics one would use for

9:54

really cutting edge, delivering amazing

9:57

cutting edge research, we were able

9:59

to do.

9:59

But in a way, what chat GPT

10:02

and the large models and the public reaction

10:04

to that confirmed is that AI

10:07

has entered a new era. And

10:09

by the way, those of you, it was a little bit surprising

10:12

for all of us at the coal face, including

10:14

I think open AI, how viral that

10:16

went, because we all had

10:18

us and some other startups

10:21

like Anthropic and OpenAI. We all had these

10:24

large language models. They're all roughly the same

10:27

capabilities. And it was surprising,

10:29

not so much what the technology was, because we all

10:31

understood that, but the public's appetite

10:33

for that, I would say. And obviously

10:35

the buzz that generated. And I think that's

10:37

indicative of something we've all been feeling for the

10:40

last, I would say, two, three years, which is these

10:42

systems are reaching a level of maturity

10:45

now and sophistication, where

10:48

it can come, really come out of the

10:50

research phase and the lab

10:53

and go into powering incredible

10:55

next generation products and experiences

10:58

and also breakthroughs, things

11:00

like AlphaFold, directly being

11:02

useful for biologists. And so to

11:05

me, this is just indicative

11:08

of a new phase that AI

11:10

is in of being practically useful to

11:12

people in their everyday lives and actually

11:14

being able to solve really hard real world

11:16

problems that really matter, not just the

11:18

curiosities or fun like games. When

11:21

you recognize that shift, I think

11:23

that necessitates a change in your approach

11:26

as to how you're approaching the research and

11:29

how much focus you're having on products and those

11:31

kinds of things. And I think that's what we

11:34

all came to the realization of, which

11:36

now was the time to streamline our

11:38

AI efforts and focus the

11:41

more the obvious conclusion of that

11:43

was to do the merger.

11:44

I want to just stop there for one

11:47

second and ask a philosophical question. It

11:50

feels like the chat GPT

11:52

moment that led to this AI explosion this

11:54

year

11:55

was really rooted in the AI being

11:57

able to do something that regular people could do. I

12:01

want you to write me an email. I want you to write me a screenplay.

12:03

And maybe it's a C, maybe the output of the LLM

12:06

is a C plus,

12:07

but it's still like something I can do, right? Like people

12:09

can see it. I want you to fill out the rest of this photo.

12:12

That's how many people can imagine doing, maybe

12:14

they don't have the skills to it, but they can imagine doing it.

12:17

All the previous AI demos that we

12:19

have gotten, even yours, Alpha Fold, you're

12:21

like, this is gonna

12:22

model all the proteins in the world. Like

12:25

I can't do that. That's like, great, a computer should do

12:27

that. Like even a microbiologist might think,

12:29

that is great, I am very excited that a computer can

12:31

do that because I'm just looking at how much time it would

12:33

take us and there's no way we could ever do it. I

12:36

wanna beat the world champion at Go. I

12:39

can't do that. It's like, fine, a computer can do that. There's this

12:42

turn where the computer is starting

12:44

to do things I can do.

12:45

And it's not even do it, and they're not even like necessarily

12:48

the most complicated tasks. Like read

12:50

this webpage and deliver a summary of it to

12:52

me. But that's

12:54

the thing that unlocked everyone's brain. And

12:56

I'm wondering why you think the industry didn't see that turn

12:59

coming. Because we've been very focused

13:01

on these very difficult things that people couldn't

13:03

do. And it seems like what got

13:05

everyone is when the computer started doing things

13:08

people do all the time.

13:10

Yeah, I think that analysis is correct. I think

13:12

that is why the large language

13:15

models have really entered the public consciousness

13:17

because it's something the average

13:19

person, the Joe public can

13:21

actually understand and interact

13:24

with. And of course, language is core

13:26

to human intelligence and our

13:28

everyday lives. So I think

13:30

that does explain why chat bots specifically

13:33

have sort of gone viral in the way they

13:35

have. Even though I would say

13:37

things like alpha fold,

13:40

I mean, of course I'd be biased in saying

13:42

this, but I think it's actually had the

13:44

most unequivocally biggest

13:47

sort of beneficial effects so far in AI

13:49

on the world, because there's a million biologists

13:52

now, researchers and medical researchers have used

13:54

alpha fold. I think that's nearly every biologist in the

13:56

world, every big pharma company is using

13:58

it to advance their drug discovery. programs. I've

14:01

had dozens of Nobel Prize winner

14:03

level biologists and chemists talk

14:06

to me about how they're using alpha-fold. So

14:08

a certain set of all the wealth scientists,

14:10

let's say, they all know alpha-fold. It's

14:13

affected and massively accelerated their

14:15

important research work. But of

14:17

course, the average person in the street doesn't know

14:20

what proteins are even and

14:22

doesn't know what the importance of those things

14:25

are for things like drug discovery. Whereas

14:27

obviously, for a chatbot, everyone can understand

14:29

this is incredible. And it's very visceral to

14:32

get it to write you a poem or something

14:35

everybody can understand and process and measure

14:38

compared to what they do

14:40

or are able to do.

14:42

It seems like that is the focus of

14:44

productized AI, these

14:46

chatbot interfaces or these

14:49

generative products that are going to make stuff

14:51

for people. That's

14:53

where the risk has been focused. Even a conversation

14:56

about risk has dramatically escalated.

14:58

I want to make sure we talk about that at length. But

15:00

even the conversation about risk has escalated

15:03

because people can now see, oh, these schools can do

15:05

stuff. Did you perceive the

15:07

same level of scrutiny when you

15:09

were working on alpha-fold? It

15:11

doesn't seem like anyone thought, oh, alpha-fold is going to

15:14

destroy humanity.

15:15

No, but there was a lot of scrutiny,

15:17

but just again, it was in a very specialized area

15:20

with renowned experts. And actually, we did talk

15:23

to over 30 experts

15:25

in the field from top

15:27

biologists to bioethicists to biosecurity

15:30

people. And actually, our partners, we

15:32

partnered with the European Bioinformatics Institute

15:34

to release the alpha-fold database of all the protein

15:37

structures. And they guided us

15:39

as well on how this could be safely put

15:41

out there. So there was a lot of scrutiny.

15:43

And the overwhelming conclusion from the people we

15:46

consulted was that the benefits far outweighed

15:48

any risks, although we did make some small

15:50

adjustments based on their feedback about which

15:53

structures to release. There was a lot of scrutiny,

15:55

but again, it's just in a very expert

15:58

sort of domain. And with the... Just

16:00

going back to your first question on the generative models,

16:02

I do think we are right at

16:05

the beginning of an incredible

16:07

new era that's going to play out over the next five,

16:09

10 years, not only in advancing science

16:12

with AI, but in terms of

16:14

the types of products we can build

16:16

to improve people's everyday lives,

16:18

billions of people in their everyday lives and help

16:21

them to be more efficient and to enrich their

16:23

lives. I think what we're seeing today

16:25

with these chatbots is literally

16:27

just scratching the surface. There's

16:29

a lot more types of AI than generative AI. Generative

16:32

AI is now the in thing, but I think

16:34

that planning and deep reinforcement

16:37

learning and problem solving and

16:39

reasoning, those kinds of capabilities

16:41

are going to come back in in the

16:43

next wave after this, along

16:45

with the current capabilities of the current systems.

16:48

I think in a year or two's time, if

16:50

we were to talk again, we're going to be

16:53

talking about

16:54

entirely new types of products

16:56

and experiences and services that

16:59

we've never seen before capabilities. I'm

17:01

very excited about building those things,

17:03

actually. That's one of the reasons I'm very excited

17:05

about leading Google DeepMind now

17:07

in this new era and focusing on

17:10

building these AI-powered

17:12

next generation products.

17:14

Let's stay in the weeds of Google

17:16

DeepMind itself for one more turn. Sunar

17:19

Pashai comes to you and says, all right, I'm the CEO of

17:21

Alphabet and the CEO of Google. I can just make this call.

17:23

I'm going to bring DeepMind into Google, merge you with Google

17:26

Brain. You're going to be the CEO. How did you

17:28

react to that prompt? It wasn't

17:30

like that. It was much more of

17:32

a conversation between the leaders of

17:35

the various different relevant groups and

17:37

Sunar about the

17:39

inflection point that we're seeing, the

17:41

maturity of the systems and

17:43

what could be possible

17:44

with those in the product

17:46

space and how to improve

17:48

experiences for our users, our billions

17:50

of users, and how exciting

17:52

that might be and what that all requires

17:55

in totality, both the change in focus,

17:57

a change in the approach to research. the

18:00

combination of resources that are required,

18:02

like compute resources. So there

18:04

was a sort of big collection of factors

18:07

to take into account that we all discussed

18:09

as a leadership group. And then

18:12

conclusions from that then result in

18:14

actions, including the merger

18:17

and also what the plans are then for

18:19

the next couple of years and what

18:21

the focus should be of that merged unit.

18:25

Do you perceive a difference being a

18:27

CEO inside of Google versus being a CEO

18:30

inside of Alphabet?

18:31

It's still early days, but I

18:33

think it's been pretty similar because

18:36

although DeepMind was an Alphabet

18:38

company, it was very unusual for

18:41

another bet, as they call it, Alphabet, which

18:44

is that we already were very

18:46

closely integrated and

18:48

collaborating with many of the

18:50

Google product area teams and

18:53

groups. We have an applied team at

18:55

DeepMind whose job it was to

18:58

translate our research work into

19:00

features and products by collaborating with

19:02

the Google product teams. And so we've

19:05

had hundreds of successful launches already actually

19:07

over the last few years, just quiet ones behind the scenes.

19:09

So in fact, many of the services

19:11

or devices or systems that you

19:14

use every day at Google will have some

19:16

kind of DeepMind technology under the hood as

19:19

a component. So we already had that

19:21

integrative structure. And then, of course,

19:24

what we were famous for was doing the scientific

19:26

advances and gaming advances. But

19:28

behind the scenes, there was a lot of bread and butter work going

19:31

on that was affecting all parts

19:33

of Google. We were different from other bets where

19:35

they have to make a business outside

19:37

of Google and become

19:40

an independent business. That was never the

19:42

goal or the remit for us, even as

19:44

an independent company. And

19:47

now within Google, we're just more

19:49

tightly integrated in terms of the product services.

19:51

And I see that as an advantage because we can

19:53

actually go deeper and do more exciting and

19:56

ambitious things in a much closer collaboration

19:58

with these other companies.

19:59

product teams than we could from outside

20:02

of Google. But we still retain some

20:05

sort of latitude to pick the processes

20:07

and the systems that optimize our

20:09

mission of producing the

20:11

most capable, in general, AI systems in

20:13

the world. There's been a bunch of reporting

20:16

that this is actually a bit of a culture clash. You're

20:18

now in charge of both. How have you structured the

20:20

group? How is Google DeepMind structured under

20:23

you as CEO? And how are you managing that

20:25

culture integration? Actually,

20:27

it turns out the culture is a lot more

20:29

similar

20:29

than perhaps has been reported externally.

20:32

And in the end, it's actually been surprisingly

20:35

smooth and pleasant because you're talking

20:37

about two world-class research groups,

20:40

two of the best AI research

20:42

organizations in the world, incredible

20:44

talent on both sides, storied

20:46

histories. As we were thinking about

20:48

the merger and planning it, we were looking

20:51

at we had some document where we listed,

20:54

I guess, the top 10 breakthroughs from each group.

20:57

And when you take that in totality, it's

21:00

like 80%, 90% over the

21:02

last decade of the breakthroughs that underpin

21:05

the modern AI industry, from deep reinforcement

21:08

learning to transformers, of course. Transformer

21:10

is a type of neural network architecture. So

21:12

it's an incredible set of people and talent.

21:15

And there's massive respect for both

21:17

groups on both sides. And there was actually a lot

21:19

of collaboration on a project-based

21:22

level ongoing over the last decade.

21:24

So of course, we all know each other very well.

21:27

I just think it's a question of focus

21:30

and a bit of coordination across both

21:32

groups, actually, and more, in

21:34

terms of what are we going to focus on,

21:38

other places that make sense to collaborate,

21:41

two separate teams to collaborate on, and

21:43

maybe deduplicate some efforts that

21:45

basically are overlapping. So

21:47

fairly obvious stuff, to be honest. But

21:50

it's important, moving into this new phase

21:52

now, where we're kind of into more of an engineering

21:55

phase of AI. And that requires

21:57

huge resources, both compute, engineering,

21:59

other things. And even as a

22:02

company the size of Google, we've got to pick

22:04

out bets carefully and be clear about

22:06

which arrows we're going to put our

22:08

wood behind and then focus on those and

22:11

then massively deliver on those things. So it's

22:13

just, I think it's part of the natural course

22:15

of evolution as to where we are in

22:18

the AI journey. So that thing

22:20

you talked about, we're going to combine these groups, we're

22:22

going to pick what we're doing, we're going to deduplate some efforts.

22:25

Those are structure questions. Have you decided on a

22:27

structure yet?

22:28

And what do you think that structure will be?

22:30

We have. I mean, the structure is still evolving.

22:32

We're only a couple of months into it. We wanted

22:34

to make sure we didn't break anything that

22:36

was working. Both teams are incredibly productive

22:39

doing super amazing research, but

22:42

also plugging in to very

22:44

important product things that are going

22:46

on. You keep saying

22:48

both teams. Do you think of it as both two

22:51

teams or are you trying to make one team? No,

22:53

no, it's for sure. It's one unified team.

22:55

I like to call it a super unit.

22:58

And I'm very excited about that. But obviously

23:00

we're still sort of combining that and forming

23:02

the new culture and forming the new grouping,

23:05

including the organizational structures as

23:07

complex things, putting two big

23:09

research groups together like this. But I

23:12

think over by the end of the summer, you

23:14

know, we'll be a single unified entity.

23:16

And I think that'd be very exciting. And we're already feeling

23:19

even a couple of months in the benefits

23:21

and the strengths of that with projects like

23:23

Gemini that you may have heard of, which is our

23:25

next generation multimodal large

23:28

models. Very, very exciting work going

23:30

on there, combining all the

23:32

best ideas from across both

23:34

world class research groups. Gemini,

23:37

by the way, is Google's next generation AI

23:39

language model. Think of it like GPT-4 compared

23:42

to GPT-3.

23:45

We have to take a quick break. We'll be back in a moment.

23:55

This

23:57

episode is brought to you by Shopify.

24:00

That's the sound of switching your business to Shopify,

24:03

the global commerce platform that supercharges

24:05

your selling. Harness the best converting checkout

24:07

and same intuitive features, trusted

24:10

apps, and powerful analytics used by

24:12

the world's leading brands. Stop

24:14

leaving sales on the table. Discover

24:16

why millions trust Shopify to build,

24:18

grow, and run their business. Sign

24:20

up today for your $1 per month trial period

24:23

at shopify.com slash tech23.

24:26

Hello, I'm Esther Perel. I'm

24:28

a psychotherapist and host of the podcast,

24:30

Where Should We Begin? Relationships

24:33

expectations are at an all-time

24:35

high, and yet the norms

24:38

are less and less clear, and we are

24:40

literally making up the new norms

24:42

as we go. Whether it's your

24:44

work relationships, friendships, or

24:47

romantic relationships, I invite

24:49

you to enter into my office

24:52

and listen in on my sessions, where

24:55

I help people explore the challenges

24:57

and choices in their relationships.

25:01

You will listen intensely to

25:03

them, but you will actually see yourselves.

25:07

And in the process, you become unstuck

25:10

and empowered in your own relationships. Join

25:14

me in my office every Monday

25:16

morning for a new episode. Listen

25:19

and follow Where Should We Begin on your

25:22

favorite podcast app.

25:28

We're back with Dennis DeSantis. You

25:33

have a lot of decisions to make, right? Well, you're describing

25:35

as much of complicated decisions than out in the world. How should we regulate this,

25:37

which is another set of very complicated decisions? And

25:41

you're a chess champion. You're

25:43

a person who's made the games. So what's

25:45

your framework for making decisions? I

25:48

suspect it's more rigorous than other ones I might

25:50

hear about. I think it probably is.

25:51

And I think

25:54

if you play a game like chess that seriously, effectively,

25:56

professionally, since, you know, in all

25:59

my childhood, since the age of 10, I think

26:01

it's very formative for your brain. So

26:04

chess is a sort of problem solving

26:06

and strategizing. I find it a very

26:08

useful framework for many things

26:11

and decision making. Chess is basically

26:13

decision making under pressure with

26:15

an opponent. It's very complex and I think it's a

26:17

great thing. I advocate it being taught

26:19

at school, actually, part of the school curriculum, because I think

26:21

it's a really fantastic training ground

26:24

for problem solving and decision making. But

26:27

then I think actually the overarching approach

26:29

is more of like the scientific method. So

26:31

I think all my training is doing my PhDs

26:34

and postdocs and so on. Obviously,

26:36

I did it in neuroscience, so I was learning about the brain,

26:39

but it also taught me how to do

26:41

rigorous sort of hypothesis testing and hypothesis

26:44

generation and then update

26:46

based on empirical evidence. So

26:49

I think the whole scientific method, as

26:51

well as the chess planning, both can

26:54

be translated into the business domain.

26:57

You have to be smart about how to translate that, so

26:59

you can't be academic about these things. And

27:01

often in the real world, in business, there's

27:03

a lot of uncertainty and hidden

27:05

information that you don't know. So

27:08

in chess, obviously, all the information is there

27:10

for you on the board. You can't just sort of directly

27:12

translate those skills, but I think

27:14

in the background, they can be very helpful

27:17

if applied in the right way. Give me an example. Make that real

27:19

for people. How do you combine those two in some decision you've made? There's

27:22

so many decisions I make every day. It's hard

27:24

to come up with one now, but I tend

27:26

to try and plan out and scenario

27:28

plan many, many years in advance. I

27:31

tell you, the way I try to approach things is

27:33

I have an end goal. I'm quite good at imagining

27:36

things, so that's a different skill. Visualizing

27:38

or imagining what an end state,

27:41

a perfect end state would look like, whether that's

27:43

organizational or it's product-based or it's

27:45

research-based. And then

27:47

I kind of work back from the end point

27:50

and then figure out what all the steps would

27:52

be required and in what order

27:54

to make that outcome as likely

27:57

as possible. So that's a little

27:59

bit chess-like. right? In the sense of like you

28:01

have some plan that you would like to get

28:03

to checkmate, you know, your opponent, but

28:06

your many moves away from that. So what

28:08

are the incremental things one must do to

28:10

improve your position in order to increase

28:13

the likelihood of that final outcome?

28:16

And I found that extremely useful

28:18

to do that search process from, you

28:21

know, the end goal back to the state

28:23

that the current state that you find yourself

28:25

in.

28:26

Let's put that next to some products, right? You said there's a lot

28:28

of deep mind technology in a lot of Google products. I

28:31

think obviously the ones that we can all look

28:33

at are barred in the new search

28:35

generative experience. Obviously there's AI

28:37

and Google photos and all the stuff, but focused

28:40

on sort of the LLM moment, it's barred in the search generative

28:42

experience. Those can't be the end

28:44

state, right? Like that, that's not,

28:46

they're not finished and Gemini is coming and we'll probably

28:48

improve both of those and all that will happen. When

28:51

you think about the end state of those products, what do you

28:53

see?

28:53

The AI systems around Google are

28:56

also not just in the, in the consumer facing

28:58

things, but also under the hood, right? That you may not

29:00

realize. So even for example, one of

29:02

the things we applied our AI systems to

29:05

very initially was the cooling systems

29:07

in Google's data centers, enormous data

29:09

centers, and actually reducing the energy

29:11

they use by nearly 30% that the

29:14

cooling systems use, which is obviously huge if you

29:16

multiply that by all of the data centers

29:18

and computers they have there. So there's

29:21

actually a lot of things under the hood where AI is being used

29:23

to improve the efficiency of those systems

29:25

all the time. But you're right, the current

29:27

products are not the end state. They're just, they're

29:30

actually just way points, I would say in the

29:32

case of chat bots and,

29:34

and those kinds of systems, I think ultimately

29:36

they will become these incredible universal

29:39

personal assistants that you kind of use

29:42

multiple times during the day for

29:44

many, many things across really useful

29:46

and helpful things across your daily lives, from

29:50

what books to read to, you know, recommendations

29:52

on, on maybe live events and things like that,

29:54

to booking your travel, to

29:57

planning, you know, trips for you to a system.

30:00

existing year in your everyday work.

30:02

I think we're still far away from that

30:05

with the current chatbots. I

30:07

think we know what's missing,

30:09

things like planning and reasoning and memory.

30:12

We're working really hard on those things.

30:15

I think what you'll see in maybe

30:17

a couple of years' time, today's chatbots

30:20

will look trivial by comparison to

30:23

what's coming in the next few years.

30:25

My background is as a person who's reported on

30:27

computers, I think of computers as somewhat modular systems.

30:29

You look at a phone, it's got a screen, it's got a chip, it's got

30:31

a cell antenna, whatever.

30:33

Should I look at AI systems that way?

30:35

There's an LLM, which is a very convincing

30:38

human language interface, and behind

30:40

it might be alpha fold that's actually doing

30:43

the protein folding. Is that how you think about stitching

30:45

these things together, or is it a different

30:47

evolutionary pathway?

30:49

Actually, there's a whole branch

30:51

of research going into what's called tool

30:53

use. This is the idea that

30:56

these large language models or large multi-modal

30:58

models, they're expert at language,

31:00

of course, and maybe a few other capabilities

31:03

like math and possibly coding. But

31:05

when you ask them to do something specialized,

31:08

like fold a protein or play

31:10

a game of chess or something like this,

31:13

actually what they end up doing is calling a tool,

31:15

which could be another AI system, that

31:18

then provides the

31:20

solution or the answer to that particular problem.

31:23

Then that's transmitted back to the user via

31:25

language or pictorially

31:28

through the central large language model

31:30

system. It may be

31:33

invisible to the user because to the

31:35

user it just looks like one big AI system that

31:37

has many capabilities. But under the hood,

31:40

it could be that actually the AI

31:42

system is broken down into smaller ones that

31:44

have specializations. I

31:47

actually think that probably is going to be the next

31:49

generation of systems will use those

31:53

kinds of capabilities. Then you can think

31:55

of the central system as almost

31:58

a switch statement that you... you

32:00

effectively prompt with language, and

32:02

it roots your query or your question

32:05

or whatever it is you're asking it to

32:07

the right tool

32:09

to solve that question for you or

32:12

provide the solution for you. And

32:14

then transmit that back in a very understandable

32:16

way, again, using through the interface,

32:19

the best interface really of natural

32:21

language. Does that get you closer to

32:23

an AGI? I know that's like in your Twitter

32:25

bio, right? This is where you are headed is AGI.

32:28

Does that

32:29

process get you closer to an AGI? Or does that

32:31

get you to some sort of maximum

32:33

state and you got to do something else? I think

32:36

that is on the critical path to

32:38

AGI.

32:39

And that's another reason, by the way,

32:41

I'm very excited about this new role and

32:44

actually doing more products and things. Because

32:46

I actually think the product roadmap

32:48

from here and the research roadmap

32:50

from here towards something like AGI

32:53

or human level AI is very

32:55

complimentary, right? So I think the kinds

32:57

of capabilities one would need to push in

33:00

order to build those kinds

33:02

of products that are useful in your everyday

33:04

life, like a universal assistant, requires

33:07

pushing on some of these capabilities like planning

33:09

and memory and reasoning that I think are

33:11

vital for us to get to AGI. So

33:14

I actually think there's a really neat feedback

33:16

loop now between products and research

33:19

where they can effectively help

33:22

each other.

33:22

I feel like I had a lot of car CEOs in the show

33:25

at the beginning of it. I asked all of them,

33:26

when do you think we're going to get self-driving cars? And they all said

33:29

five years. And they've been saying five years for five years, right?

33:31

Yes. I'm going to ask you the same version

33:34

of that question about AGI, but I feel like that the number

33:36

has gotten smaller recently with people I've talked to. How

33:39

many years until you think we have AGI?

33:40

I think there's a lot of uncertainty over

33:43

how many more breakthroughs are required

33:46

to get to AGI, big, big breakthroughs,

33:48

innovative breakthroughs, versus just scaling

33:51

up existing solutions. And I think

33:54

it very much depends on that in terms

33:56

of timeframe, obviously if there are a lot

33:58

of breakthroughs still required.

33:59

those are a lot harder to do and take

34:02

a lot longer. But right now, I would

34:05

not be surprised if we approach

34:07

something like AGI or AGI like in

34:10

the next decade. In the next decade. I'm

34:12

gonna come back to you in 10 years when I see if that happens.

34:14

Sure.

34:17

That's not a straight line though. You call it the critical path. That's

34:19

not a straight line, right? There's breakthroughs along the

34:21

way that might upset

34:23

the train and send you along a different

34:25

path, you think?

34:26

Yeah, research is never a straight line, right? If

34:28

it is, then it's not real research. If

34:31

you knew the answer before you started it, then that's

34:33

not research. So research and blue sky research

34:35

always, sort of at the frontier,

34:38

always has uncertainty around it. And

34:40

that's why you can't really predict timelines with

34:42

any certainty. But what you can look

34:44

at is trends and we can sort of look at

34:46

the quality of ideas and projects that are being

34:49

worked on today, look at how they're progressing.

34:52

That could go either way over the next five to 10 years

34:55

where we might asymptote, we might hit a

34:57

brick wall with current techniques and

35:00

scaling we may find, I

35:02

wouldn't be surprised if that happened either that we may find

35:04

that just scaling the existing systems resulted

35:07

in diminishing returns in terms

35:09

of the performance of the system. And actually

35:11

that would then signal some

35:13

new innovations were really required to

35:15

make further progress. And so at the moment,

35:17

I think nobody knows which regime

35:19

we're in. So the answer to that is you have

35:22

to push on both as hard as possible.

35:24

So both the scaling and the engineering

35:27

of existing systems and existing ideas,

35:30

as well as investing heavily

35:32

into exploratory research

35:35

directions that you think might

35:37

deliver innovations that might solve

35:40

some of the weaknesses in

35:42

the current systems. And that's one advantage

35:44

of being a large research organization with a lot

35:46

of resources is we can bet on both

35:49

of those things maximally, both

35:51

of those directions. So I think in

35:53

a way, I'm kind of agnostic to that question

35:56

of like, do we need more breakthroughs or will existing

35:58

systems just scale all the way? the way, my

36:00

view is it's an empirical question. One

36:03

should push both as hard as possible

36:05

and then the results

36:08

will speak for themselves.

36:10

This is a real tension. When

36:12

you were deep-minded and alphabet and you were very

36:14

research-focused, and you had the luxury

36:17

of that, and then the research was moved back

36:19

into Google, and Google's engineers would turn

36:21

it into products. You can see how that relationship worked.

36:23

Now you're inside of Google. Google is under a lot of

36:25

pressure as a company to win this battle.

36:28

Those are product concerns. Those are make it real

36:30

for people and go win in the market. There's

36:33

a leaked memo that went around. It was purportedly from inside Google.

36:35

The company had no moat in it. Open

36:38

source AI models or leaked models

36:40

would run on people's laptops and they would outpace

36:42

the company because the history of open computing

36:45

would outpace a closed source competitor. Was

36:47

that memo real?

36:48

I think that memo was real. Engineers

36:52

at Google often write various documents

36:54

and sometimes they get leaked and go viral.

36:56

I think that's just a sort of kind

36:59

of thing that happens. I

37:02

wouldn't take it too seriously. These

37:04

are just opinions. I think it's

37:07

interesting to listen to them and then you've

37:09

got to chart your own course. I haven't read

37:11

that specific memo in detail, but I

37:13

disagree with the conclusions from that. There's

37:16

obviously open source and publishing.

37:20

We've done tons of that in the history of DeepMind.

37:22

We gave, our fold was open sourced. We

37:25

obviously believe in open source

37:27

and supporting research and open research.

37:30

That's a key thing of the scientific discourse,

37:33

which we've been a huge part of. As Google,

37:35

of course, publishing transformers and other things

37:38

and tensorflow. You look at all the things we've

37:40

done, we do a huge amount in that

37:42

space. I also think there are

37:44

other considerations that need to be had as well. Obviously

37:47

commercial ones, but also safety questions

37:50

as well about access to these very powerful

37:52

systems. If bad actors can access

37:54

it who maybe are that technical, so they couldn't

37:56

have built it themselves, but they can certainly

37:58

reconfigure a system. that is out there,

38:01

what do you do about those things? And

38:03

I think that's been quite theoretical till

38:05

now, but I think that that is really

38:08

important from here all the

38:10

way to AGI as these systems

38:13

become more general, more sophisticated,

38:15

more powerful. That question is

38:17

going to be very important about how does

38:19

one stop bad actors just using

38:22

these systems for things they weren't intended

38:24

for, but for malicious purposes. That's

38:26

something we need to increasingly come

38:28

up with. But just to back to your question,

38:31

is look at the history of what Google

38:33

and DeepMind have done in terms of coming

38:35

up with new innovations and breakthroughs and multiple,

38:38

multiple breakthroughs over the last decade or

38:40

more. I would bet on us

38:42

and I'm certainly very confident that that will

38:44

continue and actually be even

38:47

more true over the next decade in

38:49

terms of us producing the next

38:51

key breakthroughs, just like we did

38:54

in the past. Do you think that's the moat? We invented

38:57

most of this stuff, so we're going to invent the most of the next

38:59

stuff. Yeah, I don't really think about those moats,

39:01

but I'm incredibly competitive

39:03

person. That's maybe another thing I got from

39:05

Jess and many researchers

39:07

are. Of course, they're doing it to discover

39:10

knowledge and ultimately that's what we're here

39:12

for, is to improve the human

39:14

condition. But also we want to be first to do

39:16

these things and do them responsibly

39:18

and boldly. We have some of the world's best

39:20

researchers. I think we have the biggest collection of

39:23

great researchers in the world, anywhere in the world. And

39:26

incredible track record and there's no reason

39:28

why that shouldn't continue in the future.

39:31

And in fact, I think with our new organization

39:33

and environment might be conducive to

39:36

even more and faster paced

39:38

breakthroughs than we've done in the past.

39:40

You're leading me towards risk and regulation.

39:42

I want to talk about that, but I want to start with just sort

39:44

of a different spin on it. You're

39:46

talking about all the work that has to be done. You talk about deep mind,

39:49

reinforcement learning, how that works. We just

39:51

ran a gigantic story in collaboration with

39:53

New York Magazine. It's on the cover of New York Magazine

39:55

about the taskers who are actually

39:57

doing the training, who are actually labeling.

39:59

the data, there's a lot of labor

40:02

conversation with AI along the way. Hollywood

40:05

writers are on strike right now because they don't want chat GPT

40:07

to write a bunch of scripts. I think that's appropriate.

40:10

But then there's a new class of labor that's being developed

40:12

where a bunch of people around the world

40:15

are sitting in front of computers and saying, yeah, that's a stop sign.

40:18

No, that's not a stop sign. Yep, that's clothes you can wear. No,

40:20

that's not clothes you can wear. Is that

40:22

a forever state? Is that just a new class of

40:24

work that needs to be done for these systems to operate? Or

40:26

does that come to an end? It's hard

40:28

to say. I think it's definitely

40:31

a moment in time and

40:33

the current systems and what they're requiring

40:35

at the moment. We've been very careful

40:37

just to say from our part, and I think you quoted

40:40

some of our researchers in that article to

40:42

be very careful to pay living wages

40:45

and be very responsible about

40:47

how we

40:48

do that kind of work and which partners we use.

40:51

And we also use internal teams as well. So

40:53

I think we've been actually, I'm very proud of how

40:56

responsible we've been on that type of work.

40:59

But going forwards, I

41:01

think there may be ways that these systems,

41:04

especially once you have millions and millions

41:06

of users, effectively can

41:08

bootstrap themselves. Or

41:11

one could imagine AI systems

41:13

that are capable of actually sort

41:16

of conversing with themselves or critiquing themselves.

41:18

This would be a bit like turning language systems

41:21

into a game-like setting, which of course we're

41:23

very expert in. And we've been thinking about where these

41:25

reinforcement learning systems, different versions

41:28

of them can actually sort of rate each

41:30

other in some way. And it may not be

41:32

as good as a human rater, but it's actually

41:35

a useful way to sort of do some

41:37

of the bread and butter rating and then maybe just calibrate

41:39

it by checking those ratings

41:42

with a human rater at the end, rather

41:44

than getting human raters to rate everything. So

41:46

I think there are lots of innovations

41:48

I can see coming down the line that

41:51

will help with this and potentially

41:54

mean that there's less requirement for

41:56

this all to be done by human raters. But

41:58

you think there's always human raters.

41:59

the mix. Even as you get closer to AGI,

42:02

it seems like you need someone to

42:04

tell the computer if it's doing a good job or not.

42:06

Let's take Alpha Zero as an example. Our

42:09

general games playing system that ended up learning

42:11

itself how to play any two-player game,

42:14

including chess and go. It's interesting.

42:16

What happened there is we set up the system

42:18

so that it could play against itself

42:21

tens of millions of times. In fact,

42:24

it built up its own knowledge base. It started

42:26

from random, played itself, bootstrapped

42:28

itself, trained better versions of itself, and

42:30

played those off each other in mini tournaments.

42:33

But at the end, you still want to test it

42:35

against the human world champion or something like

42:37

this, or an external computer

42:39

program that was built in a conventional

42:41

way so that you can just calibrate your

42:44

own metrics, which are telling you these systems

42:47

are improving according to these objectives

42:49

or these metrics. But you don't know for sure

42:52

until you calibrate it with an external benchmark

42:54

or measure. Depending on what that is,

42:57

a human rater or human benchmark, human expert

42:59

is often the best thing to calibrate

43:02

your internal testing against. You

43:05

make sure that your internal tests are

43:07

actually mapping reality. Again,

43:09

that's something quite exciting about products

43:12

for researchers because when you put your

43:14

research into products and millions of people are using it every

43:16

day, that's when you get real

43:18

world feedback. There's no way around that.

43:21

That's the reality. That's the best test of

43:25

any theories or any system that

43:26

you've built. Do you think that work is rewarding

43:29

or appropriate, the labeling of data

43:31

for AI system? There's just something about

43:34

that, which is I'm going to tell a computer how to understand

43:36

the world so that it might go off in the future and displace

43:39

other people. There's a loop in there

43:41

that seems like it's worth more just moral

43:43

or philosophical consideration. Have you spent

43:45

time thinking about that?

43:47

Yeah, I do think about that. I don't really see

43:49

it like that. I think that what raters

43:51

are doing is they're part of the development

43:54

cycle of making these systems safer,

43:57

more useful for everybody, and

43:59

more

43:59

helpful and more reliable. So

44:02

I think it's a critical component. In

44:04

many industries, we have safety testing, you

44:07

know, technologies and products. And today,

44:10

that's the best we can do for

44:12

AI systems, right is to is to have

44:14

human raters. I think in the future,

44:17

next few years, I think we need more,

44:19

you know, we need a lot more research. And I've been calling

44:22

for this. And we're doing this ourselves, but it needs

44:24

more than just one organization to do this is

44:26

great, robust evaluation benchmarks

44:29

for capabilities, right, so that

44:31

we know, if a system passes

44:34

these benchmarks, then it has

44:36

certain properties and it's safe, and it's reliable

44:38

in these particular ways, right. And right now,

44:41

I think we're in the space of many researchers

44:43

in academia and civil society and elsewhere, we have

44:46

a lot of good suggestions for what those tests

44:48

could be. But I don't think they are robust

44:51

or practical yet. I think they're basically

44:53

theoretical and philosophical in nature. And

44:55

I think they need to be made practical so

44:58

that we can measure our systems, you

45:00

know, empirically against those tests.

45:02

And then that gives us some assurances about about

45:05

how the system will perform. And I

45:07

think it once we have those, then

45:09

the need for this sort of human

45:12

rating testing feedback

45:14

will be reduced. I just think that's

45:17

required in the volumes is required now,

45:19

because we don't have these kinds of independent

45:22

benchmarks yet, partly because we

45:24

haven't rigorously defined what

45:26

those properties are. I mean, it's almost neuroscience

45:30

and psychology and philosophy area

45:32

as well, right. A lot of these terms have not been

45:34

defined properly, even for you know, the

45:36

human brain.

45:40

All right, one more sharp break, we'll be right back.

45:54

We're back with Dennis Asabas, the CEO of

45:56

Google DeepMind. You've

45:58

signed a letter from in the center for AI

46:01

safety, open AI, Sam Altman,

46:03

others who have also signed this letter, that warns

46:06

against the risk from AI.

46:09

And yet you're pushing on, right? Like

46:12

Google's in the market, you've got to win, you've described yourself

46:14

as competitive. There's a tension there,

46:16

right? Needing to win in the market with products and

46:19

oh boy, please regulate us because

46:21

raw

46:22

capitalism will drive

46:24

us off the cliff with AI if we don't stop it in some

46:26

way. How do you balance that risk? It

46:28

is a tension, it's a creative tension. What we

46:31

like to say at Google is we

46:33

wanna be bold and responsible.

46:35

And that's exactly what we're trying

46:37

to do and live out and role model. So

46:39

the bold part is being brave and optimistic

46:42

about what AI, the benefits,

46:44

the amazing benefits, incredible benefits

46:46

AI can bring to the world and to help

46:49

humanity with our biggest challenges, whether that's

46:51

disease or climate or sustainability.

46:54

I think AI has

46:56

huge part to play in helping our

46:58

scientists and medical experts

47:01

solve those problems. And we're working hard

47:03

on that and on all those areas and alpha

47:05

fold again, I point to as a poster child

47:07

for that, what we want to do there. So

47:10

that that's the bold part. And then the responsible

47:12

bit is to make sure we do that

47:14

as thoughtfully as possible with

47:17

as much foresight as possible ahead

47:19

of time, you know, try and anticipate what

47:22

the issues might be if one was

47:24

successful ahead of time, not

47:26

in hindsight, like perhaps has happened with

47:28

social media, for example, where, you

47:30

know, it is this incredible growth

47:33

story, obviously it's done a lot of good in the world,

47:35

but then it turns out 15 years later,

47:37

we realize there are some consequences,

47:40

unintended consequences as well to those

47:42

types of systems. And I would like to chart

47:44

a different path with

47:45

AI. And I think it's such a profound

47:47

and important and powerful technology. I

47:49

think we have to do that with something as potentially

47:52

as transformative as AI.

47:54

And it doesn't mean no mistakes will be made.

47:56

It's, you know, it's very new, anything new, some

47:59

things you...

47:59

You can't predict everything ahead

48:02

of time, but I think we

48:04

can try and do the best job we can. And

48:06

that's what signing that letter was for, is just to point

48:09

out that there are, we

48:11

don't know the, I don't think it's likely, I

48:13

don't think it's, I don't know on the time

48:15

scales, but it's something that we

48:17

should consider too, in the limit, is

48:19

what these systems can do and might be

48:22

able to do as we get closer to AGI. We're

48:24

nowhere near that now. So this is not a question

48:26

of today's technologies or even

48:29

the next few years, but at some point,

48:31

and given the technology is accelerating very fast,

48:34

we will need to think about those questions. And we don't

48:36

want to be thinking about them on the eve

48:39

of them happening. We need to use the time

48:41

now, the next five, 10, whatever

48:43

it is years, to do the research and

48:46

to do the analysis and to engage

48:48

with, various stakeholders,

48:50

civil society, academia, government,

48:53

to figure out as this stuff is

48:55

developing very rapidly, what the

48:57

best way is of making

48:59

sure we maximize the benefits and minimize

49:02

any risks. And that includes mostly at this stage,

49:05

doing more research into these areas, like

49:08

coming up with better evaluations and benchmarks

49:10

to rigorously test the capabilities

49:13

of these frontier systems.

49:15

You talked about tool usage for AI models,

49:18

right? You'd ask an LOM, do something that goes

49:20

off and ask off of all the full protein for you. Combining

49:23

systems like that, integrating systems like that, historically,

49:26

that's where emergent behaviors appear, things

49:28

you couldn't have predicted happen start happening. Are

49:31

you worried about that? There's not a rigorous way

49:33

to test that.

49:34

You assemble a computer, the computer starts doing

49:36

stuff you never expected.

49:38

Right, exactly. I think that's exactly the

49:40

sort of thing we should be researching and thinking

49:42

about ahead of time, is as

49:45

tool use becomes more sophisticated and you can

49:47

combine different AI systems together in different

49:49

ways, there is scope for

49:51

emergent behavior. Of course, that

49:54

emergent behavior may be very desirable and

49:56

be extremely useful, but it could

49:58

also potentially be harder.

49:59

in the wrong hands and

50:02

in the hands of bad actors, whether that's individuals

50:04

or even nation-states.

50:06

Let's say the United States and the EU and China

50:08

all agree on some framework to regulate AI.

50:10

And then North Korea or Iran

50:12

says, fuck it, no rules. And that becomes

50:14

a center of bad actor research.

50:17

How does that play out? Do you see a world

50:19

in which that's possible?

50:20

Yeah, I think that is a possible world.

50:23

This is why I've been talking to governments,

50:25

UK, US mostly, but also EU on,

50:28

I think whatever

50:30

regulations or guardrails or whatever that

50:33

is that transpires over the next few years

50:35

and tests, they ideally would

50:37

be international. And there would be international

50:40

cooperation around those safeguards

50:43

and international agreement around

50:46

deployment of these systems and other things.

50:48

Now, I don't know how

50:50

likely that is given the geopolitical

50:53

tensions around the world, but that is

50:55

by far the best state. And I think what

50:57

we should be aiming for if

50:59

we can.

51:00

If the government here passes a rule, it says, here's the here's

51:02

what Google is allowed to do. Here's what Microsoft is allowed

51:05

to do. You are in charge. You

51:07

are accountable. And you can go say, all

51:09

right, we're just not running this code in our data center.

51:11

We are not going to have these capabilities. It's not legal.

51:13

If I'm just a person with a MacBook,

51:16

would you accept some limitation

51:18

on what a MacBook could do because

51:20

the threat from AI is so scary? That's

51:22

the thing I worry about. Practically, if you have open

51:24

source models and people are going to use them for

51:26

weird things, are we going to tell

51:28

Intel to restrict what its chips can do?

51:30

How would we implement that such that

51:33

it actually affects everyone and not just

51:35

we're going to throw Dennis in jail if Google does stuff

51:37

we don't like?

51:38

Those are the big questions that are being debated

51:40

right now. And I do worry about that.

51:44

On the one hand, there are a lot of benefits of open sourcing

51:46

and accelerating scientific discourse.

51:48

And lots of advances happen there. And it gives

51:50

access to many developers. On

51:53

the other hand, there could be some negative consequences

51:55

with that. If they're about individual actors

51:58

that do bad things with that,

51:59

access

52:00

and that proliferates. And I think that's

52:03

a question for the next few years that will need to

52:05

be resolved because right now I think it's okay

52:07

because the systems are not that sophisticated or that

52:09

powerful and therefore not that risky. But

52:12

I think as the systems increase

52:14

in their power and

52:17

generality, the access question will need

52:19

to be thought about from government and how they

52:21

want to restrict that or control that or

52:24

monitor that is going to be an important question. I

52:26

don't have any answers for you because I think this is a societal

52:28

question actually that requires stakeholders

52:31

from right across society to come together

52:34

and weigh up the benefits with the

52:36

risks there.

52:37

You said we're not there yet, but Google's own work

52:39

in AI certainly had

52:41

some controversy associated with this around

52:44

responsibility, around what the models

52:46

can do or can't do. There's a famous

52:48

stochastic parrots paper from Emily

52:50

Bender and Tim Nicki-Brew and Margaret Mitchell that led

52:52

to a lot of controversy inside of Google, led to

52:55

them leaving. Did you read that paper and

52:57

think, okay, this is a correct. LMS

53:00

are going to lie to people and Google

53:02

will be responsible for that. How

53:05

do you think about that now with all of the scrutiny?

53:08

I think

53:09

the large language models and

53:12

I think this is one reason that

53:14

Google has been very responsible with this is that we know

53:16

that they hallucinate and they

53:18

can be inaccurate. That's one

53:20

of the key areas I think that has to be

53:22

improved over the next few years is factuality

53:26

and grounding and making sure

53:28

that they don't spread disinformation,

53:30

these kinds of things. That's very much top

53:32

of mind for us. We

53:35

have many ideas of how to improve

53:37

that and our old

53:39

DeepMinds Sparrow language

53:41

model was an experiment into,

53:44

which we published a couple of years ago, was an experiment

53:46

into just how good can we get

53:48

factuality and rules adherence

53:51

in these systems. Turns out we can

53:53

maybe make it an order of magnitude better, but

53:55

it sometimes comes at the expense of lucidness

53:58

or creativity on the part. of the language

54:00

model and therefore usefulness. So it's a bit of

54:02

like a Pareto frontier where if you improve one

54:04

dimension, you sort of reduce the

54:07

capability in another dimension. Ideally,

54:09

what we want to do in the next phases and the next

54:12

generations of systems is combine the best of

54:14

both worlds, keep the creativity

54:16

and lucidness and fondness of

54:18

the current systems, but improve

54:21

their factuality and reliability.

54:23

We've got a long way to go on that, but

54:25

I can see things improving and I don't

54:27

see any theoretical reason why

54:30

these systems can't get to

54:32

extremely high levels of accuracy

54:34

and reliability in the next few years.

54:37

When you're using the Google search generative

54:39

experience, do you believe what it says?

54:42

I do. I sometimes double check

54:44

things, especially in the scientific domain where

54:47

I've had very funny situations where actually all of

54:49

these models do this, where you ask them to

54:51

summarize an area of research, which I think would

54:53

be super useful if they could do that, and

54:55

then say, what are the key papers I should read?

54:58

They come up with very plausible signing

55:00

papers with very plausible author

55:02

lists, but then when you go and look into

55:05

it, it turns out they're just like the most

55:07

famous people in that field or

55:10

the titles from two different papers

55:12

combined together, but of course they're extremely plausible

55:15

as a collection of words. I think

55:18

there what needs to happen is these systems need to

55:20

understand that citations

55:22

and papers and author lists

55:25

are a unitary block rather than

55:27

a word-by-word prediction. There's

55:31

interesting cases like that where we need to improve

55:33

and that's something which is, of course, us

55:36

as wanting to advance the frontiers of science. That's

55:38

a particularly interesting use case that we would like

55:40

to improve and fix for our own

55:43

needs as well. I'd love these systems to better

55:45

summarize for me. Here are the top five

55:47

papers to read about

55:48

a particular disease or something

55:50

like that to just quickly onboard you

55:53

in that particular area. I think would be incredibly useful.

55:56

I will tell you, I clicked a link that was basically

55:58

a link to Google, my friend. and John Gruber

56:00

and SGE confidently told me that

56:02

he pioneered the use of a Mac in newspapers

56:05

and invented WebKit. And I

56:07

don't know where that came from. And is

56:10

there a level, is there a quality level, a truthfulness

56:13

level that you need to hit before you roll that out to

56:15

the mass

56:16

audience?

56:17

Yeah, we think about this all the

56:19

time, especially at Google because

56:22

of the incredibly high standards

56:24

Google holds itself to on things

56:26

like search and that we all rely

56:29

on every day in every moment of every day,

56:31

really. We want to kind of get towards

56:33

that level of reliability. Obviously we're a long, long,

56:35

long way away from that at the moment with not

56:38

just us, but anybody with their generative systems.

56:40

But that's the gold standard. And

56:43

actually things like tool use can

56:45

come in very handy here where you could in

56:48

effect build these systems so that

56:50

they fact check themselves, perhaps

56:52

even using search or other

56:55

reliable sources, cross

56:57

reference, just like a good researcher

56:59

would, cross reference your facts. Also

57:02

having a better understanding of the world, what

57:05

are research papers, what entities

57:07

are they, these kinds of things. So

57:09

they can,

57:10

these systems need to sort of get a, have a better understanding

57:13

of the media they're dealing

57:15

with. And maybe also give

57:18

the systems the ability to reason and

57:20

plan because then they could potentially turn

57:22

that on their own outputs and sort

57:24

of critique themselves. And again,

57:26

this is something we have a lot of experience in, in

57:29

games programs that are, you know,

57:31

they don't just output the first move that

57:33

you think of in chess or go, right? You

57:35

actually plan and do some, do

57:38

some search around that and

57:40

then back up. And sometimes they change

57:42

their minds and

57:43

switch to a better move. And that you could

57:45

imagine some kind of process like that with

57:47

words and language as well.

57:49

There's the concept of model collapse,

57:52

right? That we're going to train LLMs

57:54

on a bunch of LLM generated data and that's going to go

57:56

into a circle. When you talk about,

57:59

cross-referencing facts and I

58:02

think about Google, Google going out in the

58:04

web and trying to cross-reference a bunch of stuff

58:06

But maybe all that stuff has been generated by LMS

58:09

that were hallucinating in 2023. How

58:11

do you guard against that?

58:12

We're working on some pretty cool solutions

58:14

to that I think the answer is and

58:17

this is an answer to deepfakes as well is to do

58:20

some sort of encrypted watermarking Sophisticated

58:22

walk marking that can't be removed Easily

58:25

or at all and it's probably built into

58:27

the genitive models themselves So

58:29

it's part of the genitive process We hope

58:31

to release that and maybe provide

58:33

it to third parties as well as a as a generic

58:36

solution But I think that the industry

58:38

in the field needs those types of solutions where we

58:41

can mark Generated media

58:43

be that images audio perhaps even

58:45

text with you know Some kind of kite mark

58:48

that says to the user and

58:51

future AI systems that with

58:53

ease with AI generated And

58:55

I think that's a very very pressing need

58:57

right now for near-term issues

58:59

With AI like deepfakes and disinformation

59:02

and so on but I actually think

59:04

a solution is On the horizon

59:06

now.

59:07

I had Microsoft CTO and EVP of AI

59:09

Kevin Scott on the show a few weeks ago He said something

59:11

very similar. I promised him that we would do a one-hour

59:14

episode on metadata. So you're coming

59:16

for that one That

59:19

will be our most popular episode if I know this

59:21

audience a full hour on metadata

59:23

ideas Will be our most popular episode sounds perfect.

59:26

Well Dennis. Thank you so much for coming on decoder. We

59:28

have to come back soon. Thanks so much

59:32

Thanks again to Dennis

59:34

Sabas for taking the time to chat and decoder today and

59:36

thank you for listening I hope you enjoyed it as

59:39

always. I'd love to hear what you think of decoding You

59:41

can email us at decoder at the verge comm I

59:43

read all the emails We can hit us up

59:45

directly on tick tock check it out.

59:47

It's at decoder pod It's a lot of fun if

59:50

you like decoder, please share with your friends subscribe

59:52

wherever your podcast I really like the show hit

59:54

us with that five-star review Decoder is a

59:56

production of urge and part of the Vox media podcast network

59:59

today's episode is for produced by Raghum Manavallam and Jackie

1:00:01

McDermott. It was edited by Callie Wright. The

1:00:03

decoder of music is by Breakmaster Cylinder. Our editorial

1:00:06

director is Brooke Mentors and our executive director

1:00:08

of video and audio is Eleanor Donovan. We'll

1:00:10

see you next time.